DTI prediction method based on hierarchical multi-modal self-attention graph neural network

Abstract

The application discloses a DTI prediction method based on a hierarchical multimodal self-attention graph neural network, and relates to the technical field of drug research. The method comprises the following steps: extracting protein sequence features and protein 2-mer sequence features by using a convolution module; extracting drug SMILES features and drug molecular graph features by using the convolution module in combination with a graph neural network; performing multimodal feature fusion on the protein sequence features, the protein 2-mer sequence features, the drug SMILES features and the drug molecular graph features; and taking a drug target pair feature vector as an input of a classifier to perform DTI prediction. The application comprehensively uses multimodal data such as SMILES, drug molecular graphs, protein sequences and 2-mer sequences to help the model better understand and depict the features and interaction relationship of drugs and proteins.

Description

Technical Field

[0001] This invention relates to the field of drug research technology, and more specifically, to a DTI prediction method based on hierarchical multimodal self-attention graph neural networks. Background Technology

[0002] Drug-target interaction (DTI) prediction is a crucial step in drug development. Traditional screening experiments require significant human and material resources, leading to high drug development costs. With the increasing size of compound and target libraries, computational methods for DTI prediction have become a practical approach that can effectively improve drug development efficiency and reduce costs. Computational DTI prediction methods are mainly divided into two categories: machine learning-based DTI prediction and deep learning-based DTI prediction. Machine learning-based DTI prediction methods utilize algorithms such as random forests, support vector machines, and logistic regression to train on known drug-target interaction datasets, learning the association patterns between drugs and targets.

[0003] Deep learning-based drug-target affinity prediction methods model DTI prediction as a binary classification task. These methods use deep learning models to automatically learn feature representations of drugs and targets and capture the complex relationships between them. With the rapid development of deep learning technology, numerous models have been proposed for DTI prediction. For example, some methods use molecular and reaction features as model inputs, introduce feature kernel functions, and perform DTI prediction by calculating similarity in the feature space. Another example is the DeepDTA model, which uses only drug samples and protein sequences as inputs and employs two CNN modules to extract drug and protein features respectively for drug-target affinity prediction, achieving significant results.

[0004] While sequence-based drug detection and treatment (DTI) prediction methods have made significant progress, sequence representations cannot provide information about the chemical structure of molecules, causing models to fail to capture the similarities between molecular structures and affecting the accuracy of DTI prediction models. In contrast to sequence-based representations, graph representations can capture the structural and topological information of drug molecules, demonstrating a significant advantage in DTI prediction tasks. Therefore, graph neural network-based DTI prediction methods are gradually emerging. For example, using a graph neural network (GNN) to learn the features of drug molecules, then employing a CNN to extract protein features from the protein sequence, and concatenating the two to obtain drug-protein pairs, which are then fed into a classification network for prediction, effectively improving the accuracy of DTI prediction.

[0005] In DTI prediction methods, various data representation methods have their limitations, which to some extent restricts the learning ability of deep learning models based on a single data representation method. Data from different modalities often contain their own unique information cues. Combining data from these modalities can provide more comprehensive and richer information, helping the model to better understand and characterize the features and relationships of the data, and further improving the model's learning and generalization abilities. Therefore, in recent years, some research has used multimodal data to improve the performance of DTI prediction models. However, multimodal DTI prediction methods still have shortcomings in multimodal data fusion. Most models neglect intramodal interactions when modeling interactions between modalities, and cannot simultaneously model intramodal and intermodal interactions.

[0006] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0007] To address the problems in related technologies, this invention proposes a DTI prediction method based on hierarchical multimodal self-attention graph neural networks, in order to overcome the aforementioned technical problems existing in the existing related technologies.

[0008] Therefore, the specific technical solution adopted by the present invention is as follows:

[0009] The DTI prediction method based on hierarchical multimodal self-attention map neural network includes the following steps:

[0010] S1. Extract protein sequence features and protein 2-mer sequence features using a convolution module;

[0011] S2. Extract drug SMILES features and drug molecule graph features using convolutional modules and graph neural networks;

[0012] S3. Multimodal feature fusion of protein sequence features, protein 2-mer sequence features, drug SMILES features, and drug molecule map features;

[0013] S4. Based on the multimodal feature fusion results, the feature vector of drug target pairs is obtained, and the feature vector of drug target pairs is used as the input of the classifier for DTI prediction.

[0014] Preferably, the extraction of protein sequence features and protein 2-mer sequence features using a convolution module includes the following steps:

[0015] S11. Encode the original protein sequence and convert the encoded original protein sequence into a marker sequence;

[0016] S12. Map the labeled sequence into the embedding space and use convolutional layers to extract protein sequence features from the labeled sequence;

[0017] S13. Divide the original protein sequence into sequence fragments and convert each sequence fragment into a 2-mer coding sequence;

[0018] S14. Map the 2-mer coding sequence into the embedding space and use a convolutional layer to extract the protein 2-mer sequence features from the 2-mer coding sequence.

[0019] Preferably, the extraction of drug SMILES features and drug molecule graph features using a convolutional module combined with a graph neural network includes the following steps:

[0020] S21. Encode and fill the drug SMILES characters to obtain drug molecule vectors, and use the drug molecule vectors as input to the graph neural network.

[0021] S22. Map the drug molecule vectors into the embedding space and use the convolution module to extract drug SMILES features from the drug molecule vectors;

[0022] S23. Convert the drug SMILES into a drug molecule graph, and obtain the initial features of the nodes and the initial features of the edges from the drug molecule graph.

[0023] The initial characteristics of a node include atomic number, number of chemical bonds, formal charge, chirality, number of hydrogen atoms connected to the atom, hybridization, whether it participates in the formation of an aromatic ring, and atomic mass.

[0024] The initial characteristics of an edge include bond type, conjugacy, bond position, and geometric heterogeneity;

[0025] S24. Based on the initial features of nodes and edges, extract drug molecule graph features using a directed message passing network.

[0026] Preferably, extracting drug molecule graph features using a directed message passing network based on the initial features of nodes and edges includes the following steps:

[0027] S241. Take the initial features of the node as the initial hidden state of the node, and associate the edge B with the initial features of the node. vw The initial features are transformed into edge B vw The initial hidden state;

[0028] S242. Using the average function, combine the initial features of the node, the initial features of its neighbors, and the edge B in the t-th layer. kv Aggregate the hidden states to obtain edge B. vw Message delivery in the (t+1)th layer;

[0029] S243, move side B vw Message transmission at level t+1 and edge B at level t vw The hidden state is used as the input to the fully connected neural network to obtain the edge B in the (t+1)th layer. vw The hidden state;

[0030] S244. Sum the hidden states of all edges connected to the node to obtain the node's transmitted message. At the same time, concatenate the node's initial features and the node's transmitted message and transmit them to the linear layer to obtain the node's final hidden state.

