Drug-target interaction prediction method, device and medium based on multi-view feature fusion

Abstract

The application discloses a drug-target interaction prediction method and device based on multi-view feature fusion and a medium, and comprises the following steps: (1) extracting drug features: first, generating a drug molecule graph, extracting molecular fragments to construct multiple subgraph structures, and then using a graph encoder to encode to obtain drug features; (2) extracting target features: first, extracting the protein pocket of the target, dividing the protein pocket into multiple subpockets to obtain multiple subgraph structures, and then using a graph encoder to encode to obtain the pocket features of the target; in addition, a self-attention module is used to extract the sequence features of the target, and a cross-attention is used to fuse the pocket features and the sequence features to obtain the target features; (3) using a bilinear attention to fuse the drug features and the target features, and calculating the interaction probability between the drug and the target. The application can effectively mine the biochemical attribute features in the drug molecules and the target protein structures, and improve the prediction accuracy of the drug-target interaction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bioinformatics, and relates to a drug-target interaction prediction method, device and medium based on multi-view feature fusion. BACKGROUND

[0002] Drugs play an irreplaceable important role in maintaining human health and treating diseases. In the treatment of diseases, whether it is antibiotics for the control of bacterial infections, or antiviral drugs for the inhibition of virus spread, drugs serve as a key tool to affect the human immune system against pathogens. For example, imatinib is a drug that effectively treats chronic myeloid leukemia, which works by targeting and inhibiting the activity of a specific enzyme in tumor cells, thereby preventing the progression of cancer. In addition, another type of anticancer drug, PD-1 inhibitor, activates the patient's immune system to enable it to recognize and attack tumor cells, and has shown wide application value in the treatment of various cancers such as melanoma and lung cancer. These examples fully demonstrate the importance of drugs in modern medicine and their profound impact on improving human health.

[0003] Drug development is a complex and lengthy process, involving multiple key steps from identifying drug targets to developing new drugs. First, the target is identified, second, potential lead compounds are screened from a large number of compounds through virtual screening technology, and finally, the binding ability and potential biological activity of these compounds to the target are evaluated using drug-target interaction prediction technology, thereby achieving the discovery and development of new drugs. Traditional virtual screening methods mainly rely on biochemical experiments, which screen a large number of compounds and verify one by one whether they interact with the target. The advantage of this method is that it can discover new molecules with biological activity, but its disadvantages are also obvious: low screening efficiency, high cost, and often unable to clearly determine the specific mechanism of action of the molecule and the target. To overcome the limitations of traditional methods, emerging technologies such as computer-aided drug design have been gradually introduced into the field of drug development, particularly in the task of drug-target interaction prediction. These technologies not only assist in virtual screening of candidate drugs efficiently, but also significantly improve the efficiency of drug discovery, while providing deeper insights into the mechanism of action of molecules and targets. According to the computational model for predicting drug-target pair interactions, there are three types:

[0004] (1) Computational model based on molecular docking. The drug-target interaction prediction model based on molecular docking uses computational simulation to predict the interaction between a drug molecule (ligand) and a specific target (usually a protein). The core of this method is to calculate the optimal spatial position and interaction mode of the drug and target binding to determine their affinity and binding mode.

[0005] (2) Ligand-based computational models. Ligand-based drug target interaction prediction models mainly rely on the structural information of known active ligands (i.e., small molecule compounds known to have interactions with the target), and predict the interactions of new ligands with the target by analyzing and comparing the common characteristics of these ligands. The core of such models is to construct a "pharmacophore" or feature map that can effectively represent drug activity, so as to predict potential drug-target binding in the absence of known targets or structural information.

[0006] (3) Deep learning-based computational models. Lee et al. used amino acid sequences and drug fingerprints extended connection fingerprints (ECFP4) as inputs, used convolutional neural networks on protein sequences of different lengths to capture local residue patterns, and proposed a DeepConvDTI model. Based on the assumption that the binding sites between drugs and targets appear in their specific substructures, Huang et al. proposed a DTI model MolTrans using the frequent subsequence algorithm (FCS) to decompose protein and drug molecules into common subsequences. Huang et al. proposed a DeepDTA model that applies a multi-layer CNN module to calculate the binding affinity between drug SMILES and protein sequences. Considering that molecular structure provides detailed information about how drugs and proteins bind at the atomic level, Torng et al. proposed a CPI-GNN model that converts each two-dimensional molecular graph into a graph with nodes composed of its atoms and edges composed of chemical bonds. Then, CNN and GNN are used to represent protein sequences and 2D drug molecular graphs, respectively, and then an attention mechanism is used to generate the final prediction.

[0007] In summary, when predicting potential drug-target interactions, it is crucial to effectively explore the biochemical attribute features hidden in the structures of drug molecules and proteins. However, existing computational models have not fully explored the characteristics of protein and drug molecular structures, so the interpretability of their prediction results still needs to be improved. SUMMARY

[0008] In view of the lack of sufficient characterization of protein structure information in current drug-target interaction prediction methods and the neglect of special substructure patterns (such as subfragments and functional groups) in molecules, the present application proposes a drug-target interaction prediction method, device and medium based on multi-view feature fusion, which can effectively explore the biochemical attribute features in drug molecules and protein structures, significantly improve the prediction accuracy of drug-target interactions, and the prediction performance is stable.

