A drug-drug interaction prediction method based on multi-channel feature fusion

Abstract

The present invention discloses a drug-drug interaction prediction method based on multi-channel feature fusion, which utilizes the diverse features of drug data, combines a multi-channel feature fusion module and a cross-attention mechanism, effectively combines drug molecular substructures, drug similarity matrices and molecular fingerprints, and then uses a KAN neural network to achieve efficient and accurate DDI prediction. This method efficiently integrates multiple feature information data of drugs, such as drug substructures, drug similarities, and drug molecular fingerprint data, to achieve efficient and accurate multi-channel feature fusion drug-drug reaction prediction; the MCF-DDI model proposed in the present invention shows excellent performance in interpretable experiments and comparative experiments, not only the evaluation indicators and stability are improved to varying degrees, but also the key substructures in drug molecules can be identified and the importance of these substructures in drug interaction prediction can be verified, providing strong theoretical support for drug research and development and clinical application.

Description

Technical Field

[0001] The present invention belongs to the technical field of bioinformatics, and in particular relates to a drug-drug interaction prediction method based on multi-channel feature fusion. Background Art

[0002] For complex diseases, combination therapy achieves synergistic effects but also increases the risk of adverse drug-drug interactions. Therefore, in order to maximize the synergistic effects of drugs while reducing the impact of adverse reactions on patients, it is crucial to accurately predict and evaluate possible interactions between drugs. In recent decades, researchers have conducted in-depth studies on the mechanisms and influencing factors of drug interactions, but due to the involvement of multiple complex biochemical properties, such as targets, enzymes, and transporters, effective prediction and management of drug interactions remains a challenging problem.

[0003] Drug-drug interaction (DDI) refers to the mutual influence that may occur when two or more drugs are used at the same time, which may manifest as enhanced efficacy or adverse reactions, or even lead to unexpected side effects. The combined use of multiple drugs may achieve the same effect as high doses at lower doses by enhancing the therapeutic effect, thereby reducing treatment costs and potential toxic side effects. However, if DDI leads to reduced efficacy or adverse reactions, patients may need higher doses or more frequent medications to achieve therapeutic effects, and may even have a direct impact on life and health. Traditional wet experimental methods, such as in vitro experiments and clinical trials, although capable of conducting a limited number of drug combination studies, are time-consuming and costly, and it is difficult to cover all possible DDI situations. In contrast, the emergence of computational methods has greatly accelerated the large-scale prediction and screening process, while being able to comprehensively consider the potential interactions of multiple drug combinations, providing healthcare professionals with a more efficient and personalized way to formulate treatment plans, thereby improving medical efficiency and success rate. At present, DDI prediction methods can be roughly divided into four types: based on unstructured data, based on similarity, based on networks, and based on multi-model or multi-view feature fusion.

[0004] In the early research of DDI, researchers usually rely on unstructured data to obtain DDI information. However, these methods are limited by the reliance on the accuracy of the data set and fail to fully capture the complex interaction patterns and relationships between drugs. In order to overcome the limitation of strong dependence between different data sets, researchers began to explore the extraction of similarity features from within or between molecules to effectively capture the association between drugs. However, these methods may lack sufficient scalability when dealing with large-scale data sets, especially when the side effect information is incomplete. With the rise of deep learning and the development of computing power, researchers prefer to use network-based methods to solve the above problems. By constructing different networks to simulate the relationship between various drugs, the association between drug molecules or between molecules can be effectively captured, thereby solving the problem of data non-scalability. The structural features of drug molecules can be represented by the corresponding simplified molecular input line entry system (SMILES) string and further converted into a molecular graph suitable for input into graph neural networks (GNN). At present, GNN-based methods are roughly divided into feature extraction from drug molecular graphs and feature extraction from DDI networks. MFFGNN considers drugs as nodes in the DDI network, combines molecular graphs and SMILES sequences, and predicts DDIs through link prediction. GMPNN-CS introduces an information transfer neural network that learns from the molecular graph representation of drugs and captures chemical substructures of different sizes and shapes for DDI prediction between drug pairs. GGI-DDI is conceptualized at three different levels of granularity, introducing a granular computing paradigm for DDI prediction. To further address the problem of data scarcity, researchers have also tried to use the rich information in biomedical networks to predict DDIs, and have achieved remarkable results in the task of DDI prediction for emerging drugs. The design limitation of traditional graph neural networks is that only a single edge can be connected between nodes, and the complex physical and chemical relationships between drugs cannot be fully utilized. To solve this problem, hypergraph neural networks are proposed for DDI prediction.

[0005] Although the above models have achieved improved results, they still have some limitations. First, existing graph- or similarity-based methods only consider the physicochemical properties or similarity features of molecules, ignoring the potential correlation between other multimodal features and DDIs events. Second, these methods often fail to fully capture the complex features within or between molecules, resulting in potential information loss. Third, traditional feature combination strategies fail to fully consider the importance of drug feature differences to DDI events.

