A method for predicting drug target associations
By employing collaborative contrastive learning and adaptive self-stepping sampling strategies, this study utilizes the GCN model to extract features from the structural data of drugs and proteins, and transforms multiple interaction networks into similarity networks. This addresses the problem in existing technologies that fail to fully leverage the complementarity of multiple interaction networks, thereby improving the prediction accuracy of drug-target associations.
Patent Information
- Application Number
- CN202410914418.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-09
- Publication Date
- 2025-11-07
- Estimated Expiration
- 2044-07-09
AI Technical Summary
Existing methods for predicting drug-target associations fail to fully utilize the complementary information among multiple interaction networks, affecting the accuracy of drug-target association predictions.
We employ a collaborative contrastive learning and adaptive self-stepping sampling strategy to extract features from the structural data of drugs and proteins using a GCN model. We also convert multiple interaction networks into similarity networks and use MLP to predict potential drug-protein interactions.
It improves the prediction accuracy of drug-target associations by learning more consistent feature representations through more full utilization of complementary knowledge among multiple interactive networks.
Smart Images
Figure CN119028427B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of computational biology, and particularly relates to a method for predicting drug-target association. BACKGROUND
[0002] Existing drug-target (protein) association prediction methods include structure-based methods, network-based methods and hybrid methods. The structure-based method first collects SMILES string data of drugs, amino acid sequence data of proteins or three-dimensional structure data of proteins, and then uses deep learning methods such as CNN and Transformer to extract features from SMILES and amino acid sequence data or to obtain the linkage relationship between drug atoms and the linkage relationship between amino acid residues according to SMILES string and protein three-dimensional structure data, and then uses methods such as GCN and GAT to extract features.
[0003] The network-based method first obtains diversified interaction and association networks between biological entities, such as drug-disease, drug-side effect, protein-disease, etc. interaction networks, and then uses similarity algorithms such as cosine similarity and Gaussian interaction similarity to convert these interaction networks into similarity networks of drugs or proteins, and then uses deep learning methods such as GCN and RWR to learn more effective feature representations of drugs and proteins from the similarity networks. The hybrid method usually constructs two encoders: a structure data encoder and a network data encoder. The features learned by the two encoders are fused through an attention mechanism or the features learned by the structure encoder are input into the network data encoder to simultaneously utilize structure type data and network type data to learn more rich information of drug and target feature representations.
[0004] Although the structure-based method and the network-based method have achieved certain results, the structure-based method does not consider the influence of the association relationship between biological entities on the drug and target association relationship, and the network-based method does not utilize the structure information of drugs and targets themselves; the hybrid method uses structure data and network data, but the existing hybrid method usually simply aggregates the features learned from multiple interaction networks, and does not fully utilize the complementary information between multiple interaction networks. It is difficult to fully utilize the complementary knowledge between multiple interaction networks to learn more consistent feature representations from each interaction network, which affects the prediction accuracy of drug-target association relationship. SUMMARY
[0005] The purpose of the present application is to provide a method for predicting drug-target association relationship to solve the above problems.
[0006] To achieve the above object, the following technical scheme is adopted in the present application: A method for predicting drug target association, the method comprising the following steps:
[0007] S1, collecting and downloading drug and protein structure data and associated interaction networks;
[0008] S2, drug and protein structure data processing, converting interaction networks;
[0009] S3, extracting and learning drug structure features and protein structure features from structure data;
[0010] S4, designing a collaborative contrast learning strategy to calculate a contrast learning loss;
[0011] S5, designing an adaptive self-step sampling strategy to obtain a high-reliability contrast learning sample pair;
[0012] S6, obtaining the final feature representation of the drug and protein, and using MLP to predict the potential drug-protein interaction.