[0031] S245. By splicing together the final hidden states of all nodes, the drug molecule graph features are obtained.

[0032] Preferably, the multimodal feature fusion of protein sequence features, protein 2-mer sequence features, drug SMILES features, and drug molecule map features includes the following steps:

[0033] S31. The protein sequence features and protein 2-mer sequence features are spliced ​​together to obtain the protein combinatorial features, and the first-level protein fusion features are calculated based on the protein combinatorial features.

[0034] S32. The drug SMILES features and drug molecule graph features are spliced ​​together to obtain drug combination features, and the first-level drug fusion features are calculated based on the drug combination features.

[0035] S33. The drug target pair features are obtained by splicing the primary drug fusion features and the primary protein fusion features, and the final fusion features are calculated based on the drug target pair features.

[0036] S34. The final fused features are split and weighted fused to obtain the final modal features.

[0037] Preferably, the process of splicing protein sequence features and protein 2-mer sequence features to obtain protein combinatorial features, and calculating primary protein fusion features based on protein combinatorial features, includes the following steps:

[0038] S311. Feature splicing is performed on protein sequence features and protein 2-mer sequence features to obtain protein combinatorial features;

[0039] S312. Perform linear transformations on the protein combinatorial features to obtain the protein query matrix, protein bond matrix, and protein value matrix.

[0040] S313. Based on the protein query matrix and the protein bond matrix, the protein attention matrix is ​​calculated;

[0041] S314. Multiply the protein attention matrix by the protein value matrix to obtain the first-level protein fusion feature.

[0042] Preferably, the process of concatenating drug SMILES features and drug molecule map features to obtain drug combination features, and calculating primary drug fusion features based on drug combination features, includes the following steps:

[0043] S321. Perform feature splicing on the drug SMILES features and drug molecule map features to obtain drug combination features;

[0044] S322. Perform a linear transformation on the drug combination features to obtain the drug query matrix, drug key matrix, and drug value matrix;

[0045] S323. Based on the drug query matrix and drug key matrix, calculate the drug attention matrix;

[0046] S324. Multiply the drug attention matrix and the drug value matrix to obtain the first-level drug fusion feature.

[0047] Preferably, the formula for calculating the primary drug fusion characteristics is:

[0048]

[0049] In the formula, Indicates primary drug fusion characteristics; Q smiles K smiles and V smiles These represent the drug SMILES query matrix, drug SMILES key matrix, and drug SMILES value matrix, respectively; Q graph K graph and V graph These represent the drug molecule graph query matrix, drug molecule graph key matrix, and drug molecule graph value matrix, respectively; A drug Represents the drug attention matrix; V drug This represents the drug value matrix.

[0050] Preferably, the process of splicing primary drug fusion features and primary protein fusion features to obtain drug target pair features, and calculating the final fusion features based on the drug target pair features, includes the following steps:

[0051] S331. The primary drug fusion features and the primary protein fusion features are spliced ​​together to obtain drug target pair features.

[0052] S332. Perform linear transformations on the drug target pair features to obtain the drug target pair query matrix, drug target pair key matrix, and drug target pair value matrix.

[0053] S333. Based on the drug target pair query matrix and the drug target pair key matrix, calculate the drug target pair attention matrix.

[0054] S324. Multiply the drug target pair attention matrix with the drug target pair value matrix to obtain the final fused feature.

[0055] Preferably, obtaining the drug target pair feature vector based on the multimodal feature fusion result, and using the drug target pair feature vector as input to the classifier for DTI prediction includes the following steps:

[0056] S41. Construct a classifier by combining a fully connected layer with the RELU activation function and a regularization layer;

[0057] S42. The final modal features are concatenated to obtain the drug target pair feature vector, and the drug target pair feature vector is loaded into the classifier for DTI prediction.

[0058] The beneficial effects of this invention are as follows:

[0059] 1. The DTI prediction method based on hierarchical multimodal self-attention graph neural network provided by this invention uses drug smiles, drug molecular graphs, protein sequences, and protein 2-mer sequences as inputs to HMSA-DTI. It captures inter-modal and intra-modal interactions through hierarchical multimodal attention mechanism, realizing deep fusion of drug multimodal features and protein multimodal features. At the same time, it is evaluated on three benchmark datasets, DrugBank, human, and C. elegans, using 10-fold cross-validation. It is also compared with other advanced DTI prediction models, thus verifying that the HMSA-DTI prediction model has significant advantages and strong competitiveness in multiple indicators.

[0060] 2. The DTI prediction method based on hierarchical multimodal self-attention graph neural network provided by this invention integrates multimodal data such as SMILES, drug molecule graphs, protein sequences and 2-mer sequences to help the model better understand and characterize the features and interactions between drugs and proteins. It uses node-level features of D-MPNN instead of graph-level features and can simultaneously model intramodal and intermodal interactions to improve the discriminability and robustness of fused features. Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 This is a flowchart of a DTI prediction method based on a hierarchical multimodal self-attention graph neural network according to an embodiment of the present invention;

[0063] Figure 2 This is a structural diagram of the HMSA-DTI model in the DTI prediction method based on hierarchical multimodal self-attention graph neural network according to an embodiment of the present invention;

[0064] Figure 3 This is a structural diagram of the cross-attention mechanism in the DTI prediction method based on a hierarchical multimodal self-attention graph neural network according to an embodiment of the present invention;

[0065] Figure 4 This is a structural diagram of the hierarchical multimodal self-attention mechanism in the DTI prediction method based on hierarchical multimodal self-attention graph neural network according to an embodiment of the present invention;

[0066] Figure 5 This is a schematic diagram of the PR curve of HMSA-DTI on the DrugBank dataset in the DTI prediction method based on hierarchical multimodal self-attention graph neural network according to an embodiment of the present invention.

[0067] Figure 6 This is a schematic diagram of the ROC curve of HMSA-DTI on the DrugBank dataset in the DTI prediction method based on hierarchical multimodal self-attention graph neural network according to an embodiment of the present invention.

[0068] Figure 7 This is a schematic diagram illustrating the influence of the K value on ACC and AUC in the DTI prediction method based on a hierarchical multimodal self-attention graph neural network according to an embodiment of the present invention.

[0069] Figure 8 This is a schematic diagram illustrating the influence of the number of D-MPNN network layers on ACC in the DTI prediction method based on hierarchical multimodal self-attention graph neural network according to an embodiment of the present invention. Detailed Implementation

[0070] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0071] According to embodiments of the present invention, a DTI prediction method based on a hierarchical multimodal self-attention graph neural network is provided.

