Method for predicting association relationship of drug targets
Through collaborative comparative learning and adaptive self-step sampling strategies, combined with GCN and Jaccard similarity algorithms, the structural characteristics and similarity network characteristics of drugs and proteins are extracted, and the failure to fully utilize biological entity association relationships and structural information in the prior art is solved. The problem of this paper significantly improves the prediction accuracy of drug-target association relationships.
Patent Information
- Application Number
- CN202410914418.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-09
- Publication Date
- 2024-11-26
- Estimated Expiration
- 2044-07-09
AI Technical Summary
The existing drug target association prediction methods fail to make full use of the association relationship between biological entities and the structural information of the drug and the target itself, and the mixed methods fail to effectively utilize the complementary information between multiple interactive networks, resulting in drug-target The accuracy of association relationship prediction is insufficient.
Using collaborative contrast learning and adaptive self-step sampling strategy, the structural characteristics and similarity network characteristics are extracted by collecting the structural data and interaction network of drugs and proteins, and predicting potential drug-protein interaction relationships through MLP, using GCN and Jaccard similarity The degree algorithm builds a similarity network, combining comparative learning loss and adaptive self-step sampling strategy to learn feature representations with higher consistency.
The prediction accuracy of drug-target association relationships is improved, and the complementarity knowledge between multiple interactive networks is fully utilized, which significantly improves the prediction effect.
Smart Images

Figure CN119028427A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of computational biology, and in particular to a method for predicting drug-target association relationships. Background Art
[0002] Existing drug target (protein) association prediction methods include those based on structure type data, network type data and hybrid methods. The method based on structure data first collects the SMILES string data of the drug, the amino acid sequence data of the protein or the three-dimensional structure data of the protein, and then uses deep learning methods such as CNN and Transformer to extract features from the SMILES and amino acid sequence data, or obtains the link relationship between drug atoms and the link relationship between amino acid residues based on the SMILES string and the three-dimensional structure data of the protein, and then uses GCN, GAT and other methods to extract features.
[0003] The network-type data-based method first obtains a variety of interactions and association networks between biological entities, such as drug-disease, drug-side effect, protein-disease interaction networks, and then uses similarity algorithms such as cosine similarity and Gaussian interaction similarity to convert these interaction networks into drug or protein similarity networks. Then, deep learning methods such as GCN and RWR are used to learn more effective feature representations of drugs and proteins from the similarity networks. The hybrid method usually constructs two encoders: a structural data encoder and a network data encoder. The features learned by the two encoders are fused through the attention mechanism or the features learned by the structural encoder are input into the network data encoder to simultaneously utilize the structural type data and the network type data, so as to learn the feature representations of drugs and targets with richer information to the greatest extent possible.
[0004] Although structure-based and network-based methods have achieved certain results, structure-based methods do not consider the impact of the association between biological entities on the association between drugs and targets, while network-based methods do not utilize the structural information of drugs and targets themselves; although hybrid methods utilize structural data and network data, existing hybrid methods usually simply aggregate the features learned from multiple interactive networks and do not fully utilize the complementary information between multiple interactive networks. It is difficult to fully utilize the complementary knowledge between multiple interactive networks and learn more consistent feature representations from each interactive network, which affects the prediction accuracy of drug-target associations. Summary of the invention
[0005] The purpose of the present invention is to provide a method for predicting drug-target association relationships in response to the above problems.
[0006] To achieve the above object, the present invention adopts the following technical scheme: a method for predicting drug-target association, the method comprising the following steps:
[0007] S1. Collect and download drug and protein structure data and associated interaction networks;
[0008] S2, drug and protein structure data processing, conversion of interaction networks;
[0009] S3, extract and learn the structural features of drugs and proteins from structural data;
[0010] S4. Design collaborative contrastive learning strategy and calculate contrastive learning loss;
[0011] S5. Design an adaptive self-paced sampling strategy to obtain highly reliable contrastive learning sample pairs;
[0012] S6. Obtain the final feature representation of drugs and proteins, and use MLP to predict potential drug-protein interactions.