[0013] In step S1, the drug and protein structure data includes:
[0014] ① drug SMILES string data;
[0015] ② protein amino acid sequence data;
[0016] ③ protein amino acid three-dimensional structure data;
[0017] The interaction network of the drug and the protein includes:
[0018] a, drug-disease association network;
[0019] b, drug-side effect association network;
[0020] c, drug-drug interaction network;
[0021] d, drug-protein association network;
[0022] e, protein-disease interaction network;
[0023] f, protein-protein interaction network;
[0024] The drug-protein association network dataset includes four biological entities and six interactions, wherein the four biological entities include 708 drugs, 1512 proteins, 5603 diseases, and 4192 side effects; and the six interactions include 1923 drug-protein interactions, 199214 drug-disease associations, 10036 drug-drug interactions, 80164 drug-side effect associations, 1923 protein-drug interactions, 1596745 protein-disease associations, and 7363 protein-protein interactions.
[0025] Step S2 specifically includes the following steps:
[0026] S21, converting a SMILES string into an atomic interaction graph by using a dgl tool;
[0027] S22, converting a protein amino acid three-dimensional structure into an amino acid residue graph by calculating the distance between amino acid residues;
[0028] S23, converting each interaction network into a similarity network by using a Jaccard similarity algorithm;
[0029] In step S21, the dgl tool is an open source tool. In step S22, the protein amino acid three-dimensional structure includes three-dimensional coordinate information of each amino acid residue. By calculating the distance between any two amino acid residues, two amino acids with a distance less than a specific threshold are regarded as having a link relationship, and thus an amino acid residue contact graph is obtained. In step S23, a similarity network is constructed based on each interaction network by using a Jaccard similarity algorithm. The similarity algorithm formula is:
[0030]
[0031] Wherein A represents the adjacency matrix representation of any one interaction network.
[0032] A i represents the i-th row of the adjacency matrix A.
[0033] S ij represents the similarity value between nodes i and j learned from the adjacency matrix A.
[0034] Step S3 specifically includes the following steps:
[0035] S31, extracting drug structure features on the atomic interaction graph by using a GCN model;
[0036] S32, extracting protein structure features on the amino acid residue graph by using a GCN model;
[0037] In step S31 and step S32, the GCN learns the feature representation of the drug and the protein from the atomic interaction graph and the amino acid residue contact map. The formula for learning the feature representation of the drug and the protein is:
[0038]
[0039] wherein: represents the association matrix between atoms of any one drug or the association matrix between amino acid residues in the protein;
[0040] n represents the number of atoms in the drug or the number of amino acid residues in the protein;
[0041] represents the initial feature of the drug or the protein;
[0042] is a diagonal matrix, and each diagonal element has a value of the degree of the atom of the drug or the amino acid residue of the protein;
[0043] is a learnable weight parameter, and f is the feature dimension of the output atom of the drug or the amino acid residue of the protein;
[0044] is a nonlinear activation function;
[0045] After updating the feature representation of each atom, a readout function is used to combine the features of all atoms to obtain the feature representation of the drug, and the formula of the readout function is:
[0046] h = Readout (X’);
[0047] The feature representation of the drug d is fused by the formula of the readout function. i The feature representation of the structure of the drug is obtained by fusing all the atom information. The feature representation of the protein p is fused. i The feature representation of the structure of the protein is obtained by fusing all the amino acid residue information.