[0072] This invention proposes a DTI prediction method based on a hierarchical multimodal self-attention graph neural network. This method consists of four parts: protein feature extraction, drug feature extraction, multimodal feature fusion, and DTI prediction. Protein feature extraction uses protein sequences and 2-mer sequences as inputs, extracting protein features through a convolutional module. The drug feature extraction module uses smils and molecular graphs as inputs, extracting drug features through a convolutional module and a graph neural network, respectively. To fully utilize the interaction information between protein and drug multimodal features to improve feature representation capability, the four features are fused, and finally, these features are concatenated as input to a classifier to predict the interaction between drugs and proteins.

[0073] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, the DTI prediction method based on a hierarchical multimodal self-attention graph neural network according to an embodiment of the present invention includes the following steps:

[0074] S1. Use the convolution module to extract protein sequence features and protein 2-mer sequence features.

[0075] It should be noted that, in order to extract the global and local structural features of proteins to better describe their properties, the protein sequence is used to represent the global structure of the protein, and a convolutional module is used to extract features as a global feature representation; at the same time, k-mer sequences are used to describe the local structure of the protein, and features are extracted through a convolutional module as a protein local feature table.

[0076] The extraction of protein sequence features and protein 2-mer sequence features using a convolutional module includes the following steps:

[0077] S11. Encode the original protein sequence and convert the encoded original protein sequence into a marker sequence.

[0078] S12. Map the labeled sequence into the embedding space and use convolutional layers to extract protein sequence features from the labeled sequence.

[0079] It should be noted that the original protein sequence is first encoded, and then padded with 0s to obtain a marker sequence of uniform length. The labeled sequences are mapped to the embedding space, and then protein sequence features are extracted through two consecutive convolutional layers. Its expression is:

[0080] X seq =CNN seq (Emb(x token ));

[0081] In the formula, X seq Representing protein sequence features; CNN seq Indicates a CNN layer; Emb represents an embedding layer; L seq The spatial dimension representing the sequence features; d protein The embedding dimension of the sequence features; x token This represents a labeled sequence.

[0082] S13. The original protein sequence is divided into sequence fragments, and each sequence fragment is converted into a 2-mer coding sequence.

[0083] S14. Map the 2-mer coding sequence into the embedding space and use a convolutional layer to extract the protein 2-mer sequence features from the 2-mer coding sequence.

[0084] It should be noted that, in order to obtain the local feature representation of the protein, this invention uses a sliding window with a window size of 2 and a stride of 1 to segment the protein sequence into sequence fragments of length 2, and maps them to numbers from 1 to 484 to obtain a 2-mer coding sequence; the 2-mer coding sequence is mapped to the embedding space, and two convolutional layers are used for feature extraction to obtain the protein 2-mer sequence features. Its expression is:

[0085] X mer =CNN mer (Emb(x mer ));

[0086] In the formula, CNN mer Represents a convolutional layer; Emb represents an embedding layer; x mer Indicates the 2-mer sequence characteristics of the encoded protein; L mer d represents the spatial dimension of a 2-mer sequence. protein This represents the embedding dimension of the 2-mer sequence.

[0087] It should be noted that, as Figure 2 As shown, this invention maps and embeds the four inputs into the space respectively, and then uses convolutional blocks (CNN) and directed message passing neural networks (D-MPNN) to extract features to obtain four features. Then, a hierarchical multimodal self-attention mechanism is used to fuse multimodal features to capture drug target interaction information, and finally outputs the prediction score through a classifier.

[0088] S2. Use convolutional modules and graph neural networks to extract drug SMILES features and drug molecule graph features.

[0089] It should be noted that drug molecules are typically represented using SMILES or molecular graphs. SMILES is a method of representing chemical molecules using strings. It consists of a series of characters, each representing an atom, bond, or specific chemical information. Due to its advantages of small storage space and ease of recognition, it is commonly used in computational chemistry, cheminformatics, and bioinformatics. Molecular graphs, on the other hand, are a method of representing chemical molecules using graph structures. The nodes of the graph represent atoms in the molecule, while the chemical bonds between atoms are represented as edges. This representation can provide information about the topological structure of the molecule. This invention considers using both SMILES and molecular graphs as drug representation methods to improve feature robustness.

[0090] The process of extracting drug SMILES features and drug molecule graph features using convolutional modules and graph neural networks includes the following steps:

[0091] S21. Encode and fill the drug SMILES characters to obtain drug molecule vectors, and use the drug molecule vectors as input to the graph neural network.

[0092] S22. Map the drug molecule vector into the embedding space and use the convolution module to extract drug SMILES features from the drug molecule vector.

[0093] It should be noted that when using SMILES to represent drug molecules as input to a neural network, they need to be converted into vector representations. This invention uses integers from 1 to 64 to encode the characters in SMILES, then pads the ends with 0 to obtain drug molecule vectors of uniform length, and maps them to the embedding space; finally, CNN blocks are used to extract drug sequence features. Its expression is:

[0094] X smiles =CNN smiles (Emb(x smiles ));

[0095] In the formula, CNN smiles Represents the spatial dimension of a molecule; Emb represents the embedding layer; d drug The embedding dimension of a molecule: X smiles This represents the coded drug SMILES characteristics.

[0096] S23. Convert the drug SMILES into a drug molecule graph, and obtain the initial features of the nodes and the initial features of the edges from the drug molecule graph.

[0097] It should be noted that a drug molecule can be represented as a graph G = (A, B), where A is the set of nodes (atoms) and B is the set of edges (bonds). Nodes represent atoms in the drug molecule, and edges represent chemical bonds between atoms. This invention uses eight properties as the initial characteristics x of the nodes: atomic number, degree (number of chemical bonds), formal charge, chirality, number of hydrogen atoms connected to the atom, hybridization, whether it participates in the formation of an aromatic ring, and atomic mass. v The initial features of the edges are e, including bond type (single / double / triple / aromatic), conjugacy, bond position (whether it is inside a ring), and geometric heterogeneity. vw Furthermore, molecular graph features are extracted using a directed message passing network (D-MPNN). Unlike traditional message passing neural networks (MPNN), D-MPNN considers not only node features but also the directionality of edges.

[0098] The initial characteristics of nodes include atomic number, number of chemical bonds, formal charge, chirality, number of hydrogen atoms connected to the atom, hybridization, whether they participate in the formation of aromatic rings, and atomic mass; the initial characteristics of edges include bond type, conjugation, bond position, and geometric isomerism.

[0099] S24. Based on the initial features of nodes and edges, extract drug molecule graph features using a directed message passing network.

