KR20240082270A - Protein language model based on protein structure - Google Patents

Abstract

The disclosed technology relates to determining the pathogenicity of nucleotide variants. In particular, the disclosed technology relates to the step of designating a specific amino acid at a specific position of the protein as a gap amino acid and designating the amino acid remaining at the remaining position of the protein as a non-gap amino acid. The disclosed techniques further include generating a gapped spatial representation of the protein that includes the spatial configuration of non-gap amino acids and excludes the spatial configuration of gap amino acids; and determining the pathogenicity of a nucleotide variant based, at least in part, on the gap space representation and the representation of alternative amino acids produced by the nucleotide variant at a particular position.

Description

Protein language model based on protein structure

priority application

This application claims priority and benefit from the following US application. The priority application is incorporated herein by reference for all purposes.

U.S. Regular Patent Application No. 17/533,091, entitled “Protein Structure-Based Protein Language Models,” filed November 22, 2021 (Attorney Docket No. ILLM 1050-2/IP-2164-US), priority claimed U.S. Provisional Patent Application No. 63/253,122, filed October 6, 2021 (Attorney Docket No. ILLM 1050-1/IP-2164-PRV), U.S. Provisional Patent Application No. 63/281,579, filed November 19, 2021 (Attorney Docket No. ILLM 1050-1/IP-2164-PRV) No. ILLM 1060-1/IP-2270-PRV) and U.S. Provisional Patent Application No. 63/281,592, filed November 19, 2021 (Attorney Docket No. ILLM 1061-1/IP-2271-PRV); and

U.S. Regular Patent Application Ser. No. 17/953,286, entitled “Predicting Variant Pathogenicity From Evolutionary Conservation Using Three-Dimensional (3D) Protein Structure Voxels,” filed September 26, 2022 (Attorney Docket No. ILLM 1060-2/IP-2270) -US) Priority Claim U.S. Provisional Patent Application No. 63/253,122, filed October 6, 2021 (Attorney Docket No. ILLM 1050-1/IP-2164-PRV), U.S. Provisional Patent Application No. 63/281,579, filed 2021 19 November 2021 (Attorney Docket ILLM 1060-1/IP-2270-PRV) and U.S. Provisional Patent Application No. 63/281,592, filed November 19, 2021 (Attorney Docket ILLM 1061-1/IP-2271) -PRV); and

U.S. Regular Patent Application No. 17/953,293 Title “Combined And Transfer Learning of a Variant Pathogenicity Predictor Using Gapped and Non-Gapped Protein Samples,” Filed September 26, 2022 (Attorney Docket No. ILLM 1061-2/IP- 2271-US) Priority Claimed U.S. Provisional Patent Application No. 63/253,122, filing date October 6, 2021 (Attorney Docket No. ILLM 1050-1/IP-2164-PRV), U.S. Provisional Patent Application No. 63/281,579, filing date and U.S. Provisional Patent Application No. 63/281,592, filed November 19, 2021 (Attorney Docket No. ILLM 1060-1/IP-2270-PRV), filed November 19, 2021 (Attorney Docket No. ILLM 1061-1/IP- 2271-PRV).

Technology field

The disclosed technology relates to artificial intelligence type computer and digital data processing systems and corresponding data processing methods and intelligence emulation products (i.e., knowledge-based systems, inference systems, and knowledge acquisition systems); Includes systems for inference with uncertainty (e.g., fuzzy logic systems), adaptive systems, machine learning systems, and artificial neural networks. In particular, the disclosed technology relates to analyzing multi-channel voxelized data using deep convolutional neural networks.

References

The following are incorporated by reference for all purposes as if fully set forth herein:

Sundaram, L. et al. Predicting the clinical impact of human mutation with deep neural networks. Nat. Genet. 50, 1161-1170 (2018)];

Jaganathan, K. et al. Predicting splicing from primary sequence with deep learning. Cell 176, 535-548 (2019)];

U.S. Patent Application Serial No. 62/573,144, filed October 16, 2017 and entitled "Training a Deep Pathogenicity Classifier Using Large-Scale Benign Training Data" (Attorney Docket No. ILLM 1000-1/IP-1611-PRV) );

U.S. Patent Application No. 62/573,149, filed October 16, 2017 and entitled “Pathogenicity Classifier Based on Deep Convolutional Neural Networks (CNNs)” (Attorney Docket No. ILLM 1000-2/IP-1612-PRV) ;

U.S. Patent Application Serial No. 62/573,153, filed October 16, 2017 and entitled "Deep Semi-Supervised Learning That Generates Large-Scale Pathogenic Training Data" (Attorney Docket No. ILLM 1000-3/IP-1613- PRV);

U.S. Patent Application Serial No. 62/582,898, filed November 7, 2017 and entitled "Pathogenicity Classification of Genomic Data Using Deep Convolutional Neural Networks (CNNs)" (Attorney Docket No. ILLM 1000-4/IP-1618- PRV);

U.S. Patent Application No. 16/160,903, filed October 15, 2018 and entitled “Deep Learning-Based Techniques for Training Deep Convolutional Neural Networks” (Attorney Docket No. ILLM 1000-5/IP-1611-US) ;

U.S. Patent Application Serial No. 16/160,986, entitled “Deep Convolutional Neural Networks for Variant Classification,” filed October 15, 2018 (Attorney Docket No. ILLM 1000-6/IP-1612-US);

U.S. Patent Application Serial No. 16/160,968, filed October 15, 2018 and entitled "Semi-Supervised Learning for Training an Ensemble of Deep Convolutional Neural Networks" (Attorney Docket No. ILLM 1000-7/IP-1613- US);

U.S. Patent Application Serial No. 16/407,149, filed May 8, 2019 and entitled "Deep Learning-Based Techniques for Pre-Training Deep Convolutional Neural Networks" (Attorney Docket No. ILLM 1010-1/IP-1734- US);

U.S. Patent Application Serial No. 17/232,056, filed April 15, 2021 and entitled "Deep Convolutional Neural Networks to Predict Variant Pathogenicity Using Three-Dimensional (3d) Protein Structures" (Attorney Docket No. ILLM 1037-2/ IP-2051-US).

U.S. Patent Application Serial No. 63/175,495, filed April 15, 2021 and entitled “Multi-Channel Protein Voxelization to Predict Variant Pathogenicity Using Deep Convolutional Neural Networks” (Attorney Docket No. ILLM 1047-1/IP-2142) -PRV);

U.S. Patent Application Serial No. 63/175,767, entitled “Efficient Voxelization for Deep Learning,” filed April 16, 2021 (Attorney Docket No. ILLM 1048-1/IP-2143-PRV);

U.S. Patent Application Serial No. 17/468,411, filed September 7, 2021 and entitled “Artificial Intelligence-Based Analysis of Protein Three-Dimensional (3d) Structures” (Attorney Docket No. ILLM 1037-3/IP-2051A) -US).

The subject matter discussed in this section should not be assumed to be prior art merely as a result of its references within this section. Similarly, it should not be assumed that issues related to the subject matter mentioned in this section or provided as background have been previously recognized in the prior art. The subject matter in this section merely represents different approaches, which approaches themselves may also correspond to implementations of the claimed technology.

In a broad sense, genomics, also referred to as functional genomics, aims to characterize the function of all genomic elements in an organism using genome-scale analyzes such as genome sequencing, transcriptome profiling, and proteomics. Genomics emerged as a data-driven science and it works by discovering novel properties from the exploration of genome-scale data rather than testing preconceived models and hypotheses. Applications of genomics include finding associations between genotypes and phenotypes, discovering biomarkers for patient stratification, predicting the function of genes, and charting biochemically active genomic regions such as transcriptional enhancers. It includes doing.

Genomics data is too large and too complex to be examined through visual studies of pairwise correlations alone. Instead, analysis tools are required to support the discovery of unexpected relationships, to derive new hypotheses and models, and to make predictions. Unlike some algorithms where assumptions and domain expertise are hard-coded, machine learning algorithms are designed to automatically detect patterns in data. Therefore, machine learning algorithms are suitable for data-driven science, and especially genomics. However, the performance of a machine learning algorithm can strongly depend on how the data is represented, i.e., how each variable (also called a feature) is computed. For example, to classify a tumor as malignant or benign from a fluorescence microscopy image, a preprocessing algorithm can detect cells, identify cell types, and generate a list of cell counts for each cell type. there is.

The machine learning model can take the estimated cell counts, which are examples of hand-crafted features, as input features for classifying the tumor. The central problem is that classification performance is highly dependent on the quality and relevance of these features. For example, relevant visual features such as cell morphology, distances between cells, or localization within an organ are not captured in cell counts, and this incomplete representation of the data can reduce classification accuracy.

Deep learning, a subdivision of machine learning, addresses these issues by embedding the calculation of features in the machine learning model itself to produce an end-to-end model. These results were realized through the development of deep neural networks, i.e. machine learning models containing successive elementary operations, which compute increasingly complex features by taking as input the results of preceding operations. Deep neural networks can improve prediction accuracy by discovering relevant features of high complexity, such as the cell morphology and spatial organization of cells in the example above. Construction and training of deep neural networks has been made possible by an explosion of data, algorithmic advances, and substantial increases in computational capacity, particularly through the use of graphical processing units (GPUs).

The goal of supervised learning is to obtain a model that takes features as input and returns predictions for the so-called target variable. An example of a supervised learning problem is determining whether an intron is splice out by considering features on the RNA, such as the presence of a canonical splice site sequence, the location of the splicing fork, or the length of the intron. It is to predict (target). Training a machine learning model refers to learning its parameters, which usually involves minimizing the loss function on the training data with the goal of making accurate predictions on unseen data.

For many supervised learning problems in computational biotechnology, the input data can be represented as a table with a number of columns or features, each containing numerical or categorical data potentially useful for making predictions. While some input data are naturally represented as features in a table (e.g., temperature or time), other input data (such as deoxyribonucleic acid (DNA) sequences in k -mer counts) need to be adapted to a tabular representation. It needs to be converted first using a process called feature extraction. For intron splicing prediction problems, the presence or absence of canonical splice site sequences, the location of splicing forks, and intron length can be preprocessed features collected in a tabular format. Tabular data is the standard for a wide range of supervised machine learning models, ranging from simple linear models such as logistic regression to more flexible nonlinear models such as neural networks and many others.

Logistic regression is a binary classifier, a supervised learning model that predicts a binary target variable. Specifically, logistic regression predicts the probability of a positive class by calculating a weighted sum of input features mapped to the [0,1] interval using a sigmoid function, a type of activation function. A parameter of logistic regression, or other linear classifiers using different activation functions, is the weight of the weighted sum. A linear classifier fails when the class, for example of spliced-out or non-spliced-out introns, cannot be well distinguished by a weighted sum of the input features. To improve prediction performance, new input features can be added manually by transforming or combining existing features in new ways, for example, by taking squares or pairwise products.

Neural networks automatically learn these nonlinear feature transformations using hidden layers. Each hidden layer can be thought of as a number of linear models whose outputs are transformed by a non-linear activation function, such as a sigmoid function or the more popular rectified-linear unit (ReLU). Together, these layers organize input features into related complex patterns, which facilitate the task of distinguishing between the two classes.

Deep neural networks use many hidden layers, and a layer is considered fully connected when each neuron receives input from all neurons in the preceding layer. Neural networks are typically trained using stochastic gradient descent, an algorithm suitable for training models on very large data sets. Implementations of neural networks using modern deep learning frameworks enable rapid prototyping with different architectures and data sets. Fully connected neural networks can be used in a number of genetics applications, including determining exons splice in for a given sequence from sequence features such as sequence conservation or the presence of binding motifs for splice factors. predicting the percentage of; prioritizing potential disease-causing genetic variants; and predicting cis -regulatory elements within a given genomic region using features such as chromatin marks, gene expression, and evolutionary conservation.

For effective prediction, local dependencies in spatial and longitudinal data must be considered. For example, shuffling the pixels of a DNA sequence or image severely destroys informative patterns. These local dependencies establish spatial or longitudinal data separate from the tabular data, for which the ordering of features is arbitrary. We consider the problem of classifying genomic regions as bound versus unbound by specific transcription factors, where binding regions are defined as high-confidence binding events in chromatin immunoprecipitation followed by sequencing (ChIP-seq) data. By recognizing sequence motifs, transcription factors bind to DNA. Fully connected layers based on sequence derived features such as the number of k -mer instances within the sequence or position weight matrix (PWM) matching can be used for this task. Therefore, because k -mer or PWM instance frequencies are robust to shifting motifs within a sequence, such models can generalize well to sequences with the same motif located at different positions. However, they will not recognize patterns in which transcription factor binding relies on the combination of multiple motifs with well-defined spacing. Additionally, the number of possible k -mers increases exponentially with k -mer length, which poses both storage and overfitting problems.

A convolutional layer is a special form of a fully connected layer, where the same fully connected layer is applied locally to all sequence positions, for example in a 6 bp window. This approach can also be seen as scanning sequences using multiple PWMs, for example for the transcription factors GATA1 and TAL1. By using the same model parameters across positions, the total number of parameters is drastically reduced, and the network can detect motifs in positions that were unseen during training. Each convolutional layer scans a sequence through multiple filters by generating a scalar value at every position that quantifies the match between the filters and the sequence. As in a fully connected neural network, a non-linear activation function (usually ReLU) is applied to each layer. Next, a pooling operation is applied, which aggregates activations in adjacent bins across the position axis, typically taking the maximum or average activation for each channel. Pooling reduces the effective sequence length and coarsens the signal. The subsequent convolutional layer constitutes the output of the previous layer and can detect whether the GATA1 motif and TAL1 motif were present in some distance range. Finally, the output of the convolutional layer can be used as input to a fully connected neural network to perform the final prediction task. Accordingly, different types of neural network layers (eg, fully connected layers and convolutional layers) can be combined within a single neural network.

A convolutional neural network (CNN) can predict various molecular phenotypes based on DNA sequence alone. Applications include classifying transcription factor binding sites and predicting molecular phenotypes such as chromatin features, DNA contact maps, DNA methylation, gene expression, translation efficiency, RBP binding, and microRNA (miRNA) targets. . In addition to predicting molecular phenotypes from sequences, convolutional neural networks can be applied to many more technical tasks traditionally handled by hand-written bioinformatics pipelines. For example, convolutional neural networks can predict the specificity of guide RNAs, denoise ChIP-seq, improve Hi-C data resolution, predict experiments of origin from DNA sequences, and gene Mutants can be called.

Convolutional neural networks have also been employed to model long-range dependencies in the genome. Although interacting regulatory elements may be located distal on the unfolded linear DNA sequence, these elements are often proximal in the actual 3D chromatin conformation. Therefore, modeling molecular phenotypes from linear DNA sequences, despite being a rough approximation of chromatin, would be improved by allowing for long-range dependencies and allowing the model to implicitly learn aspects of 3D organization, such as promoter-enhancer looping. You can. This is achieved using dilated convolution with a receptive field of up to 32 kb. Extended convolution also allows splice sites to be predicted from the sequence using a 10 kb receptive field, thereby enabling integration of gene sequences across distances as long as a typical human intron (see Jaganathan, K. et al. Predicting splicing from primary sequence with deep learning. Cell 176, 535-548 (2019).

Different types of neural networks can be characterized by their parameter sharing schemes. For example, fully connected layers have no parameter sharing, while convolutional layers impose translation invariance by applying the same filter at every location of their input. Recurrent neural networks (RNNs) are an alternative to convolutional neural networks for processing sequential data, such as DNA sequences or time series, that implement different parameter sharing schemes. Recurrent neural networks apply the same operation to each sequence element. The operation takes as input a memory of previous sequence elements and a new input. It updates the memory and selectively emits output that is passed on to subsequent layers or used directly as model predictions. By applying the same model at each sequence element, the recurrent neural network is invariant to the positional index in the processed sequence. For example, recurrent neural networks can detect open reading frames in a DNA sequence regardless of their location in the sequence. This task requires recognition of a predetermined sequence of inputs, such as a start codon followed by an in-frame stop codon.

The main advantage of recurrent neural networks over convolutional neural networks is that they can theoretically pass information through infinitely long sequences in memory.

Additionally, recurrent neural networks can naturally process sequences of widely varying lengths, such as mRNA sequences. However, convolutional neural networks combined with various tricks (e.g., dilated convolutions) can reach similar or even better performance than recurrent neural networks for sequence modeling tasks such as audio synthesis and machine translation. Recurrent neural networks can aggregate the output of convolutional neural networks to predict single cell DNA methylation status, RBP binding, transcription factor binding, and DNA accessibility. Additionally, because recurrent neural networks apply sequential operations, they cannot be easily parallelized and are therefore much slower to compute than convolutional neural networks.

Each human has a unique genetic code, but most of the human genetic code is common to all humans. In some cases, the human genetic code may contain outliers, called genetic variants, which may be common among individuals in a relatively small group of human populations. For example, a particular human protein may contain a particular sequence of amino acids, while variants of that protein may differ by as much as one amino acid within that particular sequence that is otherwise identical.

Genetic variants can be pathogenic, leading to disease. Although most of such genetic variants have been depleted from the genome by natural selection, the ability to identify which genetic variants are likely to be pathogenic allows researchers to focus on these genetic variants to identify the corresponding diseases and their diagnosis, treatment, or cure. It can be helpful in gaining understanding. The clinical interpretation of millions of human genetic variants remains unclear. Some of the most frequent pathogenic variants are single nucleotide missense mutations that change amino acids in proteins. However, not all missense mutations are pathogenic.

Models that can predict molecular phenotypes directly from biological sequences can be used as in silico perturbation tools to probe associations between genetic and phenotypic variation, identify quantitative trait loci, and perform variant prioritization. It has emerged as a new method for ranking. This approach is critical given that most of the variants identified by genome-wide association studies of complex phenotypes are non-coding, making it difficult to estimate their effect and contribution to the phenotype. Additionally, linkage disequilibrium results in blocks of variants being co-inherited, which makes it difficult to pinpoint individual causal variants. Therefore, sequence-based deep learning models that can be used as interrogation tools to assess the impact of such variants provide a promising approach to find potential drivers of complex phenotypes. One example includes predicting the effect of short insertions or deletions (indels) and non-coding single nucleotide variants indirectly from differences between two variants in terms of transcription factor binding, chromatin accessibility, or gene expression prediction. Other examples include predicting the creation of novel splice sites from the sequence or quantitative effects of genetic variants on splicing.

An end-to-end deep learning approach for variant effect prediction is applied to predict pathogenicity of missense variants from sequence conservation data and protein sequences (Sundaram, L. et al., referred to herein as “PrimateAI”) al. Predicting the clinical impact of human mutations with deep neural networks. Nat. 50, 1161-1170 (2018). PrimateAI uses deep neural networks trained on variants of known pathogenicity by data augmentation using cross-species information. In particular, PrimateAI uses the sequences of wild-type and variant proteins to compare differences and determine the pathogenicity of mutations using trained deep neural networks. Such approaches utilizing protein sequences for pathogenicity prediction are promising because they can avoid circularity problems and overfitting to prior knowledge. However, compared to an adequate number of data to effectively train a deep neural network, the number of clinical data available in ClinVar is relatively small. To overcome this data shortage, PrimateAI uses common human variants and variants from primates as benign data, but simulated variants based on trinucleotide context were used as unlabeled data.

PrimateAI outperforms previous methods when trained directly on sequence alignment. PrimateAI learns key protein domains, conserved amino acid positions, and sequence dependencies directly from training data consisting of approximately 120,000 human samples. PrimateAI substantially exceeds the performance of other variant pathogenicity prediction tools in distinguishing benign and pathogenic de novo mutations in candidate developmental disorder genes and reproducing prior knowledge from ClinVar. These results suggest that PrimateAI is an important step forward for variant classification tools that can reduce the reliance of clinical reports on prior knowledge.

Central to protein biology is the understanding of how structural elements give rise to observed functions. The plethora of protein structural data enables the development of computational methods to systematically derive the rules governing structural-functional relationships. However, the performance of these methods is critically dependent on the choice of protein structure representation.

Protein sites are microenvironments within a protein structure that are distinguished by their structural or functional roles. A site can be defined by a three-dimensional (3D) location and local neighbors around that location where the structure or function resides. Central to rational protein engineering is the understanding of how the structural arrangement of amino acids creates functional properties within protein regions. Determination of the structural and functional roles of individual amino acids within a protein provides information to aid engineers and modify protein function. Identifying functionally or structurally important amino acids allows for focused engineering efforts, such as site-directed mutagenesis, to alter targeted protein functional properties. Alternatively, this knowledge can help avoid engineering designs that would defeat desired functionality.

Because it has been established that structure is much more conserved than sequence, the growing body of protein structural data provides the opportunity to systematically study the underlying patterns that govern structural-functional relationships using data-driven approaches. A fundamental aspect of any computational protein analysis is how protein structural information is expressed. The performance of a machine learning method often depends more on the choice of data representation than the machine learning algorithm employed. A good representation efficiently captures the most critical information, while a poor representation creates a noisy distribution without an underlying pattern.

The recent success of protein structural overabundance and deep learning algorithms provides an opportunity to develop tools to automatically extract task-specific representations of protein structures. Therefore, an opportunity arises to predict variant pathogenicity using multi-channel voxelized representations of 3D protein structures as input to deep neural networks.

In the drawings, like reference numbers generally refer to like parts throughout the different views. Additionally, the drawings are not necessarily drawn to scale, but instead focus generally on illustrating the principles of the disclosed technology. In the following description, various implementations of the disclosed technology are described with reference to the following drawings.
1 is a flow diagram illustrating the process of a system for determining pathogenicity of a variant in accordance with various embodiments of the disclosed technology.
2 schematically depicts an exemplary reference amino acid sequence of a protein and an alternative amino acid sequence of a protein according to one embodiment of the disclosed technology.
FIG. 3 shows the classification of amino acid atoms by amino acid in the reference amino acid sequence of FIG. 2 according to one embodiment of the disclosed technology.
FIG. 4 illustrates amino acid-specific properties of 3D atomic coordinates of alpha-carbon atoms classified in FIG. 3 based on amino acids according to one implementation of the disclosed technology.
Figure 5 schematically shows a process for determining a distance value for each voxel according to one implementation of the disclosed technology.
Figure 6 shows an example of a distance channel for each 21 amino acids according to one implementation of the disclosed technology.
7 is a schematic diagram of a distance channel tensor according to one implementation of the disclosed technology.
Figure 8 shows one-hot encoding of reference and alternative amino acids from Figure 2 according to one implementation of the disclosed technology.
Figure 9 is a schematic diagram of voxelized one-hot encoded reference amino acids and voxelized one-hot encoded variant/alternative amino acids according to one implementation of the disclosed technology.
FIG. 10 schematically illustrates a connection process for connecting the distance channel tensor and the reference allele tensor of FIG. 7 on a voxel-by-voxel basis according to one implementation of the disclosed technology.
FIG. 11 schematically illustrates a concatenation process for concatenating the distance channel tensor of FIG. 7, the reference allele tensor of FIG. 10, and the alternative allele tensor on a voxel-by-voxel basis, according to one implementation of the disclosed technology.
FIG. 12 is a flow diagram illustrating the process of a system for determining pan-amino acid conservation frequencies of nearest atoms and assigning them to voxels (voxelization) according to one implementation of the disclosed technology.
Figure 13 depicts voxel-to-nearest amino acid according to one implementation of the disclosed technology.
Figure 14 shows an exemplary multiple sequence alignment of reference amino acid sequences across 99 species according to one implementation of the disclosed technology.
Figure 15 shows an example of determining a pan-amino acid conservation frequency sequence for a specific voxel according to one implementation of the disclosed technology.
FIG. 16 illustrates each pan-amino acid conservation frequency determined for each voxel using the positional frequency logic described in FIG. 15 according to one implementation of the disclosed technology.
17 shows a voxelized per-voxel evolution profile according to one implementation of the disclosed technology.
18 shows an example of an evolutionary profile tensor according to one implementation of the disclosed technology.
FIG. 19 is a flow diagram illustrating the process of a system for determining the conservation frequency per amino acid of the nearest atom and assigning it to a voxel (voxelization) according to one implementation of the disclosed technology.
20 illustrates various examples of voxelized annotation channels associated with distance channel tensors according to one implementation of the disclosed technology.
21 illustrates various combinations and permutations of input channels that can serve as input to a pathogenicity classifier for determining pathogenicity of a target variant according to one implementation of the disclosed technology.
22 illustrates various methods of calculating the disclosed distance channel according to various implementations of the disclosed technology.
23 shows different examples of evolution channels according to various implementations of the disclosed technology.
24 illustrates different examples of annotation channels according to various implementations of the disclosed technology.
25 illustrates different examples of structured trust channels according to various implementations of the disclosed technology.
Figure 26 illustrates an example processing architecture of a pathogenicity classifier according to one implementation of the disclosed technology.
Figure 27 illustrates an example processing architecture of a pathogenicity classifier according to one implementation of the disclosed technology.
Figures 28, 29, 30, 31a, and 31b use PrimateAI as a benchmark model to demonstrate the classification superiority of the disclosed PrimateAI 3D compared to PrimateAI.
32A and 32B illustrate the disclosed efficient voxelization process according to various implementations of the disclosed technology.
Figure 33 illustrates how atoms are associated with voxels containing atoms according to one implementation of the disclosed technology.
FIG. 34 illustrates generating a voxel-to-atom mapping from an atom-to-voxel mapping to identify the closest atom on a voxel-by-voxel basis according to one implementation of the disclosed technology.
35A and 35B illustrate how without using the disclosed efficient voxelization, the disclosed efficient voxelization has a runtime complexity of O(#atoms) versus a runtime complexity of O(#atoms * #voxels).
Figure 36 depicts an example computer system that can be used to implement the disclosed techniques.
Figure 37 depicts one implementation of determining variant pathogenicity for a target alternate amino acid based on processing of gapped protein spatial representation.
Figure 38 shows an example of spatial representation of a protein.
Figure 39 shows an example of a gap space representation of the protein illustrated in Figure 38.
Figure 40 shows an example of an atomic space representation of the protein illustrated in Figure 38.
Figure 41 shows an example of a gap atom space representation of the protein described in Figure 38.
Figure 42 depicts one implementation of a pathogenicity classifier that determines variant pathogenicity for a target replacement amino acid based on processing of the gap protein spatial representation and the replacement amino acid representation of the target replacement amino acid.
Figure 43 shows one implementation of training data used to train a pathogenicity classifier.
Figure 44 depicts one implementation of generating a gap space representation for a reference protein sample using reference amino acids as gap amino acids.
Figure 45 depicts one implementation of training a pathogenicity classifier on benign protein samples.
Figure 46 shows one implementation of training a pathogenicity classifier on a pathogenic protein sample.
Figure 47 shows how certain unreachable amino acid classes are masked during training.
Figure 48 depicts one implementation of determining the final pathogenicity score.
Figure 49A shows that variant pathogenicity determinations were made for target replacement amino acids that fill the vacancy created by the reference gap amino acid at a given position in the protein.
Figure 49B shows that each variant pathogenicity determination was made for an amino acid of each amino acid class that fills the vacancy created by a reference gap amino acid at a given position in the protein.
Figure 50 depicts one embodiment of determining variant pathogenicity for multiple alternative amino acids based on processing of gap protein spatial representations.
Figure 51 depicts one implementation of a pathogenicity classifier that determines variant pathogenicity for multiple alternative amino acids based on processing of gap protein spatial representations.
Figure 52 shows one implementation of simultaneously training a pathogenicity classifier for benign and pathogenic protein samples.
Figure 53 depicts one embodiment of determining variant pathogenicity for multiple alternative amino acids based on processing gap protein spatial representations and generating evolutionary conservation scores for multiple alternative amino acids in response.
Figure 54 illustrates an evolutionarily conservative determinant in operation according to one implementation.
Figure 55 illustrates one implementation of determining pathogenicity based on predicted evolution scores.
Figure 56 shows one implementation of training data used to train an evolutionary conservation determinant.
Figure 57 depicts one implementation of simultaneously training evolutionary conservation determinants for benign and pathogenic protein samples.
Figure 58 shows various implementations of real label encoding used to train evolutionary conservation determinants.
Figure 59 shows an example location-specific frequency matrix (PSFM).
Figure 60 shows an example location-specific score matrix (PSSM).
Figure 61 shows one implementation of generating PSFM and PSSM.
Figure 62 shows example PSFM encoding.
Figure 63 shows example PSSM encoding.
Figure 64 illustrates two datasets on which the models disclosed herein can be trained.
65A and 65B illustrate one implementation of joint learning of the model disclosed herein.
Figures 66A and 66B illustrate one implementation of using transfer learning to train a model disclosed herein using the two datasets shown in Figure 64.
Figure 67 illustrates one implementation of generating training data and labels to train the model disclosed herein.
Figure 68 depicts one embodiment of a method for determining pathogenicity of nucleotide variants.
Figure 69 depicts one implementation of a system for predicting structural tolerance of amino acid substitutions.
Figures 70A, 70B and 70C show performance results demonstrating objective indicators of non-obviousness and creativity.

The following discussion is presented to enable any person skilled in the art to make and use the disclosed technology, and is presented in relation to specific applications and requirements thereof. Various modifications to the disclosed implementations will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other implementations and applications without departing from the spirit and scope of the disclosed technology. Accordingly, the disclosed techniques are not intended to be limited to the embodiments shown but are to be accorded the broadest scope consistent with the principles and features disclosed herein.