[0009] To achieve the above technical purposes, the technical scheme adopted by the present application is as follows:

[0010] A drug-target interaction prediction method based on multi-view feature fusion, comprising:

[0011] Step 1, extracting drug features and target features;

[0012] (1) Extracting drug features: first, generate the corresponding 2D molecular graph according to the SMILES sequence of the drug, and construct the graph structure of the drug according to the characteristics of each atom in the drug molecule and the chemical bond relationship between them; then extract the molecular fragments of the drug to obtain the subgraph structure corresponding to each molecular fragment; finally, use a GraphTransformer as a molecular graph encoder to encode the subgraph structure corresponding to each molecular fragment of the drug to obtain the drug features;

[0013] (2) Extracting target features: first, extract the protein pocket of the target, and construct the graph structure of the target according to the residues in the protein pocket and the chemical bonds between them; then divide the target into multiple sub-pockets to obtain the subgraph structure corresponding to each sub-pocket; then use a GraphTransformer as a pocket encoder to encode the subgraph structure corresponding to each sub-pocket of the target to obtain the pocket features of the target; in addition, use a self-attention module to extract the sequence features of the target from the one-dimensional amino acid sequence of the target; finally, use cross-attention to fuse the pocket features and sequence features of the target to obtain the target features;

[0014] Step 2, using a bilinear attention network as a classifier to deeply fuse the drug features and target features, and calculating the interaction probability between the drug and the target through a multi-layer perception.

[0015] Further, in the graph structure of the drug, each node represents an atom, and its feature vector includes the following attributes: the type of the atom, the degree of the atom, the formal charge of the atom, the number of free radical electrons of the atom, the hybridization state of the atom, the aromaticity of the atom, the total number of hydrogen atoms connected to the atom, and the chirality of the atom; the edge between the nodes represents the chemical bond between the atoms, and its features include the following attributes: the type of the bond, whether the bond is a conjugated bond, whether the bond belongs to a ring structure, and whether the bond has stereochemical characteristics;

[0016] In the graph structure of the target, each node represents a different residue in the three-dimensional structure of the target protein, and the edge between the nodes represents the chemical bond between the corresponding residues; the feature vector of each node residue includes the residue type, the residue self-distance, and the residue dihedral angle, and the features of the edge between the node residues include: the residue connectivity, the residue CA distance, the residue center distance, and the residue maximum distance.

[0017] Further, the RDKit method library is used to generate a 2D molecular graph of the drug according to the SMILES sequence of the drug; and the Macfrag method is used to process the 2D molecular graph of the drug to extract molecular fragments of the drug that meet the preset chemical significance.

[0018] Further, by combining the computational geometry method and the protein clustering technology, the protein pocket is extracted from the three-dimensional structure of the target protein, and the specific steps include:

[0019] First, the protein space is voxelized to construct a grid structure of the protein;

[0020] Then, the convex hull of the protein is calculated, and the surface thereof is divided into a plurality of triangles, each triangle representing a potential binding site;

[0021] Next, the protein pocket is defined by the volume generated by the vertices of the triangle, and the nearest empty voxels to the protein atoms are identified to determine the possible ligand atom positions as candidate protein pockets;

[0022] Finally, according to the overlap between each candidate protein pocket and the biochemical and physical characteristics of each candidate protein pocket, a plurality of protein pockets with high confidence are selected from all candidate protein pockets as the final extracted protein pockets. Further, GraphTransformer is used as a molecular graph encoder to encode the subgraph structure corresponding to each molecular fragment of the drug, specifically:

[0023] (1) the feature of the i-th node in the subgraph structure corresponding to each molecular fragment of the drug is represented as x i , and the feature of the edge between the i-th node and the j-th node is represented as y ij ;

[0024] (2) the node feature x i and the edge feature y ij in the input subgraph structure are linearly transformed as follows:

[0025]

[0026] wherein, and are linear layer weight parameters in GraphTransformer, and are linear layer bias parameters in GraphTransformer; represents the node feature obtained by linearly transforming x i , represents the edge feature obtained by linearly transforming y ij ;

[0027] (3) Use the eigenvectors of Laplacian operator as node position encoding:

[0028]

[0029] wherein, represents the Laplacian matrix of the subgraph structure; I is an N-dimensional unit matrix with diagonal line 1, and A is an N*N adjacency matrix, which is constructed by the connection relationship between each node in the subgraph structure; D is the degree matrix, which represents the degree of each node; Λ, U represent the diagonal matrix of eigenvalue and the orthogonal matrix of eigenvector respectively; According to the distribution of eigenvalue after Laplacian matrix decomposition, the eigenvectors corresponding to the smallest non-zero eigenvalues are selected as the multi-dimensional position encoding λ i of the node; is a weight matrix; is a bias vector; represents the initial position encoding of node i, which is obtained by linear transformation of the Laplacian eigenvector λ i ; represents the node feature added with position information;

[0030] (4) The node feature and edge feature are fused through the multi-head attention mechanism of GraphTransformer;

[0031] wherein, the detailed updating steps of the lth layer of GraphTransformer network are as follows:

[0032]

[0033] wherein, respectively represent the features of node i and node j at the lth layer, represents the edge between node i and node j at the lth layer; Q k,l , K k,l , V k,l and E k,l are the parameters of the kth head linear layer at the lth layer, respectively represent the query, key and value obtained by the normalized function Norm() of the kth head node feature and edge feature at the lth layer, represents the edge feature of node i and node j at the lth layer; is the attention weight calculated by the node feature and edge feature together;

