CN116959555A - Method and system for compound-protein affinity prediction based on protein three-dimensional structure - Google Patents

Abstract

The invention provides a method for predicting compound-protein affinity based on a protein three-dimensional structure, which comprises the following steps: s1, a compound feature extraction step, namely obtaining updated atomic features and aggregated node features by using a deep graph convolution network and a multi-head attention algorithm; s2, protein characteristic extraction, namely enabling sequence characteristics and structural characteristics of the protein to represent more complete protein information through a characteristic aggregation algorithm and a co-evolution strategy; s3, predicting the affinity of the compound and the protein, and obtaining a predicted affinity value according to the atomic characteristics, the polymerization node characteristics and the sequence characteristics. The invention also discloses a corresponding system comprising: compound extractor, protein extractor and affinity predictor. The invention uses a discretization distance matrix and a torsion angle matrix as protein three-dimensional structure representation, introduces a co-evolution updating mechanism to update the characteristics between the protein three-dimensional structure and the sequence, and utilizes the characteristic of aggregation nodes to improve the affinity prediction precision.

Description

Method and system for predicting compound-protein affinity based on protein three-dimensional structure

Technical Field

The present invention relates to the field of drug research and development, and in particular to a method and system for predicting compound-protein affinity based on protein three-dimensional structure.

Background Art

Accurately predicting the binding affinity of compounds to proteins is a major challenge in the virtual drug candidate screening process. Using computational methods to conduct effective high-throughput virtual screening of lead compounds can greatly accelerate the process of drug development by reducing R&D time and experimental workload. Especially in recent years, with the development of technology, more and more data sharing projects have been vigorously promoted, and a large number of new biomedical data sources have become accessible. For example, PubChem currently contains more than 111 million compound structures and 271 million bioactivity data points (cited https://doi.org/10.1093/nar/gkaa971), which greatly increases the prediction potential of compound-protein binding affinity (CPA). However, further improving the accuracy of predicted CPA remains the main challenge facing virtual drug screening technology before it can be applied in practice. In order to improve the accuracy of CPA prediction, the development of computational methods basically follows two paths: structure-free or structure-based. Although these two methods have helped to improve the accuracy of CPA prediction, both methods also face the challenge of further improving their CPA prediction accuracy.

Non-structural parameter models regard compounds (drugs) as atomic combination information and proteins (targets) as residue combination information, and calculate the distances between pairs of atoms and residues to form a [number of atoms]-[number of residues] pairwise interaction matrix. Under a certain threshold, the compound atoms are considered to be in contact with the corresponding protein residues (i.e., non-covalent interactions). In the pairwise interaction matrix, each element is a binary value, indicating whether the corresponding atom-residue pair interacts. These non-structural parameter models rely on the sequences of drugs and targets to learn interactions from paired matrices, with the goal of predicting the affinity between drugs and targets. For example, multi-layer one-dimensional convolutional neural networks (1d-CNNs) are used to extract features from protein sequences, and the features represented by these matrices are used as inputs to another downstream deep learning (DL) model to ultimately learn the affinity of drugs and targets. However, this interaction is constantly changing with time and environmental changes, because the structure of the protein is dynamic, and the pairing information obtained is only a certain state captured by the electron microscope. Therefore, this structure-free parameter model that ignores the binding site information in the three-dimensional structure and only extracts information on a single type of protein feature (considering only sequence features) is likely to have a great limitation in CPA prediction.

Previous studies have shown that the predicted pairwise non-covalent traction results can be used in structure-free models to further improve their accuracy in CPA prediction. However, simply using the area under the curve (AUC Area Under the Curve) as an evaluation metric to measure the performance of pairwise matrix prediction can be misleading in the interpretation of the results. As an evaluation metric, AUC is the area under the receiver characteristic curve (ROC), and its Y-axis and X-axis are measures of sensitivity, which are used to evaluate the ability of the model in true positive prediction or true negative prediction. It is plotted by considering different thresholds for distinguishing positive and negative values in binary classification. As shown in Figure 16, there are very many negative values and only a few positive (molecular compound-protein interaction) values in the compound-protein pairing matrix, for example, in a dataset with 5 positive pairs and 11,595 negatives. Even if the model predicts an all-negative matrix, the AUC value is still high (over 0.95) in this case. The predicted pairwise interaction matrix is an intermediate result generated during the model training process, which is inappropriately input into the original model for CPA prediction. Mathematically speaking, any intermediate result produced during the training of a machine learning model is just a representation of the input features (i.e., compound sequence, protein sequence, and pairwise interaction matrix). In theory, when this representation is fed back together with its upstream features to predict the final response (i.e., CPA), the response is still a representation of their original input features. In our experiments, we found that the act of feeding the intermediate results into the unstructured model together with the original features does not benefit the CPA prediction, indicating that the unstructured model cannot improve the accuracy of CPA prediction without the input of new information. Therefore, models built on this approach may find it difficult to further improve their accuracy.

Regarding structure-based methods for predicting CPA, one of them is molecular docking technology. Docking is a bioinformatics modeling process between target protein fragments and ligands, which takes into account the potential binding sites and three-dimensional structures of protein-drug complexes during the prediction process. However, only a few protein 3D structures are currently available, which greatly limits the application of this method. To overcome this limitation, more and more methods based on machine learning (ML) algorithms are being proposed, which rely not only on sequence data of drugs and targets, but also on spatial three-dimensional information (e.g., three-dimensional spatial coordinates, the distance between drugs and targets in ligands, and torsion angle information between protein residues) in order to obtain the ability to predict CPA more accurately. Although attractive in theory, it is very difficult to develop a practical paradigm to use protein three-dimensional structure information for CPA prediction.

At present, the main obstacle to applying protein three-dimensional information to computational models is the lack of suitable methods or models to reasonably utilize protein three-dimensional information. Some studies directly introduced three-dimensional structural information into traditional models, but failed to achieve good prediction results. Most of the existing methods for compound-protein affinity prediction also use sequence information to predict affinity values, but the prediction effect on some data sets has not been significantly improved. It is still an urgent need to develop a practical multimodal model to reasonably utilize protein three-dimensional structural information and sequence information to further improve the accuracy of CPA prediction.

