CN111798919B - Tumor neoantigen prediction method, prediction device and storage medium - Google Patents

Abstract

The invention relates to a tumor neoantigen prediction method, a prediction device and a storage medium based on chromatin advanced conformation and deep sparse learning, wherein the method invents a deep neural network prediction model based on group selection, and trains the model through training data to obtain an immunogenicity prediction value of an object to be predicted (namely a potential tumor neoantigen peptide); wherein each sample used in training the deep neural network prediction model includes chromatin 3D conformation information and features generated based on the polypeptide amino acid sequence. Compared with the prior art, the method has the advantages of high prediction precision, convenience in prediction and the like.

Description

A tumor neoantigen prediction method, prediction device and storage medium

technical field

The present invention relates to the field of neoantigen prediction in personalized tumor immunotherapy, in particular to a tumor neoantigen prediction method, prediction device and storage medium based on high-level chromatin conformation and deep sparse learning.

Background technique

At present, the conventional treatment of cancer patients mainly relies on non-individualized surgical resection, radiotherapy and chemotherapy, and targeted drug therapy. Transiently prolongs the survival of cancer patients.

In recent years, the method of tumor immunotherapy by targeting the patient's tumor cells through the tumor patient's own immune system has entered people's field of vision. In individualized tumor immunotherapy, tumor patient-specific target molecules that play a key role are called tumor neoantigens. The essence of tumor neoantigen is protein, which is produced by mutation of tumor genome. Because it contains non-synonymous mutation, it is different from abnormally expressed tumor autologous protein antigen. In vivo, tumor neoantigens can be recognized as foreign antigens by the autoimmune system and are not affected by central tolerance, thereby enabling the autoimmune system to specifically target the patient's tumor cells. Therefore, preparing tumor neoantigens into vaccines or polypeptide preparations for tumor immunotherapy can selectively kill tumor cells with high safety and remarkable effect. In this strategy, accurate and efficient individualized selection of tumor neoantigens with good expected curative effect among many peptides that may distinguish tumors from normal tissues is particularly critical. However, the current tumor neoantigen selection technology still has many technical problems, such as heavy selection workload and low precision.

The vigorous development of genomics in the past two decades has provided strong support for tumor research. By comparing the genomes of tumor cells and normal cells, people have discovered many genetic variations closely related to the occurrence and development of tumors, and partially revealed the molecular mechanisms of these genetic variations in the occurrence and development of tumors. And provide strong technical support to guide clinical treatment. In terms of somatic mutations in tumor genomes, it has been found that a single mutation in chromatin often cannot cause tumors. Numerous genetic and epigenetic variations can be found in almost every tumor patient's tumor cells after testing, including several to hundreds of Co-existence of several gene mutations, chromosomal translocations accompanied by gene mutations, and chromosome copy number variation at multiple locations are all very common. More and more evidence shows that these concomitant (simultaneous or successive) genetic variations have their inherent laws, some gene mutations are often accompanied by other gene mutations rather than random occurrence, and this concomitant gene mutation has its inherent genetic The basis of the genetic structure is still not very clear, but clarifying the relevant mechanism will lay a theoretical foundation for in-depth understanding of the molecular mechanism of tumor occurrence and development, especially for understanding the causal relationship of genetic events in tumor development, and provide an effective means for the accurate selection of tumor neoantigens , and then provide a certain basis for tumor diagnosis and treatment.

Contents of the invention

The purpose of the present invention is to provide a tumor neoantigen prediction method, prediction device and storage medium based on high-level chromatin conformation and deep sparse learning with high prediction accuracy and convenient prediction in order to overcome the above-mentioned defects in the prior art.

The purpose of the present invention can be achieved through the following technical solutions:

A tumor neoantigen prediction method based on high-level chromatin conformation and deep sparse learning, the method processes the object to be predicted through a trained deep neural network prediction model based on group selection, and obtains the tumor neoantigen corresponding to the object to be predicted Antigen immunogenicity information;

Wherein, each sample used for training the deep neural network prediction model includes chromatin 3D conformation information and features generated based on polypeptide amino acid sequences.

Further, the order of magnitude of features of each sample is thousands of levels.

Each feature of the sample belongs to a certain group, and in the training of the neural network model, all the features in a certain group are either selected or all eliminated.