[0034] (5) The features learned by multiple attention heads are integrated through Concat splicing operation to update the node feature and edge feature:

[0035]

[0036] wherein, is the feature of the (l+1)-th layer node i, i is the feature of all neighbor nodes of node i, is the feature of the edge between the (l+1)-th layer node i and node j, represents the linear layer parameters in the multi-head attention mechanism, h gt represents the number of attention heads;

[0037] (6) The updated node and edge features after multiple fusion and update of the node features and edge features by the multi-head attention mechanism are input into a feedforward network with a residual connection and a normalization operation:

[0038]

[0039] wherein, and and and are weight matrices used for converting the (l+1)-th layer node features and edge features to hidden layers; are respectively and node features and edge features updated by the residual and normalization operations; ReLU represents an activation function in a neural network;

[0040] The node features including edge features obtained by the above process are features including edge features of the n nodes of each subgraph structure features constituting the subgraph structure are represented as Finally, the features of all subgraph structures of the drug constitute the feature R of the drug d ;

[0041] The subgraph structures corresponding to each sub-pocket of the target are encoded using the GraphTransformer as a pocket encoder to obtain the pocket feature of the target, which is the same as the method of encoding the subgraph structures corresponding to each molecular fragment of the drug using the GraphTransformer as a molecular graph encoder.

[0042] Further, the use of a self-attention module to extract sequence features of the target from the one-dimensional amino acid sequence of the target specifically includes:

[0043] (1) The one-dimensional amino acid sequence of the target is one-hot encoded to obtain P s ∈R n×d ; wherein n represents the sequence length and d represents the encoding dimension;

[0044] (2) The encoding P s is input into the self-attention module to learn the sequence encoding Ps The correlation between the internal, the process as shown below:

[0045]

[0046] In the formula, Target sequence features obtained by learning the correlation between the internal sequences by self-attention mechanism, Q s , K s , V s Respectively, the parameters of the self-attention mechanism; Respectively, the query weight matrix, key weight matrix, and value weight matrix in the self-attention module; Attention s Representing the self-attention mechanism, softmax is the activation function, d k The scaling factor.

[0047] Further, the pocket features and sequence features of the target are fused using cross-attention to obtain target features, represented as:

[0048]

[0049] In the formula, W Q , W K , W P Respectively, the query weight matrix, key weight matrix, and value weight matrix, used to convert the sequence data into query vector Q P , key vector K P And value vector V P ; R p The target feature obtained by fusion; Target sequence features extracted by the self-attention module, Target pocket features obtained by encoding the subgraph structure corresponding to each sub-pocket of the target by the pocket encoder; Attention c Representing the cross-attention mechanism.

[0050] Further, step 2 uses a bilinear attention network as a classifier to deeply fuse drug features and target features, and calculates the interaction probability between drugs and targets through a multi-layer perceptron, including:

[0051] First, calculate the bilinear attention mapping to generate an attention map A;

[0052]

[0053] In the formula, q T Is a learnable weight vector used to calculate the interaction strength; M d , are learnable parameter matrices for mapping drug features R d and target features R p into shared space, respectively;

[0054] Then, the joint features f of the target features and the drug features are calculated based on the attention map A:

[0055]

[0056] wherein, σ represents an activation function, M d ', M p ' are the same as M d and are parameter matrices for mapping drug features R d and target features R p into shared space, respectively;

[0057] Finally, the final attention score a is calculated by a multi-layer perception (MLP):

[0058]

[0059] wherein, W represents a learnable weight matrix, b represents a learnable bias vector, and Sigmoid represents an activation function; and the final attention score a is calculated as follows: which is the interaction probability between the drug and the target.

[0060] An electronic device includes a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to enable the processor to implement the drug-target interaction prediction method based on multi-view feature fusion according to any one of the preceding embodiments.

[0061] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the drug-target interaction prediction method based on multi-view feature fusion according to any one of the preceding embodiments.

[0062] Advantages

[0063] This invention proposes a drug-target interaction prediction method (MotifGT-DTI), device, and medium based on multi-perspective feature fusion for accurate prediction of drug-target interactions. For target proteins, this invention employs a cross-attention mechanism to fuse two closely related perspective features: the one-dimensional amino acid sequence and the three-dimensional protein structure, thereby achieving a more comprehensive feature characterization of the target protein. For drugs, considering that the complex molecular structure of drugs is composed of nodes and edges, this invention uses a GraphTransformer, which can simultaneously focus on node features and edge features, as the encoder of the molecular graph. To accurately identify key substructures that bind to proteins, this invention also employs a motif-based molecular characterization method. The final visualization results show that this method can not only predict interactions but also locate specific binding sites in drug-target interaction pairs. Experimental results show that this invention can effectively reveal the molecular fragments responsible for interactions, providing higher interpretability for the prediction results. Attached Figure Description

[0064] Figure 1 This is a framework diagram of the drug-target interaction prediction method based on multi-view feature fusion as described in the embodiments of the present invention;

[0065] Figure 2 This is a flowchart of the drug-target interaction prediction method based on multi-view feature fusion as described in the embodiments of the present invention;

[0066] Figure 3 This is an analysis of the linkage parameters between the number of attention heads and the feature dimension in GraphTransformer in this embodiment of the invention;

[0067] Figure 4 These are the results of the cold-start generalization experiment described in the embodiments of the present invention;

[0068] Figure 5 This is the effect of the TSNE dimension reduction visualization encoder in an embodiment of the present invention;

[0069] Figure 6 This is an interpretability analysis of the model in the embodiments of the present invention. Detailed Implementation

[0070] The embodiments of the present invention will be described in detail below. These embodiments are based on the technical solutions of the present invention and provide detailed implementation methods and specific operation processes to further explain the technical solutions of the present invention.

