A drug-target interaction prediction method based on hypergraph and active learning

Abstract

The application discloses a drug-target interaction prediction method based on a hypergraph and active learning, and roughly comprises the following steps: a drug-target adjacency matrix is constructed by using known drug-target correlations, drug-drug interactions and target-target interactions; drug hypergraphs and target hypergraphs are constructed according to the known drug-target correlation matrix; the initial features of drugs and targets are embedded and input into a double-path information transmission network respectively, and a contrast learning mechanism is introduced to optimize the representation; a semantic feature fusion module based on a Transformer is used to aggregate local features and global features; the final node features are input into a multilayer perceptron to predict drug-target interactions; and part of negative samples are selected from a negative sample pool according to a query strategy, and a training set is updated. The application realizes lightweight design at the model level, proposes a negative sample selection strategy based on active learning, optimizes negative samples in the training set through continuous iteration, and improves the model performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of biological information processing, and in particular to a drug-target interaction prediction method based on hypergraph and active learning. BACKGROUND

[0002] DTI prediction plays an indispensable role in modern drug development, as it can not only improve the efficiency and reduce the cost of R&D, but also provide new opportunities for drug repositioning, and is a key technology in computer-aided drug design.

[0003] With the progress of computing technology, DTI prediction methods have gradually formed three major technical routes based on ligands, structures and deep learning. Early DTI prediction mainly adopts ligand-based and structure-based strategies. Ligand-based methods are based on the principle of chemical similarity, which believes that molecules with similar structures may interact with the same target. For example, the SEA (Similarity Ensemble Approach) platform uses two-dimensional structure alignment technology, while SwissTargetPrediction combines two-dimensional and three-dimensional feature analysis. However, this kind of method often performs poorly when dealing with new molecular structures, and the prediction effect is not ideal in the case of lack of target data. In contrast, structure-based methods focus on protein three-dimensional conformation analysis. By evaluating the matching degree of drug molecules and protein binding sites, the interaction possibility is predicted. Typical methods include reverse molecular docking and reverse pharmacophore matching. However, structure-based methods face many challenges: not only do they require high-quality protein structure data, but also require strong computing power.

[0004] The rise of deep learning brings new breakthroughs to DTI prediction. This method breaks the traditional prediction framework and integrates chemical information, genomic data and interaction pairs into the computing model. Compared with traditional methods, deep learning can directly process the raw structure information of molecules and proteins, and automatically extract features through neural networks, graph convolution or attention mechanisms. This data-driven approach eliminates the dependence on artificially designed features and has the ability to handle massive data. However, deep learning methods still face the following challenges: i) In the DTI prediction task, the labels of unverified samples are uncertain. Traditional methods simply treat all unverified samples as negative samples, which may introduce false negative samples and reduce model performance; ii) Ordinary graph neural networks can only capture second-order relationships, and high-order relationship capture is insufficient, and they rely on neighbor propagation information, with limited propagation range, resulting in decreased prediction accuracy. SUMMARY

[0005] To solve the above technical problems, the present application provides a drug-target interaction prediction method based on hypergraph and active learning.

[0006] The technical scheme for solving the above problems of the present application is: a drug-target interaction prediction method based on hypergraph and active learning, comprising the following steps:

[0007] 1) Constructing a drug-target adjacency matrix using known drug-target associations, drug-drug interactions, and target-target interactions;

[0008] 2) Constructing drug hypergraphs and target hypergraphs according to the known drug-target association matrix;

[0009] 3) Embedding the initial features of drugs and targets into a dual convolution network composed of a graph convolution branch and a hypergraph convolution branch, and introducing a contrastive learning mechanism to optimize the representation;

[0010] 4) Using a semantic feature fusion module based on Transformer to aggregate local features and global features;

[0011] 5) Inputting the final node features into a multilayer perceptron to predict drug-target interactions;

[0012] 6) Selecting part of the negative samples from the negative sample pool according to the active query strategy BALD, updating the training set, and until the number of positive and negative samples in the training set is 1:1.