Further, the output of the group selection-based deep neural network prediction model includes neoantigens with activated immunogenicity and several features most correlated with the neoantigens.

Further, the group selection-based deep neural network prediction model adopts the form of a fully connected layer.

Further, the chromatin 3D conformation information is derived from the cellular chromatin 3D conformation thermodynamic matrix obtained through Hi-C (chromatin conformation capture technology) experiments.

Further, the 3D conformation information of chromatin is obtained from public Hi-C datasets.

Further, the feature generated based on the amino acid sequence of the polypeptide includes the feature of the polypeptide containing the amino acid mutation site and the high expression information of the gene containing the mutation.

The present invention also provides a tumor neoantigen prediction device based on high-level chromatin conformation and deep sparse learning, including:

The data acquisition unit acquires samples for training, each sample includes chromatin 3D conformation information and features generated based on polypeptide amino acid sequences;

A model training unit, which obtains a deep neural network prediction model based on group selection based on the sample training;

The prediction unit acquires the object to be predicted, processes the object to be predicted through the deep neural network prediction model based on group selection, and obtains tumor neoantigen immunogenicity information corresponding to the object to be predicted.

The present invention also provides a computer-readable storage medium, including a computer program that can be executed by a processor to implement the prediction method.

Compared with the prior art, the present invention has the following beneficial effects:

First, based on the inventor’s many creative studies on the high-level conformation of chromatin, the present invention proposes for the first time to examine and analyze whether the amino acid polypeptide antigen corresponding to the mutated DNA site can activate the immunogenicity of T cells from the perspective of the 3D conformation of chromatin , the 3D conformation of chromatin is added to the feature set predicted by machine learning, that is, the spatial distribution information of the DNA mutation site corresponding to the neoantigen peptide on the chromatin, which can significantly improve the prediction accuracy of whether the neoantigen is immunogenic.

Second, the present invention independently develops a classification model based on group feature selection based deep neural network (Group Feature Selection based Deep Neural Network, DNN-GFS), which is convenient for prediction, and the prediction workload is small (cut out unnecessary nodes and nodes of the input layer in the neural network) side), and has the following advantages:

1. The feature set of the present invention uses thousands of features including comprehensive chromatin three-dimensional structure information, which can better avoid the over-fitting problem of traditional deep neural networks when encountering a large number of features, thereby improving the overall The classification prediction accuracy of ;

2. Unlike the traditional deep neural network, which is like a black box to users, the present invention can perform feature selection on input features while classifying and predicting, and select the most critical features, so as to further mine the relationship between input features and output results. provides the basis for the correlation between

3. The present invention adopts the strategy of group selection, which can group the features that should be together, that is, select the features of the same group at the same time or remove the features of the same group at the same time, so that the model can be well compatible with the prior knowledge of group classification , to improve the self-learning effectiveness of the model.

Description of drawings

Fig. 1 is a principle schematic diagram of the present invention, and the problem that the present invention solves is mainly content in the dotted line frame;

Figure 2 is a schematic diagram of the real labels of immunopositive and immunonegative of 3909 peptides;

Fig. 3 is the prediction method (DNN-GFS) of the present invention and deep neural network (DNN), support vector machine (SVM), logistic regression (LR), K nearest neighbor algorithm (KNN), Neopepsee, pTuneos, DeepHLApan, NetMHCpan, NetMHC Compared with the ROC graph (ROC curve) of different prediction methods such as IEDBimmune under 5 divisions and LOO cross-validation;

Figure 4 is a comparison of the accuracy-recall rate graph (P-R curve) of different prediction methods under 5 divisions and LOO cross-validation;

Fig. 5 is the comparison of the ROC curve and the P-R curve prediction efficacy on the independent verification data set of different prediction methods;

Figure 6 is a comparison of the score distributions of positive samples and negative samples scored by different methods, which are LOO cross-validation, 5-part cross-validation, and scoring on the verification data set;

Figure 7 is a schematic diagram of the deep neural network (DNN-GFS) based on group feature selection in the present invention, wherein a is a diagram of features belonging to different large groups, b is an illustration of the DNN-GFS architecture and group feature selection effect, and c The geometric rationale for the application of different regularization terms to weighted neural network wiring and 2D projections is illustrated from three representative perspectives.