[0006] Many researchers now tend to adopt DDI prediction methods that integrate multimodal or multi-view features to address these limitations. The method of feature fusion using attention mechanism enables the model to focus on the key parts of the input data and improves interpretability through dynamic weight adjustment. When combined with the multimodal or multi-view features of drugs, a more comprehensive understanding of drug characteristics can be obtained, resulting in more accurate results. DMFDDI integrates drug molecular graphs, DDI networks, and biochemical similarity features to predict DDI, uses GNN and sparse graph convolution to obtain the topological structure information of molecular graphs and DDI networks, and finally fuses the features of different modalities through attention mechanism. Recent related studies have also demonstrated the efficiency and accuracy of these methods, including DSN-DDI and PEB-DDI. Although various feature fusion models have performed well in routine experiments or new drug testing, there is currently no model that can simultaneously solve the integration of local and global molecular features and the potential relationship information between drug molecules. In addition, traditional feature fusion strategies lack the ability to integrate multi-level feature interactions and multi-scale information. Summary of the invention

[0007] In view of the above-mentioned deficiencies in the prior art, the drug-drug interaction prediction method based on multi-channel feature fusion provided by the present invention solves the problem that the existing related methods cannot effectively integrate local and global molecular features and potential relationship information between drug molecules, and the traditional feature fusion strategy lacks multi-level feature interaction and multi-scale information fusion capabilities.

[0008] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a drug-drug interaction prediction method based on multi-channel feature fusion, comprising the following steps:

[0009] S1. Preprocessing benchmark datasets of different sizes containing different drug responses;

[0010] S2. constructing a drug-drug reaction prediction model and training it using the benchmark dataset;

[0011] The drug-drug reaction prediction model comprises a feature extraction module, a feature fusion module and a predictor connected in sequence; wherein the feature extraction module is used to extract drug substructure features, drug molecular fingerprint features and drug similarity matrix features, and the feature fusion module is used to fuse the features extracted by the feature extraction module through a cross attention mechanism to obtain fused drug features; and the predictor is used to predict the probability of drug-drug interaction for the fused drugs;

[0012] S3. Input any two target drugs into the trained drug-drug reaction prediction model and output the predicted probability of drug-drug interaction.

[0013] Further, in the step S2, in the drug-drug reaction prediction model, the drug substructure feature is extracted through a drug substructure feature extraction channel;

[0014] In the drug substructure feature extraction channel, the drug molecular structure features are extracted by using the graph neural network GNN. The specific extraction method is as follows:

[0015] S2-A1, represent the origin in the drug molecule as a node, and the chemical bonds as an edge, and construct a drug molecule graph;

[0016] S2-A2, determine the bond-order characteristics of all edges in the drug molecule graph;

[0017] S2-A3, calculate the importance score of each bond-level feature in the entire drug molecule graph;

[0018] S2-A4, according to the importance score of the bond-level features, the drug molecule graph is pooled to obtain the drug molecule graph pooling features;

[0019] S2-A5, calculate the attention weight score of each layer of substructure according to the pooling characteristics of the drug molecule graph;

[0020] S2-A6, normalize the attention weight scores of all layer substructures, and combine the bond-level features of each layer of the graph neural network GNN to determine the final bond-level features of the edges in the drug molecule graph;

[0021] S2-A7. According to the characteristics of each node in the drug molecule graph and the final bond-level characteristics of its adjacent edges, the substructure representation of each node is obtained as the drug substructure feature.

[0022] Furthermore, in step S2-A2, the key-level feature of the edge It is expressed as:

[0023]

[0024] In the formula, represents the characteristics of the starting node a, represents the characteristics of the terminal node b, Represents the characteristics of the bond between nodes a and b;

[0025] Step S2-A4, drug molecule graph pooling features It is expressed as:

[0026]

[0027] In the formula, represents the pooling layer, Indicates the current number of GNN layers, Represents the matrix composed of all key-level features in the current layer t of GNN, Represents the adjacency matrix of the current layer graph;

[0028] In step S2-A5, the attention weight score It is expressed as:

[0029]

[0030] In the formula, represents the weight vector of the tth layer, represents the tanh activation function, W represents the learnable matrix, represents the bias term;

[0031] In step S2-A6, the final bond-level feature It is expressed as:

[0032]

[0033] In the formula, represents the key-level feature representation between nodes a and b after processing at the tth layer, express Activation function, T represents the number of network layers of the substructure;

[0034] In step S2-A7, the substructure of each node is represented by It is expressed as:

[0035]

[0036] In the formula, represents the original feature representation of node a, represents a nonlinear function, Represents the set of all neighbor nodes of node a.