[0071] The embodiment provides a drug-target interaction prediction method based on multi-view feature fusion, comprising: extracting drug features and target features, then deeply fusing the drug features and the target features and calculating interaction probability between the drug and the target. For the specific implementation process, refer to Figure 1 、 Figure 2 The embodiment is divided into multiple parts and is introduced as follows:

[0072] I. Extracting drug features

[0073] 1. Constructing drug molecule graph and graph structure

[0074] In order to construct the motif subgraph of each drug molecule, first, the SMILES data of the drug is obtained from three public databases Human, Biosnap and Drugbank respectively; then, the RDKit library function is called to construct the corresponding 2D molecular graph including the information of atoms and bonds by analyzing the SMILES sequence; and then the 2D molecular graph of the drug is processed by the Macfrag method to extract the molecular fragment subgraph with important chemical meaning in the drug molecule, which is called molecular motif.

[0075] The graph structure of the drug is constructed according to the characteristics of each atom in the drug molecule and the chemical bond relationship between them; and then the subgraph structure corresponding to each molecular fragment is obtained according to the extracted drug molecular fragments. The subgraph structure of each molecular fragment can be expressed as V d represents an atomic node, E d represents the edge (chemical bond) between nodes.

[0076] For each node in the subgraph structure, its feature is expressed as a 41-dimensional vector x i , which is composed of atom type, atom degree, atom formal charge, atom free radical electron number, atom hybridization, atom aromatic number, total number of H atoms connected to the atom and atom chirality. For the edge between any node i and node j in the subgraph structure, its feature is expressed as a 10-dimensional vector y ij , including bond type, bond is conjugated, bond is ring and bond is stereospecific.

[0077] 2. Extracting drug features using GraphTransformer

[0078] When the drug chemical structure is abstracted as a 2D undirected graph, both atoms and edges are very important and indispensable for describing chemical properties. For example, in the drug motif graph in the prediction task of the present application, the nodes usually represent different atoms in the drug molecule, and the edges represent the chemical bonds between the atoms.

[0079] One of the biggest advantages of the Graphtransformer architecture is that it extends the representation of edge features on node features, and can more accurately describe the molecular structure by considering both node features and edge features. On the other hand, compared with traditional graph structure encoders such as GCN, it converges global information based on attention mechanism, breaks the limitation of local field information, and to some extent can alleviate the problem of local smoothness.

[0080] After constructing the sub-graph structure of each molecular fragment of the drug, the GraphTransformer is used as a molecular graph encoder to encode the sub-graph structure corresponding to each molecular fragment of the drug to obtain the drug characteristics. The method comprises the following steps:

[0081] (1) The feature of the i-th node in the sub-graph structure corresponding to each molecular fragment of the drug is represented as x i , and the feature of the edge between the i-th node and the j-th node is represented as y ij .

[0082] (2) Linear transformation is performed on the node feature x i and the edge feature y ij in the input sub-graph structure:

[0083]

[0084] wherein, and are linear layer weight parameters in the GraphTransformer, and are linear layer bias parameters in the GraphTransformer; represents the node feature obtained by linear transformation of x i , represents the edge feature obtained by linear transformation of y ij .

[0085] (3) In order to better learn the topological structure between nodes, the eigenvectors of the Laplacian operator are used as node position encoding:

[0086]

[0087] wherein, represents the Laplacian matrix of the graph, I is an N-dimensional unit matrix, A is an N×N adjacency matrix, D is a degree matrix, and Λ, U represent the diagonal matrix of eigenvalues and the orthogonal matrix of eigenvectors, respectively. According to the distribution of eigenvalues after Laplacian matrix decomposition, the first k smallest non-trivial eigenvectors (corresponding to the first k smallest non-zero eigenvalues) are selected as the multi-dimensional position encoding of the nodes, represented as λ i .

[0088] (4) The fusion of node and edge features is realized by the multi-head attention mechanism of GraphTransformer.

[0089] For i and node j in the l-th layer, the node feature and edge feature are normalized as query Key Value The node feature and edge feature are injected for updating to obtain accurate global attention. Specifically, the attention score between node i and node j in the k-th head in the l-th layer is calculated based on the node feature and edge feature. The detailed updating steps of the l-th layer are as follows:

[0090]

[0091] Wherein, formula 6-8 are used to calculate the node feature, wherein Q k,l , k k,l and V k,l are linear layer parameters, and formula (9) is used to calculate the edge feature, and then the node feature and edge feature are fused using formula (10) attention mechanism.

[0092] (5) In order to obtain more rich feature representation, the features learned by multiple attention heads are integrated, and the node feature and edge feature are updated, and the specific formula is as follows:

[0093]

[0094] Wherein, represents linear layer parameters, and h gt represents the number of attention heads.