The detailed description of various implementations will be better understood when read in conjunction with the accompanying drawings. To the extent that the drawings show functional block diagrams of various implementations, the functional blocks do not necessarily represent divisions between hardware circuitry. Thus, for example, one or more of the functional blocks (e.g., a module, processor, or memory) may be a single piece of hardware (e.g., a general-purpose signal processor or a block of random access memory, a hard disk, etc.) or multiple pieces of hardware. It can be implemented in hardware. Similarly, a program may be a standalone program, integrated into the operating system as a subroutine, a function within an installed software package, etc. It will be understood that the various implementations are not limited to the arrangements and instrumentalities shown in the drawings.

The processing engine and database of the drawing, designated as modules, can be implemented in hardware or software and do not need to be divided into exactly identical blocks as shown in the drawing. Some of the modules may also be implemented on different processors, computers, or servers, or may be distributed among multiple different processors, computers, or servers. It will also be understood that some of the modules may be operated in parallel or in a different order, in combination with those shown in the figures, without affecting the functionality achieved. The modules within the figures can also be thought of as flowchart steps in the method. A module also does not necessarily require all of its code to be located contiguously in memory; Some parts of the code can be separated from other parts of the code, with code from other modules or other functions interspersed.

Protein structure-based pathogenicity determination

1 is a flow diagram depicting the system's process 100 for determining the pathogenicity of a variant. At step 102, the system's sequence accessor 104 accesses the reference and alternative amino acid sequences. At 112, the system's 3D structure generator 114 generates a 3D protein structure for a reference amino acid sequence. In some embodiments, the 3D protein structure is a homology model of a human protein. In one embodiment, the so-called SwissModel homology modeling pipeline provides a public repository of predicted human protein structures. In another embodiment, so-called HHpred homology modeling predicts the structure of the target protein from the template structure using a tool called a modeler.

Proteins are expressed as a collection of atoms and their coordinates in 3D space. Amino acids can have various atoms such as carbon atoms, oxygen (O) atoms, nitrogen (N) atoms, and hydrogen (H) atoms. Atoms can be further classified as side chain atoms and backbone atoms. The backbone carbon atoms may include alpha-carbon (C _α ) atoms and beta-carbon (C _β ) atoms.

In step 122, the system's coordinate classifier 124 sorts the 3D atomic coordinates of the 3D protein structure by amino acid. In one embodiment, classification by amino acid involves assigning 3D atomic coordinates to 21 amino acid categories (including stop or gap amino acid categories). In one example, an amino acid classification of alpha-carbon atoms could list each alpha-carbon atom under each of the 21 amino acid categories. In another example, an amino acid classification of beta-carbon atoms could list each beta-carbon atom under each of the 21 amino acid categories.

In another example, an amino acid classification of oxygen atoms could list each oxygen atom under each of the 21 amino acid categories. In another example, an amino acid classification of nitrogen atoms could list each nitrogen atom under each of the 21 amino acid categories. In another example, an amino acid classification of hydrogen atoms could list each hydrogen atom under each of the 21 amino acid categories.

Those skilled in the art will understand that, in various embodiments, classification by amino acid may include a subset of the 21 amino acid categories and a subset of different atomic elements.

At step 132, the system's voxel grid constructor 134 instantiates a voxel grid. The voxel grid can have any resolution, for example 3x3x3, 5x5x5, 7x7x7, etc. The voxels within the voxel grid may be of any size, for example 1 Angstrom (Å) on each side, 2 Å on each side, 3 Å on each side, etc. Those skilled in the art will understand that these exemplary dimensions refer to cubic dimensions because voxels are cubic. Additionally, those skilled in the art will understand that these example dimensions are non-limiting and that voxels may have any cubic dimension.

At step 142, the system's voxel grid centerer 144 centers the voxel grid on reference amino acids that experience target variants at the amino acid level. In one embodiment, the voxel grid is centered on the atomic coordinates of the particular atom of the reference amino acid experiencing the target variant, e.g., the 3D atomic coordinate of the alpha-carbon atom of the reference amino acid experiencing the target variant.

street channel

Voxels within a voxel grid may have multiple channels (or features). In one implementation, the voxels within the voxel grid have a plurality of distance channels (e.g., 21 distance channels each for 21 amino acid categories (including stop or gap amino acid categories)). At step 152, the system's distance channel generator 154 creates per-amino acid distance channels for voxels in the voxel grid. Distance channels are created independently for each of the 21 amino acid categories.

For example, consider the alanine (A) amino acid category. For example, further consider that the voxel grid is of size 3×3×3 and has 27 voxels. Then, in one implementation, the alanine distance channel contains 27 distance values for each of the 27 voxels in the voxel grid. The 27 distance values in the alanine distance channel are measured from the respective centroids of the 27 voxels in the voxel grid to their respective nearest atoms in the alanine amino acid category.

In one example, the alanine amino acid category contains only alpha-carbon atoms, so the nearest atom is that alanine alpha-carbon atom that is closest to each of the 27 voxels in the voxel grid. In another example, the alanine amino acid category contains only beta-carbon atoms, and therefore the nearest atom is that alanine beta-carbon atom that is closest to each of the 27 voxels in the voxel grid.

In another example, the alanine amino acid category contains only oxygen atoms, and therefore the nearest atom is that alanine oxygen atom that is closest to each of the 27 voxels in the voxel grid. In another example, the alanine amino acid category contains only nitrogen atoms, and therefore the nearest atom is that alanine nitrogen atom that is closest to each of the 27 voxels in the voxel grid. In another example, the alanine amino acid category contains only hydrogen atoms, and therefore the nearest atom is that alanine hydrogen atom that is closest to each of the 27 voxels in the voxel grid.

Like the alanine distance channel, distance channel generator 154 creates a distance channel (i.e., a set of voxel-wise distance values) for each of the remaining amino acid categories. In another implementation, distance channel generator 154 creates distance channels for only a subset of the 21 amino acid categories.

In other embodiments, the selection of the closest atom is not limited to a specific atom type. That is, within the target amino acid category, the closest atom to the specific voxel, the atomic element of the closest atom, and the distance value for the specific voxel calculated for inclusion in the distance channel for the target amino acid category are selected.

In another implementation, the distance channels are created atomically. Instead of or in addition to having distance channels for amino acid categories, distance values can be generated for atomic element categories regardless of the amino acid to which the atoms belong. For example, consider that the atoms of amino acids in a reference amino acid sequence span seven atomic elements: carbon, oxygen, nitrogen, hydrogen, calcium, iodine, and sulfur. The voxels within the voxel grid are then configured to have seven distance channels, and thus each of the seven distance channels has 27 voxel-specific distance values that specify the distance only to the nearest atom within the corresponding atomic element category. In another implementation, distance channels may be created for only a subset of 7 atomic elements. In another implementation, atomic element categories and distance channel creation can be further stratified by variations of the same atomic element, such as alpha-carbon (C _α ) atoms and beta-carbon (C _β ) atoms.

In another implementation, distance channels can be created on a per atom type basis, for example, a distance channel for only side chain atoms and a distance channel for only backbone atoms.

The nearest atom can be searched within a predefined maximum scan radius (e.g., 6 angstroms (Å)) from the voxel center. Additionally, multiple atoms may be closest to the same voxel within the voxel grid.

The distance is calculated between the 3D coordinates of the voxel center and the 3D atomic coordinates of the atoms. Additionally, the distance channel is created as a grid of voxels centered on the same location (e.g., centered on the 3D atomic coordinates of the alpha-carbon atom of the reference amino acid experiencing the target variant).

The distance may be a Euclidean distance. Additionally, the distance may be parameterized by atomic size (or atomic influence) (e.g., using the Lennard-Jones potential and/or Van der Waals atomic radius of that atom). Additionally, the distance value may be normalized by the maximum scan radius, or by the maximum observed distance value of the closest possible atom within the target amino acid category or target atom element category or target atom type category. In some implementations, the distance between a voxel and an atom is calculated based on the polar coordinates of the voxel and the atom. Polar coordinates are parameterized by the angle between voxels and atoms. In one implementation, this angular information is used to create an angular channel for a voxel (i.e., independent from the distance channel). In some implementations, the angle between the nearest atom and a neighboring atom (e.g., a backbone atom) can be used as a feature encoded into a voxel.

Reference allele and alternative allele channels

Voxels within a voxel grid may also have reference allele and alternative allele channels. At step 162, the system's one-hot encoder 164 generates a reference one-hot encoding of a reference amino acid in a reference amino acid sequence and an alternative one-hot encoding of an alternative amino acid in an alternative amino acid sequence. The reference amino acid experiences target variants. Alternative amino acids are target variants. The reference amino acid and the alternative amino acid are located at the same position in the reference amino acid sequence and the alternative amino acid sequence, respectively. The reference amino acid sequence and the alternative amino acid sequence have the same position-specific amino acid composition with one exception. Exceptions are positions that have a reference amino acid in the reference amino acid sequence and an alternative amino acid in the alternative amino acid sequence.

At step 172, the system's linker 174 connects the reference and alternative one-hot encodings with the amino acid-specific distance channels. In another implementation, connector 174 connects the atomic element-wise distance channels with the reference and alternative one-hot encodings. In another implementation, connector 174 connects distance channels by atom type with reference and alternative one-hot encodings.

At step 182, the system's runtime logic 184 processes the linked per-amino-acid/per-atom/atom-type distance channels and the reference and alternative one-hot encodings through a pathogenicity classifier (pathogenicity determination engine) to identify target variants. determines the pathogenicity of , which is ultimately inferred as determining the pathogenicity of the basic nucleotide variant that generates the target variant at the amino acid level. Pathogenicity classifiers are trained using labeled datasets of benign and pathogenic variants, for example using backpropagation algorithms. Additional details regarding labeled datasets of benign and pathogenic variants and exemplary architecture and training of pathogenicity classifiers can be found in commonly owned U.S. patent application Ser. No. 16/160,903; No. 16/160,986; No. 16/160,968; and 16/407,149.

2 schematically depicts a reference amino acid sequence 202 of protein 200 and an alternative amino acid sequence 212 of protein 200. Protein 200 contains N amino acids. The positions of amino acids in protein 200 are labeled 1, 2, 3...N. In the illustrated example, position 16 is the position that experiences amino acid variant 214 (mutation) caused by a base nucleotide variant. For example, for the reference amino acid sequence 202, position 1 has the reference amino acid phenylalanine (F), position 16 has the reference amino acid glycine (G) 204, and position N (e.g., of sequence 202) The last amino acid) has the reference amino acid leucine (L). Although not illustrated for clarity, the remaining positions within the reference amino acid sequence 202 contain various amino acids in an order specific to protein 200. The alternative amino acid sequence 212 is identical to the reference amino acid sequence 202 except for the variant at position 16 (214), which substitutes the reference amino acid glycine (G) (204) for the alternative amino acid alanine (A) (214). Contains.

3 shows an amino acid-by-amino acid classification of the atoms of amino acids in a reference amino acid sequence 202, also referred to herein as “atomic classification 300.” Of the 20 natural amino acids listed in column 302, certain types of amino acids may be repeated in proteins. That is, certain types of amino acids may occur more than once in a protein. Proteins may also have some undetermined amino acids that are categorized by the 21st stop or gap amino acid category. The right column of Figure 3 contains counts of alpha-carbon (C _α ) atoms from different amino acids.

Specifically, FIG. 3 shows an amino acid-by-amino acid classification of the alpha-carbon (C _α ) atoms of amino acids in the reference amino acid sequence 202. Column 308 in Figure 3 lists the total number of alpha-carbon atoms observed for the reference amino acid sequence 202 in each of the 21 amino acid categories. For example, column 308 lists the 11 alpha-carbon atoms observed for the alanine (A) amino acid category. Since each amino acid has only one alpha-carbon atom, this means that alanine occurs 11 times in the reference amino acid sequence 202. In another example, arginine (R) occurs 35 times in the reference amino acid sequence 202. The total number of alpha-carbon atoms across 21 amino acid categories is 828.

FIG. 4 shows amino acid-by-amino acid properties of the 3D atomic coordinates of the alpha-carbon atom of the reference amino acid sequence 202 based on the atomic classification 300 of FIG. 3 . This is referred to herein as “atomic coordinate bucketing (400).” In Figure 4, lists 404-440 tabulate the 3D atomic coordinates of alpha-carbon atoms bucketed into each of the 21 amino acid categories.

In the depicted implementation, bucketing 400 of Figure 4 follows classification 300 of Figure 3. For example, in Figure 3, the alanine amino acid category has 11 alpha-carbon atoms, and therefore in Figure 4, the alanine amino acid category has 11 3D atomic coordinates of the corresponding 11 alpha-carbon atoms from Figure 3. have This sorting-bucketing logic flows from Figure 3 to Figure 4 for other amino acid categories as well. However, this sorting-bucketing logic is for illustrative purposes only, and in other implementations, the disclosed technique does not need to perform sorting (300) and bucketing (400) to locate the closest atoms on a voxel-by-voxel basis. , fewer, additional, or different steps may be performed. For example, in some implementations, the disclosed technology may be used to determine classification criteria (e.g., by amino acid, by atomic element, by atom type), a predefined maximum scan radius, and types of distances (e.g., Euclidean, Mahalanobis, normalized, denormalized). ) may be used to locate the closest atom by voxel using a sorting and search algorithm that returns the closest atom by voxel from one or more databases in response to a search query configured to accept query parameters such as ). In various implementations of the disclosed technology, multiple sorting and search algorithms from current or future art can similarly be used by one skilled in the art to locate the nearest atom on a voxel-by-voxel basis.

In Figure 4, 3D atomic coordinates are represented by Cartesian coordinates x, y, z, but any type of coordinate system, such as spherical or cylindrical coordinates, may be used, and the claimed subject matter is not limited in this respect. In some embodiments, one or more databases may contain information regarding 3D atomic coordinates of alpha-carbon atoms and other atoms of amino acids in proteins. Such databases may be searchable by specific proteins.

As discussed above, voxels and voxel grids are 3D entities. However, for clarity, the figures show voxels and voxel grids in a two-dimensional (2D) format, and the description discusses this. For example, a 3x3x3 voxel grid of 27 voxels is shown and described herein as a 3x3 2D pixel grid with 9 2D pixels. Those skilled in the art will understand that the 2D format is used for representational purposes only and is intended to cover its 3D counterpart (i.e., 2D pixels represent 3D voxels, and 2D pixel grids represent 3D voxel grids). Additionally, the drawings are also not to scale. For example, a voxel of size 2 Angstroms (Å) is depicted using a single pixel.

Distance calculation per voxel

5 schematically illustrates the process of determining a voxel-wise distance value, also referred to herein as “voxel-wise distance calculation 500”. In the example shown, voxel-wise distance values are calculated only for the alanine (A) distance channel. However, the same distance calculation logic is run for each of the 21 amino acid categories, resulting in 21 amino acid-specific distance channels, as discussed above with respect to Figure 1, including the beta-carbon atom and other amino acids such as oxygen, nitrogen, and hydrogen. It can be further extended to other atom types such as atomic elements. In some implementations, atoms are randomly rotated before distance calculation to make training of the pathogenicity classifier invariant to atom orientation.

In Figure 5, the voxel grid 522 has indices (1, 1), (1, 2), (1, 3), (2, 1), (2, 2), (2, 3), (3, It has 9 voxels 514, identified as 1), (3, 2), and (3, 3). The voxel grid 522 is centered, for example, at the 3D atomic coordinates 532 of the alpha-carbon atom of the glycine (G) amino acid at position 16 in the reference amino acid sequence 202 because As discussed above in connection, in the alternative amino acid sequence 212, position 16 experiences a variant that mutates the glycine (G) amino acid to an alanine (A) amino acid. Additionally, the center of voxel grid 522 coincides with the center of voxel (2, 2).

The centered voxel grid 522 is used to calculate the distance per voxel for each of the 21 distance channels for each amino acid. For example, starting with the alanine (A) distance channel, the 3D coordinates of the centroids of each of the nine voxels 514 and the 11 The distance between the 3D atomic coordinates 402 of the alanine alpha-carbon atoms is measured. The nine distance values for the nine distances between the nine voxels 514 and their respective nearest alanine alpha-carbon atoms are then used to construct the alanine distance channel. The generated alanine distance channel arranges nine alanine distance values in the same order as the nine voxels 514 in the voxel grid 522.

The above process is run for each of the 21 amino acid categories. For example, the centered voxel grid 522 can be similarly used to calculate the arginine (R) distance channel, providing the 3D coordinates of the respective centroids of the nine voxels 514 and the 3D coordinates of the 35 arginine alpha-carbon atoms. Distances between atomic coordinates 404 are measured to locate the nearest arginine alpha-carbon atom for each of the nine voxels 514. The nine distance values for the nine distances between the nine voxels 514 and their respective nearest arginine alpha-carbon atoms are then used to construct the arginine distance channel. The created arginine distance channel arranges nine arginine distance values in the same order as the nine voxels 514 in the voxel grid 522. The 21 amino acid-specific distance channels are encoded on a voxel-by-voxel basis to form a distance channel tensor.

Specifically, in the illustrated example, distance 512 is between the center of voxel (1, 1) of voxel grid 522 and the nearest alpha-carbon ( _Cα ) atom in list 402, which is the Cα ^A5 atom. . Therefore, the value assigned to voxel (1, 1) is distance (512). In another example, the Cα ^A4 atom is the closest _Cα atom to the center of voxel (1, 2). Therefore, the value assigned to voxel (1, 2) is the distance between the center of voxel (1, 2) and the Cα ^A4 atom. In another example, the Cα ^A6 atom is the closest _Cα atom to the center of voxel (2, 1). Therefore, the value assigned to voxel (2, 1) is the distance between the center of voxel (2, 1) and the Cα ^A6 atom. In another example, the Cα ^A6 atom is also the closest _Cα atom to the centers of voxels (3, 2) and (3, 3). Therefore, the value assigned to voxel (3, 2) is the distance between the center of voxel (3, 2) and Cα ^A6 , and the value assigned to voxel (3, 3) is the distance between the center of voxel (3, 3) and Cα. ^A6 is the distance between atoms. In some implementations, the distance value assigned to voxel 514 may be a normalized distance. For example, the distance value assigned to voxel (1, 1) may be distance 512 divided by maximum distance 502 (predefined maximum scan radius). In some implementations, the nearest atom distance may be the Euclidean distance and the nearest atom distance may be normalized by dividing the Euclidean distance by the maximum nearest atom distance (e.g., maximum distance 502).

As described above, for amino acids with alpha-carbon atoms, the distance may be the distance of the nearest alpha-carbon atom from the center of the corresponding voxel to the nearest alpha-carbon atom of the corresponding amino acid. Additionally, for amino acids with beta-carbon atoms, the distance may be the nearest beta-carbon atom distance from the center of the corresponding voxel to the nearest beta-carbon atom of the corresponding amino acid. Similarly, for amino acids with backbone atoms, the distance may be the nearest backbone atom distance from the center of the corresponding voxel to the nearest backbone atom of the corresponding amino acid. Similarly, for amino acids with side chain atoms, the distance may be the distance of the nearest side chain atom from the center of the corresponding voxel to the nearest side chain atom of the corresponding amino acid. In some embodiments, the distance may additionally/alternatively include the distance to the second, third, fourth nearest atom, etc.

Distance channels for each amino acid

Figure 6 shows an example of a distance channel 600 for each of 21 amino acids. Each row in FIG. 6 corresponds to one of the 21 amino acid-specific distance channels 602 to 642. The distance channel for each amino acid includes a distance value for each voxel 514 of the voxel grid 522. For example, the amino acid-specific distance channel 602 for alanine (A) includes distance values for each of the voxels 514 of the voxel grid 522. As mentioned above, voxel grid 522 is a 3D grid with a volume of 3x3x3 and contains 27 voxels. Likewise, although Figure 6 shows voxels 514 in two dimensions (e.g., 9 voxels in a 3×3 grid), each amino acid-wise distance channel represents the 27 voxel-wise distances for a 3×3×3 voxel grid. Can contain values.

Directional encoding

In some embodiments, the disclosed techniques use an orientation parameter to specify the orientation of a reference amino acid within the reference amino acid sequence 202. In some embodiments, the disclosed techniques use orientation parameters to specify the orientation of alternative amino acids within alternative amino acid sequence 212. In some embodiments, the disclosed technology uses orientation parameters to specify the location of a protein 200 that experiences a target variant at the amino acid level.

As discussed above, all distance values in the 21 amino acid distance channels 602-642 are measured from their respective nearest atoms to voxel 514 in voxel grid 522. This closest atom is from one of the reference amino acids in the reference amino acid sequence 202. These derived reference amino acids containing the closest atom can be classified into two categories: (1) those derived reference amino acids that precede the variant empirical reference amino acid 204 within the reference amino acid sequence 202 and (2) Variant experience within the reference amino acid sequence 202 A reference amino acid that follows the reference amino acid 204 from which it is derived. Reference amino acids derived within the first category may be referred to as preceding reference amino acids. The resulting reference amino acids within the second category may be referred to as trailing reference amino acids.

Directionality parameters are applied to those distance values in the 21 amino acid-specific distance channels 602 to 642 measured from those nearest atoms originating from the preceding reference amino acid. In one implementation, the directionality parameter is multiplied by that distance value. The directionality parameter can be any number, such as -1.

As a result of the application of the directionality parameter, the 21 amino acid-specific distance channels 600 contain some distance values that indicate to the pathogenicity classifier which end of the protein 200 is the starting end and which end is the terminal end. This also allows the pathogenicity classifier to reconstruct protein sequences from the 3D protein structure information supplied by the distance channel and the reference and allele channels.

distance channel tensor

7 is a schematic diagram of the distance channel tensor 700. The distance channel tensor 700 is a voxelized representation of the per-amino acid distance channel 600 from Figure 6. In the distance channel tensor 700, 21 amino acid-specific distance channels 602 to 642 are connected for each voxel, like the RGB channels of a color image. The voxelized dimensionality of the distance channel tensor 700 is 21×3×3×3 (21 represents 21 amino acid categories, and 3×3×3 represents a 3D voxel grid with 27 voxels). ; Figure 7 is a 2D depiction of dimensions 21 x 3 x 3.

One-hot encoding

Figure 8 shows one-hot encoding 800 of a reference amino acid 204 and an alternative amino acid 214. In Figure 8, the left column is a one-hot encoding (802) of the reference amino acid glycine (G) (204), where 1 is for the glycine amino acid category and 0 is for all other amino acid categories. In Figure 8, the right column is the one-hot encoding (804) of the variant/alternative amino acid alanine (A) (214), where 1 is for the alanine amino acid category and 0 is for all other amino acid categories.

Figure 9 is a schematic diagram of a voxelized one-hot encoded reference amino acid (902) and a voxelized one-hot encoded variant/alternative amino acid (912). The voxelized one-hot encoded reference amino acid 902 is a voxelized representation of the one-hot encoding 802 of the reference amino acid glycine (G) 204 from Figure 8. The voxelized one-hot encoded alternative amino acid 912 is a voxelized representation of the one-hot encoded 804 of the variant/alternative amino acid alanine (A) 214 from Figure 8. The voxelized dimensionality of the voxelized one-hot encoded reference amino acids 902 is 21×1×1×1 (21 represents 21 amino acid categories); Figure 9 is a 2D depiction with dimensions 21 x 1 x 1. Similarly, the voxelized dimensionality of the voxelized one-hot encoded alternative amino acids 912 is 21 × 1 × 1 × 1 (21 indicates 21 amino acid categories); Figure 9 is a 2D depiction with dimensions 21 x 1 x 1.

Reference allele tensor

FIG. 10 schematically illustrates a concatenation process 1000 that concatenates the distance channel tensor 700 and the reference allele tensor 1004 of FIG. 7 on a voxel-by-voxel basis. The reference allele tensor 1004 is a voxel-wise aggregation (repeat/cloning/replication) of the voxelized one-hot encoded reference amino acids 902 from Figure 9. That is, multiple copies of the voxelized one-hot encoded reference amino acid 902 are linked to each other voxel-by-voxel according to the spatial arrangement of the voxels 514 within the voxel grid 522, and thus the reference allele tensor 1004 ) has a corresponding copy of the voxelized one-hot encoded reference amino acid 910 for each of the voxels 514 in the voxel grid 522.

The concatenation process 1000 creates a concatenated tensor 1010. The voxelized dimensionality of the reference allele tensor (1004) is 21 ); Figure 10 is a 2D depiction of the reference allele tensor 1004 with dimensions 21 x 3 x 3. The voxelized dimensionality of the connected tensor 1010 is 42×3×3×3; Figure 10 is a 2D depiction of a connected tensor 1010 with dimensions 42x3x3.

Alternative allele tensor

FIG. 11 schematically illustrates a concatenation process 1100 that concatenates the distance channel tensor 700 of FIG. 7, the reference allele tensor 1004 of FIG. 10, and the alternative allele tensor 1104 on a voxel-by-voxel basis. The alternative allele tensor 1104 is a voxel-wise aggregation (repeat/cloning/replication) of the voxelized one-hot encoded alternative amino acids 912 from Figure 9. That is, multiple copies of the voxelized one-hot encoded alternative amino acids 912 are linked to each other voxel-by-voxel according to the spatial arrangement of the voxels 514 within the voxel grid 522, and thus the alternative allele tensor 1104 ) has a corresponding copy of the voxelized one-hot encoded alternative amino acid 910 for each of the voxels 514 within the voxel grid 522.

The concatenation process 1100 creates a concatenated tensor 1110. The voxelized dimensionality of the alternative allele tensor 1104 is 21 ); Figure 11 is a 2D depiction of the alternative allele tensor 1104 with dimensions 21 x 3 x 3. The voxelized dimensionality of the connected tensor 1110 is 63×3×3×3; Figure 11 is a 2D depiction of a connected tensor 1110 with dimensions 63x3x3.

In some embodiments, the runtime logic 184 processes the concatenated tensor 1110 through a pathogenicity classifier to determine the pathogenicity of the variant/alternative amino acid alanine (A) 214, which in turn determines the pathogenicity of the variant/alternative amino acid alanine (A ) (214).

evolution conservation channel

Predicting the functional consequences of a variant depends, at least in part, on whether important amino acids for the protein family have been conserved through evolution due to negative selection (i.e., amino acid changes at these sites have been deleterious in the past) and whether mutations at these sites have been It relies on the assumption that it increases the likelihood of pathogenicity (causing disease) in humans. Typically, homologous sequences of a target protein are collected and aligned, and a metric of conservation is calculated based on the weighted frequency of different amino acids observed at the target position within the alignment.

Accordingly, the disclosed technique connects the distance channel tensor 700, the reference allele tensor 1004, and the alternative allele tensor 1004 with evolutionary channels. One example of an evolutionary channel is pan-amino acid conservation frequency. Another example of an evolutionary channel is the frequency of conservation per amino acid.

In some implementations, the evolution channel is constructed using a position weight matrix (PWM). In another implementation, the evolving channel is constructed using a position specific frequency matrix (PSFM). In another implementation, evolution channels are constructed using computational tools such as SIFT, PolyPhen, and PANTHER-PSEC. In another implementation, the evolutionary channel is a conservation channel based on evolutionary preservation. Conservation is related to conservation because it also reflects the effect of negative selection that has acted to prevent evolutionary change at a given site within a protein.

Pan-amino acid evolutionary profile

FIG. 12 is a flow diagram illustrating a process 1200 of a system for determining pan-amino acid conservation frequencies of nearest atoms and assigning them to voxels (voxelization), according to one implementation of the disclosed technology. 12, 13, 14, 15, 16, 17, and 18 are discussed simultaneously.

In step 1202, the system's similar sequence finder 1204 retrieves an amino acid sequence that is similar (homologous) to the reference amino acid sequence 202. Similar amino acid sequences can be selected from multiple species such as primates, mammals, and vertebrates.

In step 1212, the system's aligner 1214 aligns the reference amino acid sequence 202 by position with similar amino acid sequences, i.e., aligner 1214 performs a multiple sequence alignment. Figure 14 shows an exemplary multiple sequence alignment (1400) of reference amino acid sequences (202) across 99 species. In some implementations, multiple sequence alignment 1400 can be, for example, a first position frequency matrix for primates 1402, a second position frequency matrix for mammals 1412, and a third position frequency matrix for primates. It can be split to produce (1422). In another implementation, a single position frequency matrix is generated across 99 species.

In step 1222, the system's pan-amino acid conservation frequency calculator 1224 uses multiple sequence alignment to determine the pan-amino acid conservation frequency of reference amino acids within the reference amino acid sequence 202.

At step 1232, the system's closest atom finder 1234 finds the closest atom for voxel 514 in voxel grid 522. In some implementations, the search for the closest atom on a voxel-by-voxel basis may not be limited to any particular amino acid category or atom type. That is, the closest atoms per voxel can be selected across amino acid categories and amino acid types, as long as they are the closest atoms to the centroid of each voxel. In other embodiments, a search for the closest atom on a voxel-by-voxel basis may be performed only by specific atom categories, such as only certain atomic elements, such as oxygen, nitrogen, and hydrogen, or only alpha-carbon atoms, or only beta-carbon atoms, or only branched chain atoms. It may be limited to only, or to only backbone atoms.

At step 1242, the system's amino acid selector 1244 selects those reference amino acids within the reference amino acid sequence 202 that contain the closest atoms identified at step 1232. Such reference amino acids may be called closest reference amino acids. 13 shows the nearest reference, which locates the closest atom 1302 to voxel 514 in voxel grid 522 and contains the closest atom 1302 to voxel 514 in voxel grid 522. An example of mapping each amino acid 1312 is shown. This is identified as “Voxel-to-nearest amino acid mapping 1300” in Figure 13.