[0037] Furthermore, in the step S2, in the drug-drug reaction prediction model, the drug molecular fingerprint feature is extracted through a molecular fingerprint feature extraction channel;

[0038] In the molecular fingerprint feature extraction channel, the drug molecular fingerprint feature is extracted through a one-dimensional convolutional network, and the extraction method is specifically as follows:

[0039] S2-B1. Converting drug molecules into molecular fingerprints:

[0040] S2-B2, input the molecular fingerprint into a one-dimensional convolutional network to generate a feature map;

[0041] The one-dimensional convolutional network includes three sequentially connected convolutional layers with the same convolution kernel size and different numbers;

[0042] S2-B3, the generated feature map is aggregated and information filtered through the pooling layer to obtain the molecular fingerprint feature.

[0043] Furthermore, in step S2-B2, in the one-dimensional convolutional network, the output sequence length of the l+1th convolutional layer is It is expressed as:

[0044]

[0045] In the formula, represents the length of the l-th layer sequence, Indicates the length of the padding. Represents the length of the convolution kernel, and s represents the step size of each movement of the convolution kernel;

[0046] For each element in the output sequence, it is represented as:

[0047]

[0048]

[0049] In the formula, represents the input of the l+1th layer, represents the feature vector value output by the l+1th layer, represents the bias term of the convolutional layer, represents the weight corresponding to the kth channel in the l+1 layer, i represents the index of the output feature map on each element, Indicates an incrementing parameter, ;

[0050] In the step S2-B3, the operation of feature aggregation and information filtering performed by the pooling layer is expressed as:

[0051]

[0052] In the formula, represents the output value of the pooling, Indicates setting parameter value.

[0053] Furthermore, in the step S2, in the drug-drug reaction prediction model, the drug similarity feature matrix is ​​extracted based on the two properties of drug molecular structure and target through a similarity matrix feature extraction channel, and the extraction method is specifically as follows:

[0054] S2-C1, Morgan fingerprint is used to extract the structural features of drug molecules, and then the Jaccard coefficient is used to calculate the similarity scores of two molecular fingerprints, and then a similarity matrix based on molecular substructure is obtained;

[0055] S2-C2, construct a drug-target heterogeneous graph network through the drug-target dataset, use the Node2Vec algorithm to learn the embedding of nodes in the network, and map each node in the drug-target heterogeneous graph network to a low-dimensional space to obtain the low-dimensional embedding of each node;

[0056] S2-C3, extract the low-dimensional embedding representation of the drug based on the node-based low-dimensional embedding, and obtain the cosine similarity of the low-dimensional embedding representation of the two drugs, and then obtain the drug-target based similarity matrix;

[0057] S2-C4, concatenate the similarity matrix based on molecular substructure and the similarity feature matrix based on drug-target to obtain the drug similarity feature matrix.

[0058] Furthermore, in step S2-C1, the similarity score I of the molecular fingerprint is expressed as:

[0059]

[0060] In the formula, and Represents the first The value of the element;

[0061] In the step S2-C2, a random walk strategy is used in the Node2Vec algorithm to balance the sampling of BFS and DFS, which is expressed as:

[0062]

[0063] In the formula, represents the current sampling node, It means that the next sampling node is assumed to be , Indicates that the next sampling node is The probability of Representation Node Jump to the next node The probability of represents the normalization constant, Represents the set of edges in a graph.

[0064] Furthermore, in step S2, in the feature fusion module of the drug-drug reaction prediction model, a multi-head attention mechanism is used to calculate the cross-attention between the drug substructure features, drug molecular fingerprint features and drug similarity matrix features, and feature fusion is performed based on them to obtain fused drug features.

[0065] Further, in the step S2, in the drug-drug response prediction model, the predictor is a KAN predictor, which is expressed as:

[0066]

[0067] In the formula, represents the function representation of the KAN predictor, … represents the number of nodes in the 0th to L-1th KAN layers, … represents the activation function connecting two adjacent KAN layers, … represents the activation value of neurons in the 0th to L-1th KAN layers, … represents the node index, and the subscript L represents the KAN layer index.

[0068] The beneficial effects of the present invention are:

[0069] (1) The present invention utilizes the diverse features of drug data, combines a multi-channel feature fusion module and a cross-attention mechanism, effectively combines drug molecular substructures, drug similarity matrices and molecular fingerprints, and then uses a KAN neural network to achieve efficient and accurate DDI prediction; This method efficiently integrates multiple feature information data of drugs, such as drug substructures, drug similarities, and drug molecular fingerprint data, and achieves efficient and accurate multi-channel feature fusion drug-drug reaction prediction;

[0070] (2) The MCF-DDI model proposed in the present invention demonstrated excellent performance in interpretable experiments and comparative experiments. Not only were the evaluation indicators and stability improved to varying degrees, but it was also able to identify key substructures in drug molecules and verify the importance of these substructures in the prediction of drug interactions. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 This is a flow chart of the drug-drug interaction prediction method based on multi-channel feature fusion provided by the present invention.

[0072] Figure 2 This is a diagram of the MCF-DDI framework provided by the present invention.

[0073] Figure 3 To calculate drug similarity through drug-target heterogeneous graph network.

[0074] Figure 4 This is a diagram of the MCF-DDI model ablation experiment provided by the present invention.