[0013] In step S1, drug and protein structure data include:
[0014] ① Drug SMILES string data;
[0015] ②Protein amino acid sequence data;
[0016] ③ Protein amino acid three-dimensional structure data;
[0017] The drug-protein interaction network 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. 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.
[0025] Step S2 specifically includes the following steps:
[0026] S21. Convert the SMILES string into an atomic interaction diagram using the dgl tool;
[0027] S22, converting the three-dimensional amino acid structure of the protein into an amino acid residue graph by calculating the distance between the amino acid residues;
[0028] S23, converting each interaction network into a similarity network by using the Jaccard similarity algorithm;
[0029] Among them, the dgl tool in step S21 is an open source tool. In step S22, the three-dimensional structure of protein amino acids contains the 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, thereby obtaining an amino acid residue contact map; in step S23, a similarity network is constructed based on each interaction network using the Jaccard similarity algorithm, and the similarity algorithm formula is:
[0030]
[0031] Where: A represents the adjacency matrix representation of any 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 is specifically divided into the following steps:
[0035] S31, extracting drug structural features on atomic interaction graph using GCN model;
[0036] S32, using the GCN model to extract protein structural features on the amino acid residue graph;
[0037] In step S31 and step S32, the formula for learning the characteristic representation of drugs and proteins from the atomic interaction graph and the amino acid residue contact graph using GCN is:
[0038]
[0039] in: Represents the correlation matrix between atoms of any drug or the correlation matrix between amino acid residues in a 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 characteristics of a drug or protein;
[0042] is a diagonal matrix, and the value of each diagonal element is the degree of the drug atom or protein amino acid residue;
[0043] is a learnable weight parameter, and f is the characteristic dimension of the output drug atoms or protein amino acid residues;
[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. The readout function formula is:
[0046] h = Readout(X′);
[0047] By reading out the function formula, the drug d i All atomic information is used to obtain the drug structure characteristics. Fusion protein p i All amino acid residue information is used to obtain protein structural features
[0048] Step S4 specifically includes the following steps:
[0049] S41, integrating the feature representations learned by each similarity network;
[0050] S42, obtain a comprehensive representation of drugs and proteins;
[0051] S43, comparing the features learned by each network with the comprehensive features;
[0052] S44, calculate contrastive learning loss;
[0053] In step S41, taking medicine as an example, for each similarity network, denoted as S m,(m=1,2,3,4), first calculate the weight of the neighbors under the similarity network, and then aggregate the features of the neighbor nodes according to the weight, that is:
[0054]
[0055] in, represents the feature representation of the output of the lth hidden layer, represents the weight matrix of the lth layer;
[0056]
[0057] in, Based on drug i and d j The attention score calculated from the l-th layer feature of N(i) represents the drug d i The neighbor set of It is a drug i With its neighbors j The normalized attention score between ;
[0058]
[0059] in, and is the learnable parameter of the lth layer, is the learnable weight vector of the lth layer, is the drug d in the mth similarity network i and drugs j Similarity value of ;
[0060] Then the output of the last layer is used 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. The same strategy is used to obtain protein features from each protein similarity network.
[0061] In step S42, taking medicine as an example, the features learned by each similarity network are first fused to obtain a comprehensive feature representation. The specific operations are as follows:
[0062]
[0063] Among them: ∏ represents the splicing operation;
[0064] Conv is a one-dimensional convolution operation;
[0065] It is a comprehensive representation of drugs obtained by fusing all similarity networks;
[0066] The same approach was used to obtain comprehensive protein features
[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 reduced by contrastive learning. The specific definition of the contrast learning loss with the comprehensive features is as follows:
[0068]
[0069] Among them, n d is the quantity of the drug;
[0070] and are respectively the drug d in the mth drug similarity network i Positive sample set and negative sample set;
[0071] is the drug d learned from the mth drug similarity network i Features;
[0072] It is a drug i The fusion characteristics of
[0073] It is the cosine similarity algorithm;
[0074] τ is the temperature parameter;
[0075] Use the same method to obtain the contrast loss between the features learned by each protein similarity network and the fusion features. Finally, in step S44, the contrast loss of each similarity network of drugs and proteins and the fusion feature is added to obtain the overall contrast loss L c .