[0048] Step S4 specifically comprises the following steps:
[0049] S41, fusing the feature representation learned by each similarity network;
[0050] S42, obtaining a comprehensive representation of the drug and the protein;
[0051] S43, comparing the feature learned by each network with the comprehensive feature;
[0052] S44, calculating a contrastive learning loss;
[0053] In step S41, taking the drug as an example, for each similarity network, denoted as S m(m = 1, 2, 3, 4), first calculate the weight of the neighbor under the similarity network, and then aggregate the features of the neighbor nodes according to the weight, that is:
[0054]
[0055] wherein, represents the feature representation of the output of the lth hidden layer, represents the weight matrix of the lth layer;
[0056]
[0057] wherein, is the attention score calculated based on the lth layer features of the drug d i and d j , N(i) represents the neighbor set of the drug d i , is the normalized attention score between the drug d i and its neighbor d j ;
[0058]
[0059] wherein, and are the lth layer learnable parameters, is the lth layer learnable weight vector, is the similarity value between the drug d i and the drug d j in the mth similarity network;
[0060] Then the output of the last layer is taken as the final feature representation learned under the similarity network, denoted as L is the number of layers, S m indicates that the feature is learned from the mth similarity network, and the same strategy is adopted to obtain the protein feature from each protein similarity network
[0061] In step S42, taking the drug as an example, first, the features learned by each similarity network are fused to obtain a comprehensive feature representation, and the specific operation is as follows:
[0062]
[0063] wherein: ∏ represents the splicing operation;
[0064] Conv is a one-dimensional convolution operation;
[0065] is the comprehensive representation of the drug obtained by fusing all similarity networks;
[0066] The comprehensive feature of the protein is obtained using the same method
[0067] In step S43, the feature representation learned by each network is compared with the comprehensive feature, and the distance between the feature learned by each network and the comprehensive feature is narrowed down using contrastive learning. For the feature representation learned in any network The contrastive learning loss of the comprehensive feature is specifically defined as follows:
[0068]
[0069] Wherein, n d is the number of drugs;
[0070] and are the positive sample set and the negative sample set of the mth drug similarity network, respectively, for drug d i ;
[0071] is the feature of drug d i learned from the mth drug similarity network;
[0072] is the fusion feature of drug d i ;
[0073] is the cosine similarity algorithm;
[0074] τ is the temperature parameter;
[0075] The contrastive loss between the feature learned by each protein similarity network and the fusion feature is obtained using the same method Finally, in step S44, the contrastive loss of each similarity network of the drug and the protein and the fusion feature is added to obtain the overall contrastive loss L c .
[0076] Step S5 is an adaptive self-step sampling strategy designed by cooperative contrastive learning, which calculates the feature similarity between the comprehensive features of drugs or proteins, and selects adaptive contrastive learning negative sample pairs for each network according to the similarity values between drugs or proteins in each network and the comprehensive feature similarity; taking the protein as an example, for any protein p i , the contrastive learning positive sample pair set in the contrastive process of the representation learned from the similarity network is The negative sample pair is obtained through the adaptive self-step sampling strategy.
[0077] Step S5 specifically includes the following steps:
[0078] S51, define two scoring functions to measure the reliability of negative samples from different angles respectively:
[0079]
[0080] wherein, represents the negative sample reliability of the mth similarity network; represents the negative sample reliability calculated according to the comprehensive representation;
[0081] sim(p i ,p j ) represents the cosine similarity algorithm;
[0082] S52, define a self-step function to determine the number of negative samples selected in each iteration of the learning process, and the self-step function is:
[0083]
[0084] wherein, T is the maximum training round;
[0085] t is the current training round;
[0086] β is a hyperparameter that controls the ratio of the number of negative samples to be selected to the number of candidate negative samples;
[0087] n p is the number of proteins;
[0088] represents the floor operation;
[0089] num t represents the number of negative samples that should be selected in the current training round;
[0090] S53, arrange the negative samples of each similarity network in descending order according to the reliability , and select the first num t negative samples for the current training round, denoted as N ns , at the same time, arrange all negative samples in descending order according to the reliability calculated by the fusion feature, and select the first num t negative samples as the negative samples of the current round, denoted as N fs , and the union of the two sets is taken as the final negative sample set N m =N ns ∪N fs .
[0091] In step S6, the joint representation of the drug-protein pair is processed and extracted by using the CNN:
[0092]
[0093] wherein, and are the fusion features of the drug d i and the protein p j ;
[0094] is the joint representation of the drug d i and the protein p j ;
[0095] Predict the probability of drug-protein interaction through the fully connected layer:
[0096]
[0097] wherein, is the weight matrix;
[0098] Use Bceloss as the objective function of drug-protein association prediction:
[0099]
[0100] wherein, is the predicted probability of the association between the drug d i and the protein p j ;
[0101] is the true label of whether the drug d i and the protein p j have an association;
[0102] dpp is the training set of the drug protein pair.