[0075] Figure 5A comparison chart of the MCF-DDI model provided by the present invention and several different variants.

[0076] Figure 6 Heat map of drug substructure importance under the setting of Experiment 1 provided by the present invention.

[0077] Figure 7 A comparison chart of the prediction scores under different masking strategies provided by the present invention. DETAILED DESCRIPTION

[0078] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0079] The present invention provides a method for predicting drug-drug interaction based on multi-channel feature fusion. Figure 1 As shown, the following steps are included:

[0080] S1. Preprocessing benchmark datasets of different sizes containing different drug responses;

[0081] S2. constructing a drug-drug reaction prediction model and training it using the benchmark dataset;

[0082] like Figure 2 As shown, the drug-drug interaction prediction model (MCF-DDI) includes a feature extraction module, a feature fusion module and a predictor connected in sequence; wherein the feature extraction module is used to extract drug substructure features, drug molecular fingerprint features and drug similarity matrix features, and the feature fusion module is used to fuse the features extracted by the feature extraction module through a cross-attention mechanism to obtain fused drug features; the predictor is used to predict the probability of drug-drug interaction for the fused drugs;

[0083] Specifically, in Figure 2 In the figure, A) is a schematic diagram of the data fusion process, showing how the features of the three input channels are fused. B) is a simplified diagram of the KAN predictor structure, describing the composition of the predictor. C) is a distribution diagram of the B-spline function piecewise fitting curve in the KAN predictor. D) is a simplified diagram of the drug-target similarity matrix calculation process. E) is a flowchart of the MCF-DDI prediction task, outlining the entire prediction process;

[0084] S3. Input any two target drugs into the trained drug-drug reaction prediction model and output the predicted probability of drug-drug interaction.

[0085] In step S1 of the embodiment of the present invention, three benchmark data sets of different sizes are used, and the size of the data set is determined according to the number of drugs included.

[0086] Exemplarily, DrugBank, ZhangDDI and ChCh-Miner benchmark data sets are collected in this embodiment. Using data sets of different sizes for model evaluation helps to measure the accuracy and generalization of the model. The first is the small-scale ZhangDDI benchmark data set, which contains 572 drugs and 95246 DDI triples, each triplet consisting of a pair of drugs and their interaction types. Then there is the medium-sized ChCh-Miner benchmark data set, including 1514 drugs and 33669 DDI triplets. Finally, the DrugBank benchmark data set contains 1706 drugs and 191808 DDI triplets, with 86 different DDI types. The DrugBank benchmark data set generates negative samples with reference to the method in GMPNN-cs. Since the ZhangDDI and ChCh-Miner benchmark data sets come with negative samples, the negative sample generation step will not be performed. For molecular fingerprint data, 881-dimensional structure vectors are extracted from the PubChem compound database. In the DrugBank database, since molecular fingerprints could not be obtained for some drugs, 1705 drugs and 191799 DDI triplets remained after data cleaning, and the molecular fingerprint data of the remaining two data sets were fully obtained.

[0087] In step S2 of the embodiment of the present invention, in the drug-drug reaction prediction model, the drug substructure feature is extracted through a drug substructure feature extraction channel;

[0088] In the drug substructure feature extraction channel, the drug molecular structure features are extracted by using the graph neural network GNN. The specific extraction method is as follows:

[0089] S2-A1, represent the origin in the drug molecule as a node, and the chemical bonds as an edge, and construct a drug molecule graph;

[0090] S2-A2, determine the bond-order characteristics of all edges in the drug molecule graph;

[0091] S2-A3, calculate the importance score of each bond-level feature in the entire drug molecule graph;

[0092] S2-A4, according to the importance score of the bond-level features, the drug molecule graph is pooled to obtain the drug molecule graph pooling features;

[0093] S2-A5, calculate the attention weight score of each layer of substructure according to the pooling characteristics of the drug molecule graph;

[0094] S2-A6, normalize the attention weight scores of all layer substructures, and combine the bond-level features of each layer of the graph neural network GNN to determine the final bond-level features of the edges in the drug molecule graph;

[0095] S2-A7. According to the characteristics of each node in the drug molecule graph and the final bond-level characteristics of its adjacent edges, the substructure representation of each node is obtained as the drug substructure feature.

[0096] In the above step S2-A1, when determining the bond-level features, in traditional GNNs, information propagation is achieved through information exchange between nodes, and the representation of the entire graph is obtained by iteratively updating the node states and aggregating this information. However, this method usually requires more attention to the representation of the bond-level features between atoms. In this embodiment, bond is used for message passing to obtain the bond-level features of the edge. It is expressed as:

[0097]

[0098] In the formula, represents the characteristics of the starting node a, represents the characteristics of the terminal node b, A feature representing the bond between nodes a and b.

[0099] In the above steps S2-A4, the drug molecule graph pooling features It is expressed as:

[0100]

[0101] In the formula, represents the pooling layer, Indicates the current number of GNN layers, Represents the matrix composed of all key-level features in the current layer t of GNN, Represents the adjacency matrix of the current level graph.