In step 1252, the system's voxelizer 1254 voxelizes the pan-amino acid conservation frequency of the closest reference amino acid. Figure 15 shows an example of determining the pan-amino acid conservation frequency sequence for the first voxel (1, 1) in the voxel grid 522, also referred to herein as “per-voxel evolutionary profile determination 1500.”

Referring to Figure 13, the closest reference amino acid that was mapped to the first voxel (1, 1) is the aspartic acid (D) amino acid at position 15 in the reference amino acid sequence 202. A multiple sequence alignment of the reference amino acid sequence 202 with 99 homologous amino acid sequences from, for example, 99 species is then analyzed at position 15. Such position-specific and interspecies analyzes determine how many instances of an amino acid are found from each of the 21 amino acid categories at positions 15 across 100 aligned amino acid sequences (i.e., the reference amino acid sequence (202) plus 99 homologous amino acid sequences). indicates.

In the example shown in Figure 15, the aspartic acid (D) amino acid is found at position 15 in 96 of the 100 aligned amino acid sequences. Therefore, the aspartic acid amino acid category (1504) is assigned a pan-amino acid conservation frequency of 0.96. Similarly, in the example shown, the valine(V) acid amino acid is found at position 15 in four of the 100 aligned amino acid sequences. Therefore, the valine acid amino acid category (1514) is assigned a pan-amino acid conservation frequency of 0.04. Since no instances of amino acids from other amino acid categories are detected at position 15, the remaining amino acid categories are assigned a pan-amino acid conservation frequency of 0. In this way, each of the 21 amino acid categories is assigned a respective pan-amino acid conservation frequency, which can be encoded in the pan-amino acid conservation frequency sequence 1502 for the first voxel (1, 1).

FIG. 16 shows the respective Voxels 514 in the voxel grid 522 using the position frequency logic described in FIG. 15 , also referred to herein as “voxel-evolution profile mapping 1600.” Pan-amino acid conservation frequencies (1612 to 1692) are shown.

The per-voxel evolution profile 1602 is then used by the voxelizer 1254 to generate the voxelized per-voxel evolution profile 1700 shown in FIG. 17 . Often, each of the voxels 514 within the voxel grid 522 has a different pan-amino acid conservation frequency sequence and therefore a different voxelized per-voxel evolution profile because the voxels have different closest atoms and, therefore, different This is because it is regularly mapped to the nearest reference amino acid. Of course, when two or more voxels have the same nearest atom and thereby the same closest reference amino acid, the same pan-amino acid conservation frequency sequence and the same voxelized per-voxel evolution profile are assigned to each of the two or more voxels.

FIG. 18 shows an example of an evolution profile tensor 1800 in which voxelized voxel-wise evolution profiles 1700 are connected to each other voxel-wise according to the spatial arrangement of voxels 514 in the voxel grid 522. The voxelized dimensionality of the evolutionary profile tensor 1800 is 21×3×3×3 (21 represents 21 amino acid categories, and 3×3×3 represents a 3D voxel grid with 27 voxels). ; Figure 18 is a 2D depiction of an evolutionary profile tensor 1800 with dimensions 21 x 3 x 3.

At step 1262, concatenator 174 concatenates evolution profile tensor 1800 with distance channel tensor 700 on a voxel-by-voxel basis. In some implementations, the evolutionary profile tensor 1800 is concatenated voxel-by-voxel with the connector tensor 1110, creating an additional concatenated tensor (not shown) of dimension 84×3×3×3.

At step 1272, the runtime logic 184 processes the additional concatenated tensor of dimensions 84×3×3×3 through a pathogenicity classifier to determine the pathogenicity of the target variant, which in turn generates the target variant at the amino acid level. is inferred as a pathogenic determinant of the basic nucleotide variant.

Evolutionary profile per amino acid

FIG. 19 is a flow diagram illustrating the system's process 1900 for determining the conservation frequency per amino acid of the nearest atom and assigning it to a voxel (voxelization). In Figure 19, steps 1202 and 1212 are the same as in Figure 12.

In step 1922, the system's conservation frequency per amino acid calculator 1924 uses multiple sequence alignments to determine the conservation frequency per amino acid of reference amino acids within the reference amino acid sequence 202.

At step 1932, the system's nearest atom finder 1934 finds, for each voxel 514 in the voxel grid 522, the 21 closest atoms across each of the 21 amino acid categories. Each of the 21 closest atoms are different from each other because they are selected from different amino acid categories. This leads to the selection of the 21 unique closest reference amino acids for a specific voxel, which in turn leads to the creation of a 21 unique position frequency matrix for a specific voxel, which in turn leads to the creation of a 21 unique position frequency matrix per 21 unique amino acids for a specific voxel. This leads to a decision on the frequency of preservation.

At step 1942, the system's amino acid selector 1944 determines, for each voxel 514 within the voxel grid 522, a reference amino acid sequence 202 containing the 21 closest atoms identified at step 1932. Select 21 reference amino acids. Such reference amino acids may be called closest reference amino acids.

At step 1952, the voxelizer 1954 of the system voxelizes the conservation frequencies per amino acid of the 21 closest reference amino acids identified for a particular voxel at step 1942. The 21 closest reference amino acids are necessarily located at 21 different positions within the reference amino acid sequence 202 because they correspond to different base nearest atoms. Therefore, for a particular voxel, a 21 position frequency matrix can be generated for the 21 closest reference amino acids. A 21-position frequency matrix can be generated across multiple species where the homologous amino acid sequences are aligned by position with the reference amino acid sequence 202, as discussed above with respect to FIGS. 12-15.

Then, using the 21 position frequency matrix, 21 position specific conservation scores can be calculated for the 21 closest reference amino acids identified for a particular voxel. These 21 position-specific conservation scores represent the pan-amino acid conservation frequency sequence 1502 for a particular voxel, similar to the pan-amino acid conservation frequency sequence 1502 in Figure 12, except that sequence 1502 has many zero entries. While shaping the frequency of preservation; Conservation Frequency Per Amino Acid Each element (feature) in the sequence has a constant value (e.g., a floating point number) because the 21 closest reference amino acids across the 21 amino acid categories necessarily have different position frequency matrices and thus This is because they have different positions, which yields a conservation frequency per different amino acid.

The above process is run for each of the voxels 514 within the voxel grid 522, and the resulting per-voxel amino acid conservation frequencies are useful for pathogenicity determinations, similar to the pan-amino acid conservation frequencies discussed in relation to Figures 12-18. voxelized, tensorized, concatenated, and processed.

annotation channel

20 shows various examples of voxelized annotation channels 2000 coupled with distance channel tensors 700. In some embodiments, the voxelized annotation channel provides one-hot indicators for different protein annotations, e.g., whether an amino acid (residue) is part of a transmembrane region, signal peptide, active site, or any other binding site. or whether the residue is subject to posttranslational modification, PathRatio (see Pei P, Zhang A: A Topological Measurement for Weighted Protein Interaction Network. CSB 2005, 268-278.), etc. Additional examples of annotation channels can be found in the specific implementation sections below and in the claims.

The voxelized annotation channels are arranged voxel-wise (e.g., annotation channels (2002, 2004, 2006)), or voxel-wise, such that a voxel can have the same annotation sequence as the voxelized reference allele and alternative allele sequences. may have their own annotation sequence, such as a voxelized per-voxel evolution profile 1700 (e.g., annotation channels 2012, 2014, 2016 (as shown in different colors)).

Annotation channels are voxelized, tensorized, concatenated, and processed for pathogenicity determination similar to the pan-amino acid conservation frequencies discussed in conjunction with Figures 12-18.

Structural Reliability Channel

The disclosed technique can also associate various voxelized structural reliability channels with the distance channel tensor 700. Some examples of structural reliability channels include: GMQE score (provided by SwissModel); B-factor; The temperature factor column of the homology model (indicating how well residues satisfy (physical) constraints in the protein structure); A normalized number of alignments of the template protein to the residue closest to the center of the voxel (alignment provided by HHpred, e.g., a voxel is aligned among the six template structures, with three alignments indicating that the feature has the value 3/6=0.5). closest to the residue in the template structure); minimum, maximum and average TM scores; and the predicted TM score of the template protein structure that aligns to the residue closest to the voxel (continuing the example above, assuming the three template structures have TM scores 0.5, 0.5 and 1.5, the minimum is 0.5 and the average is 2 /3, maximum is 1.5). TM scores can be provided for each protein template by HHpred. Additional examples of structural reliability channels can be found in the Specific Implementations section below and in the claims.

The voxelized structural confidence channels are arranged voxel-wise such that voxels can have the same structural confidence sequence, such as voxelized reference allele and alternative allele sequences, or voxels can have identical structural confidence sequences, such as voxelized per-voxel evolutionary profiles (1700). Each can have its own structural reliability hierarchy.

Structural confidence channels are voxelized, tensorized, concatenated, and processed for pathogenicity determination similar to the pan-amino acid conservation frequencies discussed in conjunction with Figures 12-18.

pathogenicity classifier

Figure 21 shows different combinations and permutations of input channels that can be provided as inputs 2102 to a pathogenicity classifier 2108 for pathogenicity determination 2106 of the target variant. One of the inputs 2102 may be a distance channel 2104 created by a distance channel generator 2272. Figure 22 shows a different method of calculating the distance channel 2104. In one implementation, the distance channel 2104 is generated based on the distance 2202 between voxel centers and atoms across a plurality of atomic elements, regardless of amino acid. In some implementations, distance 2202 is normalized by the maximum scan radius to produce normalized distance 2202a. In another implementation, the distance channel 2104 is created based on the distance 2212 between the voxel center and the alpha-carbon atom in amino acids. In some implementations, distance 2212 is normalized by the maximum scan radius to produce normalized distance 2212a. In another implementation, the distance channel 2104 is created based on the distance 2222 between the voxel center and the beta-carbon atom in amino acids. In some implementations, distance 2222 is normalized by the maximum scan radius to produce normalized distance 2222a. In another implementation, the distance channel 2104 is generated based on the distance 2232 between the voxel center and the side chain atoms in amino acids. In some implementations, distance 2232 is normalized by the maximum scan radius to produce normalized distance 2232a. In another implementation, the distance channel 2104 is generated based on the distance 2242 between the voxel centroid and the backbone atoms in amino acids. In some implementations, distance 2242 is normalized by the maximum scan radius to produce normalized distance 2242a. In another implementation, the distance channel 2104 is generated based on the distance 2252 (one feature) between the voxel center and its respective nearest atom, regardless of atom type and amino acid type. In another implementation, the distance channel 2104 is generated based on the distance between the voxel centroid and the atom from the non-standard amino acid 2262 (one feature). In some implementations, the distance between a voxel and an atom is calculated based on the polar coordinates of the voxel and the atom. Polar coordinates are parameterized by the angle between voxels and atoms. In one implementation, this angular information is used to create an angular channel for a voxel (i.e., independent from the distance channel). In some implementations, the angle between the nearest atom and a neighboring atom (e.g., a backbone atom) can be used as a feature encoded into a voxel.

Another of the inputs 2102 may be a feature 2114 representing missing atoms within a specified radius.

Another of the inputs 2102 may be a one-hot encoding 2124 of a reference amino acid. Another of the inputs 2102 may be a one-hot encoding 2134 of variant/alternative amino acids.

Another of the inputs 2102 may be the evolution channel 2144 created by the evolution profile generator 2372, shown in FIG. 23. In one implementation, evolution channels 2144 can be created based on pan-amino acid conservation frequencies 2302. In another implementation, evolution channels 2144 can be created based on pan-amino acid conservation frequencies 2312.

Another of the inputs 2102 may be a feature 2154 representing a missing residue or a missing evolutionary profile.

Another of the inputs 2102 may be annotation channel 2164 created by annotation generator 2472, shown in FIG. 24. In one implementation, annotation channel 2154 can be generated based on molecular processing annotation 2402. In another implementation, annotation channel 2154 may be created based on region annotation 2412. In another implementation, annotation channel 2154 may be created based on site annotation 2422. In another implementation, annotation channel 2154 can be generated based on amino acid modification annotation 2432. In another implementation, annotation channel 2154 may be generated based on secondary structure annotation 2442. In another implementation, annotation channel 2154 may be created based on laboratory information annotation 2452.

Another of the inputs 2102 may be the structural reliability channel 2174 generated by the structural reliability generator 2572, shown in FIG. 25. In one implementation, structural confidence 2174 may be generated based on global model quality estimation (GMQE) 2502. In another implementation, structural confidence 2174 may be generated based on a qualitative model energy analysis (QMEAN) score 2512. In another implementation, structural reliability 2174 may be generated based on temperature factor 2522. In another implementation, structural confidence 2174 may be generated based on template modeling score 2542. Examples of mold modeling scores 2542 include minimum mold modeling scores 2542a, average mold modeling scores 2542b, and maximum mold modeling scores 2542c.

Those skilled in the art will understand that any permutation and combination of input channels may lead to input for processing through a pathogenicity classifier 2108 to determine the pathogenicity of the target variant 2106. In some implementations, only a subset of input channels may be connected. Input channels can be connected in any order. In one implementation, input channels can be concatenated into a single tensor by a tensor generator (input encoder) 2110. This single tensor can then be provided as input to a pathogenicity classifier 2108 for determining the pathogenicity of the target variant 2106.

In one implementation, pathogenicity classifier 2108 uses a convolutional neural network (CNN) with multiple convolutional layers. In other implementations, the pathogenicity classifier 2108 includes a long short-term memory network (LSTM), a bi-directional LSTM (Bi-LSTM), and a gated recurrent unit (GRU). ) uses a recurrent neural network (RNN) such as In another implementation, pathogenicity classifier 2108 uses both CNNs and RNNs. In another implementation, pathogenicity classifier 2108 uses a graph convolutional neural network that models dependencies in graph structured data. In another implementation, pathogenicity classifier 2108 uses a variational autoencoder (VAE). In another implementation, pathogenicity classifier 2108 uses a generative adversarial network (GAN). In another implementation, pathogenicity classifier 2108 may also be a language model based on self-attention, such as implemented by Transformer and BERT, for example.

In another implementation, the pathogenicity classifier 2108 can be classified into 1D convolution, 2D convolution, 3D convolution, 4D convolution, 5D convolution, dilated or atrous convolution, transposed convolution, separation by depth. You can use possible convolutions, point-wise convolutions, 1×1 convolutions, group convolutions, flat convolutions, spatial and cross-channel convolutions, shuffle grouped convolutions, spatial separable convolutions, and deconvolutions. . It includes one or more loss functions, such as logistic regression/log loss, multi-class cross-entropy/softmax loss, binary cross-entropy loss, Mean-squared error loss, L1 loss, L2 loss, smooth L1 loss, and Huber loss can be used. It supports arbitrary parallelism, efficiency, and compression schemes, such as TFRecords, compressed encoding (e.g., PNG), sharding, parallel calls for map transformations, batching, prefetching, model parallelism, data parallelism, and synchronous/asynchronous stochastic gradient descent. You can use the law (SGD). It includes an upsampling layer, a downsampling layer, recurrent connections, gates and gated memory units (e.g. LSTM or GRU), residual blocks, residual connections, highway connections, skip connections, peephole connections, activation functions (e.g. rectification) nonlinear transform functions such as linear unit (ReLU), leaky ReLU, exponential liner unit (ELU), sigmoid and hyperbolic tangent (tanh), batch normalization layer, regularization layer, dropout, pooling layer. (e.g., maximum or average pooling), a global average pooling layer, a damping mechanism, and a Gaussian error linear unit.

The pathogenicity classifier 2108 is trained using a backpropagation-based gradient update technique. Exemplary gradient descent techniques that can be used to train pathogenicity classifier 2108 include stochastic gradient descent, batch gradient descent, and mini-batch gradient descent. Some examples of gradient descent optimization algorithms that can be used to train a pathogenicity classifier 2108 are Momentum, Nesterov accelerated gradient, Adagrad, Adadelta, RMSprop, Adam, AdaMax, Nadam, and AMSGrad. In other implementations, pathogenicity classifier 2108 can be trained by unsupervised learning, semi-supervised learning, self-learning, reinforcement learning, multi-task learning, multi-modal learning, transfer learning, knowledge distillation, etc.

Figure 26 shows an example processing architecture 2600 of pathogenicity classifier 2108, according to one implementation of the disclosed technology. Processing architecture 2600 includes a cascade of processing modules 2606, 2610, 2614, 2618, 2622, 2626, 2630, 2634, 2638, and 2642, each of which performs a 1D convolution (1×1×1 CONV); May include 3D convolution (3×3×3 CONV), ReLU nonlinearity, and batch normalization (BN). Other examples of processing modules include a perfect contact (FC) layer, a dropout layer, a smoothing layer, and a final softmax layer that generates exponentially normalized scores for target variants belonging to the benign and pathogenic classes. In Figure 26, "64" indicates the number of convolutional filters applied by a particular processing module. In Figure 26, the size of the input voxel 2602 is 15×15×15×8. 26 also shows the respective volumetric dimensions of intermediate inputs 2604, 2608, 2612, 2616, 2620, 2624, 2628, 2632, 2636, 2640 generated by processing architecture 2600.

Figure 27 shows an example processing architecture 2700 of pathogenicity classifier 2108, according to one implementation of the disclosed technology. The processing architecture 2700 includes processing modules 2708, 2714, 2720, 2726, 2732, 2738, 2744, such as 1D convolution (CONV 1D), 3D convolution (CONV 3D), ReLU nonlinearity, and batch normalization (BN). , 2750, 2756, 2762, 2768, 2774, 2780). Other examples of processing modules include a fully connected (dense) layer, a dropout layer, a smoothing layer, and a final softmax layer that generates exponentially normalized scores for target variants belonging to the benign and pathogenic classes. In Figure 27, “64” and “32” indicate the number of convolutional filters applied by a particular processing module. In Figure 27, the size of the input voxel 2704 supplied by the input layer 2702 is 7x7x7x108. 27 also shows intermediate inputs 2710, 2716, 2722, 2728, 2734, 2740, 2746, 2752, 2758, 2764, 2770, 2776, 2782 and intermediate outputs 2706 generated by processing architecture 2700. , 2712, 2718, 2724, 2730, 2736, 2742, 2748, 2754, 2760, 2766, 2772, 2778, 2784).

Those skilled in the art will understand that other current and future artificial intelligence, machine learning, and deep learning models, datasets, and training techniques may be incorporated into the disclosed variant pathogenicity classifier without departing from the spirit of the disclosed technology.

Performance outcomes as objective indicators of originality and non-obviousness.

The variant pathogenicity classifier disclosed herein makes pathogenicity predictions based on 3D protein structure and is referred to as “PrimateAI 3D”. “Primate AI” is a commonly owned and previously disclosed variant pathogenicity classifier that generates pathogenicity prediction-based protein sequences. For more information about PrimateAI, see commonly owned U.S. patent application Ser. No. 16/160,903; No. 16/160,986; No. 16/160,968; and 16/407,149 and Sundaram, L. et al. Predicting the clinical impact of human mutation with deep neural networks. Nat. Genet. 50, 1161-1170 (2018)].

Figures 28, 29, 30, and 31a use PrimateAI as a benchmark model to demonstrate the classification superiority of PrimateAI 3D over PrimateAI. The performance results in Figures 28, 29, 30, 31A and 31B are generated from a classification task that accurately distinguishes benign variants from pathogenic variants across multiple validation sets. PrimateAI 3D is trained on a training set that is different from the multiple validation sets. PrimateAI 3D is trained on common human variants and variants from primates, which are used as benign datasets, but simulated variants based on trinucleotide contexts, which are used as unlabeled or pseudo-pathogenic datasets.

New Developmental Delay Disorder (new DDD) is one example of a validation set used to compare the classification accuracy of Primate AI 3D against Primate AI. The new DDD validation set labels variants from individuals with DDD as pathogenic and labels the same variants from healthy relatives of individuals with DDD as benign. A similar labeling scheme is used with the autism spectrum disorder (ASD) validation set shown in Figures 31A and 31B.

BRCA1 is another example of a validation set used to compare the classification accuracy of Primate AI 3D against Primate AI. The BRCA1 validation set labels synthetically generated reference amino acid sequences simulating proteins from the BRCA1 gene as benign variants and synthetically altered allelic amino acid sequences simulating proteins from the BRCA1 gene as pathogenic variants. A similar labeling scheme is used with different validation sets of the TP53 gene, the TP53S3 gene and its variants, and other genes and their variants shown in Figures 31A and 31B.

Figure 28 identifies the performance of the benchmark PrimateAI model with blue horizontal bars and the performance of the PrimateAI 3D model disclosed with orange horizontal bars. The green horizontal bar depicts the pathogenicity predictions derived by combining the respective pathogenicity predictions of the disclosed PrimateAI 3D model and the benchmark PrimateAI model. In the legend, “ens10” denotes an ensemble of 10 PrimateAI 3D models, each trained with a different seed training dataset and randomly initialized with different weights and biases. Additionally, “7×7×7×2” describes the size of the voxel grid used to encode the input channels during training of an ensemble of 10 PrimateAI 3D models. For a given variant, an ensemble of 10 PrimateAI 3D models each generates 10 pathogenicity predictions, which are subsequently combined (e.g., by averaging) to produce the final pathogenicity prediction for the given variant. This logic applies similarly to ensembles of different group sizes.

Also, in Figure 28, the y-axis has different validation sets and the x-axis has p-values. Larger p-values, i.e. longer horizontal bars, indicate greater accuracy in distinguishing benign variants from pathogenic variants. As evidenced by the p-value in Figure 28, PrimateAI 3D outperforms PrimateAI across most validation sets (the only exception being the tp53s3_A549 validation set). That is, the orange horizontal bars for PrimateAI 3D are consistently longer than the blue horizontal bars for PrimateAI.

Also, in Figure 28, the “Mean” category along the y-axis calculates the average of the p-values determined for each of the validation sets. Even in the average category, PrimateAI 3D outperforms PrimateAI.

In Figure 29, PrimateAI is represented by the blue horizontal bar, the ensemble of 20 PrimateAI 3D models trained with a voxel grid of size 3×3×3 is represented by the red horizontal bar, and the ensemble of 20 PrimateAI 3D models is represented by the red horizontal bar. The ensemble of 10 PrimateAI 3D models trained with a voxel grid of size is represented by the purple horizontal bar, and the ensemble of 20 PrimateAI 3D models trained with a voxel grid of size 7×7×7×2 is represented by the brown horizontal bar. The ensemble of 20 PrimateAI 3D models trained on a voxel grid of size 17 × 17 × 17 × 2 is represented by a purple horizontal bar.

Also, in Figure 29, the y-axis has different validation sets and the x-axis has p-values. As before, larger p-values, i.e. longer horizontal bars, indicate greater accuracy in distinguishing benign variants from pathogenic variants. As evidenced by the p-values in Figure 20, different configurations of PrimateAI 3D outperform PrimateAI across most validation sets. That is, the red, purple, brown, and pink horizontal bars for PrimateAI 3D are mostly longer than the blue horizontal bars for PrimateAI.

Also, in Figure 29, the “Mean” category along the y-axis calculates the average of the p-values determined for each of the validation sets. Even in the average category, PrimateAI 3D's different configurations outperform PrimateAI.

In Figure 30, the red vertical bar represents PrimateAI, and the cyan vertical bar represents PrimateAI 3D. In Figure 30, the y-axis has p-values and the x-axis has different validation sets. In Figure 30, without exception, PrimateAI 3D consistently outperforms PrimateAI across all validation sets. That is, the cyan vertical bars for PrimateAI 3D are always longer than the red vertical bars for PrimateAI.

Figures 31A and 31B identify the performance of the benchmark PrimateAI model with blue vertical bars and the performance of the PrimateAI 3D model disclosed with orange vertical bars. Green vertical bars depict pathogenicity predictions derived by combining the respective pathogenicity predictions of the disclosed PrimateAI 3D model and the benchmark PrimateAI model. 31 and 31B, the y-axis has p-values and the x-axis has different validation sets.

As evidenced by the p-values in Figures 31A and 31B, PrimateAI 3D outperforms PrimateAI across most validation sets (the only exception being the tp53s3_A549_p53NULL_Nutlin-3 validation set). That is, the orange vertical bars for PrimateAI 3D are consistently longer than the blue vertical bars for PrimateAI.

Additionally, in Figures 31A and 31B, a separate "Mean" chart calculates the average of the p-values determined for each of the validation sets. Even in the average chart, PrimateAI 3D outperforms PrimateAI.

Average statistics may be biased by outliers. To address this, separate “method rank” charts are also shown in FIGS. 31A and 31B. Higher ranks indicate poorer classification accuracy. Also in the method rank chart, PrimateAI 3D outperforms PrimateAI by having more counts of subranks 1 and 2 for PrimateAI where all are 3.

28, 29, 30, 31A and 31B, it is also clear that combining PrimateAI 3D with PrimateAI produces excellent classification accuracy. That is, a protein may be fed to PrimateAI as an amino acid sequence to produce a first output, the same protein may be fed to PrimateAI 3D as a 3D, voxelized protein structure to produce a second output, and the first and second The output can be aggregated and combined or analyzed to generate a final pathogenicity prediction for the variants experienced by the protein.

Efficient voxelization

FIG. 32 is a flow diagram illustrating an efficient voxelization process 3200 that efficiently identifies nearest atoms on a voxel-by-voxel basis.

Now, we revisit the street channel. As discussed above, the reference amino acid sequence 202 may contain different types of atoms, such as alpha-carbon atoms, beta-carbon atoms, oxygen atoms, nitrogen atoms, hydrogen atoms, etc. Accordingly, as discussed above, the distance channels can be arranged by nearest alpha-carbon atom, nearest beta-carbon atom, nearest oxygen atom, nearest nitrogen atom, nearest hydrogen atom, etc.

For example, in Figure 6, each of the nine voxels 514 has 21 amino acid-specific distance channels to the nearest alpha-carbon atom. Figure 6 shows that each of the nine voxels 514 also has 21 amino acid-specific distance channels to the nearest beta-carbon atom, and that each of the nine voxels 514 has a It can be further extended to also have the nearest normal atom distance channel for the atom. In this way, each of the nine voxels 514 can have 43 distance channels.

Now, the number of distance calculations required to identify the closest atom on a voxel-by-voxel basis for inclusion in the distance channel is discussed. Consider the example of Figure 3, which shows a total of 828 alpha-carbon atoms distributed across 21 amino acid categories. To calculate the distance channels 602 to 642 for each amino acid in Figure 6, i.e., to determine 189 distance values, the distance from each of the 9 voxels 514 to each of the 828 alpha-carbon atoms is measured, 9 * 828 = 7, resulting in 452 distance calculations. For a 3D case of 828 voxels, this results in 27 * 27 = 22,356 distance calculations. When 828 beta-carbon atoms are also included, this number increases to 27 * 1656 = 44, 712 distance calculations.

This means that the runtime complexity of identifying the nearest atom on a voxel-by-voxel basis for a single protein voxelization is O(#atom * #voxel), as shown in Figure 35a. Additionally, the runtime complexity for a single protein voxelization is O(#atoms ) increases.

As a result, distance calculations can become the most computationally expensive part of the voxelization process, consuming valuable computational resources from critical runtime tasks such as model training and model inference. For example, consider the case of model training with a training dataset of 7,000 proteins. Creating distance channels for multiple voxels across multiple amino acids, atoms, and attributes may involve more than 100 voxelizations per protein, resulting in approximately 800,000 voxelizations in a single training iteration (epoch). Running 20 to 40 epochs of training while rotating the atomic coordinates in each epoch can result in up to 32 million voxels.

In addition to the high computational cost, the size of the data for 32 million voxelizations is too large to fit into main memory (e.g., >20 TB for a 15×15×15 voxel grid). Considering repeated training runs for parameter optimization and ensemble learning, the memory footprint of the voxelization process becomes too large to be stored on disk, making the voxelization process a part of model training rather than a precomputation step.

The disclosed technology provides an efficient voxelization process that achieves up to approximately 100x speedup compared to a runtime complexity of O(#atoms * #voxels). The disclosed efficient voxelization process reduces the runtime complexity for single protein voxelization to O(#atoms). In the case of different features or channels per voxel, the disclosed efficient voxelization process reduces the runtime complexity for single protein voxelization to O(#atoms * #properties). As a result, the voxelization process becomes as fast as model training, shifting the computational bottleneck from voxelization back to calculating neural network weights on processors such as GPUs, ASICs, TPUs, FPGAs, CGRAs, etc.

In some implementations of the disclosed efficient voxelization process involving large voxel grids, the runtime complexity for single protein voxelization is O(#atoms + voxels) and O(#atoms * for the case of different features or channels per voxel. #property + voxel). “+ voxel” complexity is observed when the number of atoms is infinitesimal compared to the number of voxels, for example, when there is one atom in a 100×100×100 voxel grid (i.e., 1 million voxels per atom). In such a scenario, the runtime is dominated by the overhead of a huge number of voxels, for example, allocating memory for 1 million voxels, initializing 1 million voxels to 0, etc.

The details of the disclosed efficient voxelization process are now discussed. Figures 32A, 32B, 33, 34, and 35B are discussed simultaneously.