[0100] The extraction of drug molecule graph features using a directed message passing network, based on the initial features of nodes and edges, includes the following steps:

[0101] S241, Set the initial feature x of the node v As the initial hidden state of the node And will be combined with the initial feature x of the node. v Associated edge B vw The initial features are transformed into edge B vw initial hidden state

[0102] It should be noted that the present invention will and Let B represent the hidden states of a node and an edge at time t, respectively, where t represents the network layer number and the initial feature of the node is B. vw As its initial hidden state The initial hidden state of the edge Its expression is:

[0103]

[0104] In the formula, x v and e vwLet W represent the initial features of the node and the initial features of the edge, respectively; b This represents the parameter matrix, where α represents the ReLU activation function; concat represents the concatenation operation.

[0105] S242. Use the average function to calculate the initial features x of the nodes. v Initial features x of neighboring nodes k and the edge B in the tth layer kv Hidden state Perform aggregation to obtain edge B kv Message passing in layer t+1 Its expression is:

[0106]

[0107] In the formula, Representing edge B kv Message passing in layer t+1; M represents the average aggregation function; N(v) represents the set of neighboring nodes of node v; t represents the network layer number; Representing edge B kv The hidden state; k represents the neighbor node; x k Represents the initial characteristics of neighboring nodes; x v This represents the initial characteristics of node v.

[0108] S243, move side B vw Message delivery at level t+1 and the edge B in the tth layer vw Hidden state As input to the fully connected neural network, edge B in layer t+1 is obtained. vw Hidden state Its expression is:

[0109]

[0110] In the formula, f represents a fully connected neural network.

[0111] s244. Sum the hidden states of all edges connected to the node to obtain the node's transmitted message m. v At the same time, the initial feature x of the node v and the message m passed by the node v The features are concatenated and transmitted to the linear layer to obtain the final hidden state h of the node. v .

[0112] It should be noted that after obtaining the final hidden state of all edges through T iterations, message m is obtained by aggregating the hidden states of the edges connected to node v. v Then use message m v and the initial features x of the nodesv Obtain the final hidden state h of the node v Its expression is:

[0113]

[0114] h v =α(W a (concat(x v m v ));

[0115] In the formula, W a Represents the parameter matrix; h v This represents the final hidden state of node v; concat represents the concatenation operation.

[0116] S245. By splicing together the final hidden states of all nodes, the drug molecule graph features are obtained.

[0117] It should be noted that the readout phase of message passing networks typically uses average pooling or max pooling to aggregate all node features to obtain graph features. However, the properties of drugs are closely related to their local structures, and the aggregation process in the readout phase inevitably leads to the loss of some node feature information, affecting subsequent DTI prediction. Previous studies have often fused graph-level features of drug molecule graphs with features from other modalities. This multimodal fusion method based on graph-level features results in the model being unable to simultaneously capture intermodal and intramodal interactions. Therefore, this invention uses node-level features instead of graph-level features for DTI prediction, obtaining node-level features by concatenating the hidden states of all nodes. Among them, L g d represents the number of atoms in a drug molecule. drug The atomic embedding dimension is represented, and then a hierarchical multimodal self-attention mechanism is used to fuse the multimodal features of drugs and proteins to improve the model's representation ability.

[0118] S3. Multimodal feature fusion is performed on protein sequence features, protein 2-mer sequence features, drug SMILES features, and drug molecular map features.

[0119] It's important to note that multimodal learning enhances feature representation capabilities by fusing features from multiple data sources, thereby improving the model's predictive ability. In the DTI prediction task, the predictive performance can be improved by fusing drug SMILES features, drug molecule map features, protein sequence features, and protein 2-mer features. The simplest feature fusion method is feature concatenation, but it simply concatenates features from different modalities without considering the intermodal relationships. This results in the fused features lacking high-level semantic information and intermodal interaction information. Therefore, many studies use cross-attention to achieve multimodal feature fusion.

[0120] Taking the fusion of drug SMILES features and drug molecule graph features as an example, drug SMILES describes the arrangement of atoms in a linear manner, while the molecule graph describes the topological relationships between atoms through a graph structure. The fusion of these two representations can improve the model's ability to represent drug molecules, enabling the model to extract more robust drug features. The drug feature fusion process based on cross-attention is as follows: Figure 3 As shown, the characteristics of the drug molecule map are first analyzed. Perform a linear transformation to obtain the query matrix SMILES characteristics of drugs Obtain the key matrix by performing a linear transformation Sum matrix Its expression is:

[0121] Q graph =Linear(X) graph ) = X graph W Q +b Q

[0122] K smiles =Linear(X) smiles ) = X smiles W K +b K

[0123] V smiles =Linear(X) smiles ) = X smiles W V +b V ;

[0124] In the formula, Linear represents the linear layer, and the query matrix Q graph Key matrix K smiles Sum matrix V smiles It was obtained through three different lines, W Q W K and W V Let b represent the weight matrices in these three linear layers respectively. Q b K and b V This represents the bias vector in a linear layer.

[0125] By querying matrix Q graph Bond matrix K smiles The attention matrix is ​​obtained, and its expression is:

[0126]

[0127] In the formula, A represents the attention matrix, and Q and K represent the query matrix Q. graph Bond matrix K smiles T represents the transpose sign, K T K represents smiles transpose, d k K represents smiles The embedding dimension, that is, K smiles d att The size of this dimension.

[0128] Using attention matrix A and value matrix V smiles The fused feature obtained from SMILE features and graph features is expressed as follows:

[0129]

[0130] In the formula, L s Features derived from SMILEs It represents the length of the digital vector obtained by passing the drug SMILES sequence through the embedding layer, L g Features from drug molecular maps It represents the number of atoms, a ij v represents an element in the drug attention matrix. i The fusion feature H, representing the row vectors of V in the drug value matrix, shows that the fusion feature H is simply a linear combination of the row vectors in the value matrix V. Therefore, the fusion feature based on cross-attention is only a linear combination of the drug SMILES features and lacks drug molecule graph features. This fusion method cannot fully utilize the advantages of multimodal features.

[0131] It should be noted that the feature fusion layer can not only focus on intramodal features but also achieve intermodal feature fusion, thereby focusing on the correlation between features of all modes. In order for the model to capture intramodal and intermodal interactions, this invention proposes a feature fusion method based on hierarchical multimodal self-attention to realize intramodal and intermodal interactions. Figure 4 The structural diagram of this feature fusion method is shown. This attention mechanism consists of two levels. The first level of attention fuses drug SMILES features and drug molecule map features to obtain the first-level drug fusion feature. First-order protein fusion features are obtained by fusing protein sequence features and protein 2-mer features.

[0132] The multimodal feature fusion of protein sequence features, protein 2-mer sequence features, drug SMILES features, and drug molecule map features includes the following steps:

[0133] S31. The protein sequence features and protein 2-mer sequence features are spliced ​​together to obtain the protein combinatorial features, and the first-level protein fusion features are calculated based on the protein combinatorial features.