[0103] Compared with the prior art, the advantages of the present application are that: through the cooperative contrast learning and adaptive self-step sampling strategy, more consistent feature representations of drugs and proteins are learned on each similarity network, and the potential drug-target association is predicted through the MLP; the present application can more fully utilize the complementary knowledge between multiple interaction networks by constructing the cooperative contrast learning and adaptive self-step sampling strategy, learn more consistent feature representations from each interaction network, and thus improve the prediction accuracy of the drug-target association relationship, and the use effect is good. BRIEF DESCRIPTION OF DRAWINGS
[0104] Figure 1 is the prediction flowchart of the present application;
[0105] Figure 2 is the advantage diagram of the cooperative contrast learning in the present application;
[0106] Figure 3is the effectiveness column chart of the synergistic contrastive learning and self-step sampling strategy implemented by the present application under different feature fusion methods;
[0107] Figure 4 is the visual analysis chart of the drug-target pair feature representation learned by the present application in different training stages. DETAILED DESCRIPTION
[0108] The present application will be further described in detail below in combination with the drawings and specific embodiments.
[0109] As Figures 1-4 shown, a method for predicting drug-target correlation, the method comprises the following steps:
[0110] S1, collecting and downloading drug and protein structure data and associated interaction networks;
[0111] S2, drug and protein structure data processing, converting interaction networks;
[0112] S3, extracting and learning drug structure features and protein structure features from structure data;
[0113] S4, designing a synergistic contrastive learning strategy, calculating a contrastive learning loss;
[0114] S5, designing an adaptive self-step sampling strategy, obtaining a high-reliability contrastive learning sample pair;
[0115] S6, obtaining the final feature representation of the drug and protein, and using MLP to predict the potential drug-protein interaction.
[0116] In step S1, the drug and protein structure data includes:
[0117] ① drug SMILES string data;
[0118] ② protein amino acid sequence data;
[0119] ③ protein amino acid three-dimensional structure data;
[0120] The interaction network of the drug and protein includes:
[0121] a, drug-disease correlation network;
[0122] b, drug-side effect correlation network;
[0123] c, drug-drug interaction network;
[0124] d, drug-protein correlation network;
[0125] e, protein-disease interaction network;
[0126] f. Protein-protein interaction network;
[0127] The drug-protein association network dataset includes four biological entities and six interactions, wherein the four biological entities include 708 drugs, 1512 proteins, 5603 diseases, and 4192 side effects; and the six interactions include 1923 drug-protein interactions, 199214 drug-disease associations, 10036 drug-drug interactions, 80164 drug-side effect associations, 1923 protein-drug interactions, 1596745 protein-disease associations, and 7363 protein-protein interactions.
[0128] Step S2 specifically includes the following steps:
[0129] S21. Convert the SMILES string to an atomic interaction graph by using the dgl tool;
[0130] S22. Convert the protein amino acid three-dimensional structure to an amino acid residue graph by calculating the distance between amino acid residues;
[0131] S23. Convert each interaction network to a similarity network by using the Jaccard similarity algorithm;
[0132] In step S21, the dgl tool is an open source tool. In step S22, the protein amino acid three-dimensional structure includes three-dimensional coordinate information of each amino acid residue. By calculating the distance between any two amino acid residues, two amino acids with a distance less than a specific threshold are considered to have a link relationship, and thus an amino acid residue contact graph is obtained. In step S23, a similarity network is constructed based on each interaction network by using the Jaccard similarity algorithm. The similarity algorithm formula is:
[0133]
[0134] Wherein A represents the adjacency matrix representation of any one interaction network.
[0135] A i represents the i-th row of the adjacency matrix A.
[0136] S ij represents the similarity value between nodes i and j learned from the adjacency matrix A.
