A method for predicting drug-target affinity based on multi-source protein feature fusion

Abstract

The present invention discloses a method for predicting drug-target affinity by fusing multi-source features of proteins. First, a PPI network and an SSN network are constructed, and protein features are extracted from the network. Then, protein features such as subcellular location, sequence coding, functional sites, and domains are collected for protein characterization. Then, multi-source features are fused using a variational graph autoencoder, and finally, the drug branch is combined and input into a fully connected layer for affinity prediction. The present invention constructs a PPI network and an SSN network, so that the present invention not only pays attention to the characteristics of the target protein, but also learns the biological prior knowledge between the target protein and other proteins. The present invention extracts and fuses protein features from the aspects of protein interaction, sequence similarity, and protein subcellular location for the first time to predict drug-target affinity, thereby improving the prediction accuracy. In addition, the source of protein features in the present invention does not include protein structure, thereby abandoning the dependence on protein structure.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method for predicting drug-target affinity by fusion of multi-source features of proteins. Background Art

[0002] After entering the human body, drug molecules bind to target proteins, thereby producing drug effects and achieving the purpose of treating diseases. In the process of drug development, researchers will obtain a large number of candidate molecules, and virtual screening of these molecules is very necessary. Predicting drug-target binding affinity is an important step, which indicates the binding strength between drug molecules and target proteins. By predicting drug-target binding affinity, molecules that do not meet the affinity requirements can be screened out, greatly reducing the search space for potential drug molecules, eliminating the need for medical researchers to conduct extensive clinical trials on all potential drug molecules, and accelerating the drug development process.

[0003] The traditional computational affinity prediction method is molecular docking, including rigid docking and flexible docking, which expresses the interaction between drug-target pairs through scoring functions. However, these docking methods not only require molecular structural information and a large amount of computing resources, but also require some biologically relevant pretreatment of the molecules before docking, such as hydrogenation and protonation, which may not be possible sometimes.

[0004] With the widespread application of data-driven deep learning technology in the field of image processing, the field of bioinformatics represented by drug-target binding affinity prediction tasks has also been inspired by it. The explosive growth of relevant data in the 21st century has also laid the foundation for the application of data-driven deep learning technology in affinity prediction tasks.

[0005] Existing affinity prediction models based on deep learning can be divided into two categories, namely affinity prediction models based on protein one-dimensional sequences and affinity prediction models based on protein two-dimensional structures. The drug-target affinity prediction model based on protein one-dimensional sequences represents drug molecules and target proteins as SMILES strings and protein sequences, respectively, and applies multiple one-dimensional convolutional layers to drug branches and target protein branches, respectively, thereby automatically capturing high-order features from drug-target pairs. The two types of features obtained from the last convolutional layer are concatenated as the input of the fully connected layer, which then outputs a specific value, namely the predicted affinity. The drug-target affinity prediction model based on protein two-dimensional structures represents drug molecules and target proteins as molecular graphs (atoms as nodes, chemical bonds as edges) and protein graphs (amino acids as nodes, and the interactions between amino acids as edges), respectively, thereby retaining the spatial structural information of drug molecules and target proteins. Based on this, multiple graph neural network layers are used to extract high-order features of drug-target pairs to predict affinity.

[0006] Existing affinity prediction models based on one-dimensional protein sequences and two-dimensional spatial structures only extract features from the sequence or structure of a single target protein, while ignoring the rich biological prior knowledge between proteins, such as protein interactions and homology information, and also ignoring the expression of protein characteristics such as subcellular location and family. The accuracy of affinity prediction needs to be further improved; in addition, affinity prediction models based on two-dimensional structures still need to rely on protein structure, and there are still a large number of protein structures that have not been resolved. Therefore, the present invention collects a large amount of information such as interactions and sequence similarities between proteins, and performs multi-source feature fusion, so that the target protein characterization can provide more biological information for affinity prediction, thereby achieving the goal of improving the accuracy of affinity prediction while eliminating the dependence on protein structure. Summary of the invention

[0007] It takes an average of 10 to 15 years to develop a new drug that can be marketed, and the economic cost exceeds 2 billion US dollars. Such huge time and economic costs have prompted researchers to explore an efficient method to speed up the process of new drug discovery. On this basis, due to the explosive growth of data related to drugs, receptors, etc. in the 21st century, data-driven deep learning technology has been widely used, providing a strong impetus for the discovery of new drugs.