[0076] Step S5 designs an adaptive self-paced sampling strategy for collaborative contrastive learning, calculates the feature similarity between the comprehensive features of drugs or proteins, and selects adaptive contrastive learning negative sample pairs for each network based on the similarity values and comprehensive feature similarity between drugs or proteins in each network; taking protein as an example, for any protein p i , in the similarity network In the process of comparing the learned representation with the comprehensive representation, the set of positive sample pairs for contrastive learning is Negative sample pairs are obtained through an adaptive self-paced sampling strategy.
[0077] Step S5 specifically includes the following steps:
[0078] S51. Define two score functions to measure the reliability of negative samples from different perspectives:
[0079]
[0080] in, Represents the negative sample credibility of the mth similarity network; represents the negative sample credibility calculated based on the comprehensive representation;
[0081] sim(p i ,p j ) represents the cosine similarity algorithm;
[0082] S52. Define a self-pacing function to determine the number of negative samples selected in each iterative learning process. The self-pacing function is:
[0083]
[0084] Where T is the maximum number of training rounds;
[0085] t is the current training round;
[0086] β is a hyperparameter that controls the ratio of the number of selected negative samples to the number of candidate negative samples;
[0087] n p is the amount of protein;
[0088] Indicates a round-down operation;
[0089] num t Indicates the number of negative samples that should be selected in the current training round;
[0090] S53, the negative samples of each similarity network are divided into Sort in descending order and select the first num t Negative samples are used for the current training round, denoted as N ns At the same time, all negative samples are calculated according to the credibility of the fusion features Sort in descending order, select the first num again t As the negative samples of the current round, denoted as N fs , and then take the union of the two sets as the final negative sample set N m =N ns ∪N fs .
[0091] In step S6, CNN is used to process and extract the joint representation of drug-protein pairs:
[0092]
[0093] in, and Drug d i and protein p j The fusion characteristics of
[0094] It is a drug i and protein p j The joint representation of
[0095] Predict the probability of drug-protein interaction through the fully connected layer:
[0096]
[0097] in, is the weight matrix;
[0098] Using Bceloss as the objective function for drug-protein association prediction:
[0099]
[0100] in, It is a drug i and protein p j The predicted probability of association;
[0101] Yes i and p j Is there an associated true label?
[0102] dpp is the training set of drug-protein pairs.
[0103] Compared with the existing technology, the advantages of the present invention are: learning more consistent feature representations of drugs and proteins on each similarity network through collaborative contrast learning and adaptive self-paced sampling strategies, and predicting potential drug-target associations through MLP; by constructing collaborative contrast learning and adaptive self-paced sampling strategies, the present invention can make more full use of the complementary knowledge between multiple interactive networks, learn more consistent feature representations from each interactive network, thereby improving the prediction accuracy of drug-target associations, and has a good use effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0104] Figure 1 is a prediction flow chart of the present invention;
[0105] Figure 2 is the collaborative contrastive learning advantage map of the present invention;
[0106] Figure 3It is a histogram of the effectiveness of the collaborative contrastive learning and self-paced sampling strategy implemented by the present invention under different feature fusion methods;
[0107] Figure 4 It is a visual analysis diagram of the drug-target pair feature representation learned by the present invention at different training stages. DETAILED DESCRIPTION
[0108] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0109] like Figure 1-4 As shown, a method for predicting drug-target association relationships comprises the following steps:
[0110] S1. Collect and download drug and protein structure data and associated interaction networks;
[0111] S2, drug and protein structure data processing, conversion of interaction networks;
[0112] S3, extract and learn the structural features of drugs and proteins from structural data;
[0113] S4. Design collaborative contrastive learning strategy and calculate contrastive learning loss;
[0114] S5. Design an adaptive self-paced sampling strategy to obtain highly reliable contrastive learning sample pairs;
[0115] S6. Obtain the final feature representation of drugs and proteins, and use MLP to predict potential drug-protein interactions.