[0137] Step S3 specifically includes the following steps:
[0138] S31. Extract drug structure features on the atomic interaction graph by using the GCN model;
[0139] S32, extracting protein structure features on the amino acid residue graph using the GCN model;
[0140] As shown in the drawings, Figure A shows a feature extraction method based on traditional multi-network learning, Figure B shows a feature extraction method based on traditional contrastive learning, and Figure C shows a collaborative contrastive learning feature extraction method of the present application. Figure 2
[0141] In step S31, the formula for learning the feature representation of the drug and the protein from the atomic interaction graph and the amino acid residue contact graph using the GCN is:
[0142]
[0143] wherein: represents the association matrix between any one drug atom or the association matrix between amino acid residues in the protein;
[0144] n represents the number of atoms in the drug or the number of amino acid residues in the protein;
[0145] represents the initial feature of the drug or the protein;
[0146] is a diagonal matrix, and each diagonal element has a value of the degree of the drug atom or the amino acid residue of the protein;
[0147] is a learnable weight parameter, and f is the feature dimension of the output drug atom or protein amino acid residue;
[0148] is a nonlinear activation function;
[0149] After updating the feature representation of each atom, a readout function is used to combine the features of all atoms to obtain the feature representation of the drug, and the formula of the readout function is:
[0150] h = Readout (X');
[0151] The feature representation of the drug d i is obtained by fusing all atomic information to obtain the drug structure feature representation The feature representation of the protein p i is obtained by fusing all amino acid residue information to obtain the protein structure feature representation
[0152] Step S4 specifically comprises the following steps:
[0153] S41, fusing the feature representation learned by each similarity network;
[0154] S42, obtaining a comprehensive representation of the drug and the protein;
[0155] S43, comparing the feature learned by each network with the comprehensive feature;
[0156] S44, calculating a comparison learning loss;
[0157] In step S41, taking the drug as an example, for each similarity network, denoted as S m (m = 1, 2, 3, 4), the weight of the neighbor under the similarity network is first calculated, and then the features of the neighbor nodes are aggregated according to the weight, that is:
[0158]
[0159] wherein, represents the feature representation of the output of the lth hidden layer, represents the weight matrix of the lth layer;
[0160]
[0161] wherein, is the attention score calculated based on the lth layer features of the drug d i and d j , N(i) represents the neighbor set of the drug d i , is the normalized attention score between the drug d i and its neighbor d j ;
[0162]
[0163] wherein, and are the lth layer learnable parameters, is the lth layer learnable weight vector, is the similarity value between the drug d i and the drug d j in the mth similarity network;
[0164] Then the output of the last layer is taken as the final feature representation learned under the similarity network, denoted as L is the number of layers, S m indicates that the feature is learned from the m-th similarity network, and the same strategy is adopted to obtain the protein feature from each protein similarity network
[0165] In step S42, taking the drug as an example, first, the features learned by each similarity network are fused to obtain a comprehensive feature representation, and the specific operation is as follows:
[0166]
[0167] where ∏ denotes concatenation operation;
[0168] Conv is a one-dimensional convolution operation;
[0169] is the integrated representation of the drug obtained by fusing all similarity networks;
[0170] The integrated feature of the protein is obtained using the same method
[0171] In step S43, the feature representation learned by each network is compared with the integrated feature, and the distance between the feature learned by each network and the integrated feature is narrowed down by contrastive learning. For the feature representation learned in any network The contrastive learning loss of the integrated feature is specifically defined as follows:
[0172]
[0173] where n d is the number of drugs;
[0174] and are the positive sample set and the negative sample set of drug d i in the mth drug similarity network, respectively;
[0175] is the feature of drug d i learned from the mth drug similarity network;
[0176] is the integrated feature of drug d i ;
[0177] is the cosine similarity algorithm;
[0178] τ is a temperature parameter;
[0179] The contrastive loss between the feature learned by each protein similarity network and the integrated feature is obtained using the same method Finally, in step S44, the contrastive loss of each similarity network of the drug and the protein and the integrated feature is added to obtain the overall contrastive loss L c .
[0180] In step S5, an adaptive self-step sampling strategy is designed for cooperative contrast learning, the feature similarity between comprehensive features of drugs or proteins is calculated, and adaptive contrast learning negative sample pairs are selected for each network according to the similarity value between drugs or proteins in each network and the comprehensive feature similarity; taking proteins as an example, for any protein p i In the contrast process between the representation learned from the similarity network and the comprehensive representation, the contrast learning positive sample pair set is The negative sample pair is obtained through the adaptive self-step sampling strategy.