[0095] (6) The node and edge feature after multiple iterations are input into the feedforward network with residual connection and normalization operation:

[0096]

[0097] Wherein, and are linear layer parameters. and are linear layer parameters.

[0098] After the above process, the node feature fused with the edge information of the drug molecule can be obtained, denoted as R d .

[0099] II. Extracting target feature

[0100] 1. Constructing target protein pocket graph

[0101] ​In order to construct the pocket subgraph of each target protein, first, the pid of the target protein is obtained from three public databases Human, Biosnap and Drugbank respectively, and then the protein three-dimensional structure file is obtained from the Protein Data Bank (PDB) according to the protein pid. Then the application uses a method based on computational geometry and protein clustering to extract the corresponding protein pocket, and constructs the corresponding protein pocket graph for each protein pocket. Finally, each protein is represented as a plurality of pocket subgraphs, wherein each subgraph is represented as G p =(V p, E p ), V p represents a residue, and E p represents the edge (chemical bond) between the residues.

[0102] In order to initialize the initial features of the protein pocket subgraph, each residue node is represented as a 41-dimensional vector representing the residue type, residue self-distance and residue dihedral angle. The edge between two residues is represented as a 5-dimensional vector of residue connectivity, residue CA distance, residue center distance and residue maximum distance.

[0103] 2. Extracting pocket features of the target using GraphTransformer

[0104] After constructing the subgraph structure corresponding to each sub-pocket of the target protein, the application uses GraphTransformer as a molecular graph encoder to encode the subgraph structure corresponding to each sub-pocket of the drug, and obtains the pocket features of the target, which are the three-dimensional structure features of the target.

[0105] The method of extracting the pocket features of the target using GraphTransformer is the same as the method of using GraphTransformer as a molecular graph encoder to encode the subgraph structure corresponding to each molecular fragment of the drug.

[0106] 3. Extracting one-dimensional sequence features of the target using self-attention mechanism

[0107] There is a close relationship between the one-dimensional sequence and the three-dimensional structure of the protein. For example, the one-dimensional amino acid sequence of the protein determines its three-dimensional structure, and the three-dimensional structure determines the function and properties of the protein, such as the binding site. Based on this, the embodiments of the application fuse one-dimensional structure (sequences) and three-dimensional structure (pocktes) to represent complex protein structure.

[0108] Before performing the feature fusion of the target protein, the amino acid sequence feature of the protein is first processed. Specifically, the one-dimensional amino acid sequence of each target protein can obtain P s ∈Rn×d Each amino acid is represented as a 20-dimensional vector. Then, the initialized sequence features are sent to the self-attention module to learn the correlation between the sequences, as shown below:

[0109]

[0110] In the formula, is the target sequence feature obtained by learning the correlation between the sequences through the self-attention mechanism, i.e., the one-dimensional sequence feature of the target protein; Q s , K s , and V s are parameters of the self-attention mechanism.

[0111] 4. Fusion of protein 1D and 3D features using cross-attention

[0112] Based on the excellent performance of the attention mechanism in various fields, the embodiment uses cross-attention as a feature fusion method. Cross-attention makes full use of the advantages of two information sources to capture more complex relationships between sequences and spatial structures, as follows:

[0113]

[0114] Finally, the target protein feature R p containing both sequence information and three-dimensional structure information can be obtained.

[0115] III. Using a bilinear attention network as a classifier to predict the probability of drug target interaction

[0116] The bilinear attention network (BAN) architecture is a commonly used attention mechanism in the visual question answering task. BAN can effectively capture the complex bilinear interaction between two modalities. It introduces a bilinear pooling method to express the relationship between question (text) and image (vision) features in a more fine-grained manner. Here, the embodiment uses it to capture the fine-grained interaction features between protein and drug molecules.

[0117] Specifically, the above encoded protein feature R p and drug feature R d are input into the BAN module. First, the bilinear attention mapping is calculated to generate an attention map A, which reflects the interaction between the drug molecule and the protein molecule. Then, based on the attention map A generated in step one, the joint feature f of the protein feature and the drug molecule feature is calculated, and finally the final correlation score is calculated through the MLP layer. The specific process is as follows:

[0118]

[0119]

[0120] IV. Experimental verification

[0121] 1. Experimental data

[0122] The following data were mainly used during the experiment: (1) Protein sequence data. Since the sequence structure of a protein is the basis of a complex spatial structure, and the biological function of a protein is implemented by a single or continuous amino acid fragment, the sequence structure cannot be ignored when trying to mine the biological function of a protein. In order to obtain the protein sequence, the experiment downloaded three commonly used and public datasets, Human, Biosnap, and Drugbank, which contain 1787, 1971, and 4255 protein sequence data, respectively; (2) Protein structure data. Considering that the interaction of drug-target pairs is based on a specific region in the protein (i.e., the pocket structure), the pocket structure feature is crucial for the model to mine potential drug-target pairs. In order to obtain the pocket feature, the protein structure file (.pdb file) needs to be obtained from the Protein Data Bank (PDB) database first, which contains the atomic structure coordinates of each amino acid in the protein and the connection relationship between atoms. (3) Drug molecule data. The three commonly used and public datasets, Human, Biosnap, and Drugbank, downloaded contain 2726, 4510, and 6647 drug molecule data, respectively. The drug molecules in the dataset mainly exist in the form of SMILES sequence, which can then be converted into a 2D molecular graph by calling the relevant library function. (4) Drug-target interaction data. The three commonly used and public datasets, Human, Biosnap, and Drugbank, downloaded contain the following data in each row: 1. Drug data (in SMILES form); 2. Protein amino acid sequence data; 3. Label (representing whether the drug-protein pair actually interacts). Among them, the SMILES sequence of the drug and the amino acid sequence of the protein are used as model input, and the label information is used to calculate whether the prediction ability of the model is accurate.