[0116] In step S1, drug and protein structure data include:
[0117] ① Drug SMILES string data;
[0118] ②Protein amino acid sequence data;
[0119] ③ Protein amino acid three-dimensional structure data;
[0120] The drug-protein interaction network includes:
[0121] a. Drug-disease association network;
[0122] b. Drug-side effect association network;
[0123] c. Drug-drug interaction network;
[0124] d. Drug-protein association 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. 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.
[0128] Step S2 specifically includes the following steps:
[0129] S21. Convert the SMILES string into an atomic interaction diagram using the dgl tool;
[0130] S22, converting the three-dimensional amino acid structure of the protein into an amino acid residue graph by calculating the distance between the amino acid residues;
[0131] S23, converting each interaction network into a similarity network by using the Jaccard similarity algorithm;
[0132] Among them, the dgl tool in step S21 is an open source tool. In step S22, the three-dimensional structure of protein amino acids contains the 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, thereby obtaining an amino acid residue contact map; in step S23, a similarity network is constructed based on each interaction network using the Jaccard similarity algorithm, and the similarity algorithm formula is:
[0133]
[0134] Where: A represents the adjacency matrix representation of any 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 is specifically divided into the following steps:
[0138] S31, extracting drug structural features on atomic interaction graph using GCN model;
[0139] S32, using the GCN model to extract protein structural features on the amino acid residue graph;
[0140] like Figure 2 As shown, Figure A shows a feature extraction method based on traditional multi-network learning, Figure B shows a feature extraction method based on traditional contrast learning, and Figure C shows the collaborative contrast learning feature extraction method of the present invention.
[0141] In step S31, the formula for learning the feature representation of drugs and proteins from the atomic interaction graph and the amino acid residue contact graph using GCN is:
[0142]
[0143] in: Represents the correlation matrix between atoms of any drug or the correlation matrix between amino acid residues in a 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 characteristics of a drug or protein;
[0146] is a diagonal matrix, and the value of each diagonal element is the degree of the drug atom or protein amino acid residue;
[0147] is a learnable weight parameter, and f is the characteristic dimension of the output drug atoms or protein amino acid residues;
[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. The readout function formula is:
[0150] h = Readout(X′);
[0151] By reading out the function formula, the drug d i All atomic information is used to obtain the drug structure characteristics. Fusion protein p i All amino acid residue information is used to obtain protein structural features
[0152] Step S4 specifically includes the following steps:
[0153] S41, integrating the feature representations learned by each similarity network;
[0154] S42, obtain a comprehensive representation of drugs and proteins;
[0155] S43, comparing the features learned by each network with the comprehensive features;
[0156] S44, calculate contrastive learning loss;
[0157] In step S41, taking medicine as an example, for each similarity network, denoted as S m ,(m=1,2,3,4), first calculate the weight of the neighbors under the similarity network, and then aggregate the features of the neighbor nodes according to the weight, that is:
[0158]
[0159] in, represents the feature representation of the output of the lth hidden layer, represents the weight matrix of the lth layer;
[0160]
[0161] in, Is based on drug i and d j The attention score calculated from the l-th layer feature of N(i) represents the drug d i The neighbor set of It is a drug i With its neighbors j The normalized attention score between ;
[0162]
[0163] in, and is the learnable parameter of the lth layer, is the learnable weight vector of the lth layer, is the drug d in the mth similarity network i and drugs j Similarity value of ;
[0164] Then the output of the last layer is used 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. The same strategy is used to obtain protein features from each protein similarity network.