[0134] The process of concatenating protein sequence features and protein 2-mer sequence features to obtain protein combinatorial features, and calculating primary protein fusion features based on these combinatorial features, includes the following steps:

[0135] S311. Feature splicing is performed on protein sequence features and protein 2-mer sequence features to obtain protein combinatorial features, the expression of which is:

[0136]

[0137] In the formula, Indicates protein compositional characteristics; x mer Indicates the 2-mer sequence characteristics of the encoded protein; X seq Indicates protein sequence characteristics.

[0138] S312. Perform a linear transformation on the protein combinatorial features to obtain the protein query matrix, protein bond matrix, and protein value matrix, the expressions of which are:

[0139]

[0140]

[0141]

[0142] In the formula, Q protein Represents the protein query matrix, K protein Represents the protein bond matrix and V protein Represents the protein value matrix, Q seq Q represents the query matrix obtained by passing protein sequence features through a linear layer. mer K represents the query matrix obtained by passing protein 2-mer sequence features through a linear layer. seq K represents the bond matrix obtained by passing protein sequence features through a linear layer. mer V represents the bond matrix obtained by passing protein 2-mer sequence features through a linear layer. seq V represents the value matrix obtained by passing protein sequence features through a linear layer. mer This represents the value matrix obtained by passing the protein 2-mer sequence features through a linear layer. Protein indicates protein (the subscript or superscript "protein" indicates that it belongs to the protein family).

[0143] S313. Based on the protein query matrix and protein bond matrix, the protein attention matrix is ​​calculated, and its expression is as follows:

[0144]

[0145] In the formula, A protein The attention matrix is ​​calculated using protein sequence features and protein 2-mer sequence features. Q, K, and V represent the query matrix, key matrix, and value matrix, respectively. The subscripts seq, mer, and protein represent the protein sequence, protein 2-mer sequence, and protein (e.g., Q) respectively. protein Q represents the lookup matrix of proteins obtained by fusing protein 2-mer sequences and protein sequences. seq K represents a query matrix for protein sequences. seq (representing the bond matrix of the protein sequence), T denotes the matrix transpose, d k Represents the protein bond matrix K protein The embedding dimension, V protein V represents the value matrix of proteins, and V protein After derivation, it is equivalent to V mer and V seq These represent the value matrix of the protein 2-mer sequence and the value matrix of the protein sequence, respectively.

[0146] S314. Multiplying the protein attention matrix by the protein value matrix yields the first-level protein fusion feature, expressed as follows:

[0147]

[0148] S32. The drug SMILES features and drug molecule graph features are spliced ​​together to obtain drug combination features, and the first-level drug fusion features are calculated based on the drug combination features.

[0149] The process of concatenating drug SMILES features and drug molecule graph features to obtain drug combination features, and calculating primary drug fusion features based on drug combination features, includes the following steps:

[0150] S321. Perform feature splicing on the drug SMILES features and drug molecular graph features to obtain drug combination features.

[0151] S322. Perform a linear transformation on the drug combination features to obtain the drug query matrix, drug key matrix, and drug value matrix, the expression of which is:

[0152]

[0153]

[0154]

[0155] In the formula, Q drug K drug and V drug The drug query matrix, drug bond matrix, and drug value matrix are obtained by concatenating the drug molecule graph features and drug SMILES features, respectively, and then passing them through three linear layers. Q represents the... smiles K smiles and V smiles Let Q represent the drug SMILES query matrix, drug SMILES key moment reduction matrix, and drug SMILES value matrix obtained through drug SMILES features, respectively; graph K sraph and V graph These represent the drug molecule graph query matrix, drug molecule graph key matrix, and drug molecule graph value matrix obtained through drug molecule graph features, respectively; the derivation of Q is explained. drug K drug and V drug It also includes the query matrix, key matrix, and value matrix of drug SMILES and drug molecule graphs, which is Q. drug K drug and V drug Each contains Q graph and Q smiles K graph and K smiles and V graph and V smiles .

[0156] S323. Based on the drug query matrix and drug key matrix, the drug attention matrix is ​​calculated. Its expression is:

[0157]

[0158] In the formula, A drug It represents the drug attention matrix and simultaneously considers the correlations between SMILES-SMILES, SMILES-molecular graph, molecular graph-SMILES, and molecular graph-molecular graph.

[0159] S324. Multiply the drug attention matrix and the drug value matrix to obtain the first-level drug fusion feature.

[0160] The formula for calculating the primary drug fusion characteristics is as follows:

[0161]

[0162] In the formula, Indicates primary drug fusion characteristics; Q smiles K smiles and V smilesThese represent the drug SMILEs query matrix, drug SMILEs key matrix, and drug SMILEs value matrix, respectively; Q graph K graph and V graph These represent the drug molecule graph query matrix, drug molecule graph key matrix, and drug molecule graph value matrix, respectively; A drug Represents the drug attention matrix; V drug This represents the drug value matrix.

[0163] It should be noted that the calculation formula for the first-level drug fusion feature ignores the denominator in the attention matrix. The first-level drug fusion feature can be seen from the calculation formula of the softmax function and the first-level drug fusion feature. Simultaneously, it contains both drug SMILES features and molecular map features. Similarly, the same feature fusion method was used to obtain primary protein fusion features from protein sequence features and 2-mer sequence features.

[0164] S33. The drug target pair features are obtained by splicing the primary drug fusion features and the primary protein fusion features, and the final fusion features are calculated based on the drug target pair features.

[0165] The process of concatenating primary drug fusion features and primary protein fusion features to obtain drug target pair features, and calculating the final fusion features based on the drug target pair features, includes the following steps:

[0166] S331. The primary drug fusion features and primary protein fusion features are spliced ​​together to obtain drug target pair features.

[0167] It should be noted that the second level of attention integrates the features of the first-level drug in the same way as the first level of attention. Fusion characteristics of primary proteins Fusion was performed to obtain the drug target pair characteristic H. concat Its expression is:

[0168]

[0169] S332. Perform a linear transformation on the drug target pair features to obtain the drug target pair query matrix, drug target pair key matrix, and drug target pair value matrix, the expression of which is:

[0170]

[0171]

[0172]

[0173] S333. Based on the drug target pair query matrix and the drug target pair key matrix, the drug target pair attention matrix is ​​calculated, and its expression is as follows:

[0174]

[0175] S324. Multiply the drug target pair attention matrix with the drug target pair value matrix to obtain the final fused feature, the expression of which is:

[0176] H = AV;

[0177] S34. The final fused features are split and weighted fused to obtain the final modal features.