[0123] 2. Evaluation index

[0124] In order to verify the effectiveness of the method, the method uses five evaluation indexes, designs a comparative test, an ablation experiment, and a cold start experiment to test the prediction performance of the MotifGT-DTI method.

[0125] Receiver operating characteristic (ROC) is used to evaluate the classification ability of a binary classifier when its discrimination threshold changes. This indicator describes the relationship between the true positive rate (TPR, sensitivity) and the false positive rate (FPR, 1-specificity) at different discrimination thresholds. TPR is the proportion of correctly predicted positive samples among all positive samples. FPR is the proportion of incorrectly predicted positive samples among all negative samples. Predicted values above the discrimination threshold are classified as positive samples, and vice versa. TP refers to the number of positive samples that are classified as positive, FP refers to the number of positive samples that are classified as negative, TN refers to the number of negative samples that are classified as negative, and FN refers to the number of negative samples that are classified as positive.

[0126]

[0127] The area under the ROC curve (AUC) is calculated as one of the indicators to measure the performance of the model.

[0128] In an unbalanced data set, AUPRC can effectively measure the performance of the model. It shows the relationship between the Precision and Recall of the model at different thresholds, and the larger the value, the better the overall performance of the model. Among them, Precision refers to the proportion of actual positive samples in the positive samples predicted by the model, reflecting the performance of the model in false positives. Recall refers to the proportion of correctly predicted positive samples among all actual positive samples, which evaluates the performance of the model in false negatives.

[0129]

[0130] To more comprehensively evaluate the performance of the model, this paper also uses F1-score as an evaluation indicator. F1-score is the harmonic mean of Recall and Precision, and in the case of class imbalance, it can evaluate the overall performance of the model by calculating F1Score. The formula is as follows:

[0131]

[0132] 3. Comparison with other methods

[0133] To evaluate the performance of MotifGT-DTI, it is compared with the four state-of-the-art computational methods, including TransformerCPI, MolTrans, DrugBAN and FragXsiteDTI. For a fair comparison, all models are evaluated on the same benchmark datasets and the parameters are the recommended best parameters. On the Human dataset, MotifGT-DTI achieves the highest AUROC of 0.995 and AUPRC of 0.994, and the Precision and Recall of 0.954 and 0.962, respectively, showing the best prediction performance on this dataset. On the Biosnap dataset, MotifGT-DTI also achieves the best AUROC of 0.918 and AUPRC of 0.926, and the Precision, F1 Score and Recall of 0.844, 0.852 and 0.861, respectively, which are significantly better than other models, indicating its high generalization ability and stability. On the Drugbank dataset, MotifGT-DTI again leads with AUROC 0.896 and AUPRC 0.902, and the Precision and Recall indicators also perform relatively high, further confirming its excellent performance on multiple datasets and different evaluation indicators. Overall, MotifGT-DTI model shows significant advantages in multiple evaluation indicators, especially in AUROC and AUPRC indicators, which significantly leads, reflecting its strong generalization ability and robustness in drug-target prediction tasks, and is the most competitive model on the three datasets.

[0134] 4. Verify the generalization ability of the model

[0135] In practical applications of DTI models, most drugs and proteins in the training set will not appear in the test set. Whether the model can still provide accurate predictions when faced with unseen drugs or unseen proteins is a key issue in evaluating the generalization ability and robustness of the model. Therefore, this experiment conducts a cold start experiment on the Human dataset to evaluate the performance of MotifGT-DTI in these special cases. Specifically, three different data partition strategies are used for the experiment:

[0136] Drug cold start: Remove any drug that appears in the training set from the test set.

[0137] Target cold start: Remove any protein that appears in the training set from the test set.

[0138] Pairwise cold start: Remove any drug or target that appears in the training set from the test set.

[0139] Drug cold-start and target cold-start scenarios evaluate the sensitivity of the model when facing new drugs and new targets, respectively. The paired cold-start measures the generalization ability of the model when facing completely unknown drug-target pairs. In Figure 4 It can be seen in the middle that MotifGT-DTI achieves the highest AUROC value in all scenarios, indicating its excellent ability to capture new entity features. The paired cold-start scenario brings greater challenges by introducing new drugs and new targets. Notably, MotifGT-DTI consistently exhibits the best performance, demonstrating its strong robustness under various conditions.