Summary of the invention

In order to overcome the defects of the prior art, the purpose of the present invention is to propose a reasonable characterization method for the three-dimensional structure of proteins, and to develop a multimodal end-to-end model that can simultaneously aggregate the sequence information and structural information of proteins. After characterizing and aggregating the structural information and sequence information of proteins, a co-evolutionary attention algorithm is used to optimize the extraction of the structural features and sequence features of proteins; a graph convolution layer with special residual links is introduced to avoid the over-smoothing problem encountered in deep graph convolutions, so as to extract compound features more comprehensively; an interactive attention algorithm is introduced to realize the feature interaction between proteins and compounds in the affinity prediction module, so as to learn the potential interaction information between proteins and compounds, so as to improve the prediction accuracy of affinity.

The present invention provides a method for predicting compound-protein affinity based on protein three-dimensional structure, which is used for predicting compound-protein affinity. Compound features are extracted from compounds, and the compound features include atomic features and aggregate node features. Protein features containing implicit protein three-dimensional information are extracted from proteins. Compound features and protein features are then used to predict compound-protein affinity through an affinity prediction algorithm. The method specifically includes the following steps:

S1, compound composite feature extraction step,

According to the compound characterization graph, the initial atomic features and aggregate node features are obtained. The atomic features and aggregate node features are cyclically updated using a deep graph convolutional network and a multi-head attention mechanism to output the final atomic features and aggregate node features of the compound.

S2, protein feature extraction step;

Obtain protein sequence information and protein structural features; according to the protein sequence information and protein structural features, use a protein feature aggregation algorithm to make the protein embedded sequence features carry protein structural features, thereby obtaining protein embedded sequence features and embedded structural features; cyclically update the embedded sequence features and embedded structural features by adopting a co-evolution strategy, and finally obtain updated protein sequence features and protein structural features;

S3 compound and protein affinity prediction step;

According to the atomic features and aggregation node features of the compound obtained in step S1 and the protein sequence features in step S2, the predicted affinity value is obtained through the affinity learning unit algorithm. The larger the affinity value, the greater the probability of affinity between the compound and the protein.

Preferably, the step S1 specifically comprises:

S11, obtaining atomic features and aggregation node features according to the compound characterization graph;

S12, the atomic features are input into the deep graph convolutional network, the output of the deep graph convolutional network and the aggregate node features are added according to the first weight of the compound features, and the updated atomic features are obtained by combining the first gated recurrent function (GRU) and the atomic features;

S13, the atomic features are input into the multi-head attention representation algorithm, the output of the multi-head attention representation algorithm and the aggregate node features are added according to the second weight of the compound features, and the updated aggregate node features are obtained by combining the second gated recurrent function (GRU) and the aggregate node features;

S14: After repeating steps S12-S13 K times using the updated atomic features and the updated aggregate node features as the atomic features and the aggregate node features, the atomic features and the aggregate node features are output.

Preferably, the deep graph convolutional network in step S1 has a residual connection loop for aggregating the neighbor information of atoms onto atoms and avoiding the over-smoothing problem caused by increasing the number of network layers. GRU is used to aggregate the information between the front and back layers of the network, and a multi-head attention mechanism characterization algorithm is used to obtain diverse compound features, thereby improving the accuracy of affinity prediction in the affinity prediction step.

Preferably, the step S2 is specifically:

S21, obtaining protein sequence information and protein structural features, where the protein structural features at least include a discretized distance matrix;

S22, the protein feature aggregation algorithm is specifically as follows: the protein sequence information is passed through the first embedding layer (Eembedding layer) to obtain a sequence vector, the discretized distance matrix is passed through the second embedding layer (Eembedding layer) to obtain a discrete distance matrix feature vector, and the discrete distance matrix feature vector is used as the embedded distance matrix after feature embedding;

S23, the co-evolution strategy is implemented using N-layer protein coding algorithms, where the protein coding algorithms of each layer are the same, and the protein coding algorithms of one layer are specifically:

The two results of row addition and column addition of the embedding distance matrix are fused through the gated recurrent unit to obtain the spliced embedded structure information; the spliced embedded structure information then enters the diversity convolution layer to obtain diversified protein structure features, so as to learn protein features with more diverse features;

The embedded sequence features are sequentially passed through the diversity convolution layer and the common convolution layer to obtain diversified protein sequence features; the diversified protein structure features are added to the diversified protein sequence features through the gating logic, and sequentially passed through the gated recurrent unit and the convolution layer to obtain a structural feature vector that has both structural information and sequence information;

The structural feature vector undergoes its own combination transformation (outer sum) to output the protein structural features;

The updated sequence features are embedded through a diversity convolution layer to obtain the updated sequence features after diversified embedding; the updated sequence features after diversified embedding are added with the diversified protein structure features through gated logic to obtain a sequence feature vector that integrates the structure information and sequence information; the diversified protein sequence features are further integrated with the sequence feature vector through a gated recurrent unit to output the protein sequence features;

S24, the new embedding distance matrix output by the protein coding algorithm is the required embedding distance matrix. The obtained embedding distance matrix and embedded sequence features are used as the input of the next layer of protein coding algorithm to continue updating the embedding distance matrix and embedded sequence features until N layers of protein coding algorithms are completed. At this time, the output embedding distance matrix is the structural feature of the protein, and the embedded sequence feature is the sequence feature of the protein.

Preferably, the protein structure feature in step S21 also includes a feature auxiliary matrix; the embedded sequence feature obtained in step S22 is also added to the feature auxiliary matrix through gated logic to obtain the embedded updated sequence feature. The diversity convolution layer in step S23 divides the input feature vector into four equal parts in the feature dimension, and enters four parallel ordinary convolution layers respectively, and adds the output results, and at the same time aggregates the initial features of the sequence information input through residual connection to obtain a diversified feature vector.

Preferably, the characteristic auxiliary matrix is a torsion angle matrix, which is constructed using the sine and cosine values of the dihedral angles φ and ψ between the α-carbon atom of the protein backbone chain and the amino and carboxyl groups.

Preferably, the discretized distance matrix in step S21 divides the distances between residues in the protein stable structure into M equidistant mapping intervals that are statistically substantially in line with the normal distribution, thereby realizing the discretized encoding of the distance matrix, where M is a positive integer.