Starting in Figure 32A, at step 3202, each atom (e.g., each of the 828 alpha-carbon atoms and each of the 828 beta-carbon atoms) is assigned to a voxel containing the atom (e.g., one of nine voxels 514). It is related to one). The term “contains” refers to the 3D atomic coordinates of an atom being located within a voxel. Voxels containing atoms are also referred to herein as “atom-containing voxels.”

Figures 32B and 33 describe how voxels containing specific atoms are selected. Figure 33 uses 2D atomic coordinates as a representative of 3D atomic coordinates. Note that the voxel grid 522 is regularly spaced with each of the voxels 514 having the same step size (eg, 1 Angstrom (Å) or 2 Å).

Additionally, in Figure 33, voxel grid 522 has magenta index [0, 1, 2] along the first dimension (e.g., x-axis) and cyan index along the second dimension (e.g., y-axis). It has color index [0, 1, 2]. Additionally, in Figure 33, each voxel 514 within voxel 512 is divided by the green voxel index [voxel 0, voxel 1, ..., voxel 8] and the black voxel centroid index [(1, 1), ( 1, 2), ..., (3, 3)].

Additionally, in Figure 33, the center coordinates of the voxel center along the first dimension, that is, the first dimensional voxel coordinates, are identified in orange. Additionally, in Figure 33, the center coordinates of the voxel center along the second dimension, that is, the second dimensional voxel coordinates, are identified in red.

First, in step 3202a (step 1 in Figure 33), the 3D atomic coordinates (1.7456, 2.14323) of a particular atom are quantized to generate quantized 3D atomic coordinates (1.7, 2.1). Quantization can be achieved by rounding or truncation of bits.

Next, in step 3202b (step 2 in Figure 33), the voxel coordinates (or voxel centroid or voxel center coordinates) of voxel 514 are assigned to 3D atomic coordinates quantized in dimension units. For the first dimension, the quantized atomic coordinate 1.7 is assigned to voxel 1 because it covers first dimension voxel coordinates in the range 1 to 2 and is centered at 1.5 in the first dimension. Note that voxel 1 has index 1 along the first dimension, as opposed to index 0 along the second dimension.

For the second dimension, starting from voxel 1, the voxel grid 522 is traversed along the second dimension. This results in the quantized atomic coordinate 2.5 being assigned to voxel 7 because it covers second dimension voxel coordinates in the range 2 to 3 and is centered at 2.5 in the second dimension. Note that voxel 7 has index 2 along the second dimension, as opposed to index 1 along the first dimension.

Next, in step 3202c (step 3 in Figure 33), the dimension index corresponding to the assigned voxel coordinates is selected. That is, for voxel 1, index 1 is selected along the first dimension, and for voxel 7, index 2 is selected along the second dimension. Those skilled in the art will understand that the above steps can be similarly implemented to select a dimension index along the third dimension in the case of a third dimension.

Next, in step 3202d (step 4 in FIG. 33), a cumulative sum is generated based on positional weighting of the selected dimension index by the square of the radix. The general idea behind the positional numbering system is that numerical values are expressed through increasing powers of a radix (or base), e.g. binary numbers are radix 2, ternary numbers are radix 3, etc. Octal numbers are base 8, and hexadecimal numbers are base 16. This is often referred to as a weighted numbering system because each position is weighted by the square of the cardinality. The set of significant digits for a positional numbering system is the same size as the base of the system. For example, in the decimal system there are 10 numbers 0 through 9, and in the ternary system there are 3 numbers 0, 1, and 2. In any radix system, the largest significant number is 1 less than the radix (thus, in any radix system, 8 is not a significant number less than 9). Any decimal integer can be represented exactly in any other integration-based system, and vice versa.

Returning to the example of Figure 33, the selected dimension indices 1 and 2 are converted to a single integer by multiplying them position-wise by their respective powers in base 3 and summing the results of the position-wise multiplication. Here, base 3 is chosen because 3D atomic coordinates have three dimensions (however, Figure 33 only shows 2D atomic coordinates along two dimensions for simplicity).

Since the index 2 is located in the rightmost bit (i.e., the least significant bit), it is multiplied by 3 to the power of 0, yielding 2. Since the index 1 is located in the second rightmost bit (i.e., the second least significant bit), it is multiplied by 3 to the power of 1, yielding 3. This results in a cumulative sum of 5.

Next, at step 3202e (step 5 in Figure 33), based on the accumulated sum, the voxel index of the voxel containing the particular atom is selected. That is, the accumulated sum is interpreted as the voxel index of the voxel containing a specific atom.

At step 3212, after each atom is associated with an atom-containing voxel, each atom is further associated with one or more voxels in the neighborhood of the atom-containing voxel, also referred to herein as “neighboring voxels.” Neighboring voxels may be selected based on being within a predefined radius (eg, 5 angstroms (Å)) of the atom-containing voxel. In other implementations, neighboring voxels may be selected based on their proximity to the atom-containing voxel (e.g., top, bottom, right, left neighboring voxels). The resulting associations associating each atom with its atom-containing voxel and neighboring voxels are encoded in an atom-to-voxel mapping 3402, also referred to herein as an element-to-cell mapping. In one example, the first alpha-carbon atom is associated with a first subset of voxels 3404 that include the atom-containing voxels and neighboring voxels for the first alpha-carbon atom. In another example, the second alpha-carbon atom is associated with a second subset of voxels 3406 that includes the atom-containing voxel and neighboring voxels for the second alpha-carbon atom.

Note that no distance calculations are made to determine atom-containing voxels and neighboring voxels. Atom-containing voxels are selected by a spatial arrangement of the voxels that allows assignment of quantized 3D atomic coordinates to the centers of corresponding regularly spaced voxels within a voxel grid (without using any distance calculations). Additionally, neighboring voxels are selected due to being spatially adjacent to atom-containing voxels within the voxel grid (again without using any distance calculations).

At step 3222, each voxel is mapped to the atom it was associated with at steps 3202 and 3212. In one implementation, this mapping is encoded in a voxel-to-atom mapping 3412, which is based on the atom-to-voxel mapping 3402 (e.g., by applying a voxel-based classification key on the atom-to-voxel mapping 3402). It is created by Voxel-to-atom mapping 3412 is also referred to herein as “cell-element mapping.” In one example, the first voxel includes an alpha-carbon atom associated with the first voxel in steps 3202, 3212. is mapped to a first subset of alpha-carbon atoms 3414. In another example, the second voxel is mapped at steps 3202 and 3212 to a second subset of alpha-carbon atoms 3416 that include alpha-carbon atoms associated with the second voxel.

At step 3232, for each voxel, the distance between the voxel and the atom mapped to the voxel at step 3222 is calculated. Step 3232 has a runtime complexity of O(#atoms) because the distance to a particular atom is measured only once from each voxel to which that particular atom is uniquely mapped in voxel-to-atom mapping 3412. . This is the case when no neighboring voxels are considered. If there are no neighbors, the constant argument implied by big-O notation is 1. If there are neighbors, the big-O notation is equal to the number of neighbors + 1 because the number of neighbors is constant for each voxel, and thus the runtime complexity of O(#atoms) remains the same. In contrast, in Figure 35A, the distance to a specific atom is measured redundantly as many times as there are voxels (e.g., 27 distances to a specific atom due to 27 voxels).

In Figure 35B, based on voxel-to-atom mapping 3412, each voxel is mapped to a respective subset of 828 atoms (relative to neighboring voxels), as illustrated by a respective ellipse for each voxel. does not include distance calculations). The respective subsets generally do not overlap, with some exceptions. There is minor overlap in some cases when multiple atoms are mapped to the same voxel, as shown in Figure 35B by the yellow overlap between the prime symbol "'" and the ellipse. This minimal overlap has an additive, non-multiplicative effect on the runtime complexity of O(#atoms). This overlap is the result of determining the voxel containing the atom and then considering neighboring voxels. If there are no neighboring voxels, there may be no overlap because the atom is associated with only one voxel. However, considering neighbors, each neighbor can potentially be associated with the same atom (unless there is another atom of the same amino acid that is closer).

At step 3242, for each voxel, the atom closest to the voxel is identified, based on the distance calculated at step 3232. In one implementation, this identification is encoded in a voxel-nearest atom mapping 3422, also referred to herein as a “cell-nearest element mapping.” In one example, a first voxel has its closest alpha-nearest atom mapping. Mapped to the second alpha-carbon atom as carbon atom 3424. In another example, the second voxel is mapped to the 31-th alpha-carbon atom as its closest alpha-carbon atom 3426.

Additionally, as voxel-wise distances are calculated using the techniques discussed above, the atomic type and amino acid type categorization of the atoms, and the corresponding distance values, are stored to create categorized distance channels.

Once the distances to the nearest atoms are identified using the techniques discussed above, these distances can be encoded into a distance channel for voxelization and subsequent processing by the pathogenicity classifier 2108.

computer system

36 depicts an example computer system 3600 that can be used to implement the disclosed techniques. Computer system 3600 includes at least one central processing unit (CPU) 3672 that communicates with a number of peripheral devices via a bus subsystem 3655. These peripheral devices include, for example, storage subsystem 3610, which includes memory devices and file storage subsystem 3636, user interface input device 3638, user interface output device 3676, and network interface subsystem ( 3674). Input and output devices allow user interaction with computer system 3600. Network interface subsystem 3674 provides interfaces to external networks, including interfaces to corresponding interface devices in other computer systems.

In one implementation, pathogenicity classifier 2108 is communicatively linked to storage subsystem 3610 and user interface input device 3638.

User interface input device 3638 may include a keyboard; A pointing device such as a mouse, trackball, touchpad, or graphics tablet; scanner; Touch screen integrated within the display; audio input devices such as voice recognition systems and microphones; and other types of input devices. Generally, use of the term “input device” is intended to include all possible types of devices and manners for inputting information into computer system 3600.

User interface output device 3676 may include a non-visual display, such as a display subsystem, printer, fax machine, or audio output device. The display subsystem may include a planar device such as an LED display, a cathode ray tube (CRT), a liquid crystal display (LCD), a projection device, or some other mechanism for producing a visible image. The display subsystem may also provide non-visual displays, such as audio output devices. Generally, the use of the term “output device” is intended to include all possible types of devices and manners for outputting information from computer system 3600 to a user or to another machine or computer system.

Storage subsystem 3610 stores programming and data configurations that provide the functionality of some or all of the modules and methods described herein. These software modules are typically executed by processor 3678.

Processor 3678 may be a graphics processing unit (GPU), a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), and/or a coarse-grained reconfigurable architecture (CGRA). Processor 3678 may be hosted by a deep learning cloud platform such as Google Cloud Platform™, Xilinx™, and Cirrascale™. Examples of processors 3678 include Google's Tensor Processing Unit (TPU)™, rackmount solutions such as GX4 Rackmount Series™, GX36 Rackmount Series™, NVIDIA DGX-1™, Microsoft's Stratix V FPGA™, and Graphcore's Intelligent Processor Unit. (IPU)™, Qualcomm's Zeroth Platform™ with Snapdragon processors™, NVIDIA's Volta™, NVIDIA's DRIVE PX™, NVIDIA's JETSON TX1/TX2 MODULE™, Intel's Nirvana™, Movidius VPU™, Fujitsu DPI™, ARM Includes Lambda GPU servers with DynamicIQ™, IBM TrueNorth™, and Testa V100s™.

The memory subsystem 3622 used in the storage subsystem 3610 includes a main random access memory (RAM) 3632 for storage of instructions and data during program execution and a read-only memory (ROM) where fixed instructions are stored ( 3634). File storage subsystem 3636 may provide persistent storage for program and data files and may include a hard disk drive, a floppy disk drive with associated removable media, a CD-ROM drive, an optical drive, or a removable media cartridge. You can. Modules implementing the functionality of a given implementation may be stored by file storage subsystem 3636 within storage subsystem 3610, or on another machine accessible by the processor.

Bus subsystem 3655 provides mechanisms to allow the various components and subsystems of computer system 3600 to communicate with each other as intended. Although bus subsystem 3655 is schematically depicted as a single bus, alternative implementations of the bus subsystem may use multiple buses.

Computer system 3600 itself may be a personal computer, portable computer, workstation, computer terminal, network computer, television, mainframe, server farm, broadly distributed set of loosely networked computers, or any other data processing system or user. It can be of various types, including devices. Due to the ever-changing nature of computers and networks, the description of computer system 3600 depicted in FIG. 36 is intended only as a specific example to illustrate preferred implementations of the invention. Many other configurations of computer system 3600 are possible with more or fewer components than the computer system depicted in FIG. 36.

Amino acid sequence prediction

A protein language model trained with a masked language modeling goal is supervised to output the probability of an amino acid occurring at a specific position in the protein depending on the surrounding context. Proteins are linear polymers that fold into a variety of specific shapes for their function. The incredibly diverse three-dimensional (3D) structures, determined by the combination and sequence of 20 amino acids in a protein polymer chain (the sequence of proteins), enable the sophisticated functions of proteins responsible for most biological activities. Therefore, obtaining the structure of proteins is of utmost importance for understanding the fundamental biology of health and disease and developing therapeutic molecules. Protein structures are primarily determined by sophisticated experimental techniques such as .

Computer methods have been used to predict the structure of proteins, explain the mechanisms of biological processes, and determine the properties of proteins. Furthermore, all naturally occurring proteins are the result of an evolutionary process of random variants occurring under various selection pressures. Through this process, nature has explored only a small subset of the theoretically possible protein sequence space. Advances in machine learning, especially deep learning, are promoting a revolution in scientific research paradigms. Some deep learning-based approaches, especially in the field of structure prediction, now outperform traditional methods, often in combination with high-resolution physical modeling. Challenges remain in experimental validation, benchmarking, exploiting known physics and interpreting the model, and extending it to other biomolecules and contexts.

Protein sites are microenvironments within a protein structure that are distinguished by their structural or functional roles. A site can be defined by a three-dimensional location and local neighbors around that location where the structure or function resides. Central to rational protein engineering is the understanding of how the structural arrangement of amino acids creates functional properties within protein regions. Determination of the structural and functional roles of individual amino acids within a protein provides information to aid engineers and modify protein function. Identifying functionally or structurally important amino acids allows for focused engineering efforts, such as site-directed mutagenesis, to alter targeted protein functional properties. In one embodiment, the disclosed technology relates to predicting spatial tolerance of amino acid substitutions. In such implementations, the disclosed technology includes gapping logic and substitution logic. Gapping logic is configured to remove a specific amino acid at a specific position from a protein and create an amino acid vacancy at a specific position in the protein. The substitution logic is configured to process proteins with amino acid vacancies and evaluate the tolerance of alternative amino acids that are candidates for filling or fitting the amino acid vacancies. The substitution logic is further configured to score the tolerance of the substituted amino acid based, at least in part, on the structural (or spatial) compatibility between the substituted amino acid and adjacent amino acids (e.g., right and left flanking amino acids) near the amino acid vacancy. Substitution logic assesses the degree to which an amino acid “fits” into its surrounding protein environment and shows that mutations that disrupt strong amino acid preferences are more likely to be deleterious. When the substitution logic was a convolutional neural network, during the training process, the weights of the convolutional filter were optimized to detect local spatial patterns that best captured local biochemical features to separate the 20 amino acid microenvironments. After the training process, the filters in the convolutional layer of the convolutional neural network are activated when the desired feature exists at some spatial location of the input. Structural (or spatial) fitness can be defined as a change or effect on protein function. Substituted amino acids are considered structurally (or spatially) incompatible if they cause changes in the function of the protein after being substituted at a specific position within the protein structure. Substituted amino acids are considered structurally (or spatially) compatible if they do not cause a change in the function of the protein after being substituted at a specific position in the protein structure. Structural (or spatial) suitability can be defined as spatial deviation measured by distance metrics. First, pre-insertion spatial measurements of the protein structure can be determined, for example, by measuring the distance between amino acids in the protein structure prior to amino acid substitution at a specific position. The distance may be an atomic distance based on the atomic coordinates of amino acid atoms. The distance between pairs of amino acids can be measured. The spatial measurements after insertion in the protein structure are then determined, for example, by re-measuring the distances between amino acids in the protein structure after amino acid substitutions at specific positions. A substituted amino acid is considered structurally (or spatially) unsuitable if the spatial deviation between the pre-insertion and post-insertion spatial measurements exceeds a threshold. A substituted amino acid is considered structurally (or spatially) suitable if the spatial deviation between the pre-insertion and post-insertion spatial measurements does not exceed a threshold.

In another embodiment, the disclosed techniques relate to predicting evolutionary conservation of amino acid substitutions. In such implementations, the disclosed technology includes gapping logic and substitution logic. Gapping logic is configured to remove a specific amino acid at a specific position from a protein and create an amino acid vacancy at a specific position in the protein. The substitution logic is configured to process proteins with amino acid vacancies and score the evolutionary conservation of substituted amino acids that are candidates for filling the amino acid vacancies. The substitution logic is further configured to score the evolutionary conservation of the substituted amino acid based, at least in part, on the structural (or spatial) compatibility between the substituted amino acid and adjacent amino acids (e.g., right and left flanking amino acids) near the amino acid vacancy. . In some implementations, evolutionary conservation is scored using evolutionary conservation frequency. In one implementation, the evolutionary conservation frequencies are based on the position-specific frequency matrix (PSFM). In another implementation, the evolutionary conservation frequency is based on a position-specific score matrix (PSSM). In one embodiment, the evolutionary conservation scores of substituted amino acids are ranked on a scale.

In another embodiment, the disclosed techniques relate to predicting evolutionary conservation of amino acid substitutions. In such implementations, the disclosed techniques include gapping logic and evolutionary conservation prediction logic. Gapping logic is configured to remove a specific amino acid at a specific position from a protein and create an amino acid vacancy at a specific position in the protein. The evolutionary conservation prediction logic is structured to process proteins with amino acid vacancies and rank the evolutionary conservation of substituted amino acids that are candidates for filling the amino acid vacancies.

Gap protein spatial representation-based pathogenicity determination for targeted alternative amino acids.

Figure 37 depicts one embodiment of determining 3700 variant pathogenicity for target replacement amino acids based on processing of gap protein spatial representation. Proteins are sequences of amino acids. Certain amino acids that have been removed or masked from a protein are referred to as “gap amino acids.” The resulting proteins lacking gap amino acids are referred to as “gap proteins” or “vacancy-containing proteins.”

The “spatial representation” of a protein represents the structural information about the protein's amino acids. The spatial representation of a protein may be based on the shape, location, position, pattern, and/or arrangement of amino acids within the protein. The spatial representation of a protein can be one-dimensional (1D), two-dimensional (2D), three-dimensional (3D), or n- dimensional ( n D) information.

In one embodiment, the spatial representation of the protein includes the per-amino acid distance channel discussed above, such as the per-amino acid distance channel 600 described above with respect to FIG. 6 . In another embodiment, the spatial representation of the protein comprises a distance channel tensor described above, such as distance channel tensor 700 described above with respect to FIG. 7 . In another embodiment, the spatial representation of the protein comprises an evolutionary profile tensor described above, such as an evolutionary profile tensor 1800 described above with respect to FIG. 18 . In another embodiment, the spatial representation of the protein comprises a voxelized annotation channel described above, such as voxelized annotation channel 2000 described above with respect to Figure 20. In another embodiment, the spatial representation of the protein includes a structural confidence channel discussed above. In other implementations, the spatial representation may also include other channels.

A “gap spatial representation” of a protein is a spatial representation of the protein that excludes at least one gap amino acid in the protein. In one embodiment, gap amino acids are excluded by excluding (or not considering or ignoring) one or more atoms or atom types of the gap amino acid when generating the gap space representation. For example, atoms of gap amino acids can be excluded from calculations (or selections or computations) that generate distance channels, evolution profiles, annotation channels, and/or structure confidence channels. In other embodiments, gap spatial representations can be generated by excluding gap amino acids from other feature channels as well.

Consider the following example, which generates a gap space representation of a protein by excluding gap amino acid atoms from the amino acid-specific distance channel calculation. In Figure 5, the Cα ^A5 atom belongs to the alanine amino acid at position 5 of the protein. Now assume that the alanine amino acid at the fifth position is selected as the gap amino acid. A gap space representation is created by calculating the distance channel without considering the distance 512 between the voxel center (1, 1) of the voxel grid 522 and the nearest alpha-carbon ( _Cα ) atom, which is the It is the Cα ^A5 atom, that is, the alanine amino acid at the fifth position.

It is also noted that this application uses “spatial representation of a protein” and “protein structure” interchangeably. It is also noted that this application uses “gap space representation of proteins” and “gap protein structures” interchangeably.

Referring to Figure 37, in operation 3702, protein sequence accessor 3704 accesses the protein with each amino acid at each position.

In operation 3712, the gap amino acid designator 3714 designates a specific amino acid at a specific position of the protein as a gap amino acid and designates the remaining amino acid at the remaining position of the protein as a non-gap amino acid. In one embodiment, the particular amino acid is a reference amino acid that is the major allele of the protein.

In operation 3722, gap space representation generator 3724 generates a gap space representation of the protein that includes the spatial configuration of non-gap amino acids and excludes the spatial configuration of gap amino acids. The spatial organization of non-gap amino acids is encoded as distance channels per amino acid class. The distance channel for each amino acid class has a voxel-specific distance value for a voxel among a plurality of voxels. The distance value for each voxel specifies the distance from the corresponding voxel among the plurality of voxels to the atom of the non-gap amino acid. The spatial configuration of a non-gap amino acid is determined based on the spatial proximity between the corresponding voxel and the atoms of the non-gap amino acid. By ignoring the distance from the corresponding voxel to the atom of the gap amino acid when determining voxel-wise distance values, the spatial configuration of the gap amino acid is excluded from the gap space representation. By ignoring the spatial proximity between the corresponding voxel and the atoms of the gap amino acid, the spatial configuration of the gap amino acid is excluded from the gap space representation.

The spatial organization of non-gap amino acids is encoded in an evolutionary profile channel based on the pan-amino acid conservation frequency of the amino acid with the closest atom to the voxel. In one embodiment, the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the nearest atom of the gap amino acid when determining the pan-amino acid conservation frequency. The spatial organization of non-gap amino acids is encoded in an evolutionary profile channel based on the per-amino acid conservation frequency of each amino acid with the closest atom to the voxel. In one embodiment, the spatial configuration of gap amino acids is excluded from the gap space representation by ignoring each nearest atom of the gap amino acid when determining the conservation frequency per amino acid. The spatial organization of non-gap amino acids is encoded in the annotation channel. In one embodiment, the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the annotation channel. The spatial organization of non-gap amino acids is encoded as a structural confidence channel. In one embodiment, the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the structural confidence channel. The spatial configuration of non-gap amino acids is encoded as an additional input channel. In one embodiment, the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining additional input channels.

In operation 3732, pathogenicity determinant 3734 determines the pathogenicity of a nucleotide variant based, at least in part, on the gap space representation and the representation of alternative amino acids produced by the nucleotide variant at a particular position. The representation of the replacement amino acid may be a one-hot encoding of the replacement amino acid (see, e.g., Figure 8). In some embodiments, the replacement amino acid is the same amino acid as the reference amino acid. In other embodiments, the replacement amino acid is an amino acid that is different from the reference amino acid.

Figure 38 shows an example of a spatial representation 3800 of a protein. The protein contains an amino acid sequence (3804). The aspartic acid (D) amino acid at position 22 of the amino acid sequence 3804 is selected as the gap amino acid 3802. Figure 39 shows an example of a gap space representation 3900 of the protein illustrated in Figure 38. In Figure 39, gap amino acid 3802 is removed from gap space representation 3900. Also in Figure 39, the absence of gap amino acid 3802 is shown as missing gap amino acid 3902.

Figure 40 shows an example of an atomic space representation 4000 of the protein illustrated in Figure 38. Figure 40 also shows atoms 4002 of gap amino acid 3802. Figure 41 shows an example of a gap atom space representation 4100 of the protein illustrated in Figure 38. In Figure 41, atom 4002 of gap amino acid 3802 is removed from gap atom space representation 4100. Also in Figure 41, the absence of atom 4002 of gap amino acid 3802 is shown as missing atom 4102 of gap amino acid 3802.

Please also note that this application uses the terms "pathogenicity determinant", "pathogenicity predictor", "pathogenicity classifier", "variant pathogenicity classifier", "evolutionary conservation predictor", and "evolutionary conservation determinant" interchangeably. do.

Figure 42 shows the work of the pathogenicity classifier (2108/2600/2700) to determine (4200) variant pathogenicity for a target replacement amino acid based on gap protein space representation (4202) and processing of the replacement amino acid representation (4212) of the target replacement amino acid. An implementation example is shown.

Pathogenicity classifiers 2108/2600/2700 determine the pathogenicity of nucleotide variants by processing the gap space representation 4202 and the representation 3212 of alternative amino acids as input and generating a pathogenicity score 4208 for the alternative amino acid as output. decide

Figure 43 shows one implementation of training data 4300 used to train pathogenicity classifiers 2108/2600/2700. Pathogenicity classifiers (2108/2600/2700) are trained on the benign training set (4302). The positive training set 4302 has each positive protein sample (4322, 4342, and 4362) for each reference amino acid at each position (4312, 4332, and 4352) of the proteome. The reference amino acid is the major allelic amino acid in the proteome. In one embodiment, the proteome has 10 million positions, so positive training set 4302 has 10 million positive protein samples. Each positive protein sample has a respective gap space representation generated using each reference amino acid as each gap amino acid. Each positive protein sample is represented by each reference amino acid and each alternative amino acid. In various embodiments, the proteome includes a human proteome and a non-human proteome, including a non-human primate proteome.

Figure 44 shows one embodiment of generating 4400 gap spatial representations (4322G, 4342G, and 4362G) for reference protein samples (4322, 4342, and 4362) using reference amino acids (4402, 4412, and 4422) as gap amino acids, respectively. It shows. Figure 45 shows one implementation of training a pathogenicity classifier (2108/2600/2700) on a positive protein sample (4500).

A pathogenicity classifier (2108/2600/2700) is trained on a specific positive protein sample and uses (i) a specific gap space representation of the specific positive protein sample (4322G), and (ii) the representation of a specific reference amino acid as a specific substitute amino acid (4402 ) (e.g., one-hot encoding) as input and generate a pathogenicity score for that specific reference amino acid as output to estimate the pathogenicity of a specific reference amino acid at a specific location in a specific positive protein sample. A specific gap space representation is created by using a specific reference amino acid as the gap amino acid and the amino acids remaining at the remaining positions in the specific positive protein sample as the non-gap amino acids.

Each positive protein sample has a true positive label 4506 indicating the absolute positivity of the positive protein sample. In one embodiment, the actual positive label is 0, 1, or minus 1. The pathogenicity score 4502 for a particular reference amino acid is compared to the true positive label to determine the error 4504 and training techniques (e.g., backpropagation 4512) are used to create a pathogenicity classifier 2108/ 2600/2700) coefficient is improved.

Pathogenicity classifiers (2108/2600/2700) are trained on the pathogenicity training set (4308). The pathogenicity training set 4308 consists of each pathogenic protein sample ( 4322A-N, 4342A-N and 4362A-N). In one embodiment, each combinatorially produced amino acid substitution is limited by the reachability of a single nucleotide polymorphism (SNP) to convert the reference codon of the reference amino acid to a substitute amino acid of the unreachable substitute amino acid class. Amino acid substitutions created combinatorially for a particular reference amino acid of a particular amino acid class at a particular position in the proteome include each replacement amino acid of the particular amino acid class and each other amino acid class.

In one embodiment, the proteome has 10 million positions, and for each 10 million positions there are 19 combinatorially generated amino acid substitutions, so the pathogenicity training set 4308 has 190 million pathogenic protein samples.

Each pathogenic protein sample has a respective gap space representation generated using each reference amino acid as each gap amino acid. Each pathogenic protein sample has a respective representation of each combinatorially generated amino acid substitution as each replacing amino acid produced by each combinatorially generated nucleotide variant at each position in the proteome.

Figure 46 shows one implementation of training a pathogenicity classifier (2108/2600/2700) on a pathogenic protein sample (4600). The pathogenicity classifier (2108/2600/2700) is trained on a specific pathogenic protein sample, (i) a specific gap space representation (4322G) of the specific pathogenic protein sample, and (ii) a specific combinatorially generated classifier as a specific alternative amino acid. Processes a representation 4622 of amino acid substitutions (e.g., one-hot encoding) as input, and generates as output pathogenicity scores for specific combinatorially generated amino acid substitutions, based on specific criteria at specific positions in a specific pathogenic protein sample. Estimate the pathogenicity of specific combinatorial amino acid substitutions for amino acids. A specific gap space representation is generated by using specific reference amino acids as gap amino acids and the remaining amino acids at the remaining positions in the specific pathogenic protein sample as non-gap amino acids.

Each pathogenic protein sample has an actual pathogenicity label that indicates the absolute pathogenicity of the pathogenic protein sample. In one embodiment, the actual pathogenic label is 1, 0, or minus 1 insofar as it is different (e.g., the opposite) from the actual benign label. The pathogenicity score 4602 for a particular combinatorially generated amino acid substitution determines the error 4604 and uses a training technique (e.g., backpropagation 4612) to generate a pathogenicity classifier 2108/2600 based on the error. /2700) is compared with the actual pathogenicity label (4606) to improve the coefficient.