[0181] Step S5 specifically includes the following steps:
[0182] S51, define two scoring functions to measure the reliability of negative samples from different angles:
[0183]
[0184] wherein, represents the negative sample reliability of the mth similarity network; represents the negative sample reliability calculated according to the comprehensive representation;
[0185] sim(p i ,p j ) represents the cosine similarity algorithm;
[0186] S52, define a self-step function to determine the number of negative samples selected in each iteration learning process, and the self-step function is:
[0187]
[0188] wherein, T is the maximum training round;
[0189] t is the current training round;
[0190] β is a hyperparameter for controlling the ratio of the number of negative samples to be selected to the number of candidate negative samples;
[0191] n p is the number of proteins;
[0192] represents the floor operation;
[0193] num t represents the number of negative samples to be selected in the current training round;
[0194] S53, arrange the negative samples of each similarity network in descending order of reliability , and select the first num t negative samples for the current training round, denoted as Nns At the same time, all negative samples are sorted in descending order according to the confidence values calculated by the fusion features In descending order, the first num t are selected as the negative samples of the current round, denoted as N fs The union of the two sets is taken as the final negative sample set N m = N ns ∪N fs .
[0195] In step S6, the joint representation of the drug-protein pair is processed and extracted using CNN:
[0196]
[0197] wherein, and are the fusion features of drug d i and protein p j ;
[0198] is the joint representation of drug d i and protein p j ;
[0199] The probability of drug-protein interaction is predicted by a fully connected layer:
[0200]
[0201] wherein, is the weight matrix;
[0202] Bceloss is used as the objective function for drug-protein association prediction:
[0203]
[0204] wherein, is the predicted probability of the association between drug d i and protein p j ;
[0205] is the true label of whether d i and p j are associated;
[0206] dpp is the training set of drug-protein pairs.
[0207] As Figure 3As shown, five feature fusion methods are evaluated: sum, average, weighted aggregation (w-agg), concatenation (concat), and convolutional neural network (CNN). Each method has four indicator results, from left to right, AUROC value (without the application), AUROC value (with the application), AUPR value (without the application), and AUPR value (with the application). As can be seen from the figure, under each feature fusion method, the AUROC and AUPR results of predicting drug-target association relationships by the application are higher than those without the application.
[0208] As shown, dark circles represent positive drug-target sample pairs, and light circles represent negative drug-target sample pairs. Figure A shows that there is a large amount of overlap between the features of positive and negative sample pairs at the initial training stage (epoch = 1). Figure B shows that the application gradually learns distinguishable positive and negative sample pair feature representations at the intermediate training stage, and Figure C shows that the application can learn more distinguishable positive and negative sample pair feature representations at the final training stage. Figure 4
[0209] In summary, the principle of the present embodiment is as follows: first, the drug SMILES string and the protein amino acid sequence are converted into the atom linkage relationship graph and the amino acid sequence contact graph, then the GCN is used to extract the preliminary structural feature representation of the drug and the protein from the atom linkage relationship graph and the amino acid sequence contact graph, then the various interaction networks related to the drug and the protein are converted into the similarity network of the drug or the protein, then the preliminary structural feature representation of the drug and the protein is taken as the initial feature input, the feature representation with higher consistency of the drug and the protein is learned on each similarity network through the collaborative contrast learning and the adaptive self-step sampling strategy, and finally the potential drug-target association is predicted by the MLP. By constructing the collaborative contrast learning and the adaptive self-step sampling strategy, the present application can more fully utilize the complementary knowledge between the various interaction networks, learn the feature representation with higher consistency from each interaction network, and thus improve the prediction accuracy of the drug-target association relationship.