[0178] It should be noted that the final fusion features are decomposed to obtain the drug sMILES features D. smiles Drug molecular map feature D graph Protein sequence characteristics D seq and protein 2-mer sequence characteristics D mer Its expression is;

[0179] Z smiles =maxpooling(D smiles +X smiles )

[0180] Z graph =maxpooling(D graph +X graph )

[0181] Z seq =maxpooling(D seq +X seq )

[0182] Z mer =maxpooling(D mer +X mer );

[0183] The drug SMILES feature D is obtained by using the residual structure to fuse the final features. smiles Drug molecular map feature D graph Protein sequence characteristics D seg and protein 2-mer sequence characteristics D mer The extracted drug sMILEs feature X smiles Drug molecular map features X graph Protein sequence characteristics X seq and protein 2-mer sequence characteristics X mer Weighted fusion is performed to obtain four final modal features Z. smiles Zgraph Z seq and Z mer .

[0184] S4. Based on the multimodal feature fusion results, the feature vector of drug target pairs is obtained, and the feature vector of drug target pairs is used as the input of the classifier for DTI prediction.

[0185] The process of obtaining drug target pair feature vectors based on multimodal feature fusion results and using these feature vectors as input to a classifier for DTI prediction includes the following steps:

[0186] S41. Combine the fully connected layer with the RELU activation function and the regularization layer to construct a classifier.

[0187] S42. The final modal features are concatenated to obtain the drug target pair feature vector, and the drug target pair feature vector is loaded into the classifier for DTI prediction.

[0188] It should be noted that the classifier FFN is constructed using four fully connected layers, the ReLU activation function, and a dropout regularization layer. A dropout layer and a LeakyReLU activation function are added after each fully connected layer. The model's loss function uses binary cross-entropy, and its expression is as follows:

[0189]

[0190] In the formula, t i The label represents the sample, with a value of 0 or 1; p i Let represent the predicted probability, n represent the number of samples, i represent the nth sample, and L represent the binary cross-entropy.

[0191] Concatenate the four final modal features Z smiles Z graph Z seq and Z mer Obtain drug target pair features, use these features as input to a classifier, and output an interaction score between 0 and 1. The expression for this score is:

[0192] score=FFN(conCat(Z smiles Z graph Z seq Z mer ));

[0193] In the formula, concat represents the splicing operation; score represents the drug target pair features.

[0194] The invention will be further explained below with reference to specific experimental procedures:

[0195] It should be noted that this invention used three benchmark datasets—DrugBank, human, and C. elegans—to validate the effectiveness of the HMSA-DTI model. The human dataset contains 3364 positive examples and 3387 negative examples; the C. elegans dataset contains 3893 positive examples and 3893 negative examples. For the DrugBank dataset, this invention followed the dataset construction method of HyperAttentionDTI to obtain 17511 positive examples and 17511 negative examples. The specific details of the three datasets are shown in Table 1.

[0196] Table 1: Details of the Benchmark Dataset

[0197]

[0198] To verify the performance of the HMSA-DTI model proposed in this invention, the following was selected:

[0199] Four methods—HyperAttentionDTI, GIFDTI, CoaDTI, and MHSADTI—were compared as baselines.

[0200] HyperAttentionDTI: HyperAttentionDT takes protein sequences and drug SMILES as input, uses CNN to extract features, and captures the interaction between drugs and proteins through the HyperAttention mechanism.

[0201] GIFDTI: GIFDTI first uses CNN and Transformer to extract drug and protein features, and then pools them to obtain drug and protein feature vectors; then it uses the interaction matrix between protein and drug to obtain interaction features; finally, it uses drug feature vectors, protein feature vectors and interaction features to perform DTI prediction.

[0202] CoaDTI: CoaDT first uses Transformer to obtain protein features, then obtains the r-radius subgraph based on the drug molecule graph and uses it as input to GraphSage to obtain drug molecule features, and finally uses self-attention and cross-attention for feature fusion.

[0203] MHSADTI: MHSADTI transforms drug SMILES into several r-radius subgraphs and uses a GAT network to extract drug features. Then, it uses a Transformer to extract protein sequence features and introduces an attention mechanism to capture the interaction between proteins and drugs.

[0204] It should be noted that this invention uses four metrics to measure model performance: AUC (Area Under the Receiver Operating Characteristic), Precision, Recall, and Accuracy. All baseline models and the HMSA-DTI model proposed in this invention were run on a Linux server with an Intel(R) Core(TM) i7-11700K@3.60GHz and two GeForce RTX3090 GPUs. To accurately evaluate model performance, 10% of the data was used as the test set, and the remaining 90% was divided into training and validation sets, and evaluated using a 10-fold cross-validation method.

[0205] ①Performance comparison of the DrugBank dataset

[0206] This invention first compares the predictive performance of the proposed HMSA-DTI model with four baseline models on the DrugBank dataset. The experimental results are shown in Table 2. The HMSA-DTI model...

[0207] The HMSA-DTI model significantly outperforms other models in AUC, Precision, Recall, and Accuracy metrics. Compared to the state-of-the-art baseline model GIFDTI, the HMSA-DTI model improves AUC, Precision, Recall, and Accuracy by 1.38%, 2.14%, 0.41%, and 1.84%, respectively. The superior performance of the proposed HMSA-DTI model is attributed to its elimination of the D-MPNN readout stage, preserving the node-level features of the drug molecule graph. This facilitates the full integration of multimodal drug and protein data, enhancing feature representation capabilities and thus improving DTI prediction performance. To more intuitively compare the performance of each model, this invention plots the precision-recall curve (PR) and receiver operating characteristic (ROC) curves of the baseline model and the HMSA-DTI model on the DrugBank dataset. Figures 5-6 As shown in the PR and ROC curves, the HMSA-DTI model proposed in this invention has significantly higher DTI prediction performance than other baseline models.

[0208] ②Performance comparison of the human dataset

[0209] This invention further compared the HMSA-DTI model with four baseline models on the human dataset, and the experimental results are shown in Table 3. On the human dataset, the HMSA-DTI model outperformed other models in all four metrics: AUC, Precision, Accuracy, and Recall. Compared with the HyperAttentionDTI model, which performed best in AUC, Precision, and Accuracy, the HMSA-DTI model improved by 1.32%, 0.99%, and 1.26%, respectively. Compared with the CoaDTI model, which performed best in Recall, the HMSA-DTI model improved by 1.33%. The experimental results demonstrate that the HMSA-DTI model proposed in this invention also exhibits excellent predictive performance on the human dataset.

