Drug-target interaction prediction method based on hypergraph and active learning

Abstract

The invention discloses a hypergraph and active learning-based drug-target interaction prediction method. The hypergraph and active learning-based drug-target interaction prediction method roughly comprises the following steps of constructing a drug-target adjacency matrix by using drug-target known association, drug-drug interaction and target-target interaction; constructing a drug hypergraph and a target hypergraph according to the drug-target known incidence matrix; the initial features of the drug and the target are respectively embedded and input into a dual-path information transfer network, and a contrast learning mechanism is introduced to optimize characterization; using a semantic feature fusion module based on Transform to aggregate local features and global features; inputting the final node features into a multi-layer perceptron to predict drug-target interaction; and selecting part of negative samples from the negative sample pool according to a query strategy, and updating the training set. According to the method, lightweight design is achieved on the model level, a negative sample selection strategy based on active learning is provided, negative samples in a training set are optimized through continuous iteration, and the model performance is improved.

Description

Technical Field

[0001] The present invention 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 Art

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

[0003] With the advancement of computing technology, DTI prediction methods have gradually formed three major technical routes: ligand-based, structure-based, and deep learning-based. Early DTI predictions mainly adopted ligand-based and structure-based strategies. The ligand-based method is based on the principle of chemical similarity and believes that molecules with similar structures may act on the same target. For example, the SEA (Similarity Ensemble Approach) platform uses two-dimensional structure comparison technology, while SwissTargetPrediction combines two-dimensional and three-dimensional feature analysis. However, such methods often perform poorly when dealing with new molecular structures, and the prediction effect is not ideal when target data is scarce. In contrast, structure-based methods focus on protein three-dimensional conformation analysis. The possibility of interaction is predicted by evaluating the degree of match between drug molecules and protein binding sites. Typical methods include reverse molecular docking and reverse pharmacophore matching. However, structure-based methods face many challenges: not only high-quality protein structure data but also powerful computing power is required.

[0004] The rise of deep learning has brought new breakthroughs in DTI prediction. This method breaks the traditional prediction framework and integrates chemical information, genomic data, and interaction pairs into the computational model. Compared with traditional methods, deep learning can directly process the original structural information of molecules and proteins, and automatically extract features through architectures such as neural networks, graph convolutions, or attention mechanisms. This data-driven approach gets rid of the reliance on manually designed features and has the ability to process massive amounts of data. Despite this, deep learning methods still have the following difficulties: i) In the DTI prediction task, there is uncertainty in the labels of unverified samples. Traditional methods simply treat all unverified samples as negative samples. This processing method may introduce false negative samples and reduce model performance; ii) Ordinary graph neural networks can only capture second-order relationships, and high-order relationships are not captured enough. They rely on neighbors to propagate information, and the propagation range is limited, resulting in a decrease in prediction accuracy. Summary of the invention

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

[0006] The technical solution of the present invention to solve the above problem is: a drug-target interaction prediction method based on hypergraph and active learning, comprising the following steps:

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

[0008] 2) construct drug hypergraph and target hypergraph based on the known drug-target association matrix;

[0009] 3) The initial feature embeddings of drugs and targets are input into the dual-path information transfer network respectively, and a contrastive learning mechanism is introduced to optimize the representation;

[0010] 4) Use the Transformer-based semantic feature fusion module to aggregate local features and global features;

[0011] 5) The final node features are input into a multi-layer perceptron to predict drug-target interactions;

[0012] 6) Select some negative samples from the negative sample pool according to the query strategy and update the training set until the number of positive and negative samples in the training set reaches 1:1.

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

[0014] 1-1) Collect data sets for drug-target interaction prediction research, including 2214 drugs, 1968 targets, and drug-target known association matrix Drug-Drug Interaction Matrix and target-target interaction matrix Where d and t represent drug and target, respectively, dd and tt represent drug-drug and target-target associations, respectively, and N d and N t represent the number of drugs and the number of targets, respectively;

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

[0016]

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

[0018] 2-1) Constructing the drug hypergraph is based on two definitions: Definition 1: In the drug-target bipartite graph, if drug iand drug j There is a sequence of adjacent vertices (i.e., a path) between them, and the number of drugs on the path is less than k, then it is called drug i (drug j ) is drug j (drug i ) has k-order reachable neighbors. Definition 2: In a drug-target bipartite graph, if drug j and target k There is a direct interaction, and drug i For drug j The k-order reachable neighbors of target k (drug i ) is drug i (target k )’s k-order reachable neighbors.