[0008] A key task motivated by deep learning technology is predicting drug-target affinity, which refers to quickly predicting the affinity between drug molecules and target proteins to determine whether the drug and the target can bind, thereby reducing the search space for potential drug molecules, eliminating the need for medical researchers to conduct extensive clinical trials on all potential drug molecules, and accelerating the drug development process. In view of the shortcomings of previous affinity prediction models, the present invention provides a method for drug-target affinity prediction by fusion of protein multi-source features.

[0009] The present invention is achieved through the following technical solution: A method for predicting drug-target affinity by fusion of multi-source features of proteins, comprising the following steps:

[0010] (1) Construction of protein-protein interaction network (PPI)

[0011] First, all human proteins were downloaded from the protein database SwissProt. Then, human proteins that were not recorded in the protein interaction database STRING were filtered out. A PPI network containing several proteins was constructed based on the interaction records in the STRING database, with each node representing a protein.

[0012] (2) Construction of sequence similarity network (SNN)

[0013] Using the proteins screened in step (1) as the search library, search for similar sequences of each protein to construct SSN. Each node in SSN represents a protein, and the edge between two nodes represents the similarity score between two proteins.

[0014] (3) Determine protein characteristics

[0015] The two types of connection relationships between proteins are constructed by two networks, namely the PPI network and the SSN network, which are represented as two adjacency matrices. The protein features are determined to represent the feature matrix. The protein features include sequence coding, subcellular location, and InterPro.

[0016] (4) Using variational graph autoencoders to fuse multi-source features

[0017] First, an undirected unweighted graph G = (V, E) is given to represent the PPI network or SSN, where V and E represent the set of nodes (proteins) in the graph G and the set of edges between nodes (protein interactions or sequence similarities). N = |V| represents the number of protein nodes, and then an adjacency matrix of size N×N can be introduced according to the graph G, denoted as A, where A represents whether there is interaction between proteins or whether there is sequence similarity between proteins. According to the three protein features in step (3), a feature matrix of size N×D can be generated, denoted as X, where D = 2045. Then, the encoder composed of two layers of graph convolutional layers (i.e., the formula below, GCN (X, A)) and the inner product decoder are jointly trained to compress the topological information between proteins and the biological characteristics of proteins into the potential representation:

[0018]

[0019]

[0020] in, is the mean vector μ i Matrix, logσ = GCN σ (X, A), and Z is obtained through the above calculation and used as the input of the decoder. The two-layer GCN is defined as:

[0021]

[0022] in, is the normalized adjacency matrix, W is a trainable parameter;

[0023] Through the two-layer GCN, the latent variable Z is calculated from the graph G representing the PPI network or SSN, and the decoder and encoder are jointly trained to reconstruct the adjacency matrix A:

[0024]

[0025]

[0026] Among them, A ij is the element of A, σ(·) is the Sigmoid function, and the loss function for training the variational graph autoencoder is as follows:

[0027] L=E q(Z∣X,A) [logp(A∣Z)]-KL[q(Z∣X,A)||p(Z)]

[0028] where KL[q(·)||p(·)] is the Kullback-Leibler divergence between q(·) and p(·), which is used to reduce the difference between the Gaussian distribution and the latent variable Z distribution. q(Z∣X,A)[logp(A|Z)] represents a cross entropy function, which is used to reduce the reconstruction error between adjacency matrices;

[0029] The D-dimensional feature matrix X is respectively connected to the adjacency matrix A of the PPI network PPI and the adjacency matrix A of SSN SSN Combined, we get two potential representations μ PPI and μ SSN :

[0030] μ PPI =GCN μ (X,A PPI )

[0031] μ SSN =GCN μ (X,A SSN )

[0032] Then the final latent representation is defined as:

[0033] μ=[μ PPI , μ S5N ]

[0034] (5) Affinity prediction based on fusion features

[0035] For the protein branch, since the superposition of fully connected layers can fit arbitrary functions, multiple fully connected layers are applied to the potential representation μ of the target protein. i , perform feature extraction;

[0036] For the drug branch, we mainly rely on the research work of GraphDTA [GraphDTA: predicting drug-target binding affinity with graph neural networks.] to first convert the SMILES string of the drug molecule into a molecular graph with atoms as nodes and bonds as edges. Based on the properties of atoms, such as atomic symbols, number of adjacent atoms, number of adjacent hydrogen atoms, atomic implicit values, and whether it is an aromatic structure, we calculate the feature matrix of the drug molecule. Then, we use three layers of graph convolutional layers and two layers of fully connected layers to extract the features of the drug molecule graph.