[0013] The above-mentioned drug-target interaction prediction method based on hypergraph and active learning, step 1) specifically comprises the following steps:

[0014] 1-1) Collecting data sets related to drug and target interaction prediction research, including 2214 drugs, 1968 targets, and a known drug-target association matrix Drug-drug interaction matrix and target-target interaction matrix Where d and t represent drugs and targets, respectively, d-d and t-t represent drug-drug and target-target associations, respectively, and N d and N t represent the number of drugs and targets, respectively;

[0015] 1-2) Constructing a drug-target adjacency matrix which can be defined as:

[0016]

[0017] The specific process of step 2) of the above-mentioned drug-target interaction prediction method based on hypergraph and active learning is as follows:

[0018] 2-1) Constructing drug hypergraph Based on two definitions, Definition 1: In drug-target bipartite graph, if drug i and drug j have a sequence of adjacent vertices, and the number of drugs in the path is less than k, then drug i (drug j ) is the k-order reachable neighbor of drug j (drug i ). Definition 2: In drug-target bipartite graph, if drug j and target k have direct interaction, and drug i is the k-order reachable neighbor of drug j , then target k (drug i ) is the k-order reachable neighbor of drug i (target k ).

[0019] For drug i , its k-order reachable drug set is represented as Mathematically, hypergraph can be defined on a family of sets, where each set represents a hyperedge, so here we can construct hyperedges by the k-order reachable target set of an item, and then construct high-order hyperedge groups according to the k-order reachable rule between drugs, which can be represented as:

[0020]

[0021] Further, the k-order reachable drug matrix of drug can be represented as:

[0022]

[0023] where power(M,k) is the k-th power of the given matrix M, is the association matrix of drug-target. Then the hyperedge group association matrix constructed by the k-order reachable rule between drugs can be represented as:

[0024]

[0025] Finally, the hypergraph matrix H d of drug is represented as:

[0026]

[0027] 2-2) Similarly, for target hypergraph, the k-order reachable target and k-order reachable drug of target can be defined symmetrically by similar methods. Specifically, the k-order reachable target matrix of target may be expressed as:

[0028]

[0029] Hyperedge group association matrix constructed by k-order reachable rules between targets may be expressed as:

[0030]

[0031] The hypergraph matrix H of the target is t may be expressed as:

[0032]

[0033] The above-mentioned drug-target interaction prediction method based on hypergraph and active learning, the step 3) comprises the following steps:

[0034] 3-1) Assign initial embedding vectors to drugs and targets, defined as and where h represents the embedding dimension, and the initial embedding vectors of drugs and targets are input into the bilinear convolution network respectively. First, the constructed drug-target adjacency matrix M is loaded into the network containing three graph convolution layers, and the global dependency relationship is captured through the graph convolution network. The embedding vector realizes message passing through the aggregation of the features of the neighbor nodes in each layer. The propagation process in the graph convolution layer is as follows:

[0035]

[0036] where f(·) is L2 norm normalization, and σ(·) is a nonlinear activation function LeakyReLu. is the drug / target aggregated embedding of the (l+1)th propagation layer of the graph convolution, X l is the drug / target aggregated embedding of the lth layer. is the normalized adjacency matrix, and the calculation formula is:

[0037]

[0038] where D represents the vertex degree matrix of the adjacency matrix M, and has

[0039] 3-2) Hypergraph allows a hyperedge to connect multiple nodes, and the representation of multi-element relationship can utilize the high-order relationship of all nodes connected by each hyperedge. Hypergraph convolution can capture more rich and complex structure information, and through the multi-layer convolution propagation mechanism, it can propagate information more widely in the network, supporting multi-point collaborative message propagation. The original embedding of drugs and targets is input into the three-layer hypergraph convolution module respectively, and the propagation process of drug embedding in the hypergraph convolution layer is as follows:

[0040]

[0041] where f(·) is L2 norm normalization, and σ(·) is a nonlinear activation function LeakyReLu. is the drug embedding of the l+1th propagation layer of the hypergraph convolution, is the drug embedding of the lth layer. is the drug hypergraph H d is the degree matrix of the vertices in the drug hypergraph H, and is the drug hypergraph H d is the degree matrix of the hyperedges in the drug hypergraph H, and

[0042] Similarly, the propagation process of the target embedding in the hypergraph convolution layer is:

[0043]

[0044] 3-3) The intra-layer feature fusion is achieved by weighted sum of the graph convolution embedding and the hypergraph convolution embedding in each layer, which is calculated as follows:

[0045]

[0046] where, and are the updated embeddings of the drug / target in the lth layer of the dual convolution network, and are the updated embeddings of the drug / target in the lth layer after graph convolution, and are the updated embeddings of the drug / target in the lth layer after hypergraph convolution. λ is a weight hyperparameter.