[0019] For drug i , and its k-order reachable drug set is expressed as Mathematically, a hypergraph can be defined on a family of sets, where each set represents a hyperedge. Therefore, a hyperedge can be constructed here through the set of k-order reachable targets of a certain item, and then a high-order hyperedge group can be constructed according to the k-order reachability rule between drugs, which can be expressed as:

[0020]

[0021] Then for the k-order reachable drug matrix of the drug It can be expressed as:

[0022]

[0023] Where power(M,k) is to calculate the kth power of the given matrix M. is the drug-target association matrix. Then the hyperedge group association matrix constructed by the k-order reachability rule between drugs is It can be expressed as:

[0024]

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

[0026]

[0027] 2-2) Similarly, for the target hypergraph, the k-th order reachable targets and k-th order reachable drugs of the target can be defined symmetrically in a similar way. Specifically, the k-th order reachable target matrix of the target is It can be expressed as:

[0028]

[0029] Hyperedge group incidence matrix constructed by k-order reachability rules between targets It can be expressed as:

[0030]

[0031] Then the target hypergraph matrix H t It is expressed as:

[0032]

[0033] In 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 the 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 target are input into the bilinear convolutional network respectively. First, the constructed drug-target adjacency matrix M is loaded into the network containing three layers of graph convolution. The global dependency is captured by the graph convolutional network, and the embedding vector realizes message passing through the feature aggregation of neighbor nodes in each layer. The propagation process in the graph convolution layer is as follows:

[0035]

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

[0037]

[0038] Among them, D represents the vertex degree matrix of the adjacency matrix M,

[0039] 3-2) Hypergraphs allow a hyperedge to connect multiple nodes, and the representation of multi-dimensional relationships can utilize the high-order relationships of all nodes connected by each hyperedge. Hypergraph convolution can capture richer and more complex structural information, and through the multi-layer convolution propagation mechanism, it can spread information more widely in the network and support multi-point collaborative message propagation. The original embeddings of drugs and targets are input into the three-layer hypergraph convolution module respectively. The propagation process of drug embedding in the hypergraph convolution layer is:

[0040]

[0041] Among them, f(·) is L2 norm normalization, and σ(·) is the 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. Drug Supermap H d The degree matrix of the vertices in is Drug Supermap H d The degree matrix of the super edge is

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

[0043]

[0044] 3-3) In order to combine local features with global features, we perform intra-layer feature fusion. Specifically, in each layer, the graph convolution-hypergraph convolution embedding is weighted summed, and the calculation formula is:

[0045]

[0046] in, and The updated embedding of drug / target fused in layer l of the bilinear convolutional network, and is the embedding of the drug / target after graph convolution update at layer l, and is the embedding of the drug / target after the hypergraph convolution update at layer l. λ is the weight hyperparameter.

[0047] 3-4) Finally, a contrastive learning mechanism is introduced to optimize the representation consistency between the two paths. We regard the two embeddings from the graph convolution and hypergraph convolution of the same drug / target as positive sample pairs, and generate negative samples through all possible combinations of graph convolution and hypergraph convolution embeddings of different drugs / targets. The goal is to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs. The contrastive learning objective function is expressed as:

[0048]

[0049] 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, which represents the similarity between embeddings. Ω is a small constant used to avoid the denominator being zero.