[0037] Finally, the feature vectors of the protein branch and the drug branch are combined into a new feature vector and input into two fully connected layers (the fully connected layer is a common linear transformation layer in deep learning, see [An Equivalence of Fully Connected Layer and Convolutional Layer]) for affinity prediction. The evaluation indicators for measuring the prediction accuracy are as follows:

[0038]

[0039] Where n represents the number of samples, p is the predicted affinity value, and y is the true affinity value.

[0040] Furthermore, in step (1), if the comprehensive score of the interaction between a pair of nodes is greater than a threshold ε, it is considered that there is a connected edge between the two nodes, and the threshold ε is 0.4.

[0041] Furthermore, in step (2), if the similarity score between the two nodes is less than the threshold It is considered that there is an edge between the two nodes, and the threshold Take 10 -4 .

[0042] Furthermore, in the step (3), when processing the sequence code, the protein sequence is encoded using the joint triplet method, and the 20 amino acids are divided into 7 categories according to the dipole and volume of the amino acid side chain, selenocysteine ​​(U) and cysteine ​​(C) are classified into one category, and then any three consecutive amino acids in the sequence are taken as a unit to form 7^3=343 possible unit types, and a vector of size (1×343) is constructed to represent the protein sequence. The value of the vector is calculated by calculating the frequency of the units present in the sequence, and all proteins are represented as 343-dimensional sequence codes.

[0043] Furthermore, in step (3), when processing subcellular locations, information about subcellular locations was collected from the SwissProt database, resulting in a total of 260 different subcellular locations. One-hot encoding was used to construct a 260-dimensional vector, which indicates whether a protein exists in these subcellular locations.

[0044] Furthermore, in step (3), based on the terms of protein family, domain and other information recorded in the InterPro database, only terms that appear more than 20 times are retained, and then each protein is encoded using a one-hot encoding method to construct a protein feature for InterPro. In this way, a protein can be represented as a feature vector of several dimensions.

[0045] Furthermore, the protein dimension is the sum of the protein sequence encoding dimension, the subcellular location encoding dimension, and the InterPro processing dimension.

[0046] Furthermore, in step (4), and GCN σ (X, A) shares the first layer parameters W 0 .

[0047] The technical problems solved by the present invention are as follows:

[0048] The present invention extracts and fuses protein features from aspects such as protein interactions, sequence similarity, and protein subcellular location for the first time to predict drug-target affinity, thereby improving the prediction accuracy. In addition, the source of protein features in the present invention does not include protein structure, thereby eliminating the dependence on protein structure. Specifically:

[0049] (1) The present invention constructs a protein interaction network and a protein sequence similarity network from a large number of proteins including the target protein, so that the present invention not only focuses on the characteristics of the target protein, but also learns the biological prior knowledge between the target protein and other proteins.

[0050] (2) The present invention collects protein features such as subcellular location, sequence coding, functional sites, and structural domains for protein characterization, taking into full account the impact of these multi-source features on drug-target affinity.

[0051] (3) The present invention uses a variational graph autoencoder to fuse protein interaction networks, sequence similarity networks, and protein information such as subcellular location, sequence encoding, functional sites, and structural domains, so that the target protein characterization after fusing multi-source features covers richer biological information to improve the accuracy of affinity prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The present invention will be further described below in conjunction with the accompanying drawings.

[0053] Figure 1 This is the modeling flow chart of the present invention (including multi-source feature fusion).

[0054] Figure 1 In the left branch, PPI and SSN represent the protein interaction network and sequence similarity network constructed by the present invention, respectively, and also describe the sequence encoding, subcellular location and InterPro as protein node features, which are 343, 260 and 1442 dimensions respectively. Figure 2 As shown in Figure 2, two biological networks (i.e., PPI and SSN) are combined with protein node features to generate a protein representation μ PPI and μ SSN and compare it with the drug branch ( Figure 1 The high-order features of drugs extracted from the right branch are combined to predict affinity.