[0165] In step S42, taking medicine as an example, the features learned by each similarity network are first fused to obtain a comprehensive feature representation. The specific operations are as follows:
[0166]
[0167] Among them: ∏ represents the splicing operation;
[0168] Conv is a one-dimensional convolution operation;
[0169] It is a comprehensive representation of drugs obtained by fusing all similarity networks;
[0170] The same approach was used to obtain comprehensive protein features
[0171] 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 reduced by contrastive learning. The specific definition of the contrast learning loss with the comprehensive features is as follows:
[0172]
[0173] Among them, n d is the quantity of the drug;
[0174] and are respectively the drug d in the mth drug similarity network i Positive sample set and negative sample set;
[0175] is the drug d learned from the mth drug similarity network i Features;
[0176] It is a drug i The fusion characteristics of
[0177] It is the cosine similarity algorithm;
[0178] τ is the temperature parameter;
[0179] Use the same method to obtain the contrast loss between the features learned by each protein similarity network and the fusion features. Finally, in step S44, the contrast loss of each similarity network of drugs and proteins and the fusion feature is added to obtain the overall contrast loss L c .
[0180] In step S5, an adaptive self-paced sampling strategy is designed for collaborative contrastive learning, and the feature similarity between the comprehensive features of drugs or proteins is calculated. At the same time, adaptive contrastive learning negative sample pairs are selected for each network according to the similarity values and comprehensive feature similarity between drugs or proteins in each network. Taking protein as an example, for any protein p i , in the similarity network In the process of comparing the learned representation with the comprehensive representation, the set of positive sample pairs for contrastive learning is Negative sample pairs are obtained through an adaptive self-paced sampling strategy.
[0181] Step S5 specifically includes the following steps:
[0182] S51. Define two score functions to measure the reliability of negative samples from different perspectives:
[0183]
[0184] in, Represents the negative sample credibility of the mth similarity network; represents the negative sample credibility calculated based on the comprehensive representation;
[0185] sim(p i ,p j ) represents the cosine similarity algorithm;
[0186] S52. Define a self-pacing function to determine the number of negative samples selected in each iterative learning process. The self-pacing function is:
[0187]
[0188] Where T is the maximum number of training rounds;
[0189] t is the current training round;
[0190] β is a hyperparameter that controls the ratio of the number of selected negative samples to the number of candidate negative samples;
[0191] n p is the amount of protein;
[0192] Indicates a round-down operation;
[0193] num t Indicates the number of negative samples that should be selected in the current training round;
[0194] S53, the negative samples of each similarity network are divided into Sort in descending order and select the first num t Negative samples are used for the current training round, denoted as Nns At the same time, all negative samples are calculated according to the credibility of the fusion features Sort in descending order, select the first num again t As the negative samples of the current round, denoted as N fs , and then take the union of the two sets as the final negative sample set N m =N ns ∪N fs .
[0195] In step S6, CNN is used to process and extract the joint representation of drug-protein pairs:
[0196]
[0197] in, and Drug d i and protein p j The fusion characteristics of
[0198] It is a drug i and protein p j The joint representation of
[0199] Predict the probability of drug-protein interaction through the fully connected layer:
[0200]
[0201] in, is the weight matrix;
[0202] Using Bceloss as the objective function for drug-protein association prediction:
[0203]
[0204] in, It is a drug i and protein p j The predicted probability of association;
[0205] Yes i and p j Is there an associated true label?
[0206] dpp is the training set of drug-protein pairs.
[0207] like Figure 3As shown, five feature fusion methods were evaluated: sum, avg, weighted aggregation (w-agg), concatenation (concat) and convolutional neural network (CNN). Each method has four indicator results, from left to right, AUROC value (not using the present invention), AUROC value (using the present invention), AUPR value (not using the present invention), and AUPR value (using the present invention). It can be seen from the figure that under each feature fusion method, the AUROC and AUOR results of the present invention for predicting the drug-target association relationship are higher than those not using the present invention.
[0208] like Figure 4 As shown, dark dots represent drug-target positive sample pairs, and light dots represent drug-target negative sample pairs. Figure A shows that in the initial training stage (epoch=1), there is a large overlap between the features of the positive and negative sample pairs. Figure B shows that in the intermediate training stage, the implementation of the present invention gradually learns distinguishable positive and negative sample pair feature representations, and Figure C shows that in the final training stage, the implementation of the present invention can learn more distinguishable positive and negative sample pair feature representations.