[0047] 3-4) The representation consistency between the graph convolution path and the hypergraph convolution path is optimized by introducing a contrastive learning mechanism, in which the graph convolution embedding and the hypergraph convolution embedding from the same drug or target constitute a positive sample pair, and all combinations of the graph convolution embedding and the hypergraph convolution embedding from different drugs or targets constitute a negative sample pair; the contrastive learning mechanism is realized by minimizing the distance between the positive sample pairs and maximizing the distance between the negative sample pairs, and its objective function is as follows:

[0048]

[0049] where τ is a temperature hyperparameter that controls the smoothness of the contrastive loss. exp(·) is the exponential operation in the logarithmic likelihood function, which represents the similarity between embeddings. Ω is a very small constant to avoid the case of zero denominator.

[0050] The above-mentioned drug-target interaction prediction method based on hypergraph and active learning, the process of step 4) is as follows:

[0051] 4-1) After passing through the three-layer double convolutional network, the embedding of each layer of the drug node and the target node is spliced to form a new embedding feature matrix is expressed as:

[0052]

[0053] Next, layer feature fusion is performed, and each layer embedding feature is weighted and combined to obtain a feature matrix is expressed as:

[0054]

[0055] The calculation formula is:

[0056]

[0057] wherein, w l is the learnable weight parameter of the lth layer.

[0058] 4-2) The semantic feature aggregation module based on Transformer uses a self-attention mechanism to fuse the feature vectors of drugs and targets. The attention aggregation operation of the drug obtains the drug feature The calculation formula is as follows:

[0059]

[0060] wherein, Q d ,K d ,V d are vectors obtained by mapping the drug features to the query, key and value spaces through different linear transformations, and the calculation formula is as follows:

[0061] Q d = W Q ·X d (22)

[0062] K d = W K ·X d (23)

[0063] V d = W V ·X d (24)

[0064] wherein, W Q ,W K ,W V are trainable parameter matrices.

[0065] Similarly, the target feature can be obtained

[0066] The above-mentioned drug-target interaction prediction method based on hypergraph and active learning, the process of step 5) is as follows:

[0067] 5-1) The input of the drug and target node features and into the multi-layer perception MLP, and the MLP is used to learn the feature interaction between the drug and the target node to obtain the interaction prediction probability The MLP learning prediction process is as follows:

[0068]

[0069] wherein, and are the feature vectors of the drug i and the target j respectively, and σ(·) is a nonlinear activation function LeakyReLu. W1, W2…W L is the weight matrix of the lth layer, and b1, b2…b L is the bias vector of the lth layer, and all are trainable parameter matrices.

[0070] 5-2) In the model training process, a binary cross-entropy loss function is used to calculate the node classification loss, and the binary cross-entropy loss function is defined as follows:

[0071]

[0072] wherein, N is the total number of prediction samples, y i represents the actual label, represents the prediction probability of the model.

[0073] 5-3) The final loss function of the model training is composed of three parts, as follows:

[0074]

[0075] wherein, α and β are trade-off hyperparameters, represents the L2 regularization to all trainable model parameters to prevent model overfitting.

[0076] The above-mentioned drug-target interaction prediction method based on hypergraph and active learning, the process of step 6) is as follows:

[0077] 6-1) The negative sample selection of the method uses pool-based active learning. Given a training set and a negative sample pool The pool-based active learning algorithm first selects a subset of the training set D train to train the model, and obtains the model M (0)Next, in each round t = 1, 2, …, T, the algorithm selects n samples from D neg according to the query strategy. Then the n samples are updated from D neg to D train , and a new model M train is learned based on the updated D (t) .

[0078] 6-2) A good query strategy can make the model sequence M (0) , M (1) , M (2) , …, M (T) perform well on the test set . This method implements a Bayesian model-based active learning strategy BALD (Bayesian Active Learning by Disagreement), which evaluates uncertainty by multiple Dropout and selects the most uncertain samples. The uncertainty index is based on the difference between two kinds of entropy, and the calculation formula is as follows:

[0079] Uncertainty(x) = H2(x) - H1(x) (28)

[0080] Where x represents a drug-target pair, H1(x) and H2(x) are the average entropy and multiple Dropout entropy respectively, and the calculation formula is:

[0081]

[0082] H2(x) = -probs(x)·log(1-probs(x)) (30)

[0083] Where probs(x) is the cumulative probability of x being a positive sample after multiple Dropout, is the average probability of x being a positive sample, and has:

[0084]

[0085] Where n drop is the total number of Dropout.