[0055] Figure 2 The model architecture of the Variational Graph Autoencoder (VGAE).

[0056] Figure 2The model architecture of VGAE is shown in Figure 1, which consists of two parts: an encoder consisting of two graph convolutional layers and an inner product decoder. Both μ and σ are generated by the encoder and used as mean and variance for the calculation of Z, which is used as the input of the decoder. DETAILED DESCRIPTION

[0057] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present invention.

[0058] In order to maximize the use of biological prior knowledge, the present invention collects 18,552 human proteins from the SwissProt database, which cover the limited number of target proteins required by the present invention. Then, a protein-protein interaction (PPI) network and a sequence similarity (SSN) network with sequence coding, subcellular location, domain, family, and functional site as node (protein) features are constructed for these 18,552 human proteins, and a variational graph auto-encoder (VGAE) framework is used to fuse multi-source features of proteins to obtain the potential representation of 18,552 proteins. Since the target protein is included in the 18,552 proteins, the potential representation of the target protein is extracted for affinity prediction.

[0059] The specific process is as follows Figure 1-2 A method for predicting drug-target affinity by fusion of multi-source features of proteins, comprising the following steps:

[0060] (1) Construct a protein-protein interaction network (PPI) and extract protein features from the network for affinity prediction tasks

[0061] The present invention first downloaded 20,376 human proteins from the protein database SwissProt. Then, human proteins that were not recorded in the protein interaction relationship database STRING were filtered out, thereby obtaining 18,552 proteins. At the same time, a PPI network containing 18,552 proteins was constructed based on the protein interaction relationship in the STRING database, with each node representing a protein. In addition, if the comprehensive score of the interaction between a pair of nodes is greater than the threshold ε, it is considered that there is a connected edge between the two nodes (the larger the comprehensive score, the stronger the credibility of the interaction between proteins). The selection of the threshold ε was experimentally measured and taken as 0.4.

[0062] (2) Constructing a sequence similarity network (SNN) and extracting protein features from the network for affinity prediction tasks

[0063] Using these 18,552 proteins as a search library, the SSN was constructed by searching for similar sequences of each of these 18,552 proteins using BLAST (a tool for finding local similarity regions between sequences). Like the PPI network, each node in the SSN represents a protein, but the edge between two nodes represents the similarity score between the two proteins. The smaller the similarity score, the more similar the two sequences are. Therefore, if the similarity score between two nodes (proteins) is less than a threshold It is considered that there is an edge between the two nodes. The selection of is determined by experimental measurement, and 10 -4 .

[0064] (3) Determine protein characteristics

[0065] Specifically, triplet sequence codes, subcellular locations, and InterPro terms were used as feature vectors of proteins for affinity prediction tasks.

[0066] 3.1 Sequence Coding

[0067] The protein sequence composed of a string of amino acid characters contains rich biochemical information. In order to compress sequences of different lengths into sequences of the same and fixed size, the present invention adopts the combined triplet (CT) method to encode the protein sequence. The present invention divides the 20 amino acids into 7 categories according to the dipole and volume of the amino acid side chain. It is worth noting that there is a selenocysteine ​​(U) independent of the above 20 amino acids in 18552 proteins. Since selenocysteine ​​is similar to cysteine ​​(C) in structure, except that the sulfur atom is replaced by a selenium atom, the present invention classifies U and C into one category. The classification results of the 21 amino acids are shown in the following table.

[0068] Classification Amino Acids CT1 A,G,V CT2 I,L,F,P CT3 Y,M,T,S CT4 H,N,Q,W CT5 R,K CT6 D,E CT7 C,U

[0069] Then each amino acid in the sequence can be encoded as a number from 1 to 7, i.e., CT1 to CT7. Any three consecutive amino acids in the sequence are then taken as a unit, forming 7^3=343 possible unit types. Therefore, the present invention constructs a vector of size (1×343) to represent the protein sequence, and calculates the value of the vector by calculating the frequency of the units present in the sequence. In this way, all 18,552 proteins are represented as 343-dimensional sequence codes.