[0210] ③ Performance on the C.elegans dataset

[0211] To further verify the performance of the HMSA-DTI model, this invention was compared on the C. elegans dataset. The experimental results are shown in Table 4. The proposed HMSA-DTI model performs excellently in AUC, Accuracy, and Recall, improving by 0.18%, 0.19%, and 0.46% respectively compared to the best baseline model, HyperAttentionDTI. However, it lags behind HyperAttentionDTI by 0.47% in Precision. The experimental results on the DrugBank, human, and C. elegans benchmark datasets demonstrate that the proposed HMSA-DTI model outperforms other baseline models in multiple metrics, making it a highly competitive model.

[0212] Furthermore, ablation experiments were conducted on the DrugBank dataset to verify the effectiveness of each module. The results are shown in Table 5. To verify the effectiveness of the proposed hierarchical multimodal self-attention-based feature fusion method, it was replaced with a concatenation-based feature fusion method to obtain the HMSA-DTI-Concat model. Experimental results show that the HMSA-DTI model using hierarchical multimodal self-attention outperforms the HMSA-DTI-Concat model in AUC, ACC, Prec, and Recall metrics by 0.61%, 0.72%, 0.76%, and 0.64%, respectively. This indicates that the hierarchical multimodal self-attention mechanism is beneficial for the fusion of multimodal features and can improve the feature expression capabilities of drugs and proteins. To verify the impact of different combinations of drug molecule representations on prediction performance, the combination of extended connectivity fingerprint (ECFP) fingerprint and molecular graph was used to replace the combination of SMILES and molecular graph to obtain the HMSA-DTI-ECFP model.

[0213] Experimental results show that the HMSA-DTI model using a combination of SMILES and molecular diagrams...

[0214] The AUC, ACC, Prec, and Recall metrics were 0.82%, 1.36%, 1.92%, and 0.37% higher than the HMSA-DTI-ECFP model, respectively. To verify the advantages of combining global and local protein features, the 2-mer sequence of the protein in the HMSA-DTI model was removed, resulting in the HMSA-DTI-Without2mer model. Experimental results show that the HMSA-DTI model, by combining protein sequence information and 2-mer sequence information, improved the AUC, ACC, Prec, and Recall metrics by 0.71%, 0.71%, 1.3%, and 0.75%, respectively. This demonstrates that multi-source data fusion based on hierarchical multimodal self-attention can effectively enhance the representational power of fused features.

[0215] It's important to note that when representing proteins using k-mer sequences, the choice of k value affects the model's predictive performance. A smaller k value indicates that the model focuses more on the local structure of the protein, while a larger k value indicates that the model focuses more on the global structure. By comparing the impact of different k values ​​on predictive performance, the corresponding vocabulary sizes for k = 1, 2, and 3 are 22, 22, and 22, respectively. 2 =484,22 3 =10648, when k=4, the size of the vocabulary is 22. 4 =234256, which leads to an exponential increase in model complexity. Therefore, this invention only verifies the performance when k=1, 2, and 3. Experimental results are as follows: Figure 7 As shown, when k=2, the model performs best in terms of ACC (Accuracy) and AUC (Area Under Receiver Operating Characteristic). Therefore, this invention selects 2-mer sequences to represent proteins.

[0216] During the training of graph neural networks, as the number of network layers increases, the hidden states of nodes tend to converge to the same value, resulting in oversmoothing and impairing the representation ability of features. This invention evaluates the impact of the number of graph neural network layers on the model's predictive performance. Experimental results are as follows: Figure 8 As shown, when the number of network layers is equal to 2, the model performs best in terms of ACC (Accuracy) and AUC (Area Under the Receiver Operating Characteristic). Figure 8 The effect of the number of layers in a D-MPNN network on ACC (Accuracy) and AUC (Area Under the Receiver Operating Characteristic).

[0217] To verify the robustness of the model, this invention randomly selects a target protein and a drug from the DrugBank dataset as the test target and drug, respectively. Then, data related to the selected drug and target protein are removed from the DrugBank dataset, and the remaining data is used as the training set to evaluate the model's robustness. In this experiment, the drug estradiol acetate (DrugBank ID: DB13952) and the target protein γ-aminobutyric acid receptor subunit rho-3 (Uniprot ID: A8MPY1) were randomly selected as test cases. For estradiol acetate, this invention obtained 10 proteins that can interact with it from the DrugBank dataset, paired them to obtain 10 positive examples, and paired the remaining 4244 proteins with estradiol acetate to obtain 4244 drug-target pairs as negative examples. The positive and negative examples were mixed to form a test set. Then, the candidate target proteins were ranked according to the predicted interaction scores, and the top 20 target proteins with the highest interaction scores were selected. The experimental results are shown in Table 6. The Recall value on the test set was 1, indicating that the model successfully predicted all positive examples, and 5 of these positive examples appeared in the top 20. The same approach was used to process the target protein γ-aminobutyric acid receptor subunit p-3 to obtain a test set consisting of 15 positive examples and 6630 negative examples. The experimental results are shown in Table 7. The HMSA-DTI model was also able to predict all positive examples, and 3 of these positive examples appeared in the top 20.

[0218] Table 2: Performance comparison of HMSA-DTI model and baseline model on DrugBank dataset

[0219]

[0220] Table 3: Performance comparison of HMSA-DTI model and baseline model on the human dataset

[0221]

[0222] Table 4: Comparison of HMSA-DTI model and baseline model on the C.elegans dataset

[0223]

[0224] Table 5: Ablation experiments on the DrugBank dataset

[0225]

[0226] Table 6: Predicted candidate targets for the drug estradiol acetate

[0227]

[0228]

[0229] Table 7: Predicted candidate drugs targeting dihydrofolate synthase

[0230]

[0231]

[0232] In summary, leveraging the technical solutions described above, the DTI prediction method based on hierarchical multimodal self-attention graph neural networks provided by this invention uses drug SMILES, drug molecule graphs, protein sequences, and protein 2-mer sequences as inputs to HMSA-DTI. It captures inter-modal and intra-modal interactions through a hierarchical multimodal attention mechanism, achieving deep fusion of drug and protein multimodal features. Simultaneously, it is evaluated using 10-fold cross-validation on three benchmark datasets: DrugBank, human, and C. elegans, and compared with other advanced DTI prediction models, thus verifying that the HMSA-DTI prediction model has significant advantages and strong competitiveness across multiple metrics. The DTI prediction method based on hierarchical multimodal self-attention graph neural networks provided by this invention integrates multimodal data such as SMILES, drug molecule graphs, protein sequences, and 2-mer sequences to help the model better understand and characterize the features and interactions between drugs and proteins. It uses node-level features of D-MPNN instead of graph-level features and can simultaneously model intra-modal and inter-modal interactions to improve the discriminability and robustness of the fused features.