[0209] In summary, the principle of this implementation is: first, convert the drug SMILES string and protein amino acid sequence into an atomic link relationship graph and an amino acid sequence contact graph, then use GCN to extract the preliminary structural feature representation of the drug and protein from the atomic link relationship graph and the amino acid sequence contact graph, then convert the various interaction networks related to the drug and protein into a similarity network of the drug or protein, then use the preliminary structural feature representation of the drug and protein as the initial feature input, learn the more consistent feature representation of the drug and protein on each similarity network through collaborative contrast learning and adaptive self-paced sampling strategy, and finally predict the potential drug-target association through MLP. By constructing collaborative contrast learning and adaptive self-paced sampling strategies, the present invention can make fuller use of the complementary knowledge between various interaction networks, learn more consistent feature representations from each interaction network, and thus improve the prediction accuracy of the drug-target association relationship.
[0210] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.
Claims
1. A method for predicting drug-target association relationships, characterized in that: The method comprises the following steps: S1. Collect and download drug and protein structure data and associated interaction networks; S2, drug and protein structure data processing, conversion of interaction networks; S3, extract and learn the structural features of drugs and proteins from structural data; S4. Design collaborative contrastive learning strategy and calculate contrastive learning loss; S5. Design an adaptive self-paced sampling strategy to obtain highly reliable contrastive learning sample pairs; S6. Obtain the final feature representation of drugs and proteins, and use MLP to predict potential drug-protein interactions.
2. A method for predicting drug-target association according to claim 1, characterized in that: In step S1, the drug and protein structure data include: ① Drug SMILES string data; ②Protein amino acid sequence data; ③ Protein amino acid three-dimensional structure data; The interactive network of drugs and proteins includes: 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 includes 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. A method for predicting drug-target association according to claim 1, characterized in that: Step S2 specifically includes the following steps: S21. Convert the SMILES string into an atomic interaction diagram using the dgl tool; S22, converting the three-dimensional amino acid structure of the protein into an amino acid residue graph by calculating the distance between the amino acid residues; S23, converting each interaction network into a similarity network by using the Jaccard similarity algorithm; Among them, the dgl tool in step S21 is an open source tool. In step S22, the three-dimensional structure of protein amino acids contains the 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, thereby obtaining an amino acid residue contact map; in step S23, a similarity network is constructed based on each interaction network using the Jaccard similarity algorithm, and the similarity algorithm formula is: Where: A represents the adjacency matrix representation of any interaction network; A i represents the i-th row of the adjacency matrix A; S ij Represents the similarity value between nodes i and j learned from the adjacency matrix A.
4. A method for predicting drug-target association according to claim 3, characterized in that: Step S3 is specifically divided into the following steps: S31, extracting drug structural features on atomic interaction graph using GCN model; S32. Use the GCN model to extract protein structural features on the amino acid residue graph.
5. A method for predicting drug-target association according to claim 4, characterized in that: In step S31 and step S32, the formula for learning the characteristic representation of drugs and proteins from the atomic interaction graph and the amino acid residue contact graph using GCN is: in: Represents the correlation matrix between atoms of any drug or the 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; Represents the initial characteristics of a drug or protein; is a diagonal matrix, and the value of each diagonal element is the degree of the drug atom or protein amino acid residue; is a learnable weight parameter, and f is the characteristic dimension of the output drug atoms or protein amino acid residues; 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. The readout function formula is: h = Readout(X′); By reading out the function formula, the drug d i All atomic information is used to obtain the drug structure characteristics h di , fusion protein p i All amino acid residue information can be used to obtain protein structural characteristics h pi .
6. A method for predicting drug-target association according to claim 5, characterized in that: Step S4 specifically includes the following steps: S41, integrating the feature representations learned by each similarity network; S42, obtain a comprehensive representation of drugs and proteins; S43, comparing the features learned by each network with the comprehensive features; S44, calculating contrastive learning loss; In step S41, taking medicine as an example, for each similarity network, denoted as S m ,(m=1,2,3,4), first calculate the weight of the neighbors under the similarity network, and then aggregate the features of the neighbor nodes according to the weight, that is: in, represents the feature representation of the output of the lth hidden layer, represents the weight matrix of the lth layer; in, Is based on drug i and d j The attention score calculated from the l-th layer feature of N(i) represents the drug d i The neighbor set of It is a drug i With its neighbors j The normalized attention score between ; in, and is the learnable parameter of the lth layer, is the learnable weight vector of the lth layer, is the drug d in the mth similarity network i and drugs j Similarity value of ; Then the output of the last layer is used 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 m similarity networks. The same strategy is used to obtain protein features from each protein similarity network.