[0102] In the above steps S2-A5, the attention weight score It is expressed as:

[0103]

[0104] In the formula, represents the weight vector of the tth layer, represents the tanh activation function, W represents the learnable matrix, Represents the bias term.

[0105] In the above step S2-A6, the final bond-level features It is expressed as:

[0106]

[0107] In the formula, represents the key-level feature representation between nodes a and b after processing at the tth layer, express Activation function, T represents the number of network layers of the substructure;

[0108] In the above step S2-A7, the substructure of each node is represented by It is expressed as:

[0109]

[0110] In the formula, represents the original feature representation of node a, represents a nonlinear function, Represents the set of all neighbor nodes of node a.

[0111] In this embodiment, by extracting molecular graph features through graph neural network GNN, the substructure features of drugs can be effectively extracted. Since drug analysis consists of many basic substructures expressing various physical and chemical properties, designing substructure extraction channels can better extract key structural features of molecules. This method not only extracts and integrates substructure features of drug molecules at different scales, but also makes the final feature expression more comprehensive and rich. In addition, since the features learned from different layers have different levels of abstraction, the fusion of features extracted from multiple layers of GNN helps reduce the risk of overfitting.

[0112] In step S2 of the embodiment of the present invention, in the drug-drug reaction prediction model, the drug molecular fingerprint feature is extracted through a molecular fingerprint feature extraction channel;

[0113] In the molecular fingerprint feature extraction channel, the drug molecular fingerprint feature is extracted through a one-dimensional convolutional network, and the extraction method is specifically as follows:

[0114] S2-B1. Converting drug molecules into molecular fingerprints:

[0115] S2-B2, input the molecular fingerprint into a one-dimensional convolutional network to generate a feature map;

[0116] The one-dimensional convolutional network includes three sequentially connected convolutional layers with the same convolution kernel size and different numbers;

[0117] S2-B3, the generated feature map is aggregated and information filtered through the pooling layer to obtain the molecular fingerprint feature.

[0118] In this embodiment, since the molecular fingerprint obtained from PubChem is one-dimensional, a one-dimensional convolution (1D-CNN) is used to extract features. For example, in this embodiment, the convolution kernel size in the one-dimensional convolution network is uniformly set to 8, and each convolution kernel slides on the input data to generate a series of feature maps through one-dimensional convolution. Based on this, in the above step S2-B2, in the one-dimensional convolution network, the output sequence length of the l+1th convolution layer is It is expressed as:

[0119]

[0120] In the formula, represents the length of the l-th layer sequence, Indicates the length of the padding. Represents the length of the convolution kernel, and s represents the step size of each movement of the convolution kernel;

[0121] For each element in the output sequence, it is represented as:

[0122]

[0123]

[0124] In the formula, represents the input of the l+1th layer, represents the feature vector value output by the l+1th layer, represents the bias term of the convolutional layer, represents the weight corresponding to the kth channel in the l+1 layer, i represents the index of the output feature map on each element, Indicates an incrementing parameter, .

[0125] In the above steps S2-B3, the operations of feature aggregation and information filtering performed by the pooling layer are expressed as:

[0126]

[0127] In the formula, represents the output value of the pooling, Indicates setting parameter value; when = 1, it means taking the average value in the pooling area, which is called average pooling; when →∞, it means taking the maximum value in the area, that is, maximum pooling.

[0128] In step S2 of the embodiment of the present invention, in the above-mentioned drug-drug reaction prediction model, the drug similarity feature matrix is ​​extracted based on the two properties of drug molecular structure and target through the similarity matrix feature extraction channel. In this embodiment, the similarity between drugs can measure the degree of association between different drugs, and can also be used to infer whether other drugs have similar physicochemical properties. By analyzing the different physicochemical characteristics of drugs, drug similarity can be captured more comprehensively.

[0129] Based on this, the above-mentioned drug similarity feature matrix extraction method is specifically as follows:

[0130] S2-C1, Morgan fingerprint is used to extract the structural features of drug molecules, and then the Jaccard coefficient is used to calculate the similarity scores of two molecular fingerprints, and then a similarity matrix based on molecular substructure is obtained;

[0131] S2-C2, construct a drug-target heterogeneous graph network through the drug-target dataset, use the Node2Vec algorithm to learn the embedding of nodes in the network, and map each node in the drug-target heterogeneous graph network to a low-dimensional space to obtain the low-dimensional embedding of each node;

[0132] S2-C3, extract the low-dimensional embedding representation of the drug based on the node-based low-dimensional embedding, and obtain the cosine similarity of the low-dimensional embedding representation of the two drugs, and then obtain the drug-target based similarity matrix;

[0133] S2-C4, concatenate the similarity matrix based on molecular substructure and the similarity feature matrix based on drug-target to obtain the drug similarity feature matrix.