[0070] 2) Subcellular location

[0071] Subcellular location reveals the specific location of a protein in the cell, such as plasma membrane, nucleus, nucleoplasm, etc. Proteins in different locations have different functions, and the combination of one or more functions may affect affinity. Therefore, subcellular location should also be used as a feature source for affinity prediction.

[0072] For these 18552 proteins, the present invention collected information about subcellular locations from the SwissProt database, resulting in a total of 260 different subcellular locations. Therefore, one-hot encoding is used to construct a 260-dimensional vector, which indicates whether a protein exists in these subcellular locations.

[0073] 3) InterPro terminology processing

[0074] InterPro (http: / / www.ebi.ac.uk / interpro / ) integrates multiple member databases such as CATH, Pfam, Panther, etc., and provides information on protein domains, families, and functional sites. In addition, all of this related information is expressed in different terms. On this basis, a total of 59,836 unique terms were collected for 18,552 proteins. In order to reduce the dimension, the present invention only retains terms that appear more than 20 times, resulting in a total of 1,442 unique terms. Similar to the subcellular location encoding, a one-hot encoding method is used to construct a 1,442-dimensional vector. In this way, a protein can be represented as a 2,045-dimensional feature vector (i.e., the sum of 343, 260, and 1,442).

[0075] (4) Using variational graph autoencoders to fuse multi-source features

[0076] The variational graph autoencoder is a framework consisting of an encoder and a decoder for unsupervised learning of graph-structured data, which can generate low-dimensional latent representations for the input graph data, thereby facilitating the fusion of multi-source features of 18,552 proteins.

[0077] Specifically, first, an undirected unweighted graph G = (V, E) is given to represent the PPI network or SSN, where V and E represent the set of nodes (proteins) in the graph G and the set of edges between nodes (protein interactions or sequence similarity), respectively, and N = |V| represents the number of protein nodes. According to the graph G, an adjacency matrix of size N × N can be introduced, denoted by A, indicating whether there is interaction between proteins or whether the sequences between proteins are similar. According to the three protein features in step (3), a feature matrix of size N × D can be generated, denoted by X, where D = 2045. Then, the encoder composed of two layers of graph convolutional layers (i.e., the formula below, GCN (X, A)) and the inner product decoder are jointly trained to compress the topological information between proteins and the biological characteristics of proteins into the potential representation:

[0078]

[0079]

[0080] Where μ = GCN μ (X, A) is the mean vector μ i Matrix, logσ = GCN σ (X, A), and Z is obtained through the above calculation and used as the input of the decoder. The two-layer GCN is defined as:

[0081]

[0082] in, is the normalized adjacency matrix, W is a trainable parameter; in addition, GCN μ (X, A) and GCN σ (X, A) shares the first layer parameters W 0 ;

[0083] In this way, the present invention can calculate the latent variable Z from the graph G representing the PPI network or SSN through a two-layer GCN. In order to optimize Z, the decoder and encoder are jointly trained to reconstruct the adjacency matrix A:

[0084]

[0085]

[0086] Since the goal of the present invention is to compress only the graph G containing biological information into the latent variable Z, the decoder only performs inner product operations and does not require any trainable parameters. ij is the element of A, σ(·) is the Sigmoid function, and the loss function for training the variational graph autoencoder is as follows:

[0087] L=E q(Z∣X,A)[logp(A∣Z)]-KL[q(Z∣X,A)||p(Z)]

[0088] where KL[q(·)||p(·)] is the Kullback-Leibler divergence between q(·) and p(·), which is used to reduce the difference between the Gaussian distribution and the latent variable Z distribution. q(Z∣X,A) [log p(A|Z)] represents a cross entropy function, which is used to reduce the reconstruction error between adjacency matrices; in this way, the trainable parameters in the encoder can be trained by the gradient descent algorithm to optimize the latent variable Z, thereby compressing the biological information (interactions, domains, sequences) in G into the latent variable Z.