In one implementation, the pathogenicity classifier (2108/2600/2700) is trained over 200 million training iterations. In this implementation, the 200 million training iterations include 10 million training iterations for the 10 million benign protein samples and 190 million iterations for the 190 million pathogenic protein samples. In one embodiment, the proteome has 1 to 10 million positions, so there are 1 to 10 million positive protein samples in the positive training set. In this embodiment, there are 19 combinatorially generated amino acid substitutions for each of the 1 to 10 million positions, so there are 19 to 190 million pathogenic protein samples in the pathogenicity training set.

In one implementation, the pathogenicity classifier (2108/2600/2700) is trained with 20 to 200 million training iterations. In this embodiment, the 20 to 200 million training iterations include 1 to 10 million training iterations using 1 to 10 million benign protein samples, and 1 to 10 million training iterations using 19 to 190 million pathogenic protein samples. Between 19 and 190 million iterations were performed.

Figure 47 illustrates how certain unreachable amino acid classes are masked 4700 during training. In operation 4702, the unreachable substitution amino acid class, limited by the reachability of a single nucleotide polymorphism (SNP) to convert the reference codon of the reference amino acid to a substitute amino acid of the unreachable substitution amino acid class, is masked to the actual label. In operation 4712, the masked amino acid class suffers no loss and does not contribute to the gradient update. In operation 4722, the masked amino acid class is identified in the lookup table. In operation 4723, the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

Figure 48 depicts one implementation of determining the final pathogenicity score. At operation 4802, in one implementation, the pathogenicity classifier 2108/2600/2700 generates a first pathogenicity score for a first replacement amino acid that is identical to the first reference amino acid. At operation 4812, in one implementation, the pathogenicity classifier 2108/2600/2700 generates a second pathogenicity score for a second replacement amino acid that is different from the first reference amino acid. At operation 4822, in one implementation, the final pathogenicity score for the second replacement amino acid is the second pathogenicity score for the second replacement amino acid.

In another alternative, the final pathogenicity score for the second replacement amino acid is based on a combination of the first pathogenicity score and the second pathogenicity score. In the first alternative to 4822a, in one embodiment, the final pathogenicity score for the second replacement amino acid is the ratio of the second pathogenicity score to the sum of the first pathogenicity score and the second pathogenicity score. In the second alternative to 4822b, in one embodiment, the final pathogenicity score for the second replacement amino acid is determined by subtracting the first pathogenicity score from the second pathogenicity score.

The discussion so far has dealt with the content shown in FIG. 49A. Figure 49A shows that a variant pathogenicity determination was made for a target replacement amino acid 4922 that fills the vacancy created by the reference gap amino acid 4902 at a given position in the protein 4912. In particular, this analysis is performed by spatially representing the protein 4912 and the pore in a 3D format, for example using voxelized amino acid category-wise distance calculations excluding reference gap amino acids 4902 (or atoms thereof).

The discussion now turns to Figure 49b. Figure 49B shows that each variant pathogenicity determination was made for an amino acid of each amino acid class 4916 that fills the vacancy created by the reference gap amino acid 4902 at a given position in the protein 4912. The inputs in Figures 49A and 49B are the same; Only the output is different, and the spatial representation of the protein (4912) and the vacancy in 3D format are also different. In Figure 49A only one pathogenicity score is generated; Meanwhile, in Figure 49B, a pathogenicity score is generated for each of the 20 amino acid classes/categories (e.g., using 20-way softmax classification).

Gap protein spatial representation-based pathogenicity determination for multiple substituted amino acids.

Figure 50 depicts one embodiment of determining 5000 variant pathogenicity for multiple alternative amino acids based on processing of gap protein spatial representation. In operation 5002, protein sequence accessor 3704 accesses the protein with each amino acid at each position.

In operation 5012, the gap amino acid designator 3714 designates a specific amino acid at a specific position of the protein as a gap amino acid, and designates the remaining amino acid at the remaining position of the protein as a non-gap amino acid. In one embodiment, the particular amino acid is a reference amino acid that is the major allele of the protein.

In operation 5022, gap spatial representation generator 3724 generates a gap spatial representation of the protein that includes the spatial configuration of non-gap amino acids and excludes the spatial configuration of gap amino acids. The spatial organization of non-gap amino acids is encoded as distance channels per amino acid class. The distance channel for each amino acid class has a voxel-specific distance value for a voxel among a plurality of voxels. The distance value for each voxel specifies the distance from the corresponding voxel among the plurality of voxels to the atom of the non-gap amino acid. The spatial configuration of a non-gap amino acid is determined based on the spatial proximity between the corresponding voxel and the atoms of the non-gap amino acid. By ignoring the distance from the corresponding voxel to the atom of the gap amino acid when determining voxel-wise distance values, the spatial configuration of the gap amino acid is excluded from the gap space representation. By ignoring the spatial proximity between the corresponding voxel and the atoms of the gap amino acid, the spatial configuration of the gap amino acid is excluded from the gap space representation.

The spatial organization of non-gap amino acids is encoded in an evolutionary profile channel based on the pan-amino acid conservation frequency of the amino acid with the closest atom to the voxel. In one embodiment, the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the nearest atom of the gap amino acid when determining the pan-amino acid conservation frequency. The spatial organization of non-gap amino acids is encoded in an evolutionary profile channel based on the per-amino acid conservation frequency of each amino acid with the closest atom to the voxel. In one embodiment, the spatial configuration of gap amino acids is excluded from the gap space representation by ignoring each nearest atom of the gap amino acid when determining the conservation frequency per amino acid. The spatial organization of non-gap amino acids is encoded in the annotation channel. In one embodiment, the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the annotation channel. The spatial organization of non-gap amino acids is encoded as a structural confidence channel. In one embodiment, the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the structural confidence channel. The spatial configuration of non-gap amino acids is encoded as an additional input channel. In one embodiment, the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining additional input channels.

At operation 5032, pathogenicity determinant 3734 determines the pathogenicity of each replacement amino acid at a particular position based, at least in part, on the gap space representation. Each alternative amino acid is a respective combinatorially generated alternative amino acid produced by each combinatorially generated nucleotide variant at a specific position.

Figure 51 depicts one implementation of a pathogenicity classifier (2108/2600/2700) that determines (5100) variant pathogenicity for multiple alternative amino acids based on processing (5102) of gap protein spatial representations. The pathogenicity classifier 2108/2600/2700 determines the pathogenicity of each alternative amino acid by processing the gap space representation 5102 as input and producing as output a respective pathogenicity score 1-20 for each amino acid class. . In some embodiments, each amino acid class corresponds to each of the 20 naturally occurring amino acids. In another embodiment, each amino acid class corresponds to a respective naturally occurring amino acid from a subset of the 20 naturally occurring amino acids. In one embodiment, the output is displayed with a respective ranking of each pathogenicity score 1-20 for each amino acid class.

Figure 52 shows one implementation of simultaneously training (5200) a pathogenicity classifier (2108/2600/2700) on benign and pathogenic protein samples. Pathogenicity classifiers (2108/2600/2700) are trained on the training set. The training set has each protein sample for each position in the proteome. The proteome has 10 million positions, so the training set has 10 million protein samples. Each protein sample has a respective gap space representation generated using each reference amino acid at each position in the proteome as each gap amino acid. The reference amino acid is the major allelic amino acid in the proteome.

Pathogenicity classifiers 2108/2600/2700 train on specific protein samples, process specific gap space representations 5202 of specific protein samples as input, and output respective pathogenicity scores 1-20 for each amino acid class. It estimates the pathogenicity of each alternative amino acid for a specific reference amino acid at a specific position in a specific protein sample by generating as . A specific gap space representation is created by using a specific reference amino acid as the gap amino acid and the amino acids remaining at the remaining positions in the specific protein sample as the non-gap amino acids.

Each protein sample has an actual label for each amino acid class. Each true label contains an absolute positive label for each reference amino acid class of each amino acid class, and contains a respective absolute pathogenic label for each alternative amino acid class of each amino acid class. In one embodiment, the absolute positive label is 0. The absolute pathogenicity label is identical for each alternative amino acid class. In one embodiment, the absolute pathogenicity label is 1.

In one embodiment, error 5204 is a comparison of pathogenicity scores for a reference amino acid class to an absolute positive label (e.g., pathogenicity score 8 for reference gap amino acid 5212 in Figure 52) and the respective absolute pathogenicity label. is determined based on a respective comparison of each pathogenicity score for each alternative amino acid class (e.g., pathogenicity scores 1-7 and 9-20 in Figure 52). In one implementation, the coefficients of pathogenicity classifier 2108/2600/2700 are improved based on error using training techniques (e.g., backpropagation 5224).

In one implementation, the pathogenicity classifier (2108/2600/2700) is trained with 10 million training iterations using 10 million protein samples. In some embodiments, the proteome has 1 to 10 million positions, so there are 1 to 10 million protein samples in the training set. In one implementation, the pathogenicity classifier (2108/2600/2700) is trained with 1 to 10 million training iterations using 1 to 10 million protein samples.

In one implementation, the pathogenicity classifier 2108/2600/2700 generates a baseline pathogenicity score for the first alternative amino acid of the reference amino acid class. In one implementation, the pathogenicity classifier 2108/2600/2700 generates a respective alternative pathogenicity score for each alternative amino acid of each alternative amino acid class.

In one embodiment, each final alternative pathogenicity score for each alternative amino acid is the respective alternative pathogenicity score. In one embodiment, each final alternative pathogenicity score for each alternative amino acid is based on a respective combination of the baseline pathogenicity score and each alternative pathogenicity score. In one embodiment, each final alternative pathogenicity score for each alternative amino acid is the respective ratio of each alternative pathogenicity score to the sum of the baseline pathogenicity score and each alternative pathogenicity score. In one embodiment, each final alternative pathogenicity score for each alternative amino acid is determined by subtracting each alternative pathogenicity score from each reference pathogenicity score.

In one implementation, the pathogenicity classifier 2108/2600/2700 has an output layer that generates a respective pathogenicity score. In some implementations the output layer is a normalization layer. In this embodiment, each pathogenicity score is normalized. In one implementation, the output layer is a softmax layer. In this implementation, each pathogenicity score is exponentially normalized. In another implementation, the output layer has each sigmoid unit each generating a respective pathogenicity score. In another embodiment, each pathogenicity score is not normalized.

Gap protein spatial representation and evolutionary conservation-based pathogenicity determination for multiple alternative amino acids.

Evolutionary conservation refers to the presence in different species of similar genes, parts of genes, or chromosome segments that reflect both the common origin of the species and the important functional properties of the conserved elements. Mutations occur spontaneously in each generation, randomly changing amino acids here and there in the protein. Individuals with mutations that impair the protein's important functions may have problems that reduce their ability to reproduce. Harmful mutations are lost from the gene pool because individuals carrying them reproduce less effectively. Because harmful mutations are eliminated, amino acids important for protein function are preserved in the gene pool. In contrast, harmless (or very rarely beneficial) mutations are maintained in the gene pool, creating variability in non-critical amino acids. Evolutionary conservation of proteins is identified by aligning the amino acid sequences of proteins with the same function from different taxa (centromere). Predicting the functional consequences of a variant depends, at least in part, on whether important amino acids for the protein family have been conserved through evolution due to negative selection (i.e., amino acid changes at these sites have been deleterious in the past) and whether mutations at these sites have been It relies on the assumption that it increases the likelihood of pathogenicity (causing disease) in humans. Typically, homologous sequences of a target protein are collected and aligned, and a metric of conservation is calculated based on the weighted frequency of different amino acids observed at the target position within the alignment. Figure 53 depicts one implementation of determining 5300 variant pathogenicity for multiple alternative amino acids based on processing gap protein spatial representations and generating evolutionary conservation scores for multiple alternative amino acids in response. In operation 5302, the gap amino acid designator 3714 designates a specific amino acid at a specific position of the protein as a gap amino acid and designates the remaining amino acid at the remaining position of the protein as a non-gap amino acid. In one embodiment, the particular amino acid is a reference amino acid that is the major allele of the protein.

In operation 5312, gap space representation generator 3724 generates a gap space representation of the protein that includes the spatial configuration of non-gap amino acids and excludes the spatial configuration of gap amino acids. The spatial organization of non-gap amino acids is encoded as distance channels per amino acid class. The distance channel for each amino acid class has a voxel-specific distance value for a voxel among a plurality of voxels. The distance value for each voxel specifies the distance from the corresponding voxel among the plurality of voxels to the atom of the non-gap amino acid. The spatial configuration of a non-gap amino acid is determined based on the spatial proximity between the corresponding voxel and the atoms of the non-gap amino acid. By ignoring the distance from the corresponding voxel to the atom of the gap amino acid when determining voxel-wise distance values, the spatial configuration of the gap amino acid is excluded from the gap space representation. By ignoring the spatial proximity between the corresponding voxel and the atoms of the gap amino acid, the spatial configuration of the gap amino acid is excluded from the gap space representation.

The spatial organization of non-gap amino acids is encoded in an evolutionary profile channel based on the pan-amino acid conservation frequency of the amino acid with the closest atom to the voxel. In one embodiment, the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the nearest atom of the gap amino acid when determining the pan-amino acid conservation frequency. The spatial organization of non-gap amino acids is encoded in an evolutionary profile channel based on the per-amino acid conservation frequency of each amino acid with the closest atom to the voxel. In one embodiment, the spatial configuration of gap amino acids is excluded from the gap space representation by ignoring each nearest atom of the gap amino acid when determining the conservation frequency per amino acid. The spatial organization of non-gap amino acids is encoded in the annotation channel. In one embodiment, the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the annotation channel. The spatial organization of non-gap amino acids is encoded as a structural confidence channel. In one embodiment, the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the structural confidence channel. The spatial configuration of non-gap amino acids is encoded as an additional input channel. In one embodiment, the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining additional input channels.

In operation 5322, evolutionary conservation determiner 5324 determines the evolutionary conservation at a particular position of each amino acid of each amino acid class based, at least in part, on the gap space representation.

Figure 54 illustrates an evolutionary preservation determinant 5324 in operation 5400 according to one implementation. In some implementations, the evolutionary conservation determinant 5324 has the same architecture as the pathogenicity classifier 2108/2600/2700. The evolutionary conservation determinant 5324 determines evolutionary conservation by processing the gap space representation 5402 as input and producing a respective evolutionary conservation score 5406 for each amino acid 5408 as output. Each evolutionary conservation score is ranked by scale. For the purposes of this disclosure, “classifier,” “determiner,” and “insert term here” may include one or more software modules, one or more hardware modules, or any combination thereof.

In operation 5332, the pathogenicity determinant 3734 is configured to replace each nucleotide for a particular amino acid with each amino acid 5408 in an alternative representation of the protein, at least in part based on the evolutionary conservation of each amino acid 5408. Determine the pathogenicity of the variant.

Figure 55 illustrates one implementation of determining pathogenicity based on predicted evolution scores. The classifier 5516 classifies a nucleotide variant as pathogenic 5508 if the evolutionary conservation score generated by the evolutionary conservation determinant 5324 for the corresponding amino acid substitution is below a threshold. In one embodiment, the classifier 5516 classifies a nucleotide variant as pathogenic 5508 if the evolutionary conservation score generated by the evolutionary conservation determinant 5324 for the corresponding amino acid substitution is 0 (i.e., an indication of non-conservation). Classify as

The classifier 5516 classifies a nucleotide variant as positive 5528 if the evolutionary conservation score generated by the evolutionary conservation determinant 5324 for the corresponding amino acid substitution exceeds a threshold. In one embodiment, the classifier 5516 considers a nucleotide variant positive 5528 when the evolutionary conservation score generated by the evolutionary conservation determinant 5324 for the corresponding amino acid substitution is non-zero (i.e., an indication of conservation). Classify.

Figure 56 shows one implementation of training data 5600 used to train the evolutionary conservation determinant 5324. Evolutionary conservation determinant 5324 is trained on a conservative training set and a non-conservative training set. The conservation training set has each conserved protein sample (5602) for each conserved amino acid at each position in the proteome. The non-conserved training set has each non-conserved (or non-conserved) protein sample 5608 for each non-conserved amino acid at each position. In various embodiments, the proteome includes a human proteome and a non-human proteome, including a non-human primate proteome.

Each position each has a set of conserved amino acids and a set of non-conserved amino acids. A particular set of conserved amino acids for a particular position of a particular protein within a proteome includes at least one major allelic amino acid observed at that particular position across multiple species. In one embodiment, the major allelic amino acid is a reference amino acid (e.g., the REF allele 5612 spanning the positive protein sample 5622 and the REF allele 5662 spanning the positive protein sample 5682). . A particular set of conserved amino acids includes one or more minor allele amino acids observed at a particular position across multiple species (e.g., ALT allele 5632 observed across benign protein samples (5642, 5652, 5662) and benign ALT allele (5692) observed across protein samples (5695, 5696).

A specific set of non-conserved amino acids for a specific position includes amino acids that are not included in the specific set of conserved amino acids (e.g., unobserved ALT alleles (5618) spanning pathogenic protein samples (5622A-N) and an unobserved ALT allele (5668) spanning pathogenic protein samples (5682A-N).

In one embodiment, each position each has C conserved amino acids in the set of conserved amino acids. In this embodiment, each position has NC non-conserved amino acids in the set of non-conserved amino acids, and NC = 20-C. The conservation training set has CP conserved protein samples, where CP = number of each position * C. The non-conserved training set has NCP non-conserved protein samples, where NCP = number of each position * (20-C). In one embodiment C ranges from 1 to 10. In other implementations, C varies for each location. In another implementation, C is the same for some of each position.

In one embodiment, the proteome has 1 to 10 million positions. In this embodiment, each of the 1 to 10 million positions has C conserved amino acids in the set of conserved amino acids. At each of the 1 to 10 million positions there are NC non-conserved amino acids in the set of non-conserved amino acids (NC = 20-C). The conservation training set has CP conservation protein samples, CP = 1 to 10 million * C. The non-conserved training set has NCP non-conserved protein samples, with NCP = 1 to 10 million * (20-C).

In one implementation, evolutionary conservation determinant 5324 is trained with 20 to 200 million training iterations. In this implementation, the 20 to 200 million training iterations include 1 to 10 million training iterations using 1 to 10 million conserved protein samples and 19 to 190 million non-conserved protein samples. 19 to 190 million iterations were performed using

In another embodiment, the proteome has 1 to 10 million positions, so there are 1 to 10 million protein samples in the training set. In this implementation, the evolutionary conservation determinant 5324 is trained with 1 to 10 million training iterations using 1 to 10 million protein samples.

Each conserved and non-conserved protein sample has a respective gap space representation generated using each reference amino acid as the respective gap amino acid at each position. Evolutionary conservation determinant 5324 trains on a specific conserved protein sample, processes the specific gap space representation of the specific conserved protein sample as input, and generates an evolutionary conservation score for the specific conserved amino acid as output to determine the specific conserved protein sample. Estimate the evolutionary conservation of specific conserved amino acids at a position. A specific gap space representation is created by using a specific reference amino acid at a specific position as the gap amino acid and the remaining amino acids at the remaining positions in a specific conserved protein sample as non-gap amino acids.

Each preserved protein sample has an actual preserved label. The actual conservation label is the evolutionary conservation frequency. In one embodiment, the actual retention label is 1. The evolutionary conservation for a particular conserved amino acid is compared to the actual conservation label to determine the error and training techniques are used to improve the coefficients of the evolutionary conservation determinant 5324 based on the error. In one implementation, the training technique is a loss function based gradient update technique (e.g., backpropagation).

In some embodiments, when a particular conserved amino acid is a particular reference amino acid, the actual conserved label is masked and not used to determine the error. In this embodiment, masking prevents the evolutionarily conserved determinant 5324 from overfitting to a particular reference amino acid.

Evolutionary conservation determinant 5324 trains on specific non-conserved protein samples, processes specific gap space representations of specific non-conserved protein samples as input, and generates evolutionary conservation scores for specific non-conserved amino acids as output, Estimate the evolutionary conservation of specific non-conserved amino acids at specific positions in a specific non-conserved protein sample. A specific gap space representation is created by using a specific reference amino acid at a specific position as the gap amino acid and the remaining amino acids at the remaining positions in the specific non-conserved protein sample as the non-gap amino acids.

Each non-conserved protein sample has an actual non-conserved label. The actual non-conserved label is the evolutionary conservation frequency. In one implementation, the actual non-conserved label is 0. The evolutionary conservation score for a particular non-conserved amino acid is compared to the actual non-conserved label to determine the error and use training techniques (e.g., backpropagation) to improve the coefficients of the evolutionary conservation determinant 5324 based on the error. do.

Evolutionary conservation determinant 5324 is trained on the training set. The training set has each protein sample for each position in the proteome. Each protein sample has a respective gap space representation generated using each reference amino acid at each position as each gap amino acid.

Figure 57 shows one implementation of simultaneously training 5700 evolutionary conservation determinants for benign and pathogenic protein samples. The evolutionary conservation determinant 5324 trains on a specific protein sample, processes a specific gap space representation 5722 of the specific protein sample as input, and generates a respective evolutionary conservation score 5720 for each amino acid as output. It estimates the evolutionary conservation of each amino acid of each amino acid class at a specific position in a specific protein sample. A specific gap space representation 5722 is created by using a specific reference amino acid at a specific position as the gap amino acid and the remaining amino acids at the remaining positions in the specific protein sample as non-gap amino acids.

Each protein sample has an actual label for each amino acid. Each actual label includes one or more conserved (positive) labels for one or more conserved amino acids (5732, 5702, 5712) in each amino acid, and one or more non-conserved labels for one or more non-conserved amino acids in each amino acid. Includes (pathogenic) label. Conservation labels and non-conserved labels each have an evolutionary conservation frequency. Each evolutionary conservation frequency is ranked according to its magnitude. In one embodiment, the preserved label is 1 and the non-preserved label is 0.

In one embodiment, error 5704 causes each comparison of each evolutionary conservation score for each conserved amino acid to each non-conserved amino acid and each comparison of each evolutionary conservation score for each conserved amino acid to each non-conserved amino acid. Evolutionary conservation scores are determined based on each comparison. The coefficients of the evolutionary conservation determinant 5324 are improved based on the error using training techniques (e.g., backpropagation 5744).

In one embodiment, the conserved amino acids include specific reference amino acids, and the conserved labels for specific reference amino acids are masked and not used to determine errors. Masking prevents the evolutionarily conserved determinant 5324 from overfitting to a specific reference amino acid.

A synonymous mutation is a point mutation, meaning it is simply a miscopied DNA nucleotide that changes one base pair in the RNA copy of DNA. A codon in RNA is a set of three nucleotides that code for a specific amino acid. Most amino acids have multiple RNA codons that translate to a specific amino acid. In most cases, if the third nucleotide is mutated, it will code for the same amino acid. Like grammatical synonyms, the mutated codon has the same meaning as the original codon, so the amino acid does not change, so it is called a synonymous mutation. If the amino acids do not change, the protein is not affected. Synonymous mutations do not change anything and do not change anything. This means that the genes or proteins do not change in any way and therefore have no real role in the evolution of the species. Synonymous mutations are actually quite common, but they go unnoticed because they have no effect.

Nonsynonymous mutations have a much greater impact on an individual than synonymous mutations. In non-synonymous mutations, a single nucleotide is usually inserted or deleted in the sequence during transcription when messenger RNA copies DNA. This single missing or added nucleotide causes a frameshift mutation that throws the entire reading frame of the amino acid sequence and mixes codons. This usually affects the amino acids encoded and alters the resulting protein expressed. The severity of this type of mutation depends on how early it occurs in the amino acid sequence. If this occurs at the beginning and the entire protein is altered, this can be a lethal mutation. Another way non-synonymous mutations can occur is when a point mutation changes a single nucleotide to a codon that does not translate into the same amino acid. In many cases, single amino acid changes do not have a significant effect on the protein and are still viable. If it occurs early in the sequence and the codon is changed and converted into a stop signal, no protein is produced and this can have serious consequences. Sometimes non-synonymous mutations are actually positive changes. Natural selection may favor new expressions of genes, and individuals may have developed advantageous adaptations from mutations. If that mutation occurs in the gamete, this adaptation is passed on to the next generation of offspring. Non-synonymous mutations increase the diversity of the gene pool, allowing natural selection to drive evolution and operate at the microevolutionary level.

The nucleotide triplets that code for amino acids are called codons. Each group of three nucleotides encodes one amino acid. The code is degenerate because there are 64 combinations of 4 nucleotides taken 3 at a time and only 20 amino acids (more than one codon per amino acid in most cases). One example of an unreachable alternative amino acid class is a class of alternative amino acids that are not encoded by a synonymous SNP. Another example of an unattainable alternative amino acid class is the class of alternative amino acids limited by the number of triple nucleotide mutation combinations deviated by single nucleotide polymorphisms (SNPs) at the triple nucleotide position of the initial codon.

In one embodiment, the unreachable substitution amino acid class limited by the reachability of the SNP to convert the reference codon of the reference amino acid to the substitute amino acid of the unreachable substitution amino acid class is masked to the actual label. In this implementation, the masked amino acid classes are not lost at all and do not contribute to the gradient update. In one implementation, masked amino acid classes are identified in a lookup table. In one implementation, the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

Specific sets of conserved amino acids and specific sets of non-conserved amino acids are identified based on the evolutionary conservation profiles of homologous proteins in multiple species. In one embodiment, the evolutionary conservation profile of homologous proteins is determined using a position-specific frequency matrix (PSFM). In another embodiment, the evolutionary conservation profile of a homologous protein is determined using a position-specific scoring matrix (PSSM).

Figure 58 shows various implementations of actual label encoding used to train the evolutionary conservation determinant 5324. The actual label encoding 5802 labels the conserved amino acid classes A, C, and F using their evolutionary conservation frequencies (e.g., PSFM or PSSM) and the remaining non-conserved amino acid classes using the “0 value.” Specify . The actual label encoding 5812 is the same as the actual label encoding 5802, except that the actual label encoding 5812 "masks" the REF major allele/most conserved amino acid class F, which is the REF major allele/most conserved amino acid. Class F does not contribute to the training of the evolutionary conservation determinant 5324 (e.g., by setting the loss calculated with the loss function for REF major allele/maximum conserved amino acid class F to 0).

The actual label encoding 5822 uses a “1 value” to label the conserved amino acid classes A, C, and F, and a “0 value” to label the remaining non-conserved amino acid classes. The actual label encoding 5832 is the same as the actual label encoding 5822, except that the actual label encoding 5832 "masks" the REF major allele/most conserved amino acid class F, which is the REF major allele/most conserved amino acid. Class F does not contribute to the training of the evolutionary conservation determinant 5324 (e.g., by setting the loss calculated with the loss function for REF major allele/maximum conserved amino acid class F to 0).

Figure 59 shows an example PSFM 5900. Figure 60 shows an example PSSM 6000. Figure 61 shows one implementation of generating PSFM and PSSM. Figure 62 shows an example PSFM 6200 encoding. Figure 63 shows an example PSSM 6300 encoding.

Multiple sequence alignment (MSA) is a sequence alignment of multiple homologous protein sequences to a target protein. MSA is an important step in the comparative analysis and property prediction of biological sequences because a lot of information, such as evolution and coevolution clusters, can be generated from MSA and mapped to selected target sequences or protein structures.

The sequence profile of a protein sequence X of length L is an L × 20 matrix in the form PSSM or PSFM. The columns of PSSM and PSFM are indexed by amino acid alphabet, with each row corresponding to a position in the protein sequence. PSSM and PSFM contain the substitution score and frequency of amino acids at different positions in the protein sequence, respectively. Each row of PSFM is normalized so that the sum is 1. The sequence profile of a protein sequence Therefore, the sequence profile contains more general evolutionary and structural information of the protein family to which protein sequence X belongs, providing useful information for remote homology detection and fold recognition.

The protein sequence (referred to as the query sequence, e.g. the reference amino acid sequence of the protein) can be used as a seed to search and align homologous sequences from protein databases (e.g. SWISSPROT), for example using the PSI-BLAST program. . The aligned sequences share some homologous segments and belong to the same protein family. The aligned sequences are further converted into two profiles, PSSM and PSFM, to represent homogeneous information. Both PSSM and PSFM are matrices with 20 rows and L columns, where L is the total number of amino acids in the query sequence. Each column of PSSM represents the log likelihood of a residue substitution at the corresponding position in the query sequence. The (i, j)th entry of the PSSM matrix represents the probability that the amino acid at the jth position of the query sequence will mutate into amino acid type i during the evolution process. PSFM contains weighted observed frequencies for each position in the aligned sequence. Specifically, the (i, j)th entry of the PSFM matrix represents the probability that amino acid type i exists at position j in the query sequence.

Given a query sequence, it is first submitted to PSI-BLAST to obtain a sequence profile by searching and aligning homologous protein sequences in a protein database (e.g., Swiss-Prot Database). Figure 61 shows the process of obtaining a sequence profile using the PSI-BLAST program. The parameters h and j of PSI-BLAST are generally set to 0.001 and 3, respectively. The sequence profile of a protein encapsulates homolog information associated with the query protein sequence. In PSI-BLAST, homology information is expressed in two matrices: PSFM and PSSM. Examples of PSFM and PSSM are shown in Figures 62 and 63, respectively.

In Figure 62, the (1, u)th element (1 ∈ {1, 2, ..., Li}, u ∈ {1, 2, ..., 20}) is the uth element at the 1st position of the query protein. Indicates the possibility of amino acids being present. For example, the probability of amino acid M being in the first position of the query protein is 0.36.