Preferably, the step S3 is specifically:

S31, receiving the atomic features and aggregate node features of the compound from step S1 as compound information and passing them into the affinity learning unit, the aggregate node features are updated through a linear layer to obtain updated aggregate node features; the compound features are updated through a linear layer, and the mean is calculated in the atomic dimension, and the features are concatenated with the updated aggregate node features to obtain the comprehensive features of the compound;

S32, receiving the sequence features of the protein from step S2 and passing the sequence features of the protein as protein information into the affinity learning unit, the protein sequence features are updated through the convolution layer, and the average is calculated in the residue dimension to obtain the comprehensive protein features;

The protein comprehensive features and compound comprehensive features are matrix multiplied (matmul), and the information of the two is fused through a linear layer to obtain the predicted affinity value.

The present invention also discloses a system for predicting compound-protein affinity based on protein three-dimensional structure, comprising: a compound extractor, a protein extractor and an affinity predictor;

The compound extractor uses a deep graph convolutional network and a multi-head attention algorithm to update the atomic features and aggregate node features of the input compound through an attention mechanism, and outputs the final atomic features and aggregate node features through a loop; the final compound features and aggregate node features are sent to the affinity predictor;

The protein extractor obtains embedded sequence features and embedded structural features through a protein feature aggregation algorithm according to the input protein sequence features and protein structural features; updates the embedded sequence features and embedded structural features by adopting a co-evolution strategy to obtain updated protein sequence features and protein structural features;

The affinity predictor is used to receive compound features, protein sequence features and aggregation node features, and obtain the predicted affinity value through the affinity learning unit algorithm.

Preferably, the compound extractor includes a deep graph convolution unit and a multi-head attention representation unit, the deep graph convolution unit is used to implement a deep graph convolution network, and the multi-head attention representation unit is used to implement a multi-head attention algorithm; the protein extractor includes a protein information aggregation unit and a co-evolution update unit, the protein information aggregation unit is used to implement a protein feature aggregation algorithm, and the co-evolution update unit is used to implement a co-evolution strategy; the affinity predictor includes an affinity learning unit.

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

1. When using the distance matrix to record the distance between protein residues, the discretized distance matrix is used to reduce the dimension of the data representation vector in the word embedding stage; and the torsion matrix is constructed using the torsion angles (dihedral angles) φ and ψ between the α carbon atom and the amino and carboxyl groups in the protein skeleton, which accurately represents the direction of the sequence, fully reflects the three-dimensional structure of the protein, and solves the problem of high-dimensional protein representation.

2. Introduce a co-evolution strategy to update the sequence information and structure information of proteins in a coordinated manner, strengthen the correlation between protein sequence information and structure information, and extract more comprehensive and diverse feature information in the final protein representation vector to characterize proteins;

3. A residual structure with initial information of graph nodes is introduced to avoid over-smoothing problem. On this basis, an aggregation node is introduced to aggregate the global information of compounds. In the graph convolution process, a gated update unit and a message passing mechanism of graph information are used to realize the information interaction and update between each atomic node and the aggregation node in the graph structure.

4. An affinity learning unit is introduced into the affinity prediction unit to aggregate the features of proteins and compounds, so as to learn the potential interaction information between proteins and compounds more comprehensively and improve the prediction accuracy of affinity.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of a process for predicting compound-protein affinity based on protein three-dimensional structure;

FIG2 is a flow chart showing the specific steps of compound composite feature extraction;

Figure 3 is a schematic diagram of the process of deep graph convolutional network;.

Figure 4 is a flowchart of the specific steps of the deep graph convolutional network;

FIG5 is a schematic diagram of the atomic feature extraction process;

FIG6 is a schematic diagram of the protein feature extraction process;

FIG7 is a flowchart of the specific steps of the protein feature aggregation algorithm;

FIG8 is a flowchart of the specific steps of the single-layer protein encoding algorithm;

FIG9 is a schematic diagram of the process of the diversity convolutional layer;

FIG10 is a schematic diagram of the flow chart of the single-layer protein encoding algorithm;

FIG11 is a flowchart of the specific steps of the affinity learning unit algorithm;

FIG12 is a comparison of the performance of various models on a data set (small data set) after similarity clustering of compounds based on molecular fingerprints;

FIG13 is a comparison of the performance of various models on a data set (small data set) after homology clustering of proteins based on a multiple sequence alignment strategy;.

Figure 14 is a comparison of the convergence of each model during the training process (model convergence speed and final convergence position);

FIG15 is a diagram showing the effect of various performance evaluation indicators as the number of graph convolution layers increases;

FIG16 shows the problems existing in the performance evaluation of the existing non-structural parameter model.

DETAILED DESCRIPTION

In order to better understand the technical solution of the present invention, the specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings need not be drawn to scale unless otherwise specified.

The present invention proposes a method for predicting compound-protein affinity based on protein three-dimensional structure, which is used to predict compound-protein affinity, by extracting compound features from compounds, the compound features include atomic features and aggregate node features, extracting protein features that contain implicit protein three-dimensional information from proteins, the protein three-dimensional information includes a discretized distance matrix and a torsion angle matrix, and then using the compound features and protein features through an affinity prediction algorithm to predict compound-protein affinity. In order to implement this method, an end-to-end model based on deep learning and the latest technology is used, and the model is named Fast Evolving Attention and Deep Graph Neural Network (FeatNN). A method for predicting compound-protein affinity based on protein three-dimensional structure implemented by FeatNN is shown in Figure 1, and the specific implementation steps are as follows:

S1, compound composite feature extraction step

The compound feature extraction step is implemented using the compound extractor.

Compound features include atomic features and bond features. In order to better characterize compound features, aggregate node features are also introduced. Compound features and aggregate node features are collectively referred to as compound composite features.

According to the compound characterization graph G = {atom, bond}, the initial atomic features and aggregate node features are obtained. The deep graph convolutional network and multi-head attention algorithm are used to update the atomic features and aggregate node features through the attention mechanism. The final atomic features and aggregate node features are output through a loop. At this time, the atomic features already contain the bond features, so the output atomic features can represent the compound features. The specific steps are shown in Figure 2.

The invention of this step is that the special residual connection loop of the deep graph convolutional network is used to aggregate the neighbor information of atoms onto the atoms and avoid the over-smoothing problem caused by the deepening of the network layers. GRU is used to aggregate the information between the front and back layers of the network, and a multi-head attention representation algorithm is used to obtain more diverse compound features. Aggregate node features are introduced, and an interactive evolution mechanism is used to update the aggregate node features and atomic features.