[0086] The beneficial effects of the present application are:

[0087] The application adopts a lightweight computing framework, integrates multi-dimensional information related to drug-target, captures complex correlation patterns between entities through a hypergraph architecture, and introduces a contrast learning strategy to optimize feature representation. At the same time, the feature fusion module based on Transformer can effectively integrate semantic information from different modules, realize deep feature interaction, and significantly enhance the generalization ability of the model. The active learning strategy optimizes the training process through dynamic negative sample selection. The method provides an efficient and reliable computing tool for drug research and development, and has certain scientific research value. BRIEF DESCRIPTION OF DRAWINGS

[0088] Figure 1 The overall flowchart of the application. DETAILED DESCRIPTION

[0089] The application will be further described below in conjunction with the drawings and examples.

[0090] As shown in the figure, a drug-target interaction prediction method based on hypergraph and active learning includes the following steps: Figure 1

[0091] 1) Use drug-target known association, drug-drug interaction and target-target interaction to construct drug-target adjacency matrix.

[0092] Step 1) specifically includes the following steps:

[0093] 1-1) Collect data sets in drug and target interaction prediction research, including 2214 drugs, and 1968 targets, and drug-target known association matrix Drug-drug interaction matrix And target-target interaction matrix Where d and t symbols represent drugs and targets respectively, d-d and t-t represent drug-drug and target-target association respectively, N d And N t represent the number of drugs and the number of targets respectively;

[0094] 1-2) Construct drug-target adjacency matrix It can be defined as:

[0095]

[0096] 2) Construct drug hypergraph and target hypergraph according to drug-target known association matrix.

[0097] The specific process of step 2) is as follows:

[0098] ​2-1) Constructing drug hypergraph Based on two definitions, Definition 1: In drug-target bipartite graph, if drug i and drug j have a sequence of adjacent vertices, and the number of drugs in the path is less than k, then drug i (drug j ) is the k-order reachable neighbor of drug j (drug i ). Definition 2: In drug-target bipartite graph, if drug j and target k have direct interaction, and drug i is the k-order reachable neighbor of drug j , then target k (drug i ) is the k-order reachable neighbor of drug i (target k ).

[0099] For drug i , its k-order reachable drug set is represented as Mathematically, hypergraph can be defined on a family of sets, where each set represents a hyperedge, so here we can construct hyperedges by the k-order reachable target set of an item, and then construct high-order hyperedge groups according to the k-order reachable rule between drugs, which can be represented as:

[0100]

[0101] Further, the k-order reachable drug matrix of drug can be represented as:

[0102]

[0103] where power(M,k) is the k-th power of the given matrix M, is the association matrix of drug-target. Then the hyperedge group association matrix constructed by the k-order reachable rule between drugs can be represented as:

[0104]

[0105] Finally, the hypergraph matrix H d of drug is represented as:

[0106]

[0107] 2-2) Similarly, for target hypergraph, the k-order reachable target and k-order reachable drug of target can be defined symmetrically by similar methods. Specifically, the k-order reachable target matrix of target is may be represented as:

[0108]

[0109] The hyperedge group association matrix constructed by the k-order reachable rule between targets may be represented as:

[0110]

[0111] The hypergraph matrix H of the target is t may be represented as:

[0112]

[0113] 3) The initial features of the drug and the target are embedded into a dual convolutional network composed of a graph convolution branch and a hypergraph convolution branch, respectively, and a contrast learning mechanism is introduced to optimize the representation.

[0114] Step 3) includes the following steps:

[0115] 3-1) Assign initial embedding vectors to drugs and targets, defined as and where h represents the embedding dimension, and the initial embedding vectors of the drug and the target are input into the dual convolutional network. First, the constructed drug-target adjacency matrix M is loaded into the network containing three layers of graph convolution, and the global dependency relationship is captured through the graph convolution network. The embedding vector realizes message passing through the aggregation of the features of the neighbor nodes in each layer. The propagation process in the graph convolution layer is as follows:

[0116]

[0117] where f(·) is L2 norm normalization, and σ(·) is a nonlinear activation function LeakyReLu. is the drug / target aggregated embedding of the (l+1)th propagation layer of the graph convolution, X l is the drug / target aggregated embedding of the lth layer. is the normalized adjacency matrix, and the calculation formula is:

[0118]

[0119] where D represents the vertex degree matrix of the adjacency matrix M, and has

[0120] 3-2) Hypergraph allows a hyperedge to connect multiple nodes, and the representation of multi-relational can utilize the high-order relationship of all nodes connected by each hyperedge. Hypergraph convolution can capture more rich and complex structure information, and obtain local high-order features through multi-layer convolution propagation mechanism. The original embedding of drugs and targets are input into three-layer hypergraph convolution module respectively, and the propagation process of drug embedding in hypergraph convolution layer is as follows:

[0121]

[0122] where f(·) is L2 norm normalization, and σ(·) is a nonlinear activation function LeakyReLu. is the drug embedding of the l+1th propagation layer of hypergraph convolution, is the drug embedding of the lth layer. is the drug hypergraph H d is the degree matrix of the vertex in the drug hypergraph H is the drug hypergraph H d is the degree matrix of the hyperedge in the drug hypergraph H

[0123] Similarly, the propagation process of target embedding in hypergraph convolution layer is as follows:

[0124]

[0125] 3-3) The intra-layer feature fusion is realized by weighted summation of graph convolution embedding and hypergraph convolution embedding in each layer, and the calculation method is as follows:

[0126]

[0127] where, and are the updated embeddings of drugs / targets after fusion in the lth layer of the bilinear convolution network, and are the updated embeddings of drugs / targets after graph convolution in the lth layer, and are the updated embeddings of drugs / targets after hypergraph convolution in the lth layer. λ is a weight hyperparameter.

[0128] 3-4) The representation consistency between graph convolution path and hypergraph convolution path is optimized by introducing a contrastive learning mechanism, in which the graph convolution embedding and hypergraph convolution embedding from the same drug or target constitute a positive sample pair, and all combinations of graph convolution embedding and hypergraph convolution embedding from different drugs or targets constitute a negative sample pair; the contrastive learning mechanism is realized by minimizing the distance between positive sample pairs and maximizing the distance between negative sample pairs, and the objective function is as follows:

[0129]

[0130] where τ is a temperature hyperparameter that controls the degree of smoothing of the contrastive loss. exp(·) is the exponential operation in the log-likelihood function, representing the similarity between embeddings. Ω is a very small constant to avoid division by zero.

[0131] 4) Using a Transformer-based semantic feature fusion module to aggregate local features and global features.

[0132] The process of step 4) is as follows:

[0133] 4-1) After passing through a three-layer double convolutional network, the embedding of each layer of the drug node and the target node is formed into a new embedding feature matrix by splicing is expressed as:

[0134]

[0135] Next, layer feature fusion is performed, and each layer embedding feature is combined by weighting to obtain a feature matrix is expressed as:

[0136]

[0137] The calculation formula is:

[0138]

[0139] where w l is the learnable weight parameter of the lth layer.

[0140] 4-2) The Transformer-based semantic feature aggregation module uses a self-attention mechanism to fuse the feature vectors of drugs and targets. The attention aggregation operation for drugs obtains drug features The calculation formula is as follows:

[0141]

[0142] where Q d ,K d ,V d are vectors of drug features mapped to the query, key, and value spaces through different linear transformations, and the calculation formula is as follows:

[0143] Q d = W Q · X d (53)

[0144] K d = W K · X d (54)

[0145] V d = W V · X d (55)

[0146] where W Q , W K , W V are trainable parameter matrices.

[0147] Similarly, the target feature

[0148] 5) The final node features are input into a multi-layer perception to predict drug-target interactions.

[0149] The process of step 5) is as follows:

[0150] 5-1) The drug and target node features and are input into a multi-layer perception MLP, which learns the feature interaction between drug and target nodes using the MLP to obtain interaction prediction probability The MLP learning prediction process is as follows:

[0151]

[0152] where are the feature vectors of drug i and target j respectively, and σ(·) is a nonlinear activation function LeakyReLu. W1, W2… W L are the weight matrices of the lth layer, and b1, b2… b L are the bias vectors of the lth layer, all of which are trainable parameter matrices.

[0153] 5-2) A binary cross-entropy loss function is applied to calculate the node classification loss during model training, and the binary cross-entropy loss function is defined as follows:

[0154]

[0155] where N is the total number of prediction samples, y i represents the actual label, and y represents the prediction probability of the model.