[0089] In order to perform feature fusion, the present invention trains encoders based on two biological networks (PPI network and SSN) respectively. Specifically, the present invention trains the 2045-dimensional feature matrix X and the adjacency matrix A of the PPI network respectively. PPI and the adjacency matrix A of SSN SSN Combined, we get two potential representations μ PPI and μ SSN :

[0090] μ PPI =GCN μ (X,d PPI )

[0091] μ SSN =GCN μ (X,d SSN )

[0092] Then the final latent representation is defined as:

[0093] μ=[μ PPI , μ SSN ]

[0094] (5) Affinity prediction based on fusion features

[0095] So far, the present invention has obtained the corresponding low-dimensional potential representation of each of the 18,552 proteins. Since a limited number of target proteins are included in these 18,552 proteins, the present invention extracts the potential representation of the target protein from the potential representation of all 18,552 proteins for affinity prediction.

[0096] For the protein branch, since the superposition of fully connected layers can fit arbitrary functions, multiple fully connected layers are applied to the potential representation μ of the target protein. i , perform feature extraction;

[0097] For the drug branch, we adopt the same operation based on the research work of GraphDTA [GraphDTA: predicting drug-target binding affinity with graph neural networks.], first converting the SMILES string of the drug molecule into a molecular graph with atoms as nodes and bonds as edges, and calculating the feature matrix of the drug molecule based on the properties of atoms, such as atomic symbols, number of adjacent atoms, number of adjacent hydrogen atoms, atomic implicit values, and whether it is an aromatic structure; then, we use three layers of graph convolutional layers and two layers of fully connected layers to extract the features of the drug molecule graph, such as Figure 1 The right branch is shown;

[0098] Finally, the feature vectors of the protein branch (i.e., step 4) and the drug branch are combined into a new feature vector and input into two fully connected layers (a common linear transformation layer in deep learning, see [An Equivalence of FullyConnected Layer and Convolutional Layer]) for affinity prediction. The evaluation indicators for measuring the prediction accuracy are as follows:

[0099]

[0100] Where n represents the number of samples, p is the predicted affinity value, and y is the true affinity value.

[0101] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for predicting drug-target affinity by fusion of multi-source protein features, characterized by: The following steps are involved: (1) Construction of protein-protein interaction network PPI First, all human proteins were downloaded from the protein database SwissProt. Then, human proteins that were not recorded in the protein interaction database STRING were filtered out. A PPI network containing several proteins was constructed based on the interaction records in the STRING database, with each node representing a protein. (2) Constructing a sequence similarity network (SNN) Using the proteins screened in step (1) as the search library, search for similar sequences of each protein to construct SSN. Each node in SSN represents a protein, and the edge between two nodes represents the similarity score between two proteins. (3) Determine protein characteristics The two types of connection relationships between proteins are constructed by two networks, namely the PPI network and the SSN network, which are represented as two adjacency matrices. The protein features are determined to represent the feature matrix. The protein features include sequence coding, subcellular location, and dimensional vectors determined by InterPro. (4) Using variational graph autoencoders to fuse multi-source features First, an undirected unweighted graph G = (V, E) is given to represent the PPI network or SSN, where V and E represent the set of nodes and the set of edges between nodes in the graph G, respectively, where the nodes are proteins and the edges between nodes are the interaction relationships or sequence similarities between proteins, and N = |V| represents the number of protein nodes. Based on the graph G, an adjacency matrix of size N × N can be introduced, denoted by A, where A represents whether there is interaction between proteins or whether the sequences between proteins are similar. Based on the three protein features in step (3), a feature matrix of size N × D can be generated, denoted by X, where D = 2045. Then, the encoder composed of two layers of graph convolutional layers and the inner product decoder are jointly trained to compress the topological information between proteins and the biological features of proteins into the potential representation: q(z i ∣X,A)=N(z i ∣μ i ,diag(σ i 2 )) Where μ = GCN μ (X,A) is the mean vector μ i Matrix, logσ = GCN σ (X, A), and Z is obtained through the above calculation and used as the input of the decoder. The two-layer GCN is defined as: in, is the normalized adjacency matrix, W is a trainable parameter; Through the two-layer GCN, the latent variable Z is calculated from the graph G representing the PPI network or SSN, and the decoder and encoder are jointly trained to reconstruct the adjacency matrix A: Among them, A ij is the element of A, σ(·) is the Sigmoid function, and the loss function for training the variational graph autoencoder is as follows: L=E q(Z∣X,A) [logp(A∣Z)]-KL[q(Z∣X,A)||p(Z)] where KL[q(·)||p(·)] is the Kullback-Leibler divergence between q(·) and p(·), which is used to reduce the difference between the Gaussian distribution and the latent variable Z distribution. q(Z∣X,A) [logp(A|Z)] represents a cross entropy function, which is used to reduce the reconstruction error between adjacency matrices; The D-dimensional feature matrix X is respectively connected to the adjacency matrix A of the PPI network PPI and the adjacency matrix A of SSN SSN Combined, we get two potential representations μ PPI and μ SSN : m PPI =GCN μ (X,A PPI ) m SSN =GCN μ (X,A SSN ) Then the final latent representation is defined as: μ=[μ PPI ,m SSN ] (5) Affinity prediction based on fusion features For the protein branch, multiple fully connected layers are applied to the latent representation of the target protein for feature extraction; For the drug branch, the SMILES string of the drug molecule is first converted into a molecular graph with atoms as nodes and bonds as edges. Based on the properties of the atoms, the feature matrix of the drug molecule is calculated. Then, the features of the drug molecule graph are extracted using three layers of graph convolutional layers and two layers of fully connected layers. Finally, the feature vectors of the protein branch and the drug branch are combined into a new feature vector and input into two fully connected layers for affinity prediction.