In order to clearly describe the steps of compound composite feature extraction, the following relevant definitions are given first. According to the molecular formula of the compound, it can be converted into a compound representation graph G = {atom, bond}, where atom represents atoms and bond represents bonds. Atom contains atomic features such as element name, aromatic type, vertex degree and atomic valence, and bond contains bond type and shape features. The values of atom and bond can be feature encoded separately through separate hot encoding strategies. According to the compound representation graph G, the features of each atom and bond in the compound can be obtained as shown below: For each atom For each key where i = 1, 2, ..., _Na , j = 1, 2, ..., _Nb , and the original atomic features are defined as The sum of atomic features is called the aggregate node feature, i.e., Considering the graph convolution layer l _c and multi-head attention features (k _c is the number of multi-heads), atom features or main node features, l _c th is defined as and The variable V with k _c head is defined as Where l _c =1, 2, ..., l _comp , k _c =1, 2, ..., k _comp .

The compound features can be obtained based on the compound graph representation G = {V, E}. More specifically, each node (i.e., atom) _vi is initially represented by a feature vector of length 82 It is represented by the concatenation of word embeddings representing atomic symbols, representing the atomic symbols, degree, explicit valence, implicit valence, and aromaticity of the corresponding atoms. Each edge (i.e., chemical bond) e _i E is initially composed of a feature vector of length 6 It represents the concatenation of word embeddings representing bond types, including mono, di, tri, aromatic, the bond is conjugated, and the bond is in a ring.

First, the atomic features of the compound are input into the linear layer and the deep graph convolutional network respectively. After being processed by the linear layer, the atomic features are input into the atomic aggregation function, the deep graph convolutional network, the multi-head attention representation algorithm and the first gated recurrent function (GRU) respectively. The atomic aggregation function is used to obtain the initial aggregation node features. The aggregation node features and the output results of the deep graph convolutional network are respectively processed by the linear layer and added, and the first weight of the compound features is obtained through the first activation function (Sigmoid). The aggregation node features and the output results of the multi-head attention representation algorithm are respectively processed by the linear layer and added, and the compound features are obtained through the first activation function (Sigmoid). The second weight, the aggregation node feature is processed by the linear layer and added to the output result of the deep graph convolutional network. The weight values of the two are the first weight of the compound feature and 1 minus the first weight of the compound feature, respectively. The result after the addition is output to the first gated recurrent function (GRU). The first gated recurrent function outputs the atomic feature, and the atomic feature is updated at this time; the aggregation node feature is processed by the linear layer and added to the output result of the multi-head attention representation algorithm. The weight values of the two are 1 minus the second weight of the compound feature and the second weight of the compound feature, respectively. The result after the addition is output to the second gated recurrent function (GRU). The second gated recurrent function outputs the aggregation node feature, and the aggregation node feature is updated at this time;

The updated aggregate node features and atomic features are used as input again for looping. The aggregate node features and atomic features obtained after looping K times are the aggregate node features and atomic features to be output. K is a positive integer greater than 2, and the specific value is set according to needs.

Graph convolutional networks (GCNs) can be used in biological research to predict the affinity of drugs to targets, but due to the limitation of depth, this method always has an over-smoothing problem. When the number of layers increases, the performance of GCNs deteriorates because the representations of nodes in GCNs converge to similar values. For deep graph convolutional networks, special residual connection loops are used to aggregate the neighbor information of atoms onto atoms and avoid the over-smoothing problem caused by deepening the number of network layers. Increasing the number of GCN layers can extract more information.

The purpose of the deep graph convolutional network in the present invention is to aggregate the neighbor information of atoms onto atoms, and the method used is as shown in Figure 3. For example, for the carbon atom C located in the circle, the features of the surrounding nitrogen atoms N, oxygen atoms O and bromine atoms Br atoms, as well as the bond features between them, will be aggregated into the features of the carbon atom through the deep graph convolutional network. The specific implementation steps are shown in Figure 4. The atomic node features and the adjacency matrix aggregate and update the neighbor node features of each atom on each atom through a message passing mechanism, and are supplemented by a special residual connection loop to prevent the problem of over-smoothing when deepening the number of network layers, where α and θ are the hyperparameters of the residual connection loop. The atomic feature extraction cycle is shown in Figure 5, where M represents the multi-head attention aggregation operation, and R represents the number of cycles for building the atomic diversity feature loop.

The multi-head attention representation algorithm is used to implement the multi-head attention mechanism, which is specifically implemented as the existing technology. In the present invention, it is used to increase the diversity of the features of the atomic nodes of the compound, so as to aggregate the compound information more comprehensively, learn more accurate compound representations, and improve the accuracy of affinity.

Therefore, at this point, the atomic features of the compound can represent the compound characteristics after repeated processing by the deep graph convolutional network and multi-head attention.

The function of a linear layer is to map an input vector to a specified dimension for subsequent operations between vectors or matrices. All linear layers mentioned in this application are different linear layers and are set according to actual needs.

The compound composite feature extraction step is implemented in a computer using the following algorithm 1. The number of layers of the deep graph convolutional network is set as needed, and the graph convolution is implemented in a computer using the following algorithm 2.

S2, protein feature extraction step;

According to the protein sequence features and protein structure features, the embedded sequence features and embedded structure features are obtained by the protein feature aggregation algorithm, and then the updated protein sequence features and protein structure features are obtained by the protein encoding algorithm using the co-evolution strategy, as shown in Figure 6. The protein feature extraction step is implemented in the computer using the following algorithm 3.

The inventive point of this step is to utilize the structural features of proteins, and through the protein feature aggregation algorithm and protein encoding algorithm proposed by the present invention, the protein structural features are implicitly embedded in the protein sequence features.

S21, obtaining sequence and structural features of proteins;

Protein sequence information can be directly obtained from relevant databases.

The structural feature is a three-dimensional structural feature, which includes at least a discretized distance matrix and may also include a structural auxiliary feature. In this embodiment, the structural feature includes a discretized distance matrix and a torsion angle matrix as a structural auxiliary feature.

The discretized distance matrix of a protein is the distance between residues after the protein is unfolded into a peptide chain, and the discretized distance matrix is a symmetric matrix. Because the distance between residues is usually the Euclidean distance, the specific value is a value in a continuous interval. Since a single value cannot represent the structural features well, it is not easy to understand the machine learning model. Therefore, a discrete distance matrix is proposed to map the distance matrix to a low-dimensional bin with equal width, that is, the continuous interval is divided into multiple equal interval segments, and the values in the same interval segment are set to the same value, which greatly reduces the amount of data in the distance matrix and can better reflect the distance characteristics between residues. Compared with the traditional distance matrix, the distance value change interval is small and there is no obvious distinction. It is not convenient for feature extraction when the data is limited. The present invention divides these continuous values with little distinction into equidistant mapping intervals that are basically in line with the normal distribution in statistics, and realizes the discretized encoding of the distance matrix. In this embodiment, the distance matrix is mapped to 38 low-dimensional bins of equal width between 3.25A and 50.75A (A is the unit of length angstrom, 1 meter = 10 to the 10th power angstrom), with an additional bin to store any larger distances and one bin to store any smaller distances, for a total of 40 bins.