In Figure 63, the (1, u)th element (1 ∈ {1, 2, ..., Li}, u ∈ {1, 2, ..., 20}) is the 1st position of the query protein in the evolution process. It represents the probability score that the amino acid in will be mutated to the uth amino acid. For example, during the evolution process, the score for mutating amino acid V at the first position of the query protein to H is -3, while amino acid V at the eighth position is -4.

Combined learning and transfer learning

Figure 64 illustrates two datasets on which the models disclosed herein can be trained, for example, through conjunctive learning (Figures 65A and 65B) or transfer learning (Figures 66A and 66B). The first training dataset is referred to as JigsawAI dataset 6406. The second training dataset is referred to as PrimateAI dataset 6408. JigsawAI dataset 6406 features voxel input 6412 missing a central residue identified as a gap amino acid, as discussed above. PrimateAI dataset 6408 features voxel input 6412 with no missing residues and complete input.

For the JigsawAI dataset 6406, the actual labels 6422 have missing or masked labels 6426 for gap amino acids (e.g., reference amino acids). For the PrimateAI dataset (6408), the actual label 6422 has 19 missing or masked labels (6436) for the remaining amino acids that are different from the alternative amino acid (benign or pathogenic) being analyzed. In one implementation, the number of samples 6432 in the JigsawAI dataset 6406 is 10 million (6436) and the number of samples in the PrimateAI dataset 6408 is 1 million (6438).

Figures 65A and 65B illustrate one implementation of joint learning 6500 of the model disclosed herein. In operation 6502 the spaced training set is accessed. The gap training set is also referred to herein as JigsawAI dataset 6406. The gap training set contains a sample of each gap protein for each position in the proteome. Each gap protein sample is labeled with the respective gap actual sequence. A specific gap true sequence for a specific gap protein sample has a positive label for a specific amino acid class corresponding to a reference amino acid at a specific position in the specific gap protein, and for each remaining amino acid class corresponding to an alternative amino acid at a specific position, respectively. It has a pathogenic label of

In operation 6512 the non-gap training set is accessed. The non-gap training set is also referred to herein as PrimateAI dataset 6408. The non-gap training set includes non-gap positive protein samples and non-gap pathogenic protein samples. Certain non-gap positive protein samples contain positive replacement amino acids at specific positions that are replaced by positive nucleotide variants. Certain non-gap pathogenic protein samples contain pathogenic replacement amino acids at specific positions that have been substituted by pathogenic nucleotide variants. A specific non-gap positive protein sample is labeled with a positive real sequence with a positive label for the specific amino acid class corresponding to the positive replacing amino acid and a respective masking label for each remaining amino acid class corresponding to the positive replacing amino acid and another amino acid. . A specific non-gap pathogenic protein sample is labeled with a positive authentic sequence with a pathogenic label for the specific amino acid class corresponding to the pathogenic substituted amino acid and a respective masking label for each remaining amino acid class corresponding to an amino acid different from the pathogenic substituted amino acid. .

In one embodiment, positive labels are masked for specific amino acid classes that correspond to reference amino acids at specific positions in specific gap proteins. In one embodiment, the non-gap positive protein sample is derived from common human and non-human primate nucleotide variants. In one embodiment, the non-gap pathogenic protein sample is derived from combinatorially simulated nucleotide variants.

At operation 6522, a respective gapped space representation is created for the gapped protein sample, and a separate non-gap spatial representation is created for the non-gap positive protein sample and the non-gap pathogenic protein sample.

In operation 6532, pathogenicity classifier 2108/2600/2700 is trained over one or more training cycles, and trained pathogenicity classifier 2108/2600/2700 is trained over one or more training cycles. ) parameters/coefficients/weights are generated as optimized results. Each training cycle uses as training examples a gapped space representation from each gapped space representation, and a non-gap spaced representation from each non-gap space representation.

In operation 6542, the pathogenicity of the variant is determined using the trained pathogenicity classifier (2108/2600/2700).

In one embodiment, the sample indicator indicates to the pathogenicity classifier (2108/2600/2700) whether the current training example is a gap space representation for gap protein samples or a non-gap space representation for non-gap protein samples. It is used to

In one implementation, the pathogenicity classifier 2108/2600/2700 generates output sequences per amino acid class in response to training example processing. The output sequence for each amino acid class has a pathogenicity score for each amino acid class.

In one implementation, the performance of the trained pathogenicity classifier 2108/2600/2700 is measured between training cycles on a validation set. In some implementations, the validation set includes pairs of gapped and non-gapped spatial representations for each retained protein sample.

In one embodiment, the trained pathogenicity classifier 2108/2600/2700 outputs a first amino acid class-specific output sequence for the gap space representation of the pair and a second amino acid class-specific output sequence for the non-gap space representation of the pair. Create. In some embodiments, the final pathogenicity score for a nucleotide variant causing an amino acid substitution in the retained protein sample is based on a combination of the first and second pathogenicity scores for amino acid substitutions in the output sequence by first and second amino acid classes. It is decided. In another embodiment, the final pathogenicity score is based on the average of the first and second pathogenicity scores.

In some implementations, at least some of the training cycles use the same number of gapped and non-gapped spatial representations. In another implementation, at least some of the training cycles use batches of training examples with equal numbers of gapped and non-gapped spatial representations.

In one implementation, the masked labels do not contribute to the error determination and therefore do not contribute to the training of the pathogenicity classifier (2108/2600/2700). In some implementations, masked labels are zeroed out.

In some implementations, gapped spatial representations are weighted differently than non-gap spatial representations, such that the pathogenicity classifier 2108/2600/2700 responds to a pathogenicity classifier 2108/2600/2700 that processes the non-gap spatial representation. The contribution of the gap space representation to the gradient update applied to the parameters of the pathogenicity classifier 2108/2600/2700 in response to the pathogenicity classifier 2108/2600/2700 processing the non-gap spatial representation. It varies from the contribution of the non-gap space representation to the gradient update applied to the parameters of . In one implementation, the variance is determined by predefined weights.

Figures 66A and 66B illustrate one implementation of using transfer learning 6600 to train a model disclosed herein using the two datasets shown in Figure 64. In operation 6602, pathogenicity classifier 2108/2600/2700 is first trained on the gap training set (i.e., JigsawAI data set 6406) to generate trained pathogenicity classifier 2108/2600/2700. .

In operation 6612, the trained pathogenicity classifier 2108/2600/2700 is further trained on the non-gap training set (i.e., PrimateAI data set 6408) to produce a retrained pathogenicity classifier 2108/2600. /2700).

In operation 6622, the pathogenicity of the variant is determined using the retrained pathogenicity classifier (2108/2600/2700).

In operation 6632, the performance of the trained pathogenicity classifier 2108/2600/2700 is measured between training cycles on a first validation set containing only non-gap spatial representations of the retained protein samples. In another embodiment, the performance of the retrained pathogenicity classifier 2108/2600/2700 is measured between training cycles on a second validation set comprising gapped and non-gapped spatial representations of the retained protein samples. .

At operation 6642, the retrained pathogenicity classifier 2108/2600/2700, in response to processing the pair, generates a first amino acid class-specific output sequence for the pair. In one embodiment, the final pathogenicity score for nucleotide variants causing amino acid substitutions in the corresponding maintained protein sample is determined based on the output sequence by first amino acid class.

Generate training data and training labels

Figure 67 shows one implementation of generating training data and labels (6700) to train the model disclosed herein.

Proteome accessor 6704 accesses multiple amino acid positions in a proteome with multiple proteins.

Reference designator 6714 specifies the major allele amino acid at a number of amino acid positions as a reference amino acid for multiple proteins.

Positive marker 6724 classifies these nucleotide substitutions as positive variants that substitute a specific reference amino acid for each amino acid position among a plurality of amino acid positions for a specific reference amino acid at a specific amino acid position in a specific alternative representation of a specific protein.

For each amino acid position in a number of amino acid positions, pathogenicity marker 6734 classifies this nucleotide substitution as a pathogenic variant that replaces a specific reference amino acid with an alternative amino acid at a specific amino acid position. A substitute amino acid is different from a specific reference amino acid.

Trainer 6744 trains a variant pathogenicity classifier 2108/2600/2700 on training data containing a spatial representation of the protein sample, such that the spatial representation includes the actual benign labels corresponding to the benign variants and the actual benign labels corresponding to the pathogenic variants. A pathogenic label is assigned.

In one embodiment, a variant pathogenicity classifier 2108/2600/2700 is trained to determine whether a substitution of a first amino acid for a second amino acid at a given amino acid position in a protein is pathogenic or benign. In this implementation, a variant pathogenicity classifier (2108/2600/2700) is trained to generate pathogenicity scores for substitutions. In one embodiment, a variant pathogenicity classifier (2108/2600/2700) is trained to determine whether each substitution of the first amino acid for each amino acid at a given amino acid position in the protein is pathogenic or benign. In this implementation, a variant pathogenicity classifier (2108/2600/2700) is trained to generate a respective pathogenicity score for each substitution. In some embodiments, each amino acid corresponds to each of the 20 naturally occurring amino acids. In another embodiment, each amino acid corresponds to a respective naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

In one embodiment, a variant pathogenicity classifier (2108/2600/2700) is trained to determine whether an insertion of an amino acid at a given empty amino acid position in a protein is pathogenic or benign. In this implementation, a variant pathogenicity classifier (2108/2600/2700) is trained to generate a pathogenicity score for the insertion. In one embodiment, a variant pathogenicity classifier (2108/2600/2700) is trained to determine whether each insertion of each amino acid at a given empty amino acid position in a protein is pathogenic or benign. In this implementation, a variant pathogenicity classifier (2108/2600/2700) is trained to generate a respective pathogenicity score for each insertion. In some embodiments, each amino acid corresponds to each of the 20 naturally occurring amino acids. In another embodiment, each amino acid corresponds to a respective naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

In one embodiment, a variant pathogenicity classifier (2108/2600/2700) is trained to determine whether a substitution of a first amino acid for a second amino acid at a given amino acid position in the protein is spatially tolerated by other amino acids in the protein. . In this implementation, a variant pathogenicity classifier (2108/2600/2700) is trained to generate a spatial tolerance score for substitutions. In one embodiment, the variant pathogenicity classifier 2108/2600/2700 is configured to determine whether each substitution of a first amino acid at a given amino acid position in the protein is spatially tolerated by the other amino acids in the protein. trained. In this implementation, a variant pathogenicity classifier (2108/2600/2700) is trained to generate a respective spatial tolerance score for each substitution. In some embodiments, each amino acid corresponds to each of the 20 naturally occurring amino acids. In another embodiment, each amino acid corresponds to a respective naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

In one embodiment, a variant pathogenicity classifier (2108/2600/2700) is trained to determine whether insertion of an amino acid at a given empty amino acid position in a protein is spatially permitted by other amino acids in the protein. In this implementation, a variant pathogenicity classifier (2108/2600/2700) is trained to generate a spatial tolerance score for insertions. In one embodiment, a variant pathogenicity classifier (2108/2600/2700) is trained to determine whether each insertion of each amino acid at a given empty amino acid position in a protein is spatially tolerated by other amino acids in the protein. In this implementation, a variant pathogenicity classifier (2108/2600/2700) is trained to generate a respective spatial tolerance score for each insertion. In some embodiments, each amino acid corresponds to each of the 20 naturally occurring amino acids. In another embodiment, each amino acid corresponds to a respective naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

In one embodiment, a variant pathogenicity classifier (2108/2600/2700) is trained to determine whether a substitution of a first amino acid for a second amino acid at a given amino acid position in a protein is evolutionarily conserved or non-conserved. . In this implementation, a variant pathogenicity classifier (2108/2600/2700) is trained to generate evolutionary conservation scores for substitutions. In one embodiment, a variant pathogenicity classifier (2108/2600/2700) is trained to determine whether each substitution of the first amino acid for each amino acid at a given amino acid position in the protein is evolutionarily conserved or non-conserved. do. In this implementation, a variant pathogenicity classifier (2108/2600/2700) is trained to generate a respective evolutionary conservation score for each substitution. In some embodiments, each amino acid corresponds to each of the 20 naturally occurring amino acids. In another embodiment, each amino acid corresponds to a respective naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

In one embodiment, a variant pathogenicity classifier (2108/2600/2700) is trained to determine whether an insertion of an amino acid at a given empty amino acid position in a protein is evolutionarily conserved or non-conserved. In this implementation, a variant pathogenicity classifier (2108/2600/2700) is trained to generate an evolutionary conservation score for the insertion.

In one embodiment, a variant pathogenicity classifier (2108/2600/2700) is trained to determine whether each insertion of each amino acid at a given empty amino acid position in a protein is evolutionarily conserved or non-conserved. In this implementation, a variant pathogenicity classifier (2108/2600/2700) is trained to generate a respective evolutionary conservation score for each insertion. In some embodiments, each amino acid corresponds to each of the 20 naturally occurring amino acids. In another embodiment, each amino acid corresponds to a respective naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

In various implementations, the spatial tolerance corresponds to the structural tolerance, and the spatial tolerance corresponds to the structural tolerance. In various embodiments, the number of amino acid positions ranges from 1 million to 10 million amino acid positions. In various embodiments, the number of amino acid positions ranges from 10 million to hundreds of millions of amino acid positions. In various embodiments, the number of amino acid positions ranges from hundreds of millions to billions of amino acid positions. In various embodiments, the number of amino acid positions ranges from 1 to 1 million amino acid positions.

In one embodiment, the unreachable substitution amino acid class, which is limited by the reachability of single nucleotide polymorphisms (SNPs) to convert the reference codon of the reference amino acid to the substitute amino acid of the unreachable substitution amino acid class, is masked to the actual label. In this implementation, the masked amino acid classes are not lost at all and do not contribute to the gradient update. In this implementation, masked amino acid classes are identified in a lookup table. In this implementation, a lookup table identifies a set of masked amino acid classes for each reference amino acid position.

In various embodiments, the spatial representation is a structural representation of the protein structure of the protein sample. In various implementations, the spatial representation is encoded using voxelization.

Pathogenicity Determination

Figure 68 depicts one implementation of a method 6800 for determining pathogenicity of a nucleotide variant. The method includes accessing a spatial representation of the protein in operation 6802. The spatial representation of a protein specifies the respective spatial configuration of each amino acid at each position in the protein.

The method includes removing specific spatial configurations of specific amino acids at specific positions from the spatial representation of the protein in operation 6812 to create a gap spatial representation of the protein. In one implementation, removal of specific spatial configurations is implemented (or automated) by a script.

The method includes, at operation 6822, determining the pathogenicity of a nucleotide variant based, at least in part, on a gap space representation and a representation of a replacement amino acid produced by the nucleotide variant at a particular location.

Structural resistance prediction

Figure 69 depicts one implementation of a system 6900 for predicting structural tolerance of amino acid substitutions. In operation 6902, gapping logic is configured to remove a specific amino acid at a specific position in the spatial representation of the protein and create an amino acid vacancy at a specific position in the spatial representation of the protein.

In operation 6912, the structural resistance prediction logic is configured to process the spatial representation of the protein with amino acid vacancies, and rank the structural resistance of substituted amino acids that are candidates for filling the amino acid vacancies based on amino acid co-occurrence patterns near the amino acid vacancies. Rank it.

Performance outcomes as objective indicators of originality and non-obviousness.

The variant pathogenicity classifier disclosed herein makes pathogenicity predictions based on 3D protein structure and is referred to as “PrimateAI 3D”. “Primate AI” is a commonly owned and previously disclosed variant pathogenicity classifier that generates pathogenicity prediction-based protein sequences. For more information about PrimateAI, see commonly owned U.S. patent application Ser. No. 16/160,903; No. 16/160,986; No. 16/160,968; and 16/407,149 and Sundaram, L. et al. Predicting the clinical impact of human mutation with deep neural networks. Nat. Genet. 50, 1161-1170 (2018)].

Variant pathogenicity classifiers trained using the transfer learning techniques disclosed herein (Figures 66A and 66B) are referred to as “transfer learning.” Variant pathogenicity classifiers trained using the combinational learning techniques disclosed herein (Figures 65A and 65B) are referred to as “joint learning”.

The performance results in Figures 70A, 70B and 70C are generated from a classification task that accurately distinguishes benign variants from pathogenic variants across multiple validation sets. New Delayed Developmental Disorder (New DDD) is an example of a validation set used to compare the classification accuracy of transfer learning against primate AI 3D and combinational learning against primate AI. The new DDD validation set marks variants in individuals with DDD as pathogenic and marks the same variants in healthy relatives of individuals with DDD as benign. A similar labeling scheme is used with the autism spectrum disorder (ASD) validation set.

BRCA1 is another example of a validation set used to compare the classification accuracy of transfer learning against primate AI 3D and combinational learning against primate AI. The BRCA1 validation set marks synthetically generated reference amino acid sequences that simulate proteins from the BRCA1 gene as benign variants, and synthetically altered allelic amino acid sequences that simulate proteins from the BRCA1 gene as pathogenic variants. A similar labeling scheme is used with different validation sets of the TP53 gene, TP53S3 gene and variants thereof, and other genes and variants thereof shown in Figures 70A, 70B and 70C.

In Figures 70A, 70B and 70C, the y-axis has p-values and the x-axis has different verification sets. As can be seen from the p-values in Figures 70A, 70B and 70C, combinational learning generally outperforms the other approaches, followed by transfer learning and PrimateAI 3D. Larger p-values, i.e. longer vertical bars, indicate greater accuracy in distinguishing benign variants from pathogenic variants. In Figures 70A, 70B, and 70C, the vertical bars of conjunctive learning are consistently longer than those of the other approaches.

Additionally, in Figures 70A, 70B and 70C, a separate "Mean" chart calculates the average of the p-values determined for each of the validation sets. The averages chart also shows that Combined Learning generally outperforms the other approaches, followed by Transfer Learning, and then PrimateAI 3D, in that the horizontal bars for Combined Learning are consistently longer than those for the other approaches. Able to know.

Average statistics may be biased by outliers. To address this, separate “method rank” charts are also shown in FIGS. 70A, 70B and 70C. Higher ranks indicate poorer classification accuracy. The method ranking chart also shows that combinational learning generally outperforms other approaches, followed by transfer learning and PrimateAI 3D. In a method ranking chart, having more low ranked 1s and 2s is better than having higher ranked 3s.

article

The disclosed technology may be practiced as a system, method, or article of manufacture. One or more features of an implementation may be combined with the base implementation. Implementations that are not mutually exclusive are taught as combinable. One or more features of an implementation may be combined with other implementations. The present invention periodically reminds users of these options. Omission from some implementations of citations repeating these options should not be considered as limiting the combinations taught in the preceding sections - these citations are hereby incorporated by reference into each of the following implementations.

One or more implementations and provisions of the disclosed technology or elements thereof may be implemented in the form of a computer product comprising a non-transitory computer-readable storage medium having computer-usable program code for performing the disclosed method steps. Moreover, one or more implementations and provisions of the disclosed technology or elements thereof may be implemented in the form of a device that includes a memory and at least one processor coupled to the memory and operative to perform the example method steps. . Additionally, in other aspects, one or more embodiments and provisions of the disclosed technology or elements thereof may be embodied in the form of means for performing one or more of the method steps described herein; The means may comprise (i) hardware module(s), (ii) software module(s) executing on one or more hardware processors, or (iii) a combination of hardware and software modules; Any of (i) through (iii) implements specific techniques set forth herein, and the software modules are stored on a computer-readable storage medium (or a plurality of such media).

The provisions described in this section may be combined into features. For brevity, combinations of features are not listed individually and are not repeated in each basic feature set. The reader will understand how the features identified in the provisions described in this section can be easily combined with the basic feature sets identified as implementations in other sections of this application. These provisions are not intended to be mutually exclusive, inclusive or limiting; The disclosed technology is not limited to these provisions, but rather includes all possible combinations, modifications and variations within the scope of the claimed technology and its equivalents.

Other implementations of the provisions described in this section may include a non-transitory computer-readable storage medium storing instructions executable by a processor to perform any of the provisions described in this section. Another implementation of the provisions described in this section may include a system including a memory and one or more processors operable to execute instructions stored in the memory to perform any of the provisions described in this section.

We set forth the following provisions:

Clause Set 1 (ILLM 1050-2)

1. A computer-implemented method for determining the pathogenicity of a nucleotide variant, comprising:

Accessing proteins with each amino acid at each position;

Designating a specific amino acid at a specific position of the protein as a gap amino acid, and designating the amino acid remaining at the remaining position of the protein as a non-gap amino acid;

generating a gap spatial representation of the protein including the spatial configuration of non-gap amino acids and excluding the spatial configuration of gap amino acids; and

A computer-implemented method comprising determining the pathogenicity of a nucleotide variant based, at least in part, on the gap space representation and the representation of alternative amino acids produced by the nucleotide variant at a particular position.

2. Clause 1, wherein the spatial configuration of non-gap amino acids is encoded as a distance channel per amino acid class,

The distance channel for each amino acid class has a voxel-specific distance value for voxels within a plurality of voxels,

A computer implemented method, wherein the voxel-wise distance value specifies the distance of an atom of a non-gap amino acid from a corresponding voxel within the plurality of voxels.

3. The computer implemented method of clause 2, wherein the spatial configuration of the non-gap amino acid is determined based on the spatial proximity between corresponding voxels and atoms of the non-gap amino acid.

4. The computer-implemented method of clause 2, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the distance of the atom of the gap amino acid from the corresponding voxel when determining the voxel-wise distance value.

5. The computer-implemented method of clause 4, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the spatial proximity between the corresponding voxel and the atoms of the gap amino acid.

6. The computer-implemented method of clause 1, wherein the particular amino acid is a reference amino acid that is the major allele of the protein.

7. The computer-implemented method of clause 1, wherein the pathogenicity predictor determines the pathogenicity of a nucleotide by the following steps:

Gap space representation and

Processing the representation of the replacement amino acid as input; and

Generating pathogenicity scores for alternative amino acids as output.

8. The computer-implemented method of clause 7, wherein the pathogenicity predictor is trained on a benign training set.

9. The computer-implemented method of clause 8, wherein the positive training set has each positive protein sample for each reference amino acid at each position in the proteome.

10. The computer-implemented method of clause 9, wherein the reference amino acid is a major allelic amino acid in the proteome.

11. The computer-implemented method of clause 10, wherein the proteome has 10 million positions and therefore the positive training set has 10 million positive protein samples.

12. The computer-implemented method of clause 11, wherein each positive protein sample has a respective gap space representation generated using each reference amino acid as each gap amino acid.

13. The computer-implemented method of clause 12, wherein each positive protein sample has as each replacement amino acid a respective representation of each reference amino acid.

14. The computer-implemented method of clause 13, wherein the pathogenicity predictor is trained on specific positive protein samples and estimates the pathogenicity of specific reference amino acids at specific positions in the specific positive protein samples by the following steps:

(i) Specific gap space representation of specific positive protein samples

- At this time, the specific gap space expression is

using specific reference amino acids as gap amino acids, and

Generated by using the remaining amino acids at the remaining positions in a specific positive protein sample as non-gap amino acids -; and

(ii) processing as input a representation of a specific reference amino acid as a specific replacement amino acid; and

Generating as output a pathogenicity score for a specific reference amino acid.

15. The computer-implemented method of clause 14, wherein each positive protein sample has an actual positive label indicating the absolute positivity of the positive protein sample.

16. The computer-implemented method of clause 15, wherein the actual positive label is 0.

17. The computer-implemented method of clause 16, wherein the pathogenicity score for a particular reference amino acid is compared to the actual positive label to determine the error and using training techniques to improve the coefficients of the pathogenicity predictor based on the error.

18. The computer-implemented method of clause 1, wherein the pathogenicity predictor is trained on a pathogenicity training set.

19. The computer-implemented method of clause 18, wherein the pathogenicity training set has a respective pathogenic protein sample for each combinatorially generated amino acid substitution for each reference amino acid at each position in the proteome.

20. The computer-implemented method of clause 19, wherein the combinatorially generated amino acid substitutions for a particular reference amino acid of a particular amino acid class at a particular position in the proteome comprise each replacement amino acid of the particular amino acid class and each other amino acid class. .

21. Clause 20: The proteome has 10 million positions, and for each 10 million positions there are 19 combinatorially generated amino acid substitutions, so the pathogenicity training set has 190 million pathogenic protein samples. How to implement it.

22. The computer-implemented method of clause 21, wherein each pathogenic protein sample has a respective gap space representation generated using each reference amino acid as each gap amino acid.

23. The method of clause 22, wherein each pathogenic protein sample is a respective representation of each combinatorially generated amino acid substitution as each replacement amino acid produced by each combinatorially generated nucleotide variant at each position of the proteome. , a computer implemented method.

24. The method of clause 23, wherein the pathogenicity predictor is a computer that trains on specific pathogenic protein samples and estimates the pathogenicity of specific combinatorially generated amino acid substitutions for specific reference amino acids at specific positions in the specific pathogenic protein sample by the following steps: How to implement:

(i) a specific gap space representation of a specific pathogenic protein sample, where the specific gap space representation is

using specific reference amino acids as gap amino acids, and

Generated by using the remaining amino acids at the remaining positions in a specific pathogenic protein sample as non-gap amino acids -; and

(ii) processing as input a representation of a specific combinatorially produced amino acid substitution as a specific replacement amino acid; and

Generating as output a pathogenicity score for specific combinatorially generated amino acid substitutions.

25. The computer-implemented method of clause 24, wherein each pathogenic protein sample has an actual pathogenicity label indicating the absolute pathogenicity of the pathogenic protein sample.

26. Computer-implemented method of clause 25, wherein the actual pathogenicity label is 1.

27. The computer-implemented method of clause 26, wherein the pathogenicity score for a particular combinatorially generated amino acid substitution is compared to the actual pathogenicity label to determine the error and uses training techniques to improve the coefficients of the pathogenicity predictor based on the error. .

28. Clause 27, wherein the pathogenicity predictor is trained with 200 million training iterations,

200 million training repetitions

10 million training iterations using 10 million positive protein samples, and

A computer-implemented method involving 190 million iterations using 190 million pathogenic protein samples.

29. Clause 10: The proteome has 1 to 10 million positions, so the positive training set has 1 to 10 million positive protein samples;

A computer-implemented method, wherein there are 19 combinatorially generated amino acid substitutions for each of the 1 to 10 million positions, so that the pathogenicity training set has 19 to 190 million pathogenic protein samples.

30. Clause 29, wherein the pathogenicity predictor is trained with 200 million to 200 million training iterations, and

20 to 200 million training repetitions

1 to 10 million training iterations using 1 to 10 million positive protein samples, and

A computer-implemented method comprising 19 to 190 million iterations using 19 to 190 million pathogenic protein samples.

31. The computer-implemented method of clause 6, wherein the replacement amino acid is an amino acid identical to the reference amino acid.

32. The computer-implemented method of clause 31, wherein the replacement amino acid is a different amino acid than the reference amino acid.

33. The method of clause 32, wherein the pathogenicity predictor generates a first pathogenicity score for a first replacement amino acid that is the same as the first reference amino acid, and the pathogenicity predictor generates a second pathogenicity score for a second replacement amino acid that is different from the first reference amino acid. Generating, computer-implemented method.

34. The computer-implemented method of clause 33, wherein the final pathogenicity score for the second replacement amino acid is the second pathogenicity score.

35. The computer-implemented method of clause 34, wherein the final pathogenicity score for the second replacement amino acid is based on a combination of the first pathogenicity score and the second pathogenicity score.

36. The computer implemented method of clause 35, wherein the final pathogenicity score for the second replacement amino acid is the ratio of the second pathogenicity score to the sum of the first pathogenicity score and the second pathogenicity score.

37. The computer-implemented method of clause 36, wherein the final pathogenicity score for the second replacement amino acid is determined by subtracting the first pathogenicity score from the second pathogenicity score.

38. The computer-implemented method of clause 1, wherein the spatial organization of non-gap amino acids is encoded into an evolutionary profile channel based on the pan-amino acid conservation frequency of the amino acid with the closest atom to the voxel.

39. The computer-implemented method of clause 38, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the nearest atoms of the gap amino acid when determining the pan-amino acid conservation frequency.

40. The computer-implemented method of clause 1, wherein the spatial organization of non-gap amino acids is encoded in an evolutionary profile channel based on the per-amino acid conservation frequency of each amino acid having the atom closest to the voxel.

41. The computer-implemented method of clause 40, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring each nearest atom of the gap amino acid when determining the conservation frequency per amino acid.

42. The computer-implemented method of clause 1, wherein the spatial configuration of non-gap amino acids is encoded as an annotation channel.

43. The computer-implemented method of clause 42, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the annotation channel.

44. The computer-implemented method of clause 1, wherein the spatial configuration of non-gap amino acids is encoded as a structural confidence channel.

45. The computer-implemented method of clause 44, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the structural confidence channel.

46. The computer-implemented method of clause 1, wherein the spatial configuration of non-gap amino acids is encoded as an additional input channel.

47. The computer-implemented method of clause 46, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining additional input channels.

48. The computer-implemented method of clause 9, wherein the proteome comprises a non-human proteome, including a human proteome and a non-human primate proteome.

49. Clause 7, wherein the class of unreachable substitute amino acids limited by the reachability of a single nucleotide polymorphism (SNP) to convert the reference codon of the reference amino acid to a substitute amino acid of the class of unreachable substitute amino acids is masked to the actual label. Computer implementation method.

50. The computer-implemented method of clause 1, wherein the masked amino acid class causes no loss and does not contribute to the gradient update.

51. The computer-implemented method of clause 50, wherein the masked amino acid classes are identified in a lookup table.