[0140] 5. Case Study

[0141] To evaluate the explainability of the model in the method of the present application in drug-target interaction prediction, a series of visualization experiments were conducted. Human lactate dehydrogenase A (LDH-A) is a key enzyme that promotes the growth of cancer cells, and inhibiting its activity is of great significance for cancer treatment. The complex of LDH-A and ligand 9YA was downloaded from PDB (PDBID: 5W8L), which did not appear in the training data and belongs to a typical paired cold-start case. Notably, the model in the method of the present application successfully identified this unseen protein-ligand interaction pair, demonstrating its performance in predicting data outside the dataset. The bilinear attention map visualizes each substructure, providing key insights and explanations about the molecular level of drug design work. As shown in Figure 6 (b), the attention scores of the predicted four drug structures in the attention map are the highest and are marked with dashed circles. Compared with the known true protein and ligand binding sites, MotifGT-DTI correctly predicts three of them, as shown in Figure 6 (c), and the remaining one is found to be highly related to the binding site. Specifically, the key structure of the carboxylic acid group in the binding is correctly identified (interacting with residues Arg168, His192, and Thr247 as a hydrogen bond acceptor). The thiazole structure interacting with residue His192 is also correctly identified. The biphenyl ring structure interacting with residues Arg105 and Asn137 is also correctly predicted as a drug structure that plays an important role in the binding process. Although the predicted oxazole structure is not currently a known binding site, research shows that it is a key chemical feature in LDH-A-related inhibitors and is closely related to reducing LDH-A enzyme activity and inhibiting cancer cell growth. All results indicate that MotifGT-DTI can identify functional molecular motifs that contribute to the interaction and provide explainability for the prediction results.

[0142] The above embodiments are the preferred embodiments of the present application, and those skilled in the art can make various modifications or improvements on the basis of the above embodiments without departing from the general concept of the present application, and these modifications or improvements should also belong to the scope of protection of the present application.

Claims

1. A method for predicting drug-target interaction based on multi-view feature fusion, characterized in that, Comprise: Step 1, extract drug features and target features; (1) Extract drug features: first, generate the corresponding 2D molecular graph according to the SMILES sequence of the drug, and construct the graph structure of the drug according to the characteristics of each atom in the drug molecule and the chemical bond relationship between them; then extract the molecular fragments of the drug to obtain the subgraph structure corresponding to each molecular fragment; finally, use a Graph Transformer as a molecular graph encoder to encode the subgraph structure corresponding to each molecular fragment of the drug to obtain the drug features; (2) Extract target features: first, extract the protein pocket of the target, and construct the graph structure of the target according to each residue in the protein pocket and the chemical bond between them; then divide the target into multiple sub-pockets to obtain the subgraph structure corresponding to each sub-pocket; then use a Graph Transformer as a pocket encoder to encode the subgraph structure corresponding to each sub-pocket of the target to obtain the pocket features of the target; In addition, a self-attention module is used to extract the sequence features of the target from the one-dimensional amino acid sequence of the target; Finally, use cross-attention to fuse the pocket features and sequence features of the target to obtain the target features; Step 2, use a bilinear attention network as a classifier to deeply fuse the drug features and target features, and calculate the interaction probability between the drug and the target through a multilayer perceptron.