The torsion angle matrix is constructed using existing technology. Considering that the distance information between each residue alone may not accurately describe the three-dimensional structure of the protein, the torsion angle matrix can be used as an auxiliary to accurately represent the direction of the protein sequence. The torsion angle matrix is constructed using the sine and cosine values of the dihedral angles φ and ψ between the α carbon atom of the protein backbone chain and the amino and carboxyl groups.

S22, obtaining embedded sequence features and embedded distance matrix through protein feature aggregation algorithm according to protein sequence features and three-dimensional structure features, as shown in FIG7 , the specific steps are:

The protein sequence information is embedded in the first embedding layer (Eembedding layer) to obtain sequence features.

The discrete distance matrix is embedded in the second embedding layer to obtain discrete distance matrix features, and the output is a sequence feature with discrete distance matrix information.

After the twist angle matrix is extracted through the convolution layer, the twist angle feature is obtained; the twist angle feature is output through the linear layer, and the first activation function (Sigmoid) is used to obtain the twist angle weight; the twist angle feature and the sequence feature are gated with the twist angle weight as the gated weight, added through the gated logic, and the output is the embedded updated sequence feature.

The protein feature aggregation algorithm is implemented in a computer using the following Algorithm 5.

S23, obtaining updated sequence features and updated structural features through a co-evolution strategy for the embedded sequence features and the embedded distance matrix;

The co-evolution strategy is implemented using an N-layer protein coding algorithm, where the protein coding algorithm of each layer is the same. The specific steps of one layer of the protein coding algorithm are shown in FIG8 .

The two results of the embedded distance matrix after row addition and column addition are respectively fused through the gated recurrent unit to obtain the spliced embedded structural information; the spliced embedded structural information then enters the diversity convolution layer to obtain diversified protein structural features, so as to learn protein features with more diverse features; at the same time, the spliced embedded structural information passes through the linear layer and uses the second activation function (Sigmoid) to obtain the structural information weight.

The embedded sequence features pass through the diversity convolution layer and the ordinary convolution layer in turn to obtain diversified protein sequence features; the diversified protein sequence features pass through the linear layer and use the third activation function (Sigmoid) to obtain the sequence information weight.

The diverse protein structural features and the diverse protein sequence features are added together through a gating unit using the sequence information weight as the gating weight to obtain a structural feature vector that initially aggregates the sequence information and structural information; when adding, the weight of the diverse protein sequence features is the sequence information weight, and the weight of the diverse protein structural features is 1 minus the embedded structural information weight.

The structural feature vector that initially aggregates sequence information and structural information and the diversified protein structural features updates the structural feature vector from the sequence dimension through a gated recurrent unit, and changes the dimension mapping through a convolutional layer to obtain a structural feature vector that has both structural information and sequence information; after the structural feature vector specifies the mapping dimension through the convolutional layer, it combines and transforms itself (outer sum) into a protein structural feature. This embodiment uses an outer sum function to implement it, that is, each time an element of the first input is multiplied by the second input as a row, until the elements of the first input are also traversed.

The updated sequence features are embedded through a diversity convolutional layer to obtain the updated sequence features after diversified embedding. The structural information weight is used as the gating weight. The updated sequence features after diversified embedding and the diversified protein structural features are added through the gating logic to obtain the sequence feature vector integrating the structural information and the sequence information. The updated sequence features after diversified embedding are further fused with the sequence feature vector through a gated recurrent unit to output the protein sequence features. When adding, the weight used for the diversified protein structural features is the sequence information weight, and the weight of the updated sequence features after diversified embedding is 1 minus the sequence information weight.

In this embodiment, the ordinary convolution layer performs traditional one-dimensional convolution operations; the diversity convolution layer can perform convolution operations in parallel and enhance the diversity of the learned features: the diversity convolution layer first divides the input feature vector into four equal parts using an equal division strategy in the feature dimension, and enters four parallel ordinary convolution layers respectively, and adds the output results. At the same time, the initial features of the sequence information input are aggregated through residual connections, and finally a diversified feature vector is obtained, as shown in Figure 9; the equal division strategy is to directly divide the vector into four equal parts from the feature dimension, which must meet the premise that the feature number can be divided by the number of divisions (in this embodiment, the number of divisions is four), and the following algorithm 9 is used in the computer to implement the above-mentioned equal division strategy.

The diversity convolutional layer is implemented in the computer using the following algorithm 7.

At this point, the distance matrix features output by the protein encoding algorithm are the required protein distance matrix, as shown in Figure 10. The obtained distance matrix and sequence features are used as the input of the next layer of protein encoding algorithm to continue updating the protein structure features and sequence features until all N layers of protein encoding algorithms are completed. At this time, the protein sequence features have already implied the protein structure features. Therefore, the protein sequence features at this time can represent the protein features.

Protein coding is implemented in computer using the following algorithm 4.

The following describes “S3, compound and protein affinity prediction step” with reference to FIG. 11 ;

In the compound and protein affinity prediction step, an affinity learning unit algorithm is introduced to learn the potential features of the interaction between proteins and compounds, and the features are aggregated separately to finally predict the affinity;

According to step S1, the atomic features and aggregate node features of the compound are received as compound information and transmitted to the affinity learning unit, and the aggregate node features are updated through the linear layer to obtain updated aggregate node features; the compound features are updated through the linear layer, and the mean is calculated in the atomic dimension at the same time, and the features are spliced with the updated aggregate node features to obtain the comprehensive features of the compound;

According to step S2, the sequence features of the protein are received and passed into the affinity learning unit as protein information. The protein sequence features are updated through the convolution layer, and the average is calculated in the residue dimension to obtain the comprehensive protein features.

The protein comprehensive features and compound comprehensive features are matrix multiplied (matmul), and the information of the two is fused through a linear layer to obtain the predicted affinity value. The basic process is shown in Figure 11, and the following algorithm 8 is used in the computer to implement it.