[0050] In 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 bilinear convolutional network, each layer of embedding of drug nodes and target nodes is concatenated to form a new embedding feature matrix It is expressed as:

[0052]

[0053] Next, perform layer feature fusion and perform weighted combination of each layer of embedded features to obtain the feature matrix It is expressed as:

[0054]

[0055] The calculation formula is:

[0056]

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

[0058] 4-2) The Transformer-based semantic feature aggregation module uses the 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:

[0059]

[0060] Among them, Q d ,K d ,V d The drug features are mapped to vectors in the query, key, and value space after different linear transformations. The calculation formula is as follows:

[0061] Q d =W Q ·X d (twenty two)

[0062] K d =W K ·X d (twenty three)

[0063] V d =W V ·X d (twenty four)

[0064] Among them, W Q ,W K ,W V is a trainable parameter matrix.

[0065] Similarly, the target features can be obtained

[0066] In 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) and Input into the multi-layer perceptron (MLP for short), use MLP to learn the feature interaction between drugs and target nodes, and obtain the interaction prediction probability The MLP learning prediction process is as follows:

[0068]

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

[0070] 5-2) During model training, the binary cross entropy loss function is applied to calculate the node classification loss. The binary cross entropy loss function The definition is as follows:

[0071]

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

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

[0074]

[0075] Among them, α and β are trade-off hyperparameters, Represents L2 regularization on all trainable model parameters to prevent model overfitting.

[0076] In 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 this method uses pool-based active learning. Given a training set And negative sample pool The pool-based active learning algorithm first uses D train Train the model and get model M (0) Next, in each round t=1,2,…,T, the algorithm selects from D according to the query strategy. negThen select n samples from D neg Update to D train , and based on the updated D train Learn a new model M (t) .

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

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

[0080] Where x represents the 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] Among them, probs(x) is the cumulative probability that x is a positive sample after multiple Dropouts. is the average probability that x is a positive sample, and we have:

[0084]

[0085] Among them, n drop is the total number of Dropouts.

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

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

[0088] Figure 1 It is the overall flow chart of the present invention. DETAILED DESCRIPTION

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

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

[0091] 1) Construct a drug-target adjacency matrix using known drug-target associations, drug-drug interactions, and target-target interactions.

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

[0093] 1-1) Collect data sets for drug-target interaction prediction research, including 2214 drugs, 1968 targets, and drug-target known association matrix Drug-Drug Interaction Matrix and target-target interaction matrix Where d and t represent drug and target, respectively, dd and tt represent drug-drug and target-target associations, respectively, and N d and N t represent the number of drugs and the number of targets, respectively;

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

[0095]

[0096] 2) Construct drug hypergraph and target hypergraph based on the known drug-target association matrix.

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

[0098] 2-1) Constructing the drug hypergraph is based on two definitions: Definition 1: In the drug-target bipartite graph, if drug i and drug j There is a sequence of adjacent vertices (i.e., a path) between them, and the number of drugs on the path is less than k, then it is called drug i (drug j ) is drug j (drug i ) has k-order reachable neighbors. Definition 2: In a drug-target bipartite graph, if drug j and target k There is a direct interaction, and drug i For drug j The k-order reachable neighbors of target k (drug i ) is drug i (target k )’s k-order reachable neighbors.

[0099] For drug i , and its k-order reachable drug set is expressed as Mathematically, a hypergraph can be defined on a family of sets, where each set represents a hyperedge. Therefore, a hyperedge can be constructed here through the set of k-order reachable targets of a certain item, and then a higher-order hyperedge group can be constructed according to the k-order reachability rule between drugs, which can be expressed as:

[0100]

[0101] Then for the k-order reachable drug matrix of the drug It can be expressed as:

[0102]

[0103] Where power(M,k) is to calculate the kth power of the given matrix M. is the drug-target association matrix. Then the hyperedge group association matrix constructed by the k-order reachability rule between drugs is It can be expressed as:

[0104]

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

[0106]

[0107] 2-2) Similarly, for the target hypergraph, the k-th order reachable targets and k-th order reachable drugs of the target can be defined symmetrically in a similar way. Specifically, the k-th order reachable target matrix of the target is It can be expressed as:

[0108]

[0109] Hyperedge group incidence matrix constructed by k-order reachability rules between targets It can be expressed as:

[0110]

[0111] Then the target hypergraph matrix H t It is expressed as:

[0112]

[0113] 3) The initial features of drugs and targets are embedded into the bilinear convolutional network respectively, and a contrastive learning mechanism is introduced to optimize the representation.