52. The computer-implemented method of clause 51, wherein the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

Clause Set 2

1. A computer-implemented method for determining the pathogenicity of a nucleotide variant, comprising:

Accessing proteins with each amino acid at each position;

Designating a specific amino acid of a specific amino acid class at a specific position of the protein as a gap amino acid, and designating the remaining amino acid at the remaining position of the protein as a non-gap amino acid;

generating a gap spatial representation of the protein including the spatial configuration of non-gap amino acids and excluding the spatial configuration of gap amino acids; and

A computer-implemented method comprising determining the pathogenicity of each substituted amino acid at a particular position based at least in part on the gap space representation.

2. Clause 1, wherein the spatial configuration of non-gap amino acids is encoded as a distance channel per amino acid class,

The distance channel for each amino acid class has a voxel-specific distance value for voxels within a plurality of voxels,

A computer implemented method, wherein the voxel-wise distance value specifies the distance of an atom of a non-gap amino acid from a corresponding voxel within the plurality of voxels.

3. The computer implemented method of clause 2, wherein the spatial configuration of the non-gap amino acid is determined based on the spatial proximity between corresponding voxels and atoms of the non-gap amino acid.

4. The computer-implemented method of clause 2, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the distance of the atom of the gap amino acid from the corresponding voxel when determining the voxel-wise distance value.

5. The computer-implemented method of clause 4, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the spatial proximity between the corresponding voxel and the atoms of the gap amino acid.

6. The computer-implemented method of clause 1, wherein the particular amino acid is a reference amino acid that is the major allele of the protein.

7. The computer-implemented method of clause 1, wherein each replacement amino acid is each combinatorially generated replacement amino acid produced by each combinatorially generated nucleotide variant at a particular position.

8. The computer-implemented method of clause 1, wherein the pathogenicity predictor determines the pathogenicity of each alternative amino acid by the following steps:

processing the gap space representation as input; and

Generating each pathogenicity score for each amino acid class as output.

9. The computer-implemented method of clause 8, wherein the pathogenicity predictor is trained on a training set.

10. The computer-implemented method of clause 9, wherein the training set has a respective protein sample for each position in the proteome.

11. The computer-implemented method of clause 10, wherein the proteome has 10 million positions and therefore the training set has 10 million protein samples.

12. The computer-implemented method of clause 11, wherein each protein sample has a respective gap space representation generated using each reference amino acid at each position in the proteome as each gap amino acid.

13. The computer-implemented method of clause 12, wherein the reference amino acid is a major allelic amino acid in the proteome.

14. The computer-implemented method of clause 13, wherein the pathogenicity predictor is trained on a specific protein sample and estimates the pathogenicity of each alternative amino acid for a specific reference amino acid at a specific position in the specific protein sample by the following steps:

Processing a specific gap space representation of a specific protein sample as input

- At this time, the specific gap space expression is

using specific reference amino acids as gap amino acids, and

Generated by using the remaining amino acids at the remaining positions in a specific protein sample as non-gap amino acids -; and

Generating each pathogenicity score for each amino acid class as output.

15. The computer-implemented method of clause 14, wherein each protein sample has an actual label for each amino acid class.

16. The clause 15, wherein each actual label comprises an absolute benign label for each reference amino acid class of each amino acid class, and each absolute pathogenic label for each alternative amino acid class of each amino acid class, Computer implementation method.

17. The computer-implemented method of clause 16, wherein the absolute positive label is 0.

18. The computer-implemented method of clause 17, wherein the absolute pathogenicity label is the same for each alternative amino acid class.

19. Computer-implemented method according to clause 18, wherein the absolute pathogenicity label is 1.

20. Clause 1, wherein comparison of pathogenicity scores for reference amino acid classes to absolute positive labels, and

A computer-implemented method, wherein the error is determined based on a respective comparison of each pathogenicity score for each alternative amino acid class to each absolute pathogenicity label.

21. The computer-implemented method of clause 20, wherein the coefficients of the pathogenicity predictor are improved based on error using training techniques.

22. The computer-implemented method of clause 21, wherein the pathogenicity predictor is trained in 10 million training iterations using 10 million protein samples.

23. The computer-implemented method of clause 8, wherein each amino acid class corresponds to each of the 20 naturally occurring amino acids.

24. The computer-implemented method of clause 23, wherein each amino acid class corresponds to a respective naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

25. Clause 11, the proteome has 1 to 10 million positions, so the training set has 1 to 10 million protein samples, and the pathogenicity predictor uses 1 to 10 million protein samples. A computer-implemented method, trained with 10 million training repetitions.

26. The method of clause 8, wherein the pathogenicity predictor generates a baseline pathogenicity score for a first alternative amino acid of a reference amino acid class, and the pathogenicity predictor generates a respective alternative pathogenicity score for each alternative amino acid of each alternative amino acid class. , computer implementation method.

27. The computer-implemented method of clause 26, wherein each final alternative pathogenicity score for each alternative amino acid is the respective alternative pathogenicity score.

28. The computer-implemented method of clause 27, wherein each final alternative pathogenicity score for each alternative amino acid is based on a respective combination of the baseline pathogenicity score and each alternative pathogenicity score.

29. The computer implemented method of clause 28, wherein each final alternative pathogenicity score for each alternative amino acid is a respective ratio of each alternative pathogenicity score to the sum of the baseline pathogenicity score and each alternative pathogenicity score.

30. The computer-implemented method of clause 29, wherein each final alternative pathogenicity score for each alternative amino acid is determined by subtracting each alternative pathogenicity score from each of the baseline pathogenicity scores.

31. The computer-implemented method of clause 8, wherein the pathogenicity predictor has an output layer that generates a respective pathogenicity score.

32. The computer-implemented method of clause 31, wherein the output layer is a normalization layer.

33. The computer-implemented method of clause 32, wherein each pathogenicity score is normalized.

34. The computer-implemented method of clause 31, wherein the output layer is a softmax layer.

35. The computer-implemented method of clause 34, wherein each pathogenicity score is exponentially normalized.

36. The computer-implemented method of clause 31, wherein the output layer has respective sigmoid units each generating a respective pathogenicity score.

37. The computer-implemented method of clause 31, wherein each pathogenicity score is not normalized.

38. The computer-implemented method of clause 1, wherein the spatial organization of non-gap amino acids is encoded into an evolutionary profile channel based on the pan-amino acid conservation frequency of the amino acid with the closest atom to the voxel.

39. The computer-implemented method of clause 38, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the nearest atoms of the gap amino acid when determining the pan-amino acid conservation frequency.

40. The computer-implemented method of clause 1, wherein the spatial organization of non-gap amino acids is encoded in an evolutionary profile channel based on the per-amino acid conservation frequency of each amino acid having the atom closest to the voxel.

41. The computer-implemented method of clause 40, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring each nearest atom of the gap amino acid when determining the conservation frequency per amino acid.

42. The computer-implemented method of clause 1, wherein the spatial configuration of non-gap amino acids is encoded as an annotation channel.

43. The computer-implemented method of clause 42, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the annotation channel.

44. The computer-implemented method of clause 1, wherein the spatial configuration of non-gap amino acids is encoded as a structural confidence channel.

45. The computer-implemented method of clause 44, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the structural confidence channel.

46. The computer-implemented method of clause 1, wherein the spatial configuration of non-gap amino acids is encoded as an additional input channel.

47. The computer-implemented method of clause 46, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining additional input channels.

48. The computer-implemented method of clause 10, wherein the proteome comprises a non-human proteome, including a human proteome and a non-human primate proteome.

49. Clause 8, wherein the unreachable substitute amino acid class is limited by the reachability of a single nucleotide polymorphism (SNP) to convert the reference codon of the reference amino acid to a substitute amino acid of the unreachable substitute amino acid class, wherein the unreachable substitute amino acid class is masked to the actual label. Computer implementation method.

50. The computer-implemented method of clause 1, wherein the masked amino acid class causes no loss and does not contribute to the gradient update.

51. The computer-implemented method of clause 50, wherein the masked amino acid classes are identified in a lookup table.

52. The computer-implemented method of clause 51, wherein the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

Clause set 3

1. A computer-implemented method for generating training data for training a variant pathogenicity classifier, comprising:

Accessing multiple amino acid positions in the proteome with multiple proteins;

Designating major allele amino acids at multiple amino acid positions as reference amino acids for multiple proteins;

For each amino acid position among multiple amino acid positions,

Classifying these nucleotide substitutions as positive variants that substitute a specific reference amino acid for a specific reference amino acid at a specific amino acid position in a specific alternative representation of a specific protein, and

Classifying these nucleotide substitutions as pathogenic variants in which a specific reference amino acid is replaced by a replacement amino acid at a specific amino acid position - the replacement amino acid is different from the specific reference amino acid; and

A computer-implemented method, comprising training a variant pathogenicity classifier using benign variants and pathogenic variants as training data.

2. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether a substitution of a first amino acid for a second amino acid at a given amino acid position in the protein is pathogenic or benign.

3. The computer-implemented method of clause 2, wherein a variant pathogenicity classifier is trained to generate pathogenicity scores for substitutions.

4. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether each substitution of a first amino acid for each amino acid at a given amino acid position in the protein is pathogenic or benign.

5. The computer-implemented method of clause 4, wherein a variant pathogenicity classifier is trained to generate a respective pathogenicity score for each substitution.

6. The computer-implemented method of clause 5, wherein each amino acid corresponds to each of the 20 naturally occurring amino acids.

7. The computer-implemented method of clause 6, wherein each amino acid corresponds to each naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

8. The computer-implemented method of clause 1, wherein a variant pathogenicity classifier is trained to determine whether an insertion of an amino acid at a given empty amino acid position in the protein is pathogenic or benign.

9. The computer-implemented method of clause 8, wherein a variant pathogenicity classifier is trained to generate a pathogenicity score for the insertion.

10. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether each insertion of each amino acid at a given empty amino acid position in the protein is pathogenic or benign.

11. The computer-implemented method of clause 10, wherein a variant pathogenicity classifier is trained to generate a respective pathogenicity score for each insertion.

12. The computer-implemented method of clause 11, wherein each amino acid corresponds to each of the 20 naturally occurring amino acids.

13. The computer-implemented method of clause 12, wherein each amino acid corresponds to each naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

14. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether a substitution of a first amino acid for a second amino acid at a given amino acid position in the protein is spatially tolerated by other amino acids in the protein.

15. The computer-implemented method of clause 14, wherein a variant pathogenicity classifier is trained to generate a spatial resistance score for substitutions.

16. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether each substitution of the first amino acid at a given amino acid position in the protein is spatially tolerated by the other amino acids in the protein. .

17. The computer-implemented method of clause 16, wherein the variant pathogenicity classifier is trained to generate a respective spatial resistance score for each substitution.

18. The computer-implemented method of clause 17, wherein each amino acid corresponds to each of the 20 naturally occurring amino acids.

19. The computer-implemented method of clause 18, wherein each amino acid corresponds to each naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

20. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether insertion of an amino acid at a given empty amino acid position in the protein is spatially permitted by other amino acids in the protein.

21. The computer-implemented method of clause 20, wherein a variant pathogenicity classifier is trained to generate a spatial resistance score for insertions.

22. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether each insertion of each amino acid at a given empty amino acid position in the protein is spatially tolerated by other amino acids in the protein.

23. The computer-implemented method of clause 22, wherein the variant pathogenicity classifier is trained to generate a respective spatial resistance score for each insertion.

24. The computer-implemented method of clause 23, wherein each amino acid corresponds to each of the 20 naturally occurring amino acids.

25. The computer-implemented method of clause 24, wherein each amino acid corresponds to each naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

26. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether a substitution of a first amino acid for a second amino acid at a given amino acid position in the protein is evolutionarily conserved or non-conserved.

27. The computer-implemented method of clause 26, wherein a variant pathogenicity classifier is trained to generate evolutionary conservation scores for substitutions.

28. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether each substitution of the first amino acid for each amino acid at a given amino acid position in the protein is evolutionarily conserved or non-conserved.

29. The computer-implemented method of clause 28, wherein a variant pathogenicity classifier is trained to generate a respective evolutionary conservation score for each substitution.

30. The computer-implemented method of clause 29, wherein each amino acid corresponds to each of the 20 naturally occurring amino acids.

31. The computer-implemented method of clause 30, wherein each amino acid corresponds to each naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

32. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether an insertion of an amino acid at a given empty amino acid position in the protein is evolutionarily conserved or non-conserved.

33. The computer-implemented method of clause 32, wherein a variant pathogenicity classifier is trained to generate evolutionary conservation scores for insertions.

34. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether each insertion of each amino acid at a given empty amino acid position in the protein is evolutionarily conserved or non-conserved.

35. The computer-implemented method of clause 34, wherein a variant pathogenicity classifier is trained to generate a respective evolutionary conservation score for each insertion.

36. The computer-implemented method of clause 35, wherein each amino acid corresponds to each of the 20 naturally occurring amino acids.

37. The computer-implemented method of clause 36, wherein each amino acid corresponds to each naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

38. The computer-implemented method of clause 14, wherein the spatial tolerance corresponds to the structural tolerance, and the spatial tolerance corresponds to the structural tolerance.

39. The computer-implemented method of clause 1, wherein the plurality of amino acid positions ranges from 1 million to 10 million amino acid positions.

40. The computer-implemented method of clause 1, wherein the plurality of amino acid positions ranges from 10 million to hundreds of millions of amino acid positions.

41. The computer-implemented method of clause 1, wherein the plurality of amino acid positions ranges from hundreds of millions to billions of amino acid positions.

42. The computer-implemented method of clause 1, wherein the plurality of amino acid positions ranges from 1 to 1 million amino acid positions.

43. Clause 1, wherein the class of unreachable substitute amino acids limited by the reachability of single nucleotide polymorphisms (SNPs) to convert the reference codon of the reference amino acid to the substitute amino acid of the class of unreachable substitute amino acids is masked to the actual label, Computer implementation method.

44. The computer-implemented method of clause 1, wherein the masked amino acid class causes no loss and does not contribute to the gradient update.

45. The computer-implemented method of clause 44, wherein the masked amino acid classes are identified in a lookup table.

46. The computer-implemented method of clause 45, wherein the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

Clause Set 4 (ILLM 1060-1)

1. A computer-implemented method for determining the pathogenicity of a nucleotide variant, comprising:

Designating a specific amino acid at a specific position of the protein as a gap amino acid, and designating the amino acid remaining at the remaining position of the protein as a non-gap amino acid;

generating a gap spatial representation of the protein including the spatial configuration of non-gap amino acids and excluding the spatial configuration of gap amino acids;

determining evolutionary conservation at a particular position of each amino acid of each amino acid class based at least in part on the gap space representation; and

A computer-implemented method comprising determining the pathogenicity of each nucleotide variant that each substitutes a particular amino acid for a respective amino acid in an alternative representation of the protein, based at least in part on the evolutionary conservation of each amino acid.

2. Clause 1, wherein the spatial configuration of non-gap amino acids is encoded as a distance channel per amino acid class,

The distance channel for each amino acid class has a voxel-specific distance value for voxels within a plurality of voxels,

A computer implemented method, wherein the voxel-wise distance value specifies the distance of an atom of a non-gap amino acid from a corresponding voxel within the plurality of voxels.

3. The computer implemented method of clause 2, wherein the spatial configuration of the non-gap amino acid is determined based on the spatial proximity between corresponding voxels and atoms of the non-gap amino acid.

4. The computer-implemented method of clause 2, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the distance of the atom of the gap amino acid from the corresponding voxel when determining the voxel-wise distance value.

5. The computer-implemented method of clause 4, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the spatial proximity between the corresponding voxel and the atoms of the gap amino acid.

6. The computer-implemented method of clause 1, wherein the particular amino acid is a reference amino acid that is the major allele of the protein.

7. The computer-implemented method of clause 1, wherein the evolutionary conservation predictor determines evolutionary conservation as follows:

processing the gap space representation as input; and

Generating each evolutionary conservation score for each amino acid as output.

8. The computer-implemented method of clause 7, wherein each evolutionary conservation score can be ranked according to scale.

9. The computer-implemented method of clause 7, further comprising classifying the nucleotide variant as pathogenic if the evolutionary conservation score generated by the evolutionary conservation predictor for the corresponding amino acid substitution is below a threshold.

10. The computer-implemented method of clause 7, further comprising classifying the nucleotide variant as pathogenic if the evolutionary conservation score generated by the evolutionary conservation predictor for the corresponding amino acid substitution is less than 0.

11. The computer-implemented method of clause 7, further comprising classifying the nucleotide variant as positive if the evolutionary conservation score generated by the evolutionary conservation predictor for the corresponding amino acid substitution is above a threshold.

12. The computer-implemented method of clause 7, further comprising classifying the nucleotide variant as positive if the evolutionary conservation score generated by the evolutionary conservation predictor for the corresponding amino acid substitution is non-zero.

13. The computer-implemented method of clause 7, wherein the evolutionary conservation predictor is trained on a conservative training set and a non-conservative training set.

14. Clause 13, wherein the conservation training set has each conserved protein sample for each conserved amino acid at each position in the proteome,

A computer implemented method wherein the non-conserved training set has each non-conserved protein sample for each non-conserved amino acid at each position.

15. The computer-implemented method of clause 14, wherein each position has a set of conserved amino acids and a set of non-conserved amino acids.

16. The computer-implemented method of clause 15, wherein the specific set of conserved amino acids for a specific position of a specific protein within a proteome comprises at least one major allelic amino acid observed at the specific position across a plurality of species.

17. The computer-implemented method of clause 16, wherein the particular set of conserved amino acids comprises one or more minor allelic amino acids observed at specific positions across a plurality of species.

18. The computer-implemented method of clause 17, wherein the set of specific non-conserved amino acids for a particular position includes amino acids that are not included in the set of specific conserved amino acids.

19. The computer-implemented method of clause 18, wherein the particular set of conserved amino acids and the particular set of non-conserved amino acids are identified based on the evolutionary conservation profile of the homologous protein in the plurality of species.

20. The computer-implemented method of clause 18, wherein the evolutionary conservation profile of the homologous protein is determined using a position-specific frequency matrix (PSFM).

21. The computer-implemented method of clause 18, wherein the evolutionary conservation profile of the homologous protein is determined using a position-specific scoring matrix (PSSM).

22. The computer-implemented method of clause 16, wherein the major allelic amino acid is a reference amino acid.

23. Clause 14, wherein each position has C conserved amino acids in the set of conserved amino acids,

Each position has NC non-conserved amino acids in the set of non-conserved amino acids, NC = 20-C ,

The conservation training set has CP conserved protein samples, where CP = number of each position * C ,

The non-conserved training set has NCP 4 non-conserved protein samples, where NCP = number of each position * (20- C ).

24. The computer-implemented method of clause 23, wherein C ranges from 1 to 10.

25. The computer-implemented method of clause 24, wherein C varies across each location.

26. The computer-implemented method of clause 25, wherein C is the same for some of each position.

27. The computer-implemented method of clause 14, wherein each conserved protein sample and each non-conserved protein sample has a respective gap space representation generated using each reference amino acid at each position as the respective gap amino acid.

28. The computer-implemented method of clause 27, wherein the evolutionary conservation predictor is trained on specific conserved protein samples and estimates the evolutionary conservation of specific conserved amino acids at specific positions in the specific conserved protein samples by the following steps:

Processing a specific gap space representation of a specific conserved protein sample as input.

- At this time, the specific gap space expression is

Generated using a specific reference amino acid at a specific position as the gap amino acid, and the remaining amino acid at the remaining position in a specific conserved protein sample as the non-gap amino acid -; and

Generating evolutionary conservation scores for specific conserved amino acids as output.

29. The computer-implemented method of clause 28, wherein each preserved protein sample has an actual preserved label.

30. The computer-implemented method of clause 29, wherein the actual conservation label is an evolutionary conservation frequency.

31. The computer-implemented method of clause 29, wherein the actual retention label is 1.

32. The computer-implemented method of clause 29, wherein the evolutionary conservation for a particular conserved amino acid is compared to the actual conservation label to determine the error and using training techniques to improve the coefficients of the evolutionary conservation predictor based on the error.

33. Clause 32, wherein when a particular conserved amino acid is a particular reference amino acid, the actual conserved label is masked and not used to determine the error;

A computer-implemented method in which masking prevents evolutionary conservation predictors from overfitting to a particular reference amino acid.

34. The computer-implemented method of clause 32, wherein the training technique is a loss function based gradient update technique.

35. The method of clause 27, wherein the evolutionary conservation predictor is trained on a specific non-conserved protein sample, and estimates the evolutionary conservation of a specific non-conserved amino acid at a specific position in the specific non-conserved protein sample, by the following steps: How to implement a computer:

Processing as input a specific gap space of a specific non-conserved protein sample, where the specific gap space representation is

Generated using a specific reference amino acid at a specific position as the gap amino acid, and the remaining amino acid at the remaining position in a specific non-conserved protein sample as the non-gap amino acid -; and

Generating evolutionary conservation scores for specific non-conserved amino acids as output.

36. The computer implemented method of clause 35, wherein each non-preserved protein sample has an actual non-preserved label.

37. The computer-implemented method of clause 35, wherein the actual non-conserved label is an evolutionary conservation frequency.

38. The computer-implemented method of clause 35, wherein the actual non-preserved label is 0.

39. The computer implementation of clause 35, wherein the evolutionary conservation score for a particular non-conserved amino acid is compared to the actual non-conserved label to determine the error and uses training techniques to improve the coefficients of the evolutionary conservation predictor based on the error. method.

40. The computer-implemented method of clause 7, wherein the evolutionary conservation predictor is trained on a training set.

41. The computer-implemented method of clause 40, wherein the training set has a respective protein sample for each position within the proteome.

42. The computer-implemented method of clause 41, wherein each protein sample has a respective gap space representation generated using each reference amino acid at each position as the respective gap amino acid.

43. The computer-implemented method of clause 42, wherein the evolutionary conservation predictor is trained on a specific protein sample and estimates the evolutionary conservation of each amino acid of each amino acid class at a specific position in the specific protein sample by the following steps:

Processing a specific gap space representation of a specific protein sample as input

- At this time, the specific gap space expression is

Generated using a specific reference amino acid at a specific position as a gap amino acid, and the remaining amino acid at the remaining position of a specific protein sample as a non-gap amino acid -; and

Generating each evolutionary conservation score for each amino acid as output.

44. The computer-implemented method of clause 43, wherein each protein sample has an actual label for each amino acid.

45. The clause 44, wherein each actual label comprises one or more conservative labels for one or more conserved amino acids of each amino acid and one or more non-conservative labels for one or more non-conserved amino acids of each amino acid. A computer implementation method.

46. The computer-implemented method of clause 45, wherein the conserved label and the non-conserved label each have an evolutionary conservation frequency.

47. The computer-implemented method of clause 46, wherein each evolutionary conservation frequency can be ranked according to magnitude.

48. The computer-implemented method of clause 46, wherein the preservation label is 1 and the non-preservation label is 0.

49. The computer-implemented method of clause 46, wherein the error is determined based on: each comparison of each evolutionary conservation score for each conserved amino acid to each conserved amino acid, and

Comparison of each evolutionary conservation score for each non-conserved amino acid with respect to each non-conserved amino acid.

50. The computer-implemented method of clause 49, wherein the coefficients of the evolutionary conservation predictor are improved based on error using training techniques.

51. Clause 50, wherein the conserved amino acids include a specific reference amino acid, and the conserved label for the specific reference amino acid is masked and not used to determine the error, and

A computer-implemented method in which masking prevents evolutionary conservation predictors from overfitting to a particular reference amino acid.

52. The clause 14 wherein the proteome has positions 1 to 10 million;

Each of the 1 to 10 million positions has C conserved amino acids in the set of conserved amino acids,

Positions 1 to 10 million each have NC non-conserved amino acids in the set of non-conserved amino acids, NC = 20- C ,

The conservation training set has CP dog conservation protein samples, CP = 1 to 10 million * C ,

The non-conserved training set has NCP eight non-conserved protein samples, and NCP = 1 to 10 million * (20- C ).

53. Clause 14, wherein the evolutionary conservation predictor is trained with 200 million to 200 million training iterations, and

20 to 200 million training repetitions

1 to 10 million training iterations using 1 to 10 million conserved protein samples, and

A computer-implemented method comprising 19 to 190 million iterations using 19 to 190 million non-conserved protein samples.

54. In clause 14, the proteome has 1 to 10 million positions, so the training set has 1 to 10 million protein samples, and the evolutionary conservation predictor uses 1 to 10 million protein samples to determine 1 million positions. A computer-implemented method that is trained with up to 10 million training repetitions.

55. The computer-implemented method of clause 1, wherein the spatial organization of non-gap amino acids is encoded into an evolutionary profile channel based on the pan-amino acid conservation frequency of the amino acid with the closest atom to the voxel.

56. The computer-implemented method of clause 55, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the nearest atoms of the gap amino acid when determining the pan-amino acid conservation frequency.

57. The computer-implemented method of clause 1, wherein the spatial organization of non-gap amino acids is encoded in an evolutionary profile channel based on the per-amino acid conservation frequency of each amino acid having the atom closest to the voxel.

58. The computer-implemented method of clause 57, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring each nearest atom of the gap amino acid when determining the conservation frequency per amino acid.

59. The computer-implemented method of clause 1, wherein the spatial configuration of non-gap amino acids is encoded as an annotation channel.

60. The computer-implemented method of clause 59, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the annotation channel.

61. The computer-implemented method of clause 1, wherein the spatial configuration of non-gap amino acids is encoded as a structural confidence channel.

62. The computer-implemented method of clause 61, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the structural confidence channel.

63. The computer-implemented method of clause 1, wherein the spatial configuration of non-gap amino acids is encoded as a structural confidence channel.

64. The computer-implemented method of clause 63, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining the structural confidence channel.

65. The computer-implemented method of clause 1, wherein the spatial configuration of non-gap amino acids is encoded as an additional input channel.

66. The computer-implemented method of clause 65, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the atoms of the gap amino acid when determining additional input channels.

67. The computer-implemented method of clause 14, wherein the proteome comprises a non-human proteome, including a human proteome and a non-human primate proteome.

68. Clause 1, wherein the class of unreachable substitute amino acids limited by the reachability of a single nucleotide polymorphism (SNP) to convert the reference codon of the reference amino acid to a substitute amino acid of the class of unreachable substitute amino acids is masked to the actual label, Computer implementation method.

69. The computer-implemented method of clause 1, wherein the masked amino acid class incurs no loss and does not contribute to the gradient update.

70. The computer-implemented method of clause 69, wherein the masked amino acid classes are identified in a lookup table.

71. The computer-implemented method of clause 70, wherein the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

Clause Set 5 (ILLM 1061-1)

1. A computer-implemented method for training pathogenicity predictors, comprising:

Accessing a gap training set containing each gap protein sample for each position in the proteome;

accessing a non-gap training set containing non-gap positive protein samples and non-gap pathogenic protein samples;

generating a respective gapped space representation for the gapped protein sample and generating a respective non-gapped spatial representation for the non-gap positive protein sample and the non-gap pathogenic protein sample;

Training a pathogenicity predictor over one or more training cycles, generating a trained pathogenicity predictor, each training cycle comprising a gap space representation from each gap space representation and a non-gap space from each non-gap space representation. Using expressions as training examples -; and

A computer-implemented method comprising determining the pathogenicity of a variant using a trained pathogenicity classifier.

2. The computer-implemented method of clause 1, wherein each gap protein sample is labeled with the respective gap actual sequence.

3. The computer-implemented method of clause 2, wherein the specific gap actual sequence for the specific gap protein sample has a positive label for a specific amino acid class that corresponds to a reference amino acid at a specific position in the specific gap protein.

4. The computer-implemented method of clause 3, wherein the particular gap protein sample has a respective pathogenicity label for each remaining amino acid class corresponding to an alternative amino acid at a particular position.

5. The computer implemented method of clause 1, wherein the particular non-gap positive protein sample comprises a positive replacement amino acid at a specific position that is substituted by a positive nucleotide variant.

6. The computer implemented method of clause 5, wherein the specific non-gap pathogenic protein sample comprises a pathogenic replacement amino acid at a specific position that is substituted by a pathogenic nucleotide variant.

7. The computer-implemented method of clause 6, wherein the specific non-gap positive protein sample is labeled with a positive authentic sequence having a positive label for a specific amino acid class corresponding to the positive replacement amino acid.

8. The computer-implemented method of clause 7, wherein the positive actual sequence is a respective masking label for each remaining amino acid class corresponding to an amino acid that is different from the positive replacement amino acid.

9. The computer implemented method of clause 8, wherein the specific non-gap pathogenic protein sample is labeled with the pathogenic actual sequence having a pathogenic label for a specific amino acid class that corresponds to the pathogenic alternative amino acid.

10. The computer-implemented method of clause 9, wherein the pathogenic actual sequence has a respective masked label for each remaining amino acid class corresponding to an amino acid that is different from the pathogenic substitute amino acid.

11. Clause 1, further comprising using a sample indicator to indicate to the pathogenicity predictor whether the current training example is a gap space representation for a gap protein sample or a non-gap space representation for a non-gap protein sample. Including, computer implemented methods.

12. The computer-implemented method of clause 1, further comprising masking positive labels for a particular class of amino acids corresponding to a reference amino acid at a particular position in a particular gap protein.