[0233] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A DTI prediction method based on hierarchical multimodal self-attention graph neural network, characterized in that, The DTI prediction method based on hierarchical multimodal self-attention map neural network includes the following steps: S1. Extract protein sequence features and protein 2-mer sequence features using a convolution module; S2. Extract drug SMILES features and drug molecule graph features using convolutional modules and graph neural networks; S3. Multimodal feature fusion of protein sequence features, protein 2-mer sequence features, drug SMILES features, and drug molecule map features; S4. Based on the multimodal feature fusion results, obtain the drug target pair feature vector, and use the drug target pair feature vector as the input of the classifier for DTI prediction; The multimodal feature fusion of protein sequence features, protein 2-mer sequence features, drug SMILES features, and drug molecule map features includes the following steps: S31. The protein sequence features and protein 2-mer sequence features are spliced ​​together to obtain the protein combinatorial features, and the first-level protein fusion features are calculated based on the protein combinatorial features. S32. The drug SMILES features and drug molecule graph features are spliced ​​together to obtain drug combination features, and the first-level drug fusion features are calculated based on the drug combination features. S33. The drug target pair features are obtained by splicing the primary drug fusion features and the primary protein fusion features, and the final fusion features are calculated based on the drug target pair features. S34. The final fused features are split and weighted fused separately to obtain the final modal features; The process of concatenating protein sequence features and protein 2-mer sequence features to obtain protein combinatorial features, and calculating primary protein fusion features based on these combinatorial features, includes the following steps: S311. Feature splicing is performed on protein sequence features and protein 2-mer sequence features to obtain protein combinatorial features; S312. Perform linear transformations on the protein combinatorial features to obtain the protein query matrix, protein bond matrix, and protein value matrix. S313. Based on the protein query matrix and the protein bond matrix, the protein attention matrix is ​​calculated; S314. Multiply the protein attention matrix by the protein value matrix to obtain the first-level protein fusion feature; The process of concatenating drug SMILES features and drug molecule graph features to obtain drug combination features, and calculating primary drug fusion features based on drug combination features, includes the following steps: S321. Perform feature splicing on the drug SMILES features and drug molecule map features to obtain drug combination features; S322. Perform a linear transformation on the drug combination features to obtain the drug query matrix, drug key matrix, and drug value matrix; S323. Based on the drug query matrix and drug key matrix, calculate the drug attention matrix; S324. Multiply the drug attention matrix and the drug value matrix to obtain the first-level drug fusion feature.

2. The DTI prediction method based on hierarchical multimodal self-attention graph neural network according to claim 1, characterized in that, The extraction of protein sequence features and protein 2-mer sequence features using a convolutional module includes the following steps: S11. Encode the original protein sequence and convert the encoded original protein sequence into a marker sequence; S12. Map the labeled sequence into the embedding space and use convolutional layers to extract protein sequence features from the labeled sequence; S13. Divide the original protein sequence into sequence fragments and convert each sequence fragment into a 2-mer coding sequence; S14. Map the 2-mer coding sequence into the embedding space and use a convolutional layer to extract the protein 2-mer sequence features from the 2-mer coding sequence.

3. The DTI prediction method based on hierarchical multimodal self-attention graph neural network according to claim 2, characterized in that, The extraction of drug SMILES features and drug molecule graph features using a convolutional module combined with a graph neural network includes the following steps: S21. Encode and fill the drug SMILES characters to obtain drug molecule vectors, and use the drug molecule vectors as input to the graph neural network. S22. Map the drug molecule vectors into the embedding space and use the convolution module to extract drug SMILES features from the drug molecule vectors; S23. Convert the drug SMILES into a drug molecule graph, and obtain the initial features of the nodes and the initial features of the edges from the drug molecule graph. The initial characteristics of the node include atomic number, number of chemical bonds, formal charge, chirality, number of hydrogen atoms connected to the atom, hybridization, whether it participates in the formation of an aromatic ring, and atomic mass. The initial characteristics of the edges include bond type, conjugacy, bond position, and geometric heterogeneity; S24. Based on the initial features of nodes and edges, extract drug molecule graph features using a directed message passing network.

4. The DTI prediction method based on hierarchical multimodal self-attention graph neural network according to claim 3, characterized in that, The extraction of drug molecule graph features using a directed message passing network based on the initial features of nodes and edges includes the following steps: S241. Take the initial features of the node as the initial hidden state of the node, and associate the edges with the initial features of the node. The initial features are transformed into edges The initial hidden state; S242. Use the average function to combine the initial features of the node, the initial features of the neighboring nodes, and the edges in the t-th layer. Aggregate the hidden states to obtain the edges. Message delivery in the (t+1)th layer; S243, Side Message passing at level t+1 and edges at level t The hidden states are used as input to the fully connected neural network to obtain the edges in the (t+1)th layer. The hidden state; S244. Sum the hidden states of all edges connected to the node to obtain the node's transmitted message. At the same time, concatenate the node's initial features and the node's transmitted message and transmit them to the linear layer to obtain the node's final hidden state. S245. By splicing together the final hidden states of all nodes, the drug molecule graph features are obtained.

5. The DTI prediction method based on hierarchical multimodal self-attention graph neural network according to claim 1, characterized in that, The formula for calculating the primary drug fusion characteristics is as follows: ； In the formula, Indicates primary drug fusion characteristics; , and These represent the drug SMILES query matrix, the drug SMILES key matrix, and the drug SMILES value matrix, respectively. , and These represent the drug molecule graph query matrix, drug molecule graph key matrix, and drug molecule graph value matrix, respectively. Represents the drug attention matrix; This represents the drug value matrix.

6. The DTI prediction method based on hierarchical multimodal self-attention graph neural network according to claim 5, characterized in that, The process of concatenating primary drug fusion features and primary protein fusion features to obtain drug target pair features, and calculating the final fusion features based on the drug target pair features, includes the following steps: S331. The primary drug fusion features and the primary protein fusion features are spliced ​​together to obtain drug target pair features. S332. Perform linear transformations on the drug target pair features to obtain the drug target pair query matrix, drug target pair key matrix, and drug target pair value matrix. S333. Based on the drug target pair query matrix and the drug target pair key matrix, calculate the drug target pair attention matrix. S324. Multiply the drug target pair attention matrix with the drug target pair value matrix to obtain the final fused feature.

7. The DTI prediction method based on hierarchical multimodal self-attention graph neural network according to claim 6, characterized in that, The process of obtaining drug target pair feature vectors based on multimodal feature fusion results and using these feature vectors as input to a classifier for DTI prediction includes the following steps: S41. Construct a classifier by combining a fully connected layer with the RELU activation function and a regularization layer; S42. The final modal features are concatenated to obtain the drug target pair feature vector, and the drug target pair feature vector is loaded into the classifier for DTI prediction.

Images