[0114] Step 3) comprises the following steps:

[0115] 3-1) Assign the 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 target are input into the bilinear convolutional network respectively. First, the constructed drug-target adjacency matrix M is loaded into the network containing three layers of graph convolution. The global dependency is captured by the graph convolutional network, and the embedding vector realizes message passing through the feature aggregation of neighbor nodes in each layer. The propagation process in the graph convolution layer is as follows:

[0116]

[0117] Among them, f(·) is L2 norm normalization, and σ(·) is the nonlinear activation function LeakyReLu. is the drug / target aggregate embedding of the l+1th propagation layer of the graph convolution, X l The drug / target aggregate embedding for layer l. To normalize the adjacency matrix, the calculation formula is:

[0118]

[0119] Among them, D represents the vertex degree matrix of the adjacency matrix M,

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

[0121]

[0122] Among them, f(·) is L2 norm normalization, and σ(·) is the 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. Drug Supermap H d The degree matrix of the vertices in is Drug Supermap H d The degree matrix of the super edge is

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

[0124]

[0125] 3-3) In order to combine local features with global features, we perform intra-layer feature fusion. Specifically, in each layer, the graph convolution-hypergraph convolution embedding is weighted summed, and the calculation formula is:

[0126]

[0127] in, and The updated embedding of drug / target fused in layer l of the bilinear convolutional network, and is the embedding of the drug / target after graph convolution update at layer l, and is the embedding of the drug / target after the hypergraph convolution update at layer l. λ is the weight hyperparameter.

[0128] 3-4) Finally, a contrastive learning mechanism is introduced to optimize the representation consistency between the two paths. We regard the two embeddings from the graph convolution and hypergraph convolution of the same drug / target as positive sample pairs, and generate negative samples through all possible combinations of graph convolution and hypergraph convolution embeddings of different drugs / targets. The goal is to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs. The contrastive learning objective function is expressed as:

[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, which represents the similarity between embeddings. Ω is a small constant used to avoid the denominator being zero.

[0131] 4) Use the 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 the three-layer bilinear convolutional network, each layer of embedding of drug nodes and target nodes is concatenated to form a new embedding feature matrix It is expressed as:

[0134]

[0135] Next, perform layer feature fusion and perform weighted combination of each layer of embedded features to obtain the feature matrix It is expressed as:

[0136]

[0137] The calculation formula is:

[0138]

[0139] Among them, w l is the learnable weight parameter of the lth layer.

[0140] 4-2) The Transformer-based semantic feature aggregation module uses the 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:

[0141]

[0142] Among them, Q d ,K d ,V d The drug features are mapped to vectors in the query, key, and value space after different linear transformations. The calculation formula is as follows:

[0143]

[0144] Among them, W Q ,W K ,W V is a trainable parameter matrix.

[0145] Similarly, the target features can be obtained

[0146] 5) The final node features are input into a multi-layer perceptron to predict drug-target interactions.

[0147] The process of step 5) is as follows:

[0148] 5-1) and Input into the multi-layer perceptron (MLP for short), use MLP to learn the feature interaction between drugs and target nodes, and obtain the interaction prediction probability The MLP learning prediction process is as follows:

[0149]

[0150] in, are the feature vectors of drug i and target j respectively, and σ(·) is the nonlinear activation function LeakyReLu. W1,W2…W L is the weight matrix of the lth layer, b1,b2…b L is the bias vector of the lth layer, which are all trainable parameter matrices.

[0151] 5-2) During model training, the binary cross entropy loss function is applied to calculate the node classification loss. The binary cross entropy loss function The definition is as follows:

[0152]

[0153] Among them, N is the total number of prediction samples, y i represents the actual label, represents the predicted probability of the model.

[0154] 5-3) The final loss function of model training consists of three parts, as follows:

[0155]

[0156] Among them, α and β are trade-off hyperparameters, Represents L2 regularization on all trainable model parameters to prevent model overfitting.