13. The computer implemented method of clause 1, wherein the non-gap positive protein sample is derived from common human and non-human primate nucleotide variants.

14. The computer implemented method of clause 1, wherein the non-gap pathogenic protein sample is derived from combinatorially simulated nucleotide variants.

15. Clause 1, wherein the pathogenicity predictor generates output sequences per amino acid class in response to processing training examples, and

A computer-implemented method wherein the output sequence for each amino acid class has a pathogenicity score for each amino acid class.

16. The computer-implemented method of clause 1, further comprising measuring performance of the trained pathogenicity predictor between training cycles on a validation set.

17. The computer-implemented method of clause 16, wherein the validation set includes pairs of gapped space representations and non-gapped space representations for each retained protein sample.

18. The method of clause 1, wherein the trained pathogenicity predictor generates a first amino acid class-specific output sequence for the gap space representation of the pair and a second amino acid class-specific output sequence for the non-gap space representation of the pair, and the retained protein A computer-implemented method, wherein a final pathogenicity score for a nucleotide variant causing an amino acid substitution in a sample is determined based on a combination of the first and second pathogenicity scores for amino acid substitutions in output sequences by first and second amino acid classes.

19. The computer-implemented method of clause 18, wherein the final pathogenicity score is based on an average of the first and second pathogenicity scores.

20. The computer-implemented method of clause 1, wherein at least some of the training cycles use an equal number of gapped space representations and non-gap space representations.

21. The computer-implemented method of clause 1, wherein at least some of the training cycles use batches of training examples having an equal number of gapped space representations and non-gap space representations.

22. The computer-implemented method of clause 1, wherein the masked label does not contribute to the determination of the error and thus does not contribute to the training of the pathogenicity predictor.

23. The computer-implemented method of clause 22, wherein the masked label is zeroed out.

24. Clause 1, wherein gap space representations are weighted differently than non-gap space representations, such that gap to gradient updates are applied to the parameters of the pathogenicity predictor in response to the pathogenicity predictor processing the non-gap spatial representation. A computer implemented method, wherein the contribution of the spatial representation varies from the contribution of the non-gap spatial representation to a gradient update applied to the parameters of the pathogenicity predictor in response to the pathogenicity predictor processing the non-gap spatial representation.

25. The computer-implemented method of clause 24, wherein the variation is determined by predefined weights.

26. A computer-implemented method for training pathogenicity predictors, comprising:

Starting with training a pathogenicity classifier on a gap training set to generate a trained pathogenicity classifier;

further training the trained pathogenicity classifier on the non-gap training set and generating a retrained pathogenicity classifier; and

A computer-implemented method comprising determining the pathogenicity of a variant using a retrained pathogenicity classifier.

27. The computer-implemented method of clause 26, further comprising measuring performance of the trained pathogenicity predictor between training cycles on a first validation set containing only non-gap spatial representations of the retained protein samples.

28. The method of clause 27, further comprising measuring the performance of the retrained pathogenicity predictor between training cycles on a second validation set comprising a gap space representation and a non-gap space representation of the retained protein samples. , computer implementation method.

29. The method of clause 28, wherein the retrained pathogenicity predictor generates a first amino acid class-specific output sequence for the pair in response to processing the pair, and for nucleotide variants causing amino acid substitutions in the corresponding maintained protein sample. A computer implemented method, wherein the final pathogenicity score is determined based on the output sequence for each first amino acid class.

30. A computer-implemented method for training pathogenicity predictors, comprising:

Accessing a gap training set containing each gap protein sample for each position in the proteome - each gap protein sample is labeled with its respective gap real sequence, and the specific gap real sequence for a particular gap protein sample is a specific gap protein sample. having a benign label for a specific amino acid class corresponding to a reference amino acid at a specific position in the gap protein, and having a respective pathogenic label for each remaining amino acid class corresponding to an alternative amino acid at a specific position -;

Accessing a non-gap training set comprising non-gap positive protein samples and non-gap pathogenic protein samples, wherein the specific non-gap positive protein sample contains a positive substituted amino acid at a specific position substituted by a positive nucleotide variant; , a specific non-gap pathogenic protein sample contains a pathogenic substituted amino acid at a specific position substituted by a pathogenic nucleotide variant, and a specific non-gap positive protein sample contains a positive label and a positive substitution for a specific amino acid class corresponding to the positive substituted amino acid. Labeled with a positive authentic sequence with a respective masking label for each remaining amino acid class corresponding to an amino acid and another amino acid, a specific non-gap pathogenic protein sample is labeled with a pathogenic label for a specific amino acid class corresponding to a pathogenic alternative amino acid and a pathogenic label for each remaining amino acid class. Labeled with the pathogenic actual sequence, with respective masking labels for each of the remaining amino acid classes corresponding to the substitute amino acid and the other amino acid -;

generating a respective gapped space representation for the gapped protein sample and generating a respective non-gapped spatial representation for the non-gap positive protein sample and the non-gap pathogenic protein sample;

Training a pathogenicity predictor over one or more training cycles, generating a trained pathogenicity predictor, each training cycle comprising a gap space representation from each gap space representation and a non-gap space from each non-gap space representation. Using expressions as training examples -; and

A computer-implemented method comprising determining the pathogenicity of a variant using a trained pathogenicity classifier.

Clause set 6

1. A computer-implemented method for generating training data for training a variant pathogenicity classifier, comprising:

Accessing multiple amino acid positions in the proteome with multiple proteins;

Designating major allele amino acids at multiple amino acid positions as reference amino acids for multiple proteins;

For each amino acid position among multiple amino acid positions,

Classifying these nucleotide substitutions as positive variants that substitute a specific reference amino acid for a specific reference amino acid at a specific amino acid position in a specific alternative representation of a specific protein, and

Classifying these nucleotide substitutions as pathogenic variants in which a specific reference amino acid is replaced by a replacement amino acid at a specific amino acid position - the replacement amino acid is different from the specific reference amino acid; and

A computer-implemented method comprising training a variant pathogenicity classifier on training data containing a spatial representation of a protein sample, such that the spatial representation is assigned an actual benign label corresponding to a benign variant and an actual pathogenic label corresponding to a pathogenic variant. .

2. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether a substitution of a first amino acid for a second amino acid at a given amino acid position in the protein is pathogenic or benign.

3. The computer-implemented method of clause 2, wherein a variant pathogenicity classifier is trained to generate pathogenicity scores for substitutions.

4. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether each substitution of a first amino acid for each amino acid at a given amino acid position in the protein is pathogenic or benign.

5. The computer-implemented method of clause 4, wherein a variant pathogenicity classifier is trained to generate a respective pathogenicity score for each substitution.

6. The computer-implemented method of clause 5, wherein each amino acid corresponds to each of the 20 naturally occurring amino acids.

7. The computer-implemented method of clause 6, wherein each amino acid corresponds to each naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

8. The computer-implemented method of clause 1, wherein a variant pathogenicity classifier is trained to determine whether an insertion of an amino acid at a given empty amino acid position in the protein is pathogenic or benign.

9. The computer-implemented method of clause 8, wherein a variant pathogenicity classifier is trained to generate a pathogenicity score for the insertion.

10. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether each insertion of each amino acid at a given empty amino acid position in the protein is pathogenic or benign.

11. The computer-implemented method of clause 10, wherein a variant pathogenicity classifier is trained to generate a respective pathogenicity score for each insertion.

12. The computer-implemented method of clause 11, wherein each amino acid corresponds to each of the 20 naturally occurring amino acids.

13. The computer-implemented method of clause 12, wherein each amino acid corresponds to each naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

14. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether a substitution of a first amino acid for a second amino acid at a given amino acid position in the protein is spatially tolerated by other amino acids in the protein.

15. The computer-implemented method of clause 14, wherein a variant pathogenicity classifier is trained to generate a spatial resistance score for substitutions.

16. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether each substitution of the first amino acid at a given amino acid position in the protein is spatially tolerated by the other amino acids in the protein. .

17. The computer-implemented method of clause 16, wherein the variant pathogenicity classifier is trained to generate a respective spatial resistance score for each substitution.

18. The computer-implemented method of clause 17, wherein each amino acid corresponds to each of the 20 naturally occurring amino acids.

19. The computer-implemented method of clause 18, wherein each amino acid corresponds to each naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

20. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether insertion of an amino acid at a given empty amino acid position in the protein is spatially permitted by other amino acids in the protein.

21. The computer-implemented method of clause 20, wherein a variant pathogenicity classifier is trained to generate a spatial resistance score for insertions.

22. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether each insertion of each amino acid at a given empty amino acid position in the protein is spatially tolerated by other amino acids in the protein.

23. The computer-implemented method of clause 22, wherein the variant pathogenicity classifier is trained to generate a respective spatial resistance score for each insertion.

24. The computer-implemented method of clause 23, wherein each amino acid corresponds to each of the 20 naturally occurring amino acids.

25. The computer-implemented method of clause 24, wherein each amino acid corresponds to each naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

26. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether a substitution of a first amino acid for a second amino acid at a given amino acid position in the protein is evolutionarily conserved or non-conserved.

27. The computer-implemented method of clause 26, wherein a variant pathogenicity classifier is trained to generate evolutionary conservation scores for substitutions.

28. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether each substitution of the first amino acid for each amino acid at a given amino acid position in the protein is evolutionarily conserved or non-conserved.

29. The computer-implemented method of clause 28, wherein a variant pathogenicity classifier is trained to generate a respective evolutionary conservation score for each substitution.

30. The computer-implemented method of clause 29, wherein each amino acid corresponds to each of the 20 naturally occurring amino acids.

31. The computer-implemented method of clause 30, wherein each amino acid corresponds to each naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

32. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether an insertion of an amino acid at a given empty amino acid position in the protein is evolutionarily conserved or non-conserved.

33. The computer-implemented method of clause 32, wherein a variant pathogenicity classifier is trained to generate evolutionary conservation scores for insertions.

34. The computer-implemented method of clause 1, wherein the variant pathogenicity classifier is trained to determine whether each insertion of each amino acid at a given empty amino acid position in the protein is evolutionarily conserved or non-conserved.

35. The computer-implemented method of clause 34, wherein a variant pathogenicity classifier is trained to generate a respective evolutionary conservation score for each insertion.

36. The computer-implemented method of clause 35, wherein each amino acid corresponds to each of the 20 naturally occurring amino acids.

37. The computer-implemented method of clause 36, wherein each amino acid corresponds to each naturally occurring amino acid from a subset of the 20 naturally occurring amino acids.

38. The computer-implemented method of clause 14, wherein the spatial tolerance corresponds to the structural tolerance, and the spatial tolerance corresponds to the structural tolerance.

39. The computer-implemented method of clause 1, wherein the plurality of amino acid positions ranges from 1 million to 10 million amino acid positions.

40. The computer-implemented method of clause 1, wherein the plurality of amino acid positions ranges from 10 million to hundreds of millions of amino acid positions.

41. The computer-implemented method of clause 1, wherein the plurality of amino acid positions ranges from hundreds of millions to billions of amino acid positions.

42. The computer-implemented method of clause 1, wherein the plurality of amino acid positions ranges from 1 to 1 million amino acid positions.

43. Clause 1, wherein the class of unreachable substitute amino acids limited by the reachability of single nucleotide polymorphisms (SNPs) to convert the reference codon of the reference amino acid to the substitute amino acid of the class of unreachable substitute amino acids is masked to the actual label, Computer implementation method.

44. The computer-implemented method of clause 1, wherein the masked amino acid class causes no loss and does not contribute to the gradient update.

45. The computer-implemented method of clause 44, wherein the masked amino acid classes are identified in a lookup table.

46. The computer-implemented method of clause 45, wherein the lookup table identifies a set of masked amino acid classes for each reference amino acid position.

47. The computer-implemented method of clause 1, wherein the spatial representation is a structural representation of the protein structure of the protein sample.

48. The computer-implemented method of clause 1, wherein the spatial representation is encoded using voxelization.

Clause set 7

1. A computer-implemented method for determining the pathogenicity of a nucleotide variant, comprising:

Accessing a spatial representation of a protein, wherein the spatial representation of the protein specifies each spatial configuration of each amino acid at each position in the protein;

generating a gap spatial representation of the protein by removing specific spatial configurations of specific amino acids at specific positions from the spatial representation of the protein; and

A computer-implemented method comprising determining the pathogenicity of a nucleotide variant based, at least in part, on the gap space representation and the representation of alternative amino acids produced by the nucleotide variant at a particular position.

2. The computer-implemented method of clause 1, wherein removal of specific spatial configurations is implemented by a script.

3. A computer-implemented method for determining the pathogenicity of a nucleotide variant, comprising:

Creating a gap protein by removing a specific amino acid at a specific position from the protein; and

A computer-implemented method comprising determining the pathogenicity of a nucleotide variant based, at least in part, on a gap protein and an alternative amino acid produced by the nucleotide variant at a particular position.

4. The computer-implemented method of clause 3, wherein the major allelic amino acid is a reference amino acid.

5. A system for predicting spatial tolerance of amino acid substitutions, comprising:

Gapping logic configured to remove specific amino acids at specific positions from the protein and create amino acid vacancies at specific positions in the protein; and

A system comprising substitution logic configured to process a protein with amino acid vacancies and score the tolerance of substituted amino acids that are candidates for filling the amino acid vacancies.

6. The system of clause 5, wherein the substitution logic is further configured to score the resistance of the substituted amino acid based, at least in part, on structural compatibility between the substituted amino acid and adjacent amino acids near the amino acid vacancy.

7. A computer-implemented method for determining the pathogenicity of a nucleotide variant, comprising:

Accessing proteins with each amino acid at each position;

Designating a specific amino acid of a specific amino acid class at a specific position of the protein as a gap amino acid, and designating the remaining amino acid at the remaining position of the protein as a non-gap amino acid;

generating a gap spatial representation of the protein including the spatial configuration of non-gap amino acids and excluding the spatial configuration of gap amino acids; and

Determining the pathogenicity of each alternative amino acid at a particular position based at least in part on the gap space representation.

- Here, each substitute amino acid has a respective amino acid class that is different from the specific amino acid class.

8. A system for predicting the evolutionary conservation of amino acid substitutions, comprising:

Gapping logic configured to remove specific amino acids at specific positions from the protein and create amino acid vacancies at specific positions in the protein; and

A system comprising substitution logic configured to process proteins with amino acid vacancies and score the evolutionary conservation of substituted amino acids that are candidates for filling the amino acid vacancies.

9. The system of clause 8, wherein the substitution logic is further configured to score evolutionary conservation of the substituted amino acid based, at least in part, on structural compatibility between the substituted amino acid and adjacent amino acids near the amino acid vacancy.

10. The system of clause 8, wherein evolutionary conservation is scored using evolutionary conservation frequency.

11. The system of clause 10, wherein the evolutionary conservation frequency is based on a site-specific frequency matrix (PSFM).

12. The system of clause 10, wherein the evolutionary conservation frequency is based on a site-specific score matrix (PSSM).

13. The system of clause 8, wherein the evolutionary conservation scores of substituted amino acids are ranked according to scale.

14. A system for predicting the evolutionary conservation of amino acid substitutions, comprising:

Gapping logic configured to remove specific amino acids at specific positions from the protein and create amino acid vacancies at specific positions in the protein; and

A system comprising evolutionary conservation prediction logic configured to process proteins with amino acid vacancies and rank evolutionary conservation of substituted amino acids that are candidates for filling amino acid vacancies.

15. A system for predicting the structural resistance of amino acid substitutions, comprising:

Gapping logic configured to remove specific amino acids at specific positions from the protein and create amino acid vacancies at specific positions in the protein; and

A system comprising structural resistance prediction logic configured to process proteins with amino acid vacancies and rank the structural resistance of substituted amino acids that are candidates for filling the amino acid vacancies based on patterns of amino acid co-occurrence near the amino acid vacancies.

16. A computer-implemented method for determining the pathogenicity of a nucleotide variant, comprising:

Accessing proteins with each amino acid at each position;

Designating a specific amino acid at a specific position of the protein as a gap amino acid, and designating the amino acid remaining at the remaining position of the protein as a non-gap amino acid;

generating a gap spatial representation of the protein including the spatial configuration of non-gap amino acids and excluding the spatial configuration of gap amino acids;

determining the evolutionary conservation of a replacement amino acid at a particular position based at least in part on the gap space representation and the representation of the replacement amino acid; and

A computer-implemented method comprising determining the pathogenicity of a nucleotide variant resulting in a replacement amino acid based at least in part on evolutionary conservation.

Clause set 8

1. A computer-implemented method for determining the pathogenicity of a nucleotide variant, comprising:

Accessing a spatial representation of a protein, wherein the spatial representation of the protein specifies each spatial configuration of each amino acid at each position in the protein;

generating a gap spatial representation of the protein by removing specific spatial configurations of specific amino acids at specific positions from the spatial representation of the protein; and

A computer-implemented method comprising determining the pathogenicity of a nucleotide variant based, at least in part, on the gap space representation and the representation of alternative amino acids produced by the nucleotide variant at a particular position.

2. The computer-implemented method of clause 1, wherein removal of specific spatial configurations is implemented by a script.

3. A computer-implemented method for determining the pathogenicity of a nucleotide variant, comprising:

Removing specific amino acids at specific positions from the spatial representation of the protein to create a gap spatial representation of the protein; and

Determining the pathogenicity of a nucleotide variant based, at least in part, on the gap space representation of the protein and the replacement amino acid produced by the nucleotide variant at a particular position.

4. The computer-implemented method of clause 3, wherein the major allelic amino acid is a reference amino acid.

5. A system for predicting spatial tolerance of amino acid substitutions, comprising:

gap logic, which is configured to remove specific amino acids at specific positions in the spatial representation of the protein and create amino acid vacancies at specific positions in the spatial representation of the protein; and

A system comprising substitution logic configured to process a spatial representation of a protein having amino acid vacancies and score resistance of substituted amino acids that are candidates for filling amino acid vacancies.

6. The system of clause 5, wherein the substitution logic is further configured to score the resistance of the substituted amino acid based, at least in part, on structural compatibility between the substituted amino acid and adjacent amino acids near the amino acid vacancy.

7. A computer-implemented method for determining the pathogenicity of a nucleotide variant, comprising:

Accessing proteins with each amino acid at each position;

Designating a specific amino acid of a specific amino acid class at a specific position of the protein as a gap amino acid, and designating the remaining amino acid at the remaining position of the protein as a non-gap amino acid;

generating a gap spatial representation of the protein including the spatial configuration of non-gap amino acids and excluding the spatial configuration of gap amino acids; and

Determining the pathogenicity of each alternative amino acid at a particular position based at least in part on the gap space representation.

- Here, each substitute amino acid has a respective amino acid class that is different from the specific amino acid class.

8. A system for predicting the evolutionary conservation of amino acid substitutions, comprising:

gap logic, which is configured to remove specific amino acids at specific positions in the spatial representation of the protein and create amino acid vacancies at specific positions in the spatial representation of the protein; and

A system comprising substitution logic configured to process a spatial representation of a protein having amino acid vacancies and score the evolutionary conservation of substituted amino acids that are candidates for filling amino acid vacancies.

9. The system of clause 8, wherein the substitution logic is further configured to score evolutionary conservation of the substituted amino acid based, at least in part, on structural compatibility between the substituted amino acid and adjacent amino acids near the amino acid vacancy.

10. The system of clause 8, wherein evolutionary conservation is scored using evolutionary conservation frequency.

11. The system of clause 10, wherein the evolutionary conservation frequency is based on a site-specific frequency matrix (PSFM).

12. The system of clause 10, wherein the evolutionary conservation frequency is based on a site-specific score matrix (PSSM).

13. The system of clause 8, wherein the evolutionary conservation scores of substituted amino acids are ranked according to scale.

14. A system for predicting the evolutionary conservation of amino acid substitutions, comprising:

gap logic, which is configured to remove specific amino acids at specific positions in the spatial representation of the protein and create amino acid vacancies at specific positions in the spatial representation of the protein; and

A system comprising evolutionary conservation prediction logic configured to process a spatial representation of a protein having amino acid vacancies and to rank evolutionary conservation of substituted amino acids that are candidates for filling amino acid vacancies.

15. A system for predicting the structural resistance of amino acid substitutions, comprising:

gap logic, which is configured to remove specific amino acids at specific positions in the spatial representation of the protein and create amino acid vacancies at specific positions in the spatial representation of the protein; and

A system comprising structural resistance prediction logic configured to process a spatial representation of a protein with amino acid vacancies and rank the structural resistance of substituted amino acids that are candidates for filling the amino acid vacancies based on patterns of amino acid co-occurrence near the amino acid vacancies. .

16. A computer-implemented method for determining the pathogenicity of a nucleotide variant, comprising:

Accessing proteins with each amino acid at each position;

Designating a specific amino acid at a specific position of the protein as a gap amino acid, and designating the amino acid remaining at the remaining position of the protein as a non-gap amino acid;

generating a gap spatial representation of the protein including the spatial configuration of non-gap amino acids and excluding the spatial configuration of gap amino acids;

determining the evolutionary conservation of a replacement amino acid at a particular position based at least in part on the gap space representation and the representation of the replacement amino acid; and

A computer-implemented method comprising determining the pathogenicity of a nucleotide variant resulting in a replacement amino acid based at least in part on evolutionary conservation.

Although the present invention is disclosed with reference to the preferred embodiments and examples detailed above, it will be understood that these examples are intended to be illustrative and not restrictive. It is contemplated that modifications and combinations will readily occur to those skilled in the art, and that such modifications and combinations will be within the spirit of the invention and the scope of the following claims.

Claims (30)

A computer-implemented method for evaluating nucleotide variants as benign or pathogenic, comprising:
Accessing proteins with each amino acid at each position;
Designating a specific amino acid at a specific position of the protein as a gap amino acid and designating the remaining amino acid at the remaining position of the protein as a non-gap amino acid;
generating a gapped spatial representation of the protein including the spatial configuration of non-gap amino acids and excluding the spatial configuration of gap amino acids; and
A computer-implemented method comprising evaluating a nucleotide variant as benign or pathogenic using a neural network-based pathogenicity predictor based, at least in part, on the gap space representation and the representation of alternative amino acids produced by the nucleotide variant at a particular position. The method of claim 1, wherein the spatial configuration of non-gap amino acids is encoded as a distance channel per amino acid class,
The distance channel for each amino acid class has a voxel-specific distance value for voxels within a plurality of voxels,
A computer implemented method, wherein the voxel-wise distance value specifies the distance of an atom of a non-gap amino acid from a corresponding voxel within the plurality of voxels. 3. The computer implemented method of claim 1 or 2, wherein the spatial configuration of the non-gap amino acid is determined based on the spatial proximity between corresponding voxels and atoms of the non-gap amino acid. The computer implementation of any one of claims 1 to 3, wherein the spatial configuration of the gap amino acid is excluded from the gap space representation by ignoring the distance of the atom of the gap amino acid from the corresponding voxel when determining the voxel-wise distance value. method. 5. A computer-implemented method according to any one of claims 1 to 4, wherein the spatial configuration of gap amino acids is excluded from the gap space representation by ignoring the spatial proximity between corresponding voxels and atoms of gap amino acids. The computer implemented method of any one of claims 1 to 5, wherein the particular amino acid is a reference amino acid that is a major allele of the protein. 7. The method according to any one of claims 1 to 6, wherein the neural network based pathogenicity predictor comprises: processing gap space representations and representations of alternative amino acids as input; and
A computer-implemented method for evaluating nucleotide variants as benign or pathogenic by generating as output a pathogenicity score for the replacing amino acid. 8. The computer-implemented method of any one of claims 1-7, wherein the neural network-based pathogenicity predictor is trained on a benign training set. 9. The computer-implemented method of claim 8, wherein the positive training set has each positive protein sample for each reference amino acid at each position in the proteome. 10. The computer implemented method of claim 9, wherein the reference amino acid is a major allelic amino acid of the proteome. 11. The computer-implemented method of claim 9 or 10, wherein the proteome has 10 million positions so the positive training set has 10 million positive protein samples. 12. The computer-implemented method of any one of claims 9-11, wherein each positive protein sample has a respective gap space representation generated using a respective reference amino acid as a respective gap amino acid. 13. The computer-implemented method of any one of claims 9-12, wherein each positive protein sample has as its respective replacement amino acid a respective representation of a respective reference amino acid. 14. The method of any one of claims 1 to 13, wherein the neural network-based pathogenicity predictor is trained on specific benign protein samples and evaluates specific reference amino acids at specific positions of the specific positive protein samples as benign or pathogenic by the following steps. Computer implementation method:
(i) Specific gap space representation of specific positive protein samples
- At this time, the specific gap space expression is
using specific reference amino acids as gap amino acids, and
Generated by using the remaining amino acids at the remaining positions in a specific positive protein sample as non-gap amino acids -; and
(ii) processing as input a representation of a specific reference amino acid as a specific replacement amino acid; and
Generating as output a pathogenicity score for a specific reference amino acid. 15. The computer-implemented method of claim 14, wherein each positive protein sample has a ground truth positive label indicating the absolute positivity of the positive protein sample. 16. The computer implemented method of claim 15, wherein the actual positive label is 0. 17. The computer-implemented method of claim 15 or 16, wherein the pathogenicity score for a particular reference amino acid is compared to the actual positive label to determine the error and using training techniques to improve the coefficients of the neural network-based pathogenicity predictor based on the error. . 18. The computer-implemented method of any one of claims 1-17, wherein the neural network-based pathogenicity predictor is trained on a pathogenicity training set. 19. The computer-implemented method of claim 18, wherein the pathogenicity training set has a respective pathogenic protein sample for each combinatorially generated amino acid substitution for each reference amino acid at each position in the proteome. 20. The computer-implemented method of claim 19, wherein the combinatorially generated amino acid substitutions for a particular reference amino acid of a particular amino acid class at a particular position in the proteome comprise each replacement amino acid of each amino acid class different from the particular amino acid class. 21. The method of claim 19 or 20, wherein the proteome has 10 million positions, and for each 10 million positions there are 19 combinatorially generated amino acid substitutions, so the pathogenicity training set contains 190 million pathogenic protein samples. Having, a computer implemented method. 22. The computer-implemented method of any one of claims 19-21, wherein each pathogenic protein sample has a respective gap space representation generated using a respective reference amino acid as a respective gap amino acid. 23. The method of any one of claims 19 to 22, wherein each pathogenic protein sample is each combinatorially produced as each replacement amino acid produced by each combinatorially produced nucleotide variant at each position of the proteome. A computer-implemented method, having a representation for each of the amino acid substitutions. 24. The method of any one of claims 1 to 23, wherein the neural network-based pathogenicity predictor is trained on specific pathogenic protein samples and selects specific combinatorially generated amino acid substitutions for specific reference amino acids at specific positions in the specific pathogenic protein sample. Computer-implemented method for evaluating benign or pathogenic by stage:
(i) a specific gap space representation of a specific pathogenic protein sample, where the specific gap space representation is
using specific reference amino acids as gap amino acids, and
Generated by using the remaining amino acids at the remaining positions in a specific pathogenic protein sample as non-gap amino acids -; and
(ii) processing as input a representation of a specific combinatorially produced amino acid substitution as a specific replacement amino acid; and
Generating as output a pathogenicity score for specific combinatorially generated amino acid substitutions. 25. The computer-implemented method of claim 24, wherein each pathogenic protein sample has an actual pathogenicity label indicating the absolute pathogenicity of the pathogenic protein sample. 26. The computer implemented method of claim 25, wherein the actual pathogenicity label is 1. 27. The method of claim 25 or 26, wherein the pathogenicity score for a particular combinatorially generated amino acid substitution is compared to the actual pathogenicity label to determine the error and training techniques are used to improve the coefficients of the pathogenicity predictor based on the error. Computer implementation method. 28. The method of any one of claims 1 to 27, wherein the neural network-based pathogenicity predictor is trained for 200 million training iterations,
200 million training repetitions
10 million training iterations using 10 million positive protein samples, and
A computer-implemented method involving 190 million iterations using 190 million pathogenic protein samples. A non-transitory computer-readable storage medium containing computer program instructions for evaluating nucleotide variants as benign or pathogenic, wherein the instructions, when executed by a processor, comprise:
Accessing proteins with each amino acid at each position;
Designating a specific amino acid at a specific position of the protein as a gap amino acid, and designating the amino acid remaining at the remaining position of the protein as a non-gap amino acid;
generating a gap spatial representation of the protein including the spatial configuration of non-gap amino acids and excluding the spatial configuration of gap amino acids; and
Implementing a method comprising evaluating a nucleotide variant as benign or pathogenic using a neural network-based pathogenicity predictor based, at least in part, on the gap space representation and the representation of alternative amino acids produced by the nucleotide variant at a specific position, A non-transitory computer-readable storage medium. A system comprising one or more processors coupled to memory,
The memory is loaded with computer instructions for evaluating nucleotide variants as benign or pathogenic, and when the instructions are executed on the processor:
Accessing proteins with each amino acid at each position;
Designating a specific amino acid at a specific position of the protein as a gap amino acid, and designating the amino acid remaining at the remaining position of the protein as a non-gap amino acid;
generating a gap spatial representation of the protein including the spatial configuration of non-gap amino acids and excluding the spatial configuration of gap amino acids; and
Implementing operations comprising evaluating a nucleotide variant as benign or pathogenic using a neural network-based pathogenicity predictor based, at least in part, on the gap space representation and the representation of alternative amino acids produced by the nucleotide variant at a particular position. system.

Images