2. The multi-view feature fusion-based drug-target interaction prediction method according to claim 1, characterized in that, In the graph structure of the drug, each node represents an atom, The feature vector includes the following attributes: the type of the atom, the degree of the atom, the formal charge of the atom, the free radical electron number of the atom, the hybridization state of the atom, the aromaticity of the atom, the total number of hydrogen atoms connected to the atom, and the chirality of the atom; The edge between the nodes represents the chemical bond between the atoms, and the features include the following attributes: the type of the bond, whether the bond is a conjugated bond, whether the bond belongs to a ring structure, and whether the bond has stereochemical characteristics; In the graph structure of the target, each node represents a different residue in the three-dimensional structure of the target protein, and the edge between the nodes represents the chemical bond between the corresponding residues; The feature vector of each node residue includes the residue type, residue self-distance and residue dihedral angle, and the feature of the edge between the node residues includes: residue connectivity, residue CA distance, residue center distance and residue maximum distance. 3.The multi-view feature fusion based drug-target interaction prediction method according to claim 1, characterized in that, Use the RDKit method library to generate the 2D molecular graph of the drug according to the SMILES sequence of the drug; use the Macfrag method to process the 2D molecular graph of the drug to extract the molecular fragments of the drug that meet the preset chemical meaning. 4.The method of predicting drug-target interaction based on multi-view feature fusion according to claim 1, characterized in that, By combining the computational geometry method and the protein clustering technology, the protein pocket is extracted from the three-dimensional structure of the target protein, and the specific steps include: First, perform voxelization processing on the protein space to construct the grid structure of the protein; Then calculate the convex hull of the protein, divide its surface into multiple triangles, and each triangle represents a potential binding site; Next, define the protein pocket by the volume generated by the triangle vertices, and identify the nearest empty voxels to the protein atoms to determine the possible ligand atom positions as candidate protein pockets; Finally, according to the overlap between each candidate protein pocket and the biochemical and physical characteristics of each candidate protein pocket, a number of protein pockets with high confidence are screened from all candidate protein pockets as the final extracted protein pockets. 5.The method of predicting drug-target interaction based on multi-view feature fusion according to claim 1, characterized in that, The Graph Transformer is used as a molecular graph encoder to encode the subgraph structure corresponding to each molecular fragment of the drug, specifically as follows: (1) the feature of the i-th node in the sub-graph structure corresponding to each molecular fragment of the drug is denoted as x i , and the feature of the edge between the i-th and j-th nodes i is denoted as y ij ; (2) Linearly transform the node features x in the input subgraph structure i and edge features y ij : wherein, and are linear layer weight parameters in the Graph Transformer, and are linear layer bias parameters in the Graph Transformer; denotes x i the node feature obtained by linear transformation, denotes y ij the edge feature obtained by linear transformation; (3) The eigenvectors of the Laplacian operator are used as node position encoding: wherein, L represents the Laplacian matrix of the subgraph structure; I is an N-dimensional unit matrix with diagonal line 1, A is an N*N adjacency matrix, which is constructed by the connection relationship between each node in the subgraph structure; D is a degree matrix, which represents the degree of each node; Λ, U respectively represent the diagonal matrix of the eigenvalue and the orthogonal matrix of the eigenvector; according to the distribution of the eigenvalue after the decomposition of the Laplacian matrix, the eigenvectors corresponding to the smallest non-zero eigenvalues are selected as the multi-dimensional position encoding λ of the node i ; is a weight matrix; is a bias vector; represents the initial position encoding of node i, which is obtained by linear transformation of the Laplacian eigenvector λ i ; represents the node feature added with position information; (4) The node features and edge features are fused through the multi-head attention mechanism of the Graph Transformer: The detailed updating steps of the lth layer of the Graph Transformer network are as follows: wherein, respectively represent the feature of the i-th node and the j-th node in the l-th layer, represents the edge between the i-th node and the j-th node in the l-th layer; Q k,l , K k,l , V k,l and E k,l is the k-th head linear layer parameter in the l-th layer, respectively represent the query, the key and the value obtained by the normalization function Norm() on the k-th head node feature and the edge feature in the l-th layer, represents the edge feature of the i-th node and the j-th node in the l-th layer; is the attention weight calculated by the node feature and the edge feature together; (5) The features learned by multiple attention heads are integrated through the Concat splicing operation to update the node features and edge features: wherein, is a feature of the (l+1)-th layer node i, N i is all the neighboring nodes of node i, is an edge feature between the (l+1)-th layer node i and node j, represents linear layer parameters in the multi-head attention mechanism, h gt represents the number of attention heads; (6) The node and edge features after multiple fusions and updates of the node features and edge features through the multi-head attention mechanism are input into a feedforward network with residual connection and normalization operation: wherein, and and and are weight matrices for transforming the (l+1)-th layer node and edge features to the hidden layer; respectively, and updated node and edge features after residual and normalization operations; ReLU denotes an activation function in the neural network; The node features including edge features are obtained through the above process Features of n nodes of each subgraph structure including edge features Features constituting the subgraph structure Indicated as Finally, the features of all subgraph structures of the drug constitute the features R of the drug d ; The Graph Transformer is used as a pocket encoder to encode the subgraph structure corresponding to each sub-pocket of the target to obtain the pocket features of the target, which is the same as the method of using the Graph Transformer as a molecular graph encoder to encode the subgraph structure corresponding to each molecular fragment of the drug. 6.The method of predicting drug-target interaction based on multi-view feature fusion according to claim 1, characterized in that, The sequence features of the target are extracted from the one-dimensional amino acid sequence of the target using a self-attention module, specifically including: (1) one-hot encoding of the one-dimensional amino acid sequence of the target to obtain P s ∈R n×d ; wherein n represents the sequence length, and d represents the encoding dimension; (2) The P s The self-attention module learns the sequence encoding P s The internal correlation between the processes is shown as follows: In the formula, Target sequence features obtained by learning the correlation between the sequences internally by the self-attention mechanism, Q s , K s , V s are parameters of the self-attention mechanism, respectively; are query weight matrix, key weight matrix, and value weight matrix in the self-attention module, respectively; Attention s represents the self-attention mechanism, and softmax is an activation function, d k is a scaling factor. 7.The multi-view feature fusion based drug-target interaction prediction method according to claim 1, characterized in that, The pocket features and sequence features of the target are fused using cross-attention to obtain target features, represented as: wherein W Q , W K , and W P represent a query weight matrix, a key weight matrix, and a value weight matrix, respectively, for converting the sequence data into a query vector Q P , a key vector K P , and a value vector V P ; R p is a fused target feature; is a target sequence feature extracted by a self-attention module, is a target pocket feature obtained by encoding the subgraph structures corresponding to the subpockets of the target by a pocket encoder; Attention c represents a cross-attention mechanism. 8.The method of predicting drug-target interaction based on multi-view feature fusion according to claim 1, characterized in that, Step 2 uses a bilinear attention network as a classifier to deeply fuse the drug features and target features, and calculates the interaction probability between the drug and the target through a multi-layer perceptron, including: First, calculate the bilinear attention mapping to generate an attention map A; where q T is a learnable weight vector for calculating the interaction strength; M d , is a learnable parameter matrix for mapping drug features R d and target features R p into a shared space, respectively; Then, based on the attention map A, calculate the joint features f of the target features and the drug features: where σ denotes an activation function, M d ′, M p ′ is the same as M d and the same function, denotes a parameter matrix for mapping the drug features R d and the target features R p to the shared space; Finally, the final attention score is calculated by a multi-layer perceptron, MLP In the formula, W represents a learnable weight matrix, b represents a learnable bias vector, and Sigmoid represents an activation function; and the attention score is finally calculated That is, the interaction probability between the drug and the target.

9. An electronic device comprising a memory and a processor, said memory having stored therein a computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1-8.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1-8.

Images