[0156] 5-3) The final loss function of model training is composed of three parts, as follows:

[0157]

[0158] where α and β are trade-off hyperparameters, represents L2 regularization on all trainable model parameters to prevent model overfitting.

[0159] 6) Selecting part of negative samples from the negative sample pool according to the active learning query strategy BALD, updating the training set until the ratio of the number of positive samples to the number of negative samples in the training set reaches 1:1.

[0160] The process of step 6) is as follows:

[0161] 6-1) The negative sample selection of this method uses pool-based active learning. Given a training set and a negative sample pool The pool-based active learning algorithm first selects D l based on D (0) . Next, in each round t = 1, 2, …, T, the algorithm selects n samples from D neg according to the query strategy. Then the n samples are updated from D neg to D l , and a new model f l is learned based on the updated D (t) .

[0162] 6-2) A good query strategy can make the classifier sequence f (0) , f (1) , f (2) , …, f (T) perform well on the test set . This method implements an active learning strategy BALD (Bayesian Active Learning by Disagreement) based on Bayesian model, which evaluates uncertainty by multiple Dropout and selects the most uncertain samples. The uncertainty index is based on the difference between two kinds of entropy, and the calculation formula is as follows:

[0163] Uncertainty(x) = H2(x) - H1(x) (59)

[0164] Where x represents a drug-target pair, H1(x) and H2(x) are the average entropy and the multiple Dropout entropy respectively, and the calculation formula is:

[0165]

[0166] H2(x) = -probs(x)·log(1-probs(x)) (61)

[0167] Where probs(x) is the cumulative probability of x being a positive sample after multiple Dropout, is the average probability of x being a positive sample, and has:

[0168]

[0169] where n drop is the total number of Dropouts.

[0170] Examples

[0171] [1] Li Y, Qiao G, Wang K, et al. Drug-target interaction predication via multi-channel graph neural networks [J]. Briefings in Bioinformatics, 2022, 23(1): bbab346.

[0172] Reference [1] proposes a method for predicting drug-target interactions based on multi-channel graph convolutional networks and graph attention. This method uses two independent graph attention networks to learn different interactions between nodes in topological graphs and feature graphs of different strengths, and then uses a graph convolutional network with a shared weight matrix to learn the common information of the two graphs, thereby combining topological structure and semantic features to improve the representation learning ability of predicting new drug-target interactions.

[0173] [2] Zeng X, Zhu S, Lu W, et al. Target identification among known drugs by deep learning from heterogeneous networks [J]. Chemical Science, 2020, 11(7): 1775-1797.

[0174] Reference [2] discloses a method for predicting drug-target interactions based on heterogeneous graph neural networks (HGNN). This method includes two key steps: (1) applying a deep neural network algorithm to network embedding, which embeds each vertex in the network into a low-dimensional vector space; (2) due to the lack of publicly available negative samples, the authors use a PU matrix completion algorithm as a projection scheme for vector spaces used to predict new drug-target interactions.

[0175] [3] Li J, Wang J, Lv H, et al. IMCHGAN: inductive matrix completion with heterogeneous graph attention networks for drug-target interactions prediction [J]. IEEE / ACM Transactions on Computational Biology and Bioinformatics, 2021, 19(2): 655-665.

[0176] Document [3] proposes a method that combines attention mechanism with inductive matrix completion (IMC) for predicting drug-target interactions. This method extracts feature representations of drugs and targets from the heterogeneous network of drug-target interactions through a two-level attention mechanism, and then inputs the extracted features into the inductive matrix completion model to generate the prediction results of drug-target interactions.

[0177] [4] Zhou D, Xu Z, Li W T, et al. MultiDTI: drug-target interaction prediction based on multi-modal representation learning to bridge the gap between new chemical entities and known heterogeneous network [J]. Bioinformatics, 2021, 37(23): 4485-4492.

[0178] Document [4] proposes a method for predicting drug-target interactions based on multi-modal heterogeneous network. First, the method inputs drugs outside the heterogeneous network into the trained model. Second, it uses a sequence feature extractor to represent the drugs according to their sequences. Third, it uses a multi-layer perceptron to map the sequence representation of the drugs to a trained common space. Finally, it uses a prediction layer based on the common space to predict the targets of the drugs.