[0157] 6) Select some negative samples from the negative sample pool according to the query strategy and update the training set until the ratio of the number of positive and negative samples in the training set reaches 1:1.

[0158] The process of step 6) is as follows:

[0159] 6-1) The negative sample selection of this method uses pool-based active learning. Given a training set And negative sample pool The pool-based active learning algorithm first uses D l Train the model and get the model f (0) Next, in each round t=1,2,…,T, the algorithm selects from D according to the query strategy. neg Then select n samples from D neg Update to D l , and based on the updated D l Learn a new model f (t) .

[0160] 6-2) A good query strategy can make the classifier sequence f (0) ,f (1) ,f (2) ,…,f (T) In the test set This method implements an active learning strategy based on the Bayesian model, BALD (Bayesian Active Learning by Disagreement), which evaluates uncertainty through multiple dropouts and selects the most uncertain samples. The uncertainty index is based on the difference between the two entropies, and the calculation formula is as follows:

[0161] Uncertainty(x)=H2(x)-H1(x) (59)

[0162] Where x represents the drug-target pair, H1(x) and H2(x) are the average entropy and multiple dropout entropy, respectively, and the calculation formula is:

[0163]

[0164] H2(x)=-probs(x)·log(1-probs(x)) (61)

[0165] Among them, probs(x) is the cumulative probability that x is a positive sample after multiple Dropouts. is the average probability that x is a positive sample, and we have:

[0166]

[0167] Among them, n drop is the total number of Dropouts.

[0168] Examples

[0169] [1]Li Y, Qiao G, Wang K, et al. Drug–target interaction prediction via multi-channel graph neural networks [J]. Briefings in Bioinformatics, 2022, 23(1):bbab346.

[0170] Reference [1] proposed 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 topological graphs and feature graph nodes 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 structures and semantic features to improve the representation learning ability for predicting new drug-target interactions.

[0171] [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.

[0172] Reference [2] discloses a method for predicting drug-target interactions based on heterogeneous graph neural networks (HGNN). The method consists of 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 the PU matrix completion algorithm as a projection scheme for the vector space used to predict novel drug-target interactions.

[0173] [3]Li J, Wang J, Lv H, et al. ICHGAN: 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.

[0174] Reference [3] proposed a method that combines attention mechanism with inductive matrix completion (IMC) to predict 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 prediction results of drug-target interactions.

[0175] [4] Zhou D, Xu Z, Li WT, 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.

[0176] Reference [4] proposed a method for predicting drug-target interactions based on a multimodal heterogeneous network. First, the method inputs drugs outside the heterogeneous network into the trained model. Second, a sequence feature extractor is used to characterize the drug based on its sequence. Next, a multilayer perceptron is used to map the sequence representation of the drug to the trained common space, and finally a prediction layer based on the common space is used to predict the drug's target.

[0177] In order to verify the prediction accuracy of the method of the present invention, experiments were conducted on the benchmark data set. The performance of the method of the present invention was evaluated on five indicators: the area under the receiver operating characteristic curve (AUC), the area under the precision-recall curve (AUPR), the F1 score (F1_score), the recall rate (Recall) and the precision (Precision). The method was compared with the above three documents. The comparison results are shown in Table 1:

[0178] Table 1 Comparison of performance indicators with other methods

[0179]

[0180] Conclusion: Compared with other related methods for predicting drug-target interactions, the method of the present invention achieves the best performance in all performance indicators, which shows that the method of the present invention 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: The following steps are involved: 1) Construct a drug-target adjacency matrix using known drug-target associations, drug-drug interactions, and target-target interactions; 2) construct drug hypergraph and target hypergraph based on the known drug-target association matrix; 3) The initial feature embeddings of drugs and targets are input into the bilinear convolutional network respectively, and a contrastive learning mechanism is introduced to optimize the representation; 4) Use the Transformer-based semantic feature fusion module to aggregate local features and global features; 5) The final node features are input into a multi-layer perceptron to predict drug-target interactions; 6) Select some negative samples from the negative sample pool according to the query strategy and update the training set until the ratio of the number of positive and negative samples in the training set reaches 1:

1.