Detailed ways

The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is carried out on the premise of the technical solution of the present invention, and detailed implementation and specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

In recent years, through Hi-C technology, people have been able to dig out the chromatin structure abnormalities of tumor cells more globally, and found that the long-distance regulation between chromatin may play a key role in gene regulation. In the previous work, the inventors of the present application found that in almost all tumors, the concomitant point mutations have obvious proximity in the three-dimensional conformation of chromatin, so they proposed and published the concept of "spatial mutation hotspots of tumors", Extending this concept, we believe that the concept of "cellular functional blocks driven by chromatin three-dimensional conformation" is very important and can help people examine the occurrence and development of tumors from a new perspective. Existing methods for discovering personalized neoantigen peptides in tumor immunity often focus on the sequence properties of antigen peptides, the interaction between antigen peptides and MHC molecules, and the interaction between antigen peptide-MHC molecular complex pMHC and TCR on somatic cells. interaction, etc., while ignoring the source of the antigenic peptide, that is, the special properties of the distribution of the corresponding mutated gene on the chromatin. For the first time, we will systematically analyze the distribution of genes corresponding to antigenic peptides in chromatin space, and find that neoantigens with immunogenicity (capable of activating T cells) and non-immunogenic neoantigens are different in chromatin space. Therefore, we added the spatial distribution information of neoantigens on chromatin to the feature set of the machine learning prediction algorithm, and found that the prediction accuracy of whether neoantigens are immunogenic can be significantly improved.

Based on the above, the present invention implements a tumor neoantigen prediction method based on advanced chromatin conformation and deep sparse learning, which uses a trained deep neural network prediction model based on group selection (deep sparse learning algorithm based on group feature selection) , DNN-GFS, Group Feature Selection based Deep NeuralNetwork) to process the object to be predicted (ie, potential tumor neoantigen peptide), and obtain the immunogenicity information of tumor neoantigen corresponding to the object to be predicted; wherein, the deep neural network Each sample used in the training of the network prediction model includes chromatin 3D conformation information and features generated based on the amino acid sequence of the polypeptide. The principle of this method is shown in the dotted box in Figure 1.

The feature set of each sample is composed of chromatin 3D conformation information and other features generated based on the amino acid sequence of the polypeptide, which is a feature set representing a certain polypeptide. The magnitude of the features of each sample is thousands of levels (more than 5000 features), specifically including the <x, y, z> 3D coordinates of the DNA site corresponding to the target peptide in the chromatin 3D space, the distance from the center of the nucleus (or nuclear membrane), the HLA subtype code of the MHC molecule presenting the antigen peptide, the occurrence frequency of 20 amino acids in the target peptide, the amino acid sparse code of the antigen peptide sequence, the amino acid BLOSM code of the antigen peptide sequence, the Amino acid BLOMAP encoding, amino acid side chain classification encoding of antigenic peptide sequence, amino acid side chain polarity encoding of antigenic peptide sequence, amino acid side chain charge encoding of antigenic peptide sequence, amino acid side chain hydrophilicity and hydrophobicity encoding of antigenic peptide sequence, antigenic peptide sequence Amino acid side chain molecular weight code, amino acid side chain frequency code of the antigenic peptide sequence in biological populations, based on all amino acid AAindex indicators listed in the AAindex (source: https://www.genome.jp/aaindex/) database coding. In this embodiment, each sample is a vector containing 5459 features.

In this embodiment, other features in the sample based on the amino acid sequence of the polypeptide are obtained by high-throughput whole-exome sequencing (ExonSeq) and whole-transcriptome sequencing (RNASeq). According to the results of whole exome sequencing, the mutation information of the tumor cells in the sample can be obtained, and finally the specific coordinate position of a certain mutation on the chromosome number is obtained, and those mutation positions that change the corresponding encoded amino acid can be found point; according to the results of whole transcriptome sequencing, it can be analyzed which genes are highly expressed in tumor cells. On the basis of the above results, all polypeptides containing amino acid mutation sites are enumerated first, and then the corresponding variant polypeptides with high expression are screened out. The length of the polypeptide is set to 9 by default, but it is not limited to 9.