[0134] In the above step S2-C1, the similarity score I of the molecular fingerprint is expressed as:

[0135]

[0136] In the formula, and Represents the first The value of the element;

[0137] In the above step S2-C2, if Figure 3 The figure shows the process of calculating the similarity between molecules through drug-target. The Node2Vec algorithm is used to learn the embedding of nodes in the network. While maintaining the network structure information and node similarity, each node in the graph is mapped to a low-dimensional vector space. In order to take into account both homogeneity and structure, a random walk strategy is used to balance the sampling of BFS and DFS, which is expressed as:

[0138]

[0139] In the formula, represents the current sampling node, It means that the next sampling node is assumed to be , Indicates that the next sampling node is The probability of Representation Node Jump to the next node The probability of represents the normalization constant, Represents the set of edges in a graph.

[0140] Among them, a biased random walk is used, which is a non-standard transition probability, which is expressed as:

[0141]

[0142] In the formula, Representing Vertex and The weight of the edge between is the correction factor, which is defined as:

[0143]

[0144] in, Representation Node and The shortest path distance between them, p controls the probability of revisiting the previous node, and q controls the probability of exploring adjacent nodes. After obtaining the low-dimensional embedding of each node in the network, only the low-dimensional embedding representation of the drug is extracted. Since cosine similarity only cares about the angle between two vectors, regardless of their absolute size, it can reflect the similarity between drugs through direction, and has stronger robustness in relatively sparse drug-target data. Therefore, cosine similarity is used to calculate the similarity representation between drugs, and then the drug similarity feature matrix is ​​obtained.

[0145] In step S2 of the embodiment of the present invention, in the feature fusion module of the drug-drug reaction prediction model, a multi-head attention mechanism is used to calculate the cross-attention between the drug substructure features, drug molecular fingerprint features and drug similarity matrix features, and feature fusion is performed based on them to obtain fused drug features.

[0146] Specifically, in this embodiment, before the features of each channel are input into the feature fusion module, the features of each channel are first normalized, which solves the problem of inconsistent distribution of different feature sources. In the multi-channel feature fusion module, a multi-head attention mechanism is used to calculate the cross attention between the two data, and finally fuse the drug features of the three channels. Multi-channel feature fusion can capture information of different types and levels, improve the representation and generalization ability of the model for DDIs prediction tasks, thereby more comprehensively exploring potential interaction relationships and improving the accuracy of prediction.

[0147] The multi-head attention mechanism introduces multiple independent attention heads to learn different attention weights in parallel to obtain richer representation capabilities. Each attention head focuses on different parts, thereby enhancing the model's ability to represent input data. The formula for the multi-head attention mechanism is as follows:

[0148]

[0149] In the MCF-DDI model, Q (Query), K (Key), and V (Value) are all set to the input data itself. is the weight matrix of the output transformation, Represents the number of heads. The calculation formula for each head is:

[0150]

[0151] in, , and Represent three adaptive weight matrices respectively, Represents the current head. The output of the cross-attention mechanism is fused with the original input data to obtain a new feature representation:

[0152]

[0153]

[0154] in, is a weight parameter used to control the balance between the original input and the attention output. The two results are then added together to obtain the cross-attention features of the current two channels as the fused drug features:

[0155]

[0156] In step S2 of the embodiment of the present invention, in the drug-drug reaction prediction model, the predictor is a KAN predictor. Specifically, after passing through the multi-channel fusion layer, the drug D in the DDI triplet can be obtained. x and D yThe final feature representation of the two drugs. The traditional method is to concatenate the features of the two drugs and then put them into a fully connected layer (MLP) for DDI prediction. However, in MCF-DDI, the latest KAN (Kolmogorov-Arnold Network) is used instead of the traditional fully connected layer. The design of KAN is inspired by the Kolmogorov-Arnold representation theorem. Unlike the traditional MLP that sets a fixed activation function on the node "neuron", KAN sets a learnable activation function on the edge "weight", and its nodes only sum the input signal without applying any nonlinearity. KAN replaces all linear weights by parameterizing each weight as a univariate function of the spline function. KAN not only learns the combinatorial structure (external degrees of freedom), but also approximates the univariate function (internal degrees of freedom) well, so it can not only learn features as effectively as MLP, but also optimize features through its unique learnable activation function. In small-scale AI+ scientific tasks, KAN outperforms MLP in accuracy and interpretability. Theoretically and empirically, KAN has a faster neural scaling law than MLP.

[0157] If f is a multivariate continuous function on a bounded domain, then f can be written as a finite combination of single variable continuous functions and binary addition operations. The following is a simple function representation containing two KAN layers:

[0158]

[0159] in, represents the pth element of x, and are all univariate functions with trainable parameters. To simplify, It can be expressed as:

[0160]

[0161] Among them, b(x) is a basis function similar to the residual connection, Represents the spline function. They are respectively expressed as:

[0162]

[0163] in, is a trainable parameter. Dimensional input and The KAN layer with a dimensional output can be defined as a one-dimensional function matrix:

[0164]

[0165] In order to extend KAN to a wider and deeper range, a generalized KAN layer is defined, and the network can be made deeper by stacking layers. The shape of KAN is represented by an integer array:

[0166]

[0167] in, Representative The number of layer nodes. Indicates Layer neurons, and use express The activation value of the neuron. Layer and Between the layers, there are activation function. The activation value of a neuron is the sum of all incoming activations:

[0168]

[0169]

[0170] in, represents the activation function connecting two layers, express The post-activation is represented by the matrix:

[0171]

[0172] In the formula, On behalf of KAN The function matrix corresponding to the layer.