[0179] To verify the prediction accuracy of the method of the present application, experiments were conducted on the benchmark dataset. The performance of the method of the present application was evaluated on five indicators: area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPR), F1 score (F1_score), recall (Recall), and precision (Precision), and compared with the three documents mentioned above. The comparison results are shown in Table 1:

[0180] Table 1 Comparison of the present method with other methods in various performance indicators

[0181]

[0182] Conclusion: Compared with other methods for predicting drug-target interactions, the present method is optimal in all performance indicators, which indicates that the present method has good reliability and high accuracy in predicting potential drug-target interactions.

Claims

1. A drug-target interaction prediction method based on hypergraph and active learning, characterized in that, Comprising the following steps: 1) Constructing drug-target adjacency matrix using known drug-target associations, drug-drug interactions and target-target interactions; 2) Constructing drug hypergraph and target hypergraph according to the known drug-target association matrix; 3) Inputting the initial features of drugs and targets into the dual convolution network composed of graph convolution branch and hypergraph convolution branch, and introducing a contrast learning mechanism to optimize the representation; The specific process is as follows: 3-1) Assign initial embedding vectors to drugs and targets, defined as and where h denotes the embedding dimension, the initial embedding vectors of drugs and targets are input into the bilinear convolution network, first, the constructed drug-target adjacency matrix is loaded into the network containing three layers of graph convolution, the global dependency is captured through the graph convolution network, and the embedding vectors realize message passing through the aggregation of features of neighbor nodes in each layer, and the propagation process in the graph convolution layer is as follows: wherein, is the L2 norm normalization, is the non-linear activation function LeakyReLu, is the graph convolution first l+ 1drug / target polymeric embedding of the propagation layer, is the drug / target polymeric embedding of the l layer, is the normalized adjacency matrix, calculated as: wherein denotes the adjacency matrix of the graph G, and ; 3-2) Hypergraph allows a hyperedge to connect multiple nodes, and the representation of multi-element relationship can utilize the high-order relationship of all nodes connected by each hyperedge. Hypergraph convolution can capture more rich and complex structural information, and through the multi-layer convolution propagation mechanism, local high-order features are obtained. The propagation process of drug embedding in hypergraph convolution layer is as follows: where, is the L2 norm normalization, is the non-linear activation function LeakyReLu, is the hypergraph convolution first l+ 1 propagation layer drug embedding, is the first l layer drug embedding, is the drug hypergraph degree matrix of the vertices in the drug hypergraph, where , is the drug hypergraph degree matrix of the hyperedges in the drug hypergraph, where , Similarly, the propagation process of target embedding in hypergraph convolution layer is as follows: 3-3) The intra-layer feature fusion is realized by weighted sum of graph convolution embedding and hypergraph convolution embedding in each layer, and the calculation method is as follows: wherein, and are the updated embeddings of drug / target at the bilinear convolution network layer l layer, and are the updated embeddings of drug / target at the graph convolution layer l layer, and are the updated embeddings of drug / target at the hypergraph convolution layer l layer, is a weight hyperparameter; 3-4) The representation consistency between graph convolution path and hypergraph convolution path is optimized by introducing a contrast learning mechanism, wherein the graph convolution embedding and hypergraph convolution embedding from the same drug or target constitute a positive sample pair, and all combinations of graph convolution embedding and hypergraph convolution embedding from different drugs or targets constitute a negative sample pair. The contrast learning mechanism is realized by minimizing the distance between positive sample pairs and maximizing the distance between negative sample pairs, and the objective function is as follows: wherein, is a temperature hyperparameter that controls the degree of smoothing of the contrastive loss, exp(⋅) is the exponential operation in the log-likelihood function, and is a very small constant to avoid division by zero; 4) Using a semantic feature fusion module based on Transformer to aggregate local features and global features; 5) Inputting the final node features into a multi-layer perceptron to predict drug-target interactions; 6) Selecting part of the negative samples from the negative sample pool according to the active learning query strategy BALD, updating the training set, and until the ratio of the number of positive samples to the number of negative samples in the training set reaches 1:

1.