The chromatin 3D conformation information is the chromatin 3D conformation heat map matrix of tumor cells. In this embodiment, the 3D conformation information of chromatin in the sample is obtained through Hi-C experiments or replaced by multiple Hi-C data sets in public databases.

The present invention adopts molecular dynamics (MD) method to develop a human genome three-dimensional conformation modeling method with a resolution of 500kb (bin-size). These containers are coarse-grained beads, and the complete genome is represented by a beaded structure composed of 23 polymer chains. The spatial position of the beads is influenced by chromatin connectivity, which limits linear adjacency of beads in near 3D, and chromatin activity, which ensures that active regions are close to the center of the nucleus. Chromatin activity was determined based on directly calculable compartmentalization of Hi-C substrates described above. The distances of the beads from the center of the nucleus were assigned values according to the spacing index, and molecular dynamics methods were used to optimize the conformation of chromatin from random structures by applying bias potentials to satisfy these distance constraints. For each cell line, 300 feasible conformational structures were optimized from random conformational structures to reduce possible variations and facilitate further analysis.

This prediction method uses a deep neural network prediction model based on group selection to score and classify the polypeptides encoded by the input feature set, and can find out the polypeptides most likely to activate T cell immunogenicity. In this example, the input of the deep neural network prediction model is 5,459 feature codes of a potential neoantigen peptide sequence, and the output is the score of whether the polypeptide can activate immunogenicity. The higher the score, the more immunogen that can activate T cells. sex. When the model is trained, as shown in Figure 7, all features of each sample belong to at most one group, and a group contains one or more elements to facilitate the use of group selection strategies, so that some features that should co-occur or be eliminated together , that is, group features, which can be selected or eliminated at the same time to improve the self-learning effect of the model, reduce the risk of over-fitting, and improve the calculation efficiency of the algorithm.

The deep neural network prediction model of this embodiment adopts the form of a fully connected layer, and its output includes immunogenic neoantigens and several features with the highest degree of correlation with the neoantigens, and the most critical features are screened out while predicting, which can Better clarify the relationship between features and output results.

Figure 2-Figure 6 shows the comparison diagrams of the prediction results of the above prediction method (DNN-GFS) and different classification methods such as DNN, SVM, LR, KNN, Neopepsee, pTuneos, DeepHLApan, NetMHCpan, NetMHC and IEDB immuno. This group of figures shows that in general, our method DNN-GFS is more effective in predicting neoantigens than other traditional machine learning algorithms.

After the above-mentioned prediction method is used for prediction, the predicted tumor neoantigens can be obtained by synthesizing the high-scoring and positively classified polypeptide sequences, and then mice can be used to observe the immune efficacy of the predicted tumor neoantigens.

In another embodiment, a tumor neoantigen prediction device based on high-level chromatin conformation and deep sparse learning is provided, including a data acquisition unit, a model training unit, and a prediction unit, wherein the data acquisition unit acquires samples for training, Each sample includes chromatin 3D conformation information and features generated based on the polypeptide amino acid sequence; the model training unit obtains a deep neural network prediction model based on group selection based on the sample training; the prediction unit obtains the object to be predicted, through the group-based The selected deep neural network prediction model processes the object to be predicted to obtain tumor neoantigen information corresponding to the object to be predicted.

In another embodiment, a computer-readable storage medium is provided, including a computer program, and the computer program can be executed by a processor to implement the prediction method.

In another embodiment, a web page terminal is provided. After obtaining the object to be predicted, the above prediction method is used to quickly obtain the prediction result of the tumor neoantigen.

The preferred specific embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make many modifications and changes according to the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art shall be within the scope of protection defined by the claims.