[0210] The specific embodiments described herein are merely illustrative of the spirit of the present application. Those skilled in the art of the present application can make various modifications or supplements to the described specific embodiments or replace them with similar ways, without departing from the spirit of the present application or exceeding the scope defined by the appended claims.
Claims
1. A method for predicting drug target associations, characterized in that, The method comprises the following steps: S1, collecting and downloading drug and protein structure data and associated interaction networks; S2, drug and protein structure data processing, converting interaction networks; S3, extracting and learning drug structure features and protein structure features from structure data; S4, designing a collaborative contrast learning strategy to calculate a contrast learning loss; S5, designing an adaptive self-step sampling strategy to obtain a high-reliability contrast learning sample pair; S6, obtaining the final feature representation of the drug and protein, and using MLP to predict the potential drug-protein interaction; Step S4 specifically comprises the following steps: S41, fusing the feature representation learned by each similarity network; S42, obtaining a comprehensive representation of the drug and protein; S43, comparing the feature learned by each network with the comprehensive feature; S44, calculating the contrast learning loss; In step S41, taking a drug as an example, for each similarity network, denoted as S m , (m = 1, 2, 3, 4), the weight of the neighbor under the similarity network is calculated first, and then the features of the neighbor nodes are aggregated according to the weight, that is: wherein, represents a feature representation of the output of the lth hidden layer, represents a weight matrix of the lth layer; wherein, is the attention score based on the drug d i and the first layer features of d j , N(i) denotes the neighborhood set of drug d i , is the normalized attention score between drug d i and its neighbors d j ; wherein, and is the l-th layer learnable parameter, is the l-th layer learnable weight vector, is the similarity value of drug d i and drug d j in the m-th similarity network. Then the output of the last layer is taken as the final feature representation learned under this similarity network, denoted as L is the number of layers, S m denotes that the feature is learned from m similarity networks. Using the same strategy, we obtain protein features from each protein similarity network In step S42, taking the drug as an example, first fuse the feature learned by each similarity network to obtain a comprehensive feature representation, and the specific operation is as follows: Wherein: ∏ represents the splicing operation; Conv is a one-dimensional convolution operation; is the drug representation obtained by fusing all similarity networks; Using the same method to obtain the comprehensive characteristics of proteins In step S43, the feature representation learned by each network is compared with the comprehensive feature, and the distance between the feature representation learned by each network and the comprehensive feature is reduced by using a contrastive learning loss; for the feature representation learned in any one network The contrastive learning loss of the comprehensive feature with the feature representation is specifically defined as follows: wherein n d is the number of drugs; and is a positive sample set and a negative sample set of drug d in the mth drug similar network, respectively. i is a drug d learned from the mth drug similarity network i characteristics of is a drug d i fusion features; sim(g) is a cosine similarity algorithm; τ is a temperature parameter; The contrast loss between the features learned by each protein similarity network and the fusion features is obtained using the same method Finally, in step S44, the contrast loss of each protein similarity network and the fusion features is added to obtain the overall contrast loss L c ; Step S5 specifically comprises the following steps: S51, defining two score functions to measure the reliability of negative samples from different angles: wherein, represents the negative sample confidence of the m-th similarity network; represents the negative sample confidence calculated according to the comprehensive representation. sim(p i ,p j ) denotes a cosine similarity algorithm; S52, defining a self-step function to determine the number of negative samples selected in each iteration learning process, and the self-step function is: Wherein, T is the maximum training round; t is the current training round; β is a hyperparameter for controlling the ratio of the number of negative samples to be selected to the number of candidate negative samples; n p is the amount of protein; denotes a floor operation; num t represents the number of negative samples that should be selected in the current training round; S53. Rank the negative samples of each similarity network according to their credibility. Sort in descending order and select the first num. t N negative samples are used in the current training round. ns Meanwhile, the credibility of all negative samples is calculated based on the fusion features. Sort in descending order, then select the first num again. t N is used as the negative sample for the current round. fs Then, the union of the two sets is taken as the final negative sample set N. m =N ns UN fs .