The inventive point of this step is to use the protein sequence features that imply protein structural features and the atomic features that represent compound features to fuse them using the affinity learning unit algorithm proposed by the present invention, and combine them with the aggregate node features to finally predict the affinity value.

The compound and protein affinity prediction steps were implemented in silico using the following algorithm8.

The present invention also proposes a compound-protein affinity prediction system based on protein three-dimensional structure, which is used to predict compound-protein affinity, as shown in FIG1 , and specifically includes the following modules:

a first input module, a second input module, a compound extractor, a protein extractor, and an affinity predictor;

The first input module receives the compound characterization graph, and obtains the compound atomic features and the aggregation node features according to the compound characterization graph;

A compound extractor, used for receiving compound atomic features and aggregation node features, obtaining updated compound features and aggregation node features according to the input, and sending the updated compound features and aggregation node features to an affinity predictor;

The second input module receives protein sequence features and protein structure features, the protein structure features include a discretized distance matrix and a torsion angle matrix, and sends the received three features to a protein extractor;

The protein extractor is used to receive protein sequence features and protein structure features, update the protein sequence features, and send them to the affinity predictor.

The affinity predictor is used to receive compound features, protein sequence features and aggregation node features, perform affinity prediction, and provide affinity prediction values for compounds and proteins.

Furthermore, the compound extractor includes a deep graph convolution unit and a multi-head attention representation unit. The deep graph convolution unit is used to implement a deep graph convolution network, which is used to aggregate the neighbor information of atoms onto atoms. The multi-head attention representation unit is used to extract more diverse compound features to improve the prediction accuracy of affinity.

The protein extractor includes a protein information aggregation unit and a co-evolution update unit. The protein information aggregation unit is used to implement a protein feature aggregation algorithm so that the sequence features of the protein carry the structural information of the protein. The co-evolution update unit is used to implement a co-evolution strategy so that the extracted protein sequence features can represent the global information of the protein.

The affinity predictor consists of affinity learning units.

The method of the present invention is compared with the prior art below, and the model implementing the present invention is named Fast Evolving Attention and Deep Graph Neural Network (FeatNN).

1. Build a test dataset

In the study, two data sets based on PDBbind (version 2020, with a total of 23,496 entries) and BindingDB (version February 6, 2022, with a total of 41,296 entries) were constructed. The PDBbind database provides experimental binding affinity data for a total of 23,496 biomolecular complexes in the PDB. In this embodiment, different measurement values of K _i , K _d and IC ₅₀ of the protein-ligand affinity data (12,628 data) after data cleaning are mainly used, and the training set: validation set: test set = 7:1:2 are used as small data set training samples. BindingDB contains 241,268 binding data of 8,661 proteins and 1,039,940 small molecules. In this embodiment, the IC ₅₀ measurement type of the protein-ligand affinity data (218,615 data) after data cleaning is mainly used, and the training set: validation set: test set = 7:1:2 is used as a large data set training sample.

Check the normality of the correlation data, because extreme data may affect the entire model. The BindingDB dataset has some extremely extreme correlation data, so the external data (μ-3σ, μ+3σ) are eliminated.

For the distance matrix and torsion angle matrix, protein structure data were obtained by searching in the RCSB PDB (Protein Data Bank) database using the PDB IDs provided by the two databases. The RSCB PDB files provide sufficient 3D structural data (spatial position) to extract the distance matrix needed to be used in the model, and the torsion matrix was calculated based on this 3D structural information.

For SMILES sequences, we obtained data from RCSB PDB based on ligand IDs.

2. Determine the evaluation indicators,

Indicators are used to evaluate the predictive performance of the model. The present invention uses the following indicators: R ² , RMSE (root mean square error), MAE (absolute mean error), MedAE, Pearson correlation coefficient, and Spearman rank correlation coefficient; among them, R ² , RMSE (root mean square error), and MAE (absolute mean error) describe the distance between the predicted value and the true value; the Pearson product-moment correlation coefficient and the Spearman rank correlation coefficient describe the correlation between the predicted value and the true value.

The detailed description of each indicator is as follows:

^R2 is a dimensionless score that describes the effectiveness of a model. It compares the predictions to random guessing based on the average of the true values;

RMSE is a standard value of MSE and is often used as a training loss in machine learning research;

MAE describes the absolute error between the predicted value and the true value;

MedAE is the median MAE, which describes the median value of the residue;

The Pearson correlation coefficient describes the linear correlation between two values;

The Spearman rank correlation coefficient is the rank form of the Pearson correlation coefficient, which describes the correlation between two variables (for example, when one variable increases, the other variable also increases), which is related to the monotonicity of the function;

3. Evaluation of FeatNN, BACPI, GraphDTA (GATNet, GAT_GCN, GCNNet, GINConvNet) and MONN.

FIG12 is a test on a data set after compound clustering based on the measurement standards of IC50 and KIKD. It can be seen from FIG12 that the FeatNN of the present invention is superior to other models in three indicators (ie, RMSE, Pearson and R ² ).

FIG13 is a test performed on new protein data. It can be seen from FIG13 that the FeatNN of the present invention is superior to other models in three indicators (ie, RMSE, Pearson and R ² ).

FIG14 is a comparison of the convergence of each model during the training process (model convergence speed and final convergence position); it can be seen from the figure that the convergence effect of the invented FeatNN is the best.

Figure 15 shows the advantages of deep graph convolution. The present invention also uses different depths for graph convolution. It can be seen that the performance evaluation indicators change with the deepening of the graph convolution layers.

Table 1 shows the performance evaluation table of MONN, GraphDTA (GATGCN, GCNNet, GINConvNet, GATConvNet), BACPI and FeatNN model of the present invention trained on the BindingDB (IC50 measurement standard, a total of 218615 data);

Table 1

Tables 2 and 3 show the model performance evaluation tables of the FeatNN of the present invention and the various models of the prior art after homology similarity clustering of compounds (Table 2) and proteins (Table 3) under different evaluation indicators (KIKD or IC50), which are trained on the PDBBind small data set (including KIKD and IC50 measurement standards, a total of 12628 data).

Table 2

Table 3

In order to further demonstrate that the present invention is feasible, the pseudo code for implementing the main modules of the present invention is placed below, hoping to better understand the method proposed by the present invention.

First, the main variable definitions in the code are shown in Table 4

Table 4 Common variable definitions

The symbol definition superscript is the name, and the subscript is the vector dimension information.