Claims (9)

1. A tumor neoantigen prediction method based on high-level chromatin conformation and deep sparse learning, characterized in that the method processes the object to be predicted through a trained deep neural network prediction model based on group feature selection, and obtains the same as described Immunogenicity information of tumor neoantigen corresponding to the object to be predicted; Wherein, the feature set of each sample used when the deep neural network prediction model is trained is composed of chromatin 3D conformation information and other features generated based on polypeptide amino acid sequences, and the chromatin 3D conformation information is based on molecular dynamics Each feature of the sample belongs to a certain group, and in the training of the neural network model, all the features in a certain group are either selected or all eliminated; The characteristics of each sample include the <x, y, z> 3D coordinates of the DNA site corresponding to the target peptide in the chromatin 3D space, the distance between the DNA site corresponding to the target peptide and the center of the nucleus or the nuclear membrane, and the presentation of antigenic peptides. The HLA subtype code of the MHC molecule, the occurrence frequency of 20 amino acids in the target peptide, the amino acid sparse code of the antigenic peptide sequence, the amino acid BLOSM code of the antigenic peptide sequence, the amino acid BLOMAP code of the antigenic peptide sequence, the amino acid side of the antigenic peptide sequence Chain classification coding, amino acid side chain polarity coding of antigenic peptide sequence, amino acid side chain charge coding of antigenic peptide sequence, amino acid side chain hydrophilicity coding of antigenic peptide sequence, molecular weight coding of amino acid side chain of antigenic peptide sequence, antigenic peptide sequence The frequency coding of amino acid side chains in biological populations is based on the coding of all amino acid AAindex indicators listed in the AAindex database. 2. The tumor neoantigen prediction method based on high-level chromatin conformation and deep sparse learning according to claim 1, wherein the order of magnitude of the features of each sample is thousands of levels. 3. The tumor neoantigen prediction method based on high-level chromatin conformation and deep sparse learning according to claim 1, wherein the output of the deep neural network prediction model based on group feature selection includes tumor immunity to potential neoantigens Prediction of originality and several features most correlated with this neoantigen. 4. The tumor neoantigen prediction method based on high-level chromatin conformation and deep sparse learning according to claim 1, wherein the deep neural network prediction model based on group feature selection adopts a fully connected layer form. 5. The tumor neoantigen prediction method based on high-level chromatin conformation and deep sparse learning according to claim 1, wherein the 3D conformation information of chromatin is derived from the 3D conformation of cellular chromatin obtained by Hi-C experiment Heatmap matrix. 6. The tumor neoantigen prediction method based on chromatin advanced conformation and deep sparse learning according to claim 1, wherein the chromatin 3D conformation information is obtained from a public Hi-C dataset. 7. The tumor neoantigen prediction method based on high-level chromatin conformation and deep sparse learning according to claim 1, wherein the features generated based on the polypeptide amino acid sequence include polypeptide features containing amino acid mutation sites and features containing The high expression information of the mutated gene was obtained. 8. A tumor neoantigen prediction device based on advanced chromatin conformation and deep sparse learning, characterized in that it includes: The data acquisition unit acquires samples used for training, and the feature set of each sample is composed of chromatin 3D conformation information and other features generated based on polypeptide amino acid sequences, and the chromatin 3D conformation information uses human Genome three-dimensional conformational modeling method to obtain; The model training unit obtains a deep neural network prediction model based on group feature selection based on the sample training, each feature of the sample belongs to a certain group, and in the neural network model training, the features in a certain group or all are selected , or all are eliminated; The prediction unit acquires the object to be predicted, processes the object to be predicted through the deep neural network prediction model based on group feature selection, and obtains the tumor neoantigen immunogenicity information corresponding to the object to be predicted; The characteristics of each sample include the <x, y, z> 3D coordinates of the DNA site corresponding to the target peptide in the chromatin 3D space, the distance between the DNA site corresponding to the target peptide and the center of the nucleus or the nuclear membrane, and the presentation of antigenic peptides. The HLA subtype code of the MHC molecule, the occurrence frequency of 20 amino acids in the target peptide, the amino acid sparse code of the antigenic peptide sequence, the amino acid BLOSM code of the antigenic peptide sequence, the amino acid BLOMAP code of the antigenic peptide sequence, the amino acid side of the antigenic peptide sequence Chain classification coding, amino acid side chain polarity coding of antigenic peptide sequence, amino acid side chain charge coding of antigenic peptide sequence, amino acid side chain hydrophilicity coding of antigenic peptide sequence, molecular weight coding of amino acid side chain of antigenic peptide sequence, antigenic peptide sequence The frequency coding of amino acid side chains in biological populations is based on the coding of all amino acid AAindex indicators listed in the AAindex database. 9. A computer-readable storage medium, comprising a computer program that can be executed by a processor to implement the prediction method according to any one of claims 1-7.

Images