2. The method for predicting drug-target interactions based on hypergraph and active learning according to claim 1, characterized in that: The step 1) specifically comprises the following steps: 1-1) Collect data sets for drug-target interaction prediction research, including 2214 drugs, 1968 targets, and drug-target known association matrix Drug-Drug Interaction Matrix and target-target interaction matrix Where d and t represent drug and target, respectively, dd and tt represent drug-drug and target-target associations, respectively, and N d and N t represent the number of drugs and the number of targets, respectively; 1-2) Constructing the drug-target adjacency matrix It can be defined as:

3. The method for predicting drug-target interactions based on hypergraph and active learning according to claim 2, characterized in that: The specific process of step 2) is as follows: 2-1) Constructing the drug hypergraph is based on two definitions: Definition 1: In the drug-target bipartite graph, if drug i and drug j There is a sequence of adjacent vertices (i.e., a path) between them, and the number of drugs on the path is less than k, then it is called drug i (drug j ) is drug j (drug i ) has k-order reachable neighbors. Definition 2: In a drug-target bipartite graph, if drug j and target k There is a direct interaction, and drug i For drug j The k-order reachable neighbors of target k (drug i ) is drug i (target k )’s k-order reachable neighbors. For drug i , and its k-order reachable drug set is expressed as Mathematically, a hypergraph can be defined on a family of sets, where each set represents a hyperedge. Therefore, a hyperedge can be constructed here through the set of k-order reachable targets of a certain item, and then a higher-order hyperedge group can be constructed according to the k-order reachability rule between drugs, which can be expressed as: Then for the k-order reachable drug matrix of the drug It can be expressed as: Where power(M,k) is to calculate the kth power of the given matrix M. is the drug-target association matrix. Then the hyperedge group association matrix constructed by the k-order reachability rule between drugs is It can be expressed as: Finally, the hypergraph matrix of drugs It is expressed as: 2-2) Similarly, for the target hypergraph, the k-th order reachable targets and k-th order reachable drugs of the target can be defined symmetrically in a similar way. Specifically, the k-th order reachable target matrix of the target is It can be expressed as: Hyperedge group incidence matrix constructed by k-order reachability rules between targets It can be expressed as: Then the target hypergraph matrix H t It is expressed as:

4. The method for predicting drug-target interactions based on hypergraph and active learning according to claim 3, characterized in that: The specific process of step 3) is as follows: 3-1) Assign the 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 target are input into the bilinear convolutional network respectively. First, the constructed drug-target adjacency matrix M is loaded into the network containing three layers of graph convolution. The global dependency is captured by the graph convolutional network, and the embedding vector realizes message passing through the feature aggregation of neighbor nodes in each layer. The propagation process in the graph convolution layer is as follows: Among them, f(·) is L2 norm normalization, and σ(·) is the nonlinear activation function LeakyReLu. is the drug / target aggregate embedding of the l+1th propagation layer of the graph convolution, X l The drug / target aggregate embedding for layer l. To normalize the adjacency matrix, the calculation formula is: Among them, D represents the vertex degree matrix of the adjacency matrix M, 3-2) Hypergraphs allow a hyperedge to connect multiple nodes, and the representation of multi-dimensional relationships can utilize the high-order relationships of all nodes connected by each hyperedge. Hypergraph convolution can capture richer and more complex structural information, and obtain local high-order features through a multi-layer convolution propagation mechanism. The original embeddings of drugs and targets are input into the three-layer hypergraph convolution module respectively. The propagation process of drug embedding in the hypergraph convolution layer is: Among them, f(·) is L2 norm normalization, and σ(·) is the 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. Drug Supermap H d The degree matrix of the vertices in is Drug Supermap H d The degree matrix of the super edge is Similarly, the propagation process of the target embedding in the hypergraph convolutional layer is: 3-3) In order to combine local features with global features, we perform intra-layer feature fusion. Specifically, in each layer, the graph convolution-hypergraph convolution embedding is weighted summed, and the calculation formula is: in, and The updated embedding of drug / target fused in layer l of the bilinear convolutional network, and is the embedding of the drug / target after graph convolution update at layer l, and is the embedding of the drug / target after the hypergraph convolution update at layer l. λ is the weight hyperparameter. 3-4) Finally, a contrastive learning mechanism is introduced to optimize the representation consistency between the two paths. We regard the two embeddings from the graph convolution and hypergraph convolution of the same drug / target as positive sample pairs, and generate negative samples through all possible combinations of graph convolution and hypergraph convolution embeddings of different drugs / targets. The goal is to minimize the distance between positive sample pairs and maximize the distance between negative sample pairs. The contrastive learning objective function is expressed as: 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, which represents the similarity between embeddings. Ω is a small constant used to avoid the denominator being zero.