[0173] Based on this, in this embodiment, in the KAN predictor, The KAN of the layer can be expressed as:

[0174]

[0175] In the formula, represents the function representation of the KAN predictor, … represents the number of nodes in the 0th to L-1th KAN layers, … represents the activation function connecting two adjacent KAN layers, … represents the activation value of neurons in the 0th to L-1th KAN layers, … represents the node index, and the subscript L represents the KAN layer index.

[0176] In this embodiment, the cross entropy function is selected as the loss function to minimize the difference between the predicted result and the true label, which is expressed as:

[0177]

[0178] in, Represents the total number of samples. represent Whether the true label of drug-drug reaction is generated between Representative model for The probability that two drugs will produce a drug-drug reaction.

[0179] The present invention provides an experimental example for verifying the effect of the drug-drug interaction prediction model (MCF-DDI) in the above method in achieving DDI prediction.

[0180] In this embodiment, the above-mentioned model MCF-DDI is compared with eight baseline methods, including GCN, GAT, DeepDDI, SSI-DDI, MR-GNN, GMPNN-CS, DMF-DDI, and SRR-DDI. Among them, GCN is a DDI prediction model based on graph convolutional neural network. GAT is a DDI prediction model based on graph attention mechanism. DeepDDI uses drug name and its structural information for DDI prediction. SSI-DDI proposes a DDI prediction method based on substructure interaction, which directly operates the original molecular graph of the drug to obtain richer feature extraction. MR-GNN designs a DDI prediction method through multi-resolution graph representation and dual graph structure. GMPNN-CS learns information transfer neural network of chemical substructures of different sizes and shapes from the molecular graph representation of the drug. DMF-DDI proposes a deep multimodal feature fusion framework DMF-DDI, which integrates drug molecular graph, DDI network and biochemical similarity features of drugs to predict DDI. SRR-DDI is a substructure fine representation learning model based on the self-attention mechanism. It improves the DDI prediction performance by improving the robustness of the substructure features that determine the properties of the drug. Then, the contribution of each feature channel and module to the model performance is confirmed by ablation studies. Secondly, visual analysis is used to illustrate the model's ability to identify key substructures in drug molecules. Finally, the effectiveness and efficiency of the MCF-DDI model in practical applications are further confirmed through example analysis.

[0181] In order to comprehensively evaluate the performance of MCF-DDI, multiple angles were considered. Six widely used indicators were used: accuracy (ACC), area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUP), precision (Prec), recall (Rec), and F1 score (F1). ACC evaluates all classification results uniformly, while the F1 value reflects the classification accuracy and completeness of the model in positive and negative examples in a more balanced way. AUROC is suitable for class-balanced datasets, while AUPRO reflects the generalization ability of the model on unbalanced datasets. The higher the values ​​of the six evaluation indicators, the more they reflect the superiority of the model. In order to reduce the impact of random errors and ensure the reliability of the conclusions, all experiments were repeated five times, and the mean and standard deviation of the indicators were reported.

[0182] In this embodiment, three groups of experiments are set in the effect verification example:

[0183] Experiment 1: The training set contains all the drugs that have appeared in the test set and the validation set.

[0184] Experiment 2: The training set contains only one drug in the DDI triplet of the test set.

[0185] Experiment 3: The training set does not contain any drugs in the DDI triplet of the test set.

[0186] The present invention uses DrugBank, ZhangDDI and ChCh-Miner benchmark data sets to evaluate the settings in Experiment 1. ZhangDDI and ChCh-Miner already contain negative samples in the two benchmark data sets, so they are divided into training, verification and test sets in a ratio of 7:1:2. In addition, each method is repeated five times to obtain the average value and standard deviation as the final result. All the raw data in DrugBank are regarded as positive samples, and the negative sampling strategy in GMPNN is adopted, that is, any drug in the positive triplet is randomly disturbed to generate corresponding negative samples. Finally, the stability of the model is verified by five-fold cross validation. In order to ensure the comprehensiveness of the data in the new drug test, the DrugBank benchmark data set containing more drug types and DDI triplets is used in the settings of Experiments 2 and 3 for experiments. Similarly, the GMPNN method is used for new drug pretreatment.

[0187] In ChCh-Mine, ZhangDDI and DrugBank, the learning rate is initialized to 0.001, the weight decay is set to 0.0001, and the penalty coefficient is set to 0.96. The Adam optimizer is used to optimize the model, and the model is initialized by random initialization. The batch size is set to 512, and the dropout is set to 0.1 to prevent overfitting.