7. A method for predicting drug-target association according to claim 6, characterized in that: In step S42, taking medicine as an example, the features learned by each similarity network are first fused to obtain a comprehensive feature representation. The specific operations are as follows: Among them: ∏ represents the splicing operation; Conv is a one-dimensional convolution operation; It is a comprehensive representation of drugs obtained by fusing all similarity networks; The same approach was used to obtain comprehensive protein features 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 reduced by contrastive learning. The specific definition of the contrast learning loss with the comprehensive features is as follows: Among them, n d is the quantity of the drug; and are the drug d in the mth drug similarity network i Positive sample set and negative sample set; is the drug d learned from the mth drug similarity network i Features; It is a drug i The fusion characteristics of It is the cosine similarity algorithm; τ is the temperature parameter; Use the same method to obtain the contrast loss L between the features learned by each protein similarity network and the fusion features p m Finally, in step S44, the contrast loss of each similarity network of drugs and proteins and the fusion feature are added to obtain the overall contrast loss L c .
8. The method for predicting drug-target association according to claim 1, characterized in that: In step S5, an adaptive self-paced sampling strategy is designed for collaborative contrastive learning, the feature similarity between the comprehensive features of drugs or proteins is calculated, and adaptive contrastive learning negative sample pairs are selected for each network according to the similarity values and comprehensive feature similarity between drugs or proteins in each network; Taking protein as an example, for any protein p i , in the similarity network In the process of comparing the learned representation with the comprehensive representation, the set of positive sample pairs for contrastive learning is Negative sample pairs are obtained through an adaptive self-paced sampling strategy.
9. The method for predicting drug-target association according to claim 1, characterized in that: Step S5 specifically includes the following steps: S51. Define two score functions to measure the reliability of negative samples from different perspectives: in, Represents the negative sample credibility of the mth similarity network; represents the negative sample credibility calculated based on the comprehensive representation; sim(p i ,p j ) represents the cosine similarity algorithm; S52. Define a self-pacing function to determine the number of negative samples selected in each iterative learning process. The self-pacing function is: Where T is the maximum number of training rounds; t is the current training round; β is a hyperparameter that controls the ratio of the number of selected negative samples to the number of candidate negative samples; n p is the amount of protein; Indicates a round-down operation; num t Indicates the number of negative samples that should be selected in the current training round; S53, the negative samples of each similarity network are divided into Sort in descending order and select the first num t Negative samples are used for the current training round, denoted as N ns At the same time, all negative samples are calculated according to the credibility of the fusion features Sort in descending order, select the first num again t As the negative samples of the current round, denoted as N fs , and then take the union of the two sets as the final negative sample set N m =N ns ∪N fs .
10. The method for predicting drug-target association according to claim 1, characterized in that: In step S6, CNN is used to process and extract the joint representation of drug-protein pairs: in, and Drug d i and protein p j The fusion characteristics of It is a drug i and protein p j The joint representation of Predict the probability of drug-protein interaction through the fully connected layer: in, is the weight matrix; Using Bceloss as the objective function for drug-protein association prediction: in, It is a drug i and protein p j The predicted probability of association; Yes i and p j Is there an associated true label? dpp is the training set of drug-protein pairs.
Citation Information
Patent Citations
Drug-target interaction prediction method based on supervised collaborative graph comparative learning
CN114023464A
Drug ATCCode prediction method based on graph transformation network
CN114420310A
Drug-target prediction method and system in combination with graph neural network and comparative learning
CN117612636A
Computer representations of peptides for efficient design of drug candidates
WO2023200866A1
Cited By
Drug-target interaction prediction method based on adversarial sample generation
CN121393627A