2. The method for predicting drug-target affinity by fusion of multi-source features of proteins according to claim 1, characterized in that: In step (1), if the comprehensive score of the interaction between a pair of nodes is greater than a threshold ε, it is considered that there is a connected edge between the two nodes, and the threshold ε is 0.

4.

3. The method for predicting drug-target affinity by fusion of multi-source features of proteins according to claim 1, characterized in that: In step (2), if the similarity score between two nodes is less than the threshold It is considered that there is an edge between the two nodes, and the threshold Take 10 -4 .

4. The method for drug-target affinity prediction of protein multi-source feature fusion according to claim 1, characterized in that: In the step (3), when processing the sequence code, the protein sequence is encoded using the joint triplet method. The 20 amino acids are divided into 7 categories according to the dipole and volume of the amino acid side chain, selenocysteine ​​U and cysteine ​​C are classified into one category, and then any three consecutive amino acids in the sequence are taken as a unit to form 7^3=343 possible unit types. A vector of size (1×343) is constructed to represent the protein sequence. The value of the vector is calculated by calculating the frequency of the units present in the sequence, and all proteins are represented as 343-dimensional sequence codes.

5. The method for predicting drug-target affinity by fusion of multi-source features of proteins according to claim 1, characterized in that: In step (3), when processing subcellular locations, information about subcellular locations was collected from the SwissProt database, resulting in a total of 260 different subcellular locations. One-hot encoding was used to construct a 260-dimensional vector, which indicates whether a protein exists in these subcellular locations.

6. The method for predicting drug-target affinity by fusion of multi-source features of proteins according to claim 1, characterized in that: In the step (3), only the terms that appear more than 20 times are retained according to the terms of protein family and domain information recorded in the InterPro database, and then each protein is encoded using a one-hot encoding method, so that a protein feature specific to InterPro can be constructed. In this way, a protein can be represented as a feature vector of several dimensions.

7. The method for predicting drug-target affinity by fusion of multi-source features of proteins according to claim 6, characterized in that: The protein dimension is the sum of the protein sequence encoding dimension, the subcellular location encoding dimension, and the InterPro processing dimension.

8. The method for drug-target affinity prediction of protein multi-source feature fusion according to claim 1, characterized in that: In step (4), GCN μ (X,A) and GCN σ (X,A) share the first layer parameters W0.

9. The method for predicting drug-target affinity by fusion of multi-source features of proteins according to claim 1, characterized in that: In step (5), the evaluation index for measuring the affinity prediction accuracy is as follows: Where n represents the number of samples, p is the predicted affinity value, and y is the true affinity value.

10. The method for drug-target affinity prediction of protein multi-source feature fusion according to claim 1, characterized in that: The properties of the atoms in step (5) include the atomic symbol, the number of adjacent atoms, the number of adjacent hydrogen atoms, the atomic implicit value, and whether it is an aromatic structure.

Images