5. The method for predicting drug-target interactions based on hypergraph and active learning according to claim 4, characterized in that: The step 4) comprises the following steps: 4-1) After passing through the three-layer dual-channel convolutional network, each layer of embedding of drug nodes and target nodes is concatenated to form a new embedding feature matrix It is expressed as: Next, perform layer feature fusion and perform weighted combination of each layer of embedded features to obtain the feature matrix It is expressed as: The calculation formula is: Among them, w l is the learnable weight parameter of the lth layer. 4-2) The Transformer-based semantic feature aggregation module uses the 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: Among them, Q d ,K d ,V d The drug features are mapped to vectors in the query, key, and value space after different linear transformations. The calculation formula is as follows: Q d =W Q ·X d (22) K d =W K ·X d (23) V d =W V ·X d (24) Among them, W Q ,W K ,W V is a trainable parameter matrix. Similarly, the target features can be obtained 6. The method for predicting drug-target interactions based on hypergraph and active learning according to claim 5, characterized in that: The step 5) comprises the following steps: 5-1) and Input into the multi-layer perceptron (MLP for short), use MLP to learn the feature interaction between drugs and target nodes, and obtain the interaction prediction probability The MLP learning prediction process is as follows: in, are the feature vectors of drug i and target j respectively, and σ(·) is the nonlinear activation function LeakyReLu. W1,W2…W L is the weight matrix of the lth layer, b1,b2…b L is the bias vector of the lth layer, which are all trainable parameter matrices. 5-2) During model training, the binary cross entropy loss function is applied to calculate the node classification loss. The binary cross entropy loss function The definition is as follows: Among them, N is the total number of prediction samples, y i represents the actual label, represents the predicted probability of the model. 5-3) The final loss function of model training consists of three parts, as follows: Among them, α and β are trade-off hyperparameters, Represents L2 regularization on all trainable model parameters to prevent model overfitting.

7. The method for predicting drug-target interactions based on hypergraph and active learning according to claim 6, characterized in that: The step 6) comprises the following steps: 6-1) This method uses pool-based active learning to select negative samples. Given a training set And negative sample pool The pool-based active learning algorithm first uses D train Train the model and get model M (0) Next, in each round t=1,2,…,T, the algorithm selects from D according to the query strategy. neg Then select n samples from D neg Update to D train , and based on the updated D train Learn a new model M (t) . 6-2) A good query strategy can make the model sequence M (0) ,M (1) ,M (2) ,…,M (T) In the test set This method implements an active learning strategy based on the Bayesian model, BALD (Bayesian Active Learning by Disagreement), which evaluates uncertainty through multiple dropouts and selects the most uncertain samples. The uncertainty index is based on the difference between the two entropies, and the calculation formula is as follows: Uncertainty(x)=H2(x)-H1(x)(28) Where x represents the drug-target pair, H1(x) and H2(x) are the average entropy and multiple dropout entropy, respectively, and the calculation formula is: H2(x)=-probs(x)·log(1-probs(x))(30) Among them, probs(x) is the cumulative probability that x is a positive sample after multiple Dropouts. is the average probability that x is a positive sample, and we have: Among them, n drop is the total number of Dropouts.

Images