2. The method of claim 1, wherein, The step 1) specifically comprises the following steps: 1-1) Collecting datasets on drugs and target interaction prediction studies, including 2214 drugs, and 1968 targets, and drug-target known association matrix , drug-drug interaction matrix , and target-target interaction matrix , where d and t denote drugs, targets, d-d and t-t denote drug-drug and target-target associations, respectively N d and N t represent the number of drugs and targets, respectively; 1-2) Constructing drug-target adjacency matrix may be defined as: 。 3. The method of claim 2, wherein the method is characterized by, The specific process of the step 2) is as follows: 2-1) Constructing drug supergraph Based on two definitions, Definition 1: In drug-target bipartite graph, if there is a sequence of adjacent vertices between drug i and drug j , and the number of drugs in the path is less than k, then drug i ( drug j ) is the k-order reachable neighbor of drug j ( drug i ); Definition 2: In drug-target bipartite graph, if there is a direct interaction between drug j and target k , and drug i is the k-order reachable neighbor of drug j , then target k ( drug i ) is the k-order reachable neighbor of drug i (target k ) . For drug i The k-order reachable drug set is represented as Mathematically, a hypergraph can be defined on a family of sets, where each set represents a hyperedge. Therefore, here a hyperedge can be constructed by the k-order reachable target set of an item, and then a high-order hyperedge group can be constructed according to the k-order reachable rule between drugs, which can be represented as: Further to the kth order reachable drug matrix for drugs may be expressed as: wherein To compute the k-th power of a given matrix M, is the drug-target association matrix, then the hyperedge association matrix constructed from the k-order reachability rule between drugs can be expressed as: Finally, the hypergraph matrix of drugs is represented as: 2-2) Similarly, for a target hypergraph, the k-order reachable targets and k-order reachable drugs of a target can be defined symmetrically, and the k-order reachable target matrix of a target can be defined as follows: may be represented as: Hyperedge group association matrix constructed by k-order reachability rule between targets may be expressed as: The hypergraph matrix of the target is represented as: 。 4. The method of claim 3, wherein the method is characterized by, The step 4) comprises the following steps: 4-1) After passing through a three-layer double-channel convolutional network, the embedding of each layer of the drug node and the target node is formed into a new embedding feature matrix by splicing , which is expressed as: Next, layer feature fusion is performed, and each layer embedding feature is weighted and combined to obtain a feature matrix is expressed as: The calculation formula is as follows: wherein, is the first l layer's learnable weight parameters; 4-2) Transformer-based semantic feature aggregation module, which uses self-attention mechanism to fuse the feature vectors of drugs and targets. The attention aggregation operation on drugs obtains drug features The calculation formula is as follows: wherein, The drug features are mapped to vectors in the space of queries, keys, and values by different linear transformations, computed as follows: wherein, is a trainable parameter matrix, By analogy, the target signature .

5. The method of claim 4, wherein the method is characterized by, The step 5) comprises the following steps: 5-1) and The input is fed into a multilayer perceptron (MLP), which uses the MLP to learn the feature interactions between drug and target nodes to obtain the interaction prediction probability. The MLP learning and prediction process is as follows: wherein, are the feature vectors of the drug i and target j , is a non-linear activation function LeakyReLu, , … is the weight matrix of the l th layer, , … is the bias vector of the l th layer, are trainable parameter matrices. 5-2) applying a binary cross-entropy loss function to calculate the node classification loss during the model training process, the binary cross-entropy loss function is defined as follows: where N is the total number of prediction samples, denotes the actual label, denotes the predicted probability of the model; 5-3) The final loss function of model training is composed of three parts, as follows: wherein, and are trade-off hyperparameters, denotes a regularization term on all trainable model parameters. to prevent overfitting of the model.

6. The method of claim 5, wherein the method is characterized by, The step 6) comprises the following steps: 6-1) Negative sample selection uses pool-based active learning, given a training set and a pool of negative samples The pool-based active learning algorithm first trains a model based on the training set, resulting in a model Next, in each round t= 1,2, … , T of the algorithm, it selects samples from n the pool based on a query strategy, then updates n these samples from to and learns a new model based on the updated ​ 6-2) Active learning strategy BALD (Bayesian Active Learning by Disagreement) based on Bayesian model, the uncertainty is evaluated by multiple Dropout, and the most uncertain sample is selected. The uncertainty index is based on the difference between two kinds of entropy, and the calculation formula is as follows: wherein, x denotes a drug-target pair, and are the average entropy and the multiple Dropout entropy, respectively, calculated as: where, is the probability of a positive sample after multiple Dropout x is the cumulative probability of a positive sample, is the average probability of a positive sample, where: wherein, is the total number of Dropouts.

Images