In addition, _Nres represents the number of residues in the input main sequence (cropped during training), _Ntors represents the torsion size, _{Nebd_s} represents the embedding size, h represents the hidden size, _Natom represents the number of atoms, _Nbond represents the number of bonds, ^Fatom represents the atomic features, ^Fbond represents the bond features, and nbs represents the neighbor dimension. Use uppercase operator names when they encapsulate learning parameters, for example, Linear is used to represent a linear transformation with a weight matrix W and a bias vector b.

LayerNorm is used to denote layer normalization that operates on the channel dimension, and the gain and bias of each channel can be learned. We also use uppercase names for random operators, such as those related to dropout. For functions without parameters, we use lowercase operator names. The definitions of operations and variables are shown below. represents multiplication, represents addition, ⊙ represents Hadamard product, a ^T b represents the dot product of two vectors, [·,·] _h refers to the dimension concatenation operation on the dimension of the hidden size. Use h to represent the hidden size, k to represent the number of attention heads, k _h to represent the size of the attention head, where k _h = h/k. softmax( _xi ) corresponds to the soft max function, softmax( _xi ) = exp( _xi )/ _∑i exp( _xi )., represents the softmax activation function; σ(·) represents the sigmoid activation function σ(x) = 1/(1+e ^-x ), tanh(·) corresponds to the tanh activation function tanh(x) = (e ^x -e ^-x )/(e ^x +e ^-x ), represents the hyperbolic tangent activation function, gelu represents the Gaussian error linear unit, and f(·) corresponds to the GELU activation function where erf(·) is the Gaussian error function,

The algorithms of each module are as follows:

Algorithm 1 Graph Convolution

con _neighhor = concat _neighhor (ver _neighbor , edge _neighbor )

neighbor_label=gelu(Linear(con _neighbor ))

output=theta⊙Linear(support)+(1-theta)⊙support

Algorithm 2 Compound Extractor

Algorithm 3 Protein feature extraction ProteinExtractor

Algorithm 4 Protein Encoder

Algorithm 5 Protein Aggregating

seq_embed=Embedding(Seq _init )

torsion_embed=CNN _Nts→h1 (TorMat _init )

torsion_vector=CNN _h1→h (torsion_embed)

gate=Sigmoid(Linear(torsion_vector))

Algorithm 6 Single-layer protein encoding EvoUpdating

MixKey=GRU(PairKey1,PairKey2)

Struct_Features=DivCNN(MixKey)

SeqGate=Sigmoid(Linear(Seq2Struct))

StructGate=Sigmoid(Linear(MixKey))

Seq2Struct_Vector＝SeqGate⊙Seq2Struct+(1-SeqGate)⊙Struct_Features

Struct_Vector=GRU(Seq2Struct_Vector,Struct_Features)

Struct2Seq_Mapping=DeepSparseCNN(Struct_Vector)

Seq_Vector＝StructGate⊙Struect_Features+(1-StructGate)⊙Seq_Features

Algorithm 7 Diversity Convolutional Layer DivCNN

def DivCNN(x):

x0＝CNN _kernel＝1 (x)

#x(seq,hidden_size)→x _k-head (seq,k,head_size)

#where k is the number of head,and head_size＝hidden_size/k

x _k-head =TransposeForScores(x)

return → {x _total }

Algorithm 8 Affinity Prediction AffinityLearning

#Inputs projections

#Output projection

return→{Affinity _Prediction }

Algorithm 9 TPS function TransposeForScores

def TransposeForScores({input}):

#input dimention:(N _res /N _atom ,h)

#output dimention:(k,N _res /N _atom ,ks)

output=DimentionReshape(input)

return{output}

Finally, it should be noted that the above-described embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A method for predicting compound-protein affinity based on a three-dimensional structure of a protein is used for predicting compound-protein affinity, extracting compound characteristics from the compound, wherein the compound characteristics comprise atomic characteristics and polymerization node characteristics, extracting protein characteristics which contain three-dimensional information of the protein from the protein, and predicting the compound-protein affinity by an affinity prediction algorithm through the compound characteristics and the protein characteristics, and specifically comprises the following steps:

S1, compound composite characteristic extraction step,

according to the compound characterization graph, atomic initial characteristics and aggregation node characteristics are obtained, the atomic characteristics and the aggregation node characteristics are circularly updated by utilizing a deep graph convolution network and a multi-head attention mechanism, and the final atomic characteristics and the aggregation node characteristics of the compound are output;

s2, extracting protein characteristics;

acquiring protein sequence information and protein structural characteristics; according to the protein sequence information and the protein structural characteristics, the protein embedding sequence characteristics have protein structural characteristics through a protein characteristic polymerization algorithm, so that the protein embedding sequence characteristics and the embedding structural characteristics are obtained; circularly updating the embedded sequence features and the embedded structure features by adopting a co-evolution strategy to finally obtain updated protein sequence features and protein structure features;

s3, predicting affinity of the compound and the protein;

according to the atomic characteristics and the aggregation node characteristics of the compound obtained in the step S1 and the protein sequence characteristics in the step S2, a predicted affinity value is obtained through an affinity learning unit algorithm, and the larger the affinity value is, the larger the probability that the compound and the protein are combined is.

2. The method for compound-protein affinity prediction based on protein three-dimensional structure according to claim 1, wherein: the step S1 specifically comprises the following steps:

s11, obtaining atomic characteristics and polymerization node characteristics according to a compound characterization graph;

s12, inputting the atomic characteristics into a deep graph convolution network, adding the output of the deep graph convolution network and the aggregation node characteristics according to a first weight of the compound characteristics, and combining the atomic characteristics with a first gating cyclic function (GRU) to obtain updated atomic characteristics;

s13, inputting the atomic characteristics into a multi-head attention characterization algorithm, adding the output of the multi-head attention characterization algorithm and the aggregation node characteristics according to a second weight of the compound characteristics, and combining the updated aggregation node characteristics through a second gating circulation function (GRU) and the aggregation node characteristics;

s14, performing S14; and (3) repeating the steps S12-S13K times by taking the updated atomic characteristics and the updated aggregation node characteristics as the atomic characteristics and the aggregation node characteristics, and outputting the atomic characteristics and the aggregation node characteristics.