2. The method of predicting drug target association according to claim 1, wherein, In step S1, the drug and protein structure data comprises: ① drug SMILES string data; ② protein amino acid sequence data; ③ protein amino acid three-dimensional structure data; The drug and protein interaction network comprises: a, drug-disease association network; b, drug-side effect association network; c, drug-drug interaction network; d, drug-protein association network; e, protein-disease interaction network; f, protein-protein interaction network; The drug-protein association network dataset comprises four biological entities and six interactions, wherein the four biological entities include: 708 drugs, 1512 proteins, 5603 diseases, and 4192 side effects; the six interactions include: 1923 drug-protein interactions, 199214 drug-disease associations, 10036 drug-drug interactions, 80164 drug-side effect associations, 1923 protein-drug interactions, 1596745 protein-disease associations, and 7363 protein-protein interactions.
3. The method of predicting drug target association according to claim 1, wherein, Step S2 specifically comprises the following steps: S21, converting the SMILES string to an atomic interaction graph by using the dgl tool; S22, converting the protein amino acid three-dimensional structure to an amino acid residue graph by calculating the distance between amino acid residues; S23, converting each interaction network to a similarity network by using the Jaccard similarity algorithm; In step S21, the dgl tool is an open source tool. In step S22, the three-dimensional structure of the protein amino acid contains three-dimensional coordinate information of each amino acid residue. The distance between any two amino acid residues is calculated, and two amino acids with a distance less than a specific threshold are considered to have a link relationship, and then an amino acid residue contact map is obtained. In step S23, a similarity network is constructed based on each interaction network by Jaccard similarity algorithm, and the similarity algorithm formula is: Wherein: A represents the adjacency matrix of any one interaction network; A i denotes the i-th row of the adjacency matrix A; S ij denotes the similarity value between nodes i and j learned from the adjacency matrix A.
4. The method of predicting drug target association according to claim 3, wherein, Step S3 is specifically divided into the following steps: S31, using the GCN model to extract the drug structure features on the atomic interaction graph; S32, using the GCN model to extract the protein structure features on the amino acid residue graph.
5. The method of predicting drug target association according to claim 4, wherein, In steps S31 and S32, the formula for learning the feature representation of the drug and protein from the atomic interaction graph and the amino acid residue contact graph by GCN is: wherein: represents a correlation matrix between atoms of any one drug or a correlation matrix between amino acid residues in a protein; n represents the number of atoms in the drug or the number of amino acid residues in the protein; representing initial characteristics of the drug or protein; is a diagonal matrix, each diagonal element has a value of the degree of a drug atom or a protein amino acid residue; are learnable weight parameters, f is the feature dimension of the output drug atom or protein amino acid residue; σ(g) is a nonlinear activation function; After updating the feature representation of each atom, a readout function is used to combine the features of all atoms to obtain the feature representation of the drug, and the readout function formula is: h = Readout (X'); Fusing drugs by reading out function formulas i All atomic information gives drug structure feature representation Fusing proteins p i All amino acid residue information gives protein structure feature representation 6. The method of predicting drug target association according to claim 1, wherein, In step S5, an adaptive self-step sampling strategy is designed for cooperative contrast learning, the feature similarity between the comprehensive features of drugs or proteins is calculated, and adaptive contrast learning negative sample pairs are selected for each network according to the similarity values between drugs or proteins in each network and the comprehensive feature similarity; for any one protein p i In the contrast process between the representation learned from the similarity network and the comprehensive representation, the contrast learning positive sample pair set is The negative sample pairs are obtained through the adaptive self-step sampling strategy.
7. The method of predicting drug target association according to claim 1, wherein, In step S6, the joint representation of the drug-protein pair is processed and extracted by using CNN: wherein, and are fusion characteristics of the drug d i and the protein p j respectively; is a drug d i and a protein p j in combination; The probability of drug-protein interaction is predicted by a fully connected layer: wherein is a weight matrix; Bceloss is used as the target function of drug-protein association prediction: wherein is a drug d i and a protein p j a predicted probability of association; is d i and p j whether there is an associated ground truth label; dpp is the training set of drug-protein pairs.