3. The method for compound-protein affinity prediction based on protein three-dimensional structure according to claim 2, wherein: the deep graph convolutional network in the step S1 is provided with a residual error connection loop for aggregating neighbor information of atoms onto atoms and avoiding the problem of excessive smoothness caused when the number of network layers is deepened, the front and rear interlayer information of the GRU aggregation network is used, and a multi-head attention mechanism characterization algorithm is used for acquiring diversified compound characteristics, so that accuracy of affinity prediction in the affinity prediction step is improved.

4. The method for compound-protein affinity prediction based on protein three-dimensional structure according to claim 1, wherein: the step S2 specifically comprises the following steps:

s21, acquiring protein sequence information and protein structural characteristics, wherein the protein structural characteristics at least comprise a discretization distance matrix;

s22, a protein characteristic aggregation algorithm specifically comprises the following steps: the protein sequence information obtains a sequence vector through a first embedding layer (Eemmbedding layer), and the discretization distance matrix obtains a discrete distance matrix feature vector through a second embedding layer (Eemmbedding layer); the discrete distance matrix feature vector is used as an embedded distance matrix after feature embedding update;

s23, realizing a co-evolution strategy by adopting N layers of protein coding algorithms, wherein the protein coding algorithms of each layer are the same, and one layer of protein coding algorithm specifically comprises the following steps:

the two results of the embedded distance matrix after row addition and column addition are subjected to feature fusion through a gating circulation unit to obtain spliced embedded structure information, and the spliced embedded structure information enters a diversity convolution layer to obtain diversified protein structural features;

the embedded sequence features sequentially pass through a diversity convolution layer and a common convolution layer to obtain diversified protein sequence features; adding the diversified protein structural features and the diversified protein sequence features through a gate control logic, and sequentially passing through a gate control circulation unit and a convolution layer to obtain structural feature vectors with structural information and sequence information at the same time;

The structural feature vector outputs the structural feature of the protein through self-combination transformation (outer sum);

the sequence features after embedding and updating are subjected to a diversity convolution layer to obtain diversified sequence features after embedding and updating; adding the diversified embedded updated sequence features and the diversified protein structural features through a gate control logic to obtain sequence feature vectors fusing the structural information and the sequence information; further fusing the diversified protein sequence characteristics with sequence characteristic vectors through a gating circulating unit to output protein sequence characteristics;

s24, the new embedding distance matrix output by the protein coding algorithm is the required embedding distance matrix, the obtained embedding distance matrix and the embedding sequence feature are used as the input of the next protein coding algorithm, the embedding distance matrix and the embedding sequence feature are continuously updated until the N protein coding algorithms are all completed, at the moment, the output embedding distance matrix is the structural feature of the protein, and the embedding sequence feature is the sequence feature of the protein.

5. The method for compound-protein affinity prediction based on protein three-dimensional structure according to claim 4, wherein:

The protein structural features in the step S21 also comprise a feature auxiliary matrix;

the embedded sequence features obtained in the step S22 are added with the feature auxiliary matrix through the gate logic to obtain embedded updated sequence features;

the diversity convolution layer in step S23 equally divides the input feature vector into four parts in the feature dimension, and respectively enters four parallel common convolution layers, sums the output results, and simultaneously, aggregates the initial features of the sequence information input through residual connection to obtain the diversity feature vector.

6. The method for compound-protein affinity prediction based on protein three-dimensional structure according to claim 5, wherein: the characteristic auxiliary matrix is a torsion angle matrix, and the torsion angle matrix is constructed by sine values and cosine values of dihedral angles phi and phi between alpha carbon atoms of a protein skeleton chain and amino groups and carboxyl groups.

7. The method for compound-protein affinity prediction based on protein three-dimensional structure according to claim 4, wherein:

the discretization distance matrix in the step S21 is to divide the distance between residues of the protein stabilizing structure into M equidistant mapping intervals which basically conform to normal distribution in statistics, so as to realize the discretization coding of the distance matrix, wherein M is a positive integer.

8. The method for compound-protein affinity prediction based on protein three-dimensional structure according to claim 1, wherein: the step S3 specifically comprises the following steps:

s31, receiving atomic characteristics and aggregation node characteristics of the compound from the step S1, and transmitting the atomic characteristics and the aggregation node characteristics of the compound as compound information into an affinity learning unit, wherein the aggregation node characteristics are updated through linear layer updating characteristics to obtain updated aggregation node characteristics; the compound features are updated through a linear layer, an average value is obtained in an atomic dimension, and feature stitching is carried out on the compound features and the updated aggregate node features, so that compound comprehensive features are obtained;

s32, receiving the sequence characteristics of the protein from the step S2, transmitting the sequence characteristics of the protein as protein information into an affinity learning unit, updating the characteristics of the protein sequence characteristics through a convolution layer, and meanwhile, obtaining the average value in the residue dimension to obtain the comprehensive characteristics of the protein;

protein complex features and compound complex features are combined by matrix multiplication (matmul) and linear layer fusion to obtain predicted affinity values.

9. A system for compound-protein affinity prediction based on a three-dimensional structure of a protein, comprising: a compound extractor, a protein extractor, and an affinity predictor;

The compound extractor updates the atomic characteristics and the aggregation node characteristics by using a deep graph convolution network and a multi-head attention algorithm according to the input atomic characteristics and the aggregation node characteristics of the compound, and circularly outputs the final atomic characteristics and the aggregation node characteristics; transmitting the final compound signature and the aggregate node signature to an affinity predictor;

a protein extractor, which obtains embedded sequence characteristics and embedded structure characteristics through a protein characteristic aggregation algorithm according to the input protein sequence characteristics and protein structure characteristics; updating the embedded sequence features and the embedded structure features by adopting a co-evolution strategy to obtain updated protein sequence features and protein structure features;

and the affinity predictor is used for receiving the compound characteristic, the protein sequence characteristic and the aggregation node characteristic, and obtaining a predicted affinity value through an affinity learning unit algorithm.

10. A system for compound-protein affinity prediction based on protein three-dimensional structure according to claim 9, characterized by comprising:

the compound extractor comprises a deep map convolution unit and a multi-head attention characterization unit, wherein the deep map convolution unit is used for realizing a deep map convolution network, and the multi-head attention characterization unit is used for realizing a multi-head attention algorithm;

The protein extractor comprises a protein information aggregation unit and a co-evolution updating unit, wherein the protein information aggregation unit is used for realizing a protein characteristic aggregation algorithm, and the co-evolution updating unit is used for realizing a co-evolution strategy;

the affinity predictor includes an affinity learning unit.