[0188] In experiment 1, to ensure fairness in comparison with other models, MCF-DDI uses the same experimental settings as the baseline model. For the ZhangDDI and ChCh-Miner benchmark datasets, they are divided into training, validation, and test sets in a ratio of 7:1:2. The whole process is repeated five times, and the results are averaged and standard deviations. In the DrugBank benchmark dataset, the present invention uses a five-fold cross-validation experiment, that is, one set of data is used as the test set in turn, and the remaining four sets of data are used as the training set. Finally, the average of the five sets of results is calculated as the final value, and the standard deviation is attached to verify the stability of the model.

[0189] Tables 1, 2, and 3 show the performance of MCF-DDI in the first experiment on benchmark datasets of different scales, with the best results in bold. As can be seen from the table, MCF-DDI has improved to a certain extent on different datasets and different indicators, proving its performance advantage over other models. Specifically, the accuracy on ChCh-Mine, ZhangDDI, and DrugBank reached 97.16%, 99.23%, and 98.95%, respectively, and the standard deviation obtained from five cross-validations showed higher stability.

[0190] For the ChCh-Mine dataset, it only accounts for less than 20% of the data volume of DrugBank, which greatly increases the difficulty of model learning. However, it can be seen from the results that MCF-DDI has achieved very good results in various evaluation indicators. This shows that the excellent performance of MCF-DDI on datasets of different sizes is attributed to the multi-channel fusion of local information of intermolecular substructures, global information represented by molecular fingerprints, and potential association information based on similarity between drugs. Similarly, the evaluation indicators of MCF-DDI on other datasets have also been significantly improved. These results fully demonstrate the superiority of the model.

[0191] Table 1. Comparative experimental results on the DrugBank dataset, expressed as mean ± standard deviation.

[0192]

[0193] Table 2. Comparative experimental results on the ZhangDDI dataset, expressed as mean ± standard deviation.

[0194]

[0195] Table 3. Comparative experimental results on the ChCh-Miner dataset, expressed as mean ± standard deviation.

[0196]

[0197] When evaluating the performance of the model in new drug prediction, the present invention designed Experiment 2 and Experiment 3. Tables 4 and 5 show the performance of MCF-DDI under these two settings in the DrugBank benchmark dataset. A five-fold cross-validation method was used to ensure the stability and reliability of the results, and the final results were presented in the form of mean and standard deviation.

[0198] In the setting of Experiment 2, the training set only contains one drug in the DDI triplet of the test set. According to the results in Table 4, the MCF-DDI model achieved an accuracy of 78.13%, which is significantly better than other benchmark models. In addition to accuracy, other evaluation indicators also generally showed a certain degree of improvement. In the setting of Experiment 3, the training set did not contain any drugs that appeared in the DDI triplet of the test set, which further tested the generalization ability of the model. As shown in Table 5, MCF-DDI performed better than other comparison models in all evaluation indicators. Comparing the experimental results of Experiments 1, 2 and 3, it can be seen that when it comes to new drug prediction, the performance of all models generally decreases. This is because drugs that did not appear in the training set appeared in the test set, resulting in the model only being able to learn a small number of features of these new drugs from limited data, which further tests the robustness of the model. Compared with the second-best SRR-DDI model, although it integrates drug substructure features and molecular structure similarity matrix for DDI prediction, it ignores the potential associations between drugs and the global information of drugs. MCF-DDI can fuse local and global features within the molecule and extract hidden correlation features between drugs, thereby integrating richer information in the feature extraction module, which is the key factor for its better performance.

[0199] Table 4. Comparative experimental results of Experiment 2 on the DrugBank dataset, expressed as mean ± standard deviation.

[0200]

[0201] Table 5. Comparative experimental results of Experiment 3 on the DrugBank dataset, expressed as mean ± standard deviation.

[0202]

[0203] In order to verify the effectiveness of each feature, a feature ablation experiment was conducted on the DrugBank benchmark dataset. Specifically, one drug feature or one channel was first deleted, and then the prediction results under different feature combinations were observed. Figure 1As shown in the figure, when any feature or channel is lost, various evaluation indicators will decrease to varying degrees. Specifically, the performance of any combination of two features is better than that of a single feature of the two features, and the performance is best when the three feature channels are fused. This confirms that each feature is useful for improving the model, indicating that multi-channel feature fusion can capture multi-level drug features, thereby making up for the limitations of a single channel in capturing features.

[0204] pass Figure 4 From the intuitive results, we can see that when the molecular fingerprint features of the drug are eliminated, the performance loss is the most serious, which shows that the drug molecular fingerprint can well characterize the characteristics of the drug. Secondly, when the molecular substructure feature channel is eliminated, the performance of the model also decreases significantly, which shows that it is effective to use molecular substructure to characterize the global characteristics of the drug. Although the model performance decreases less when the drug similarity feature is not used, drug similarity can reflect the mechanism and effect of the drug. Through the similarity between different drugs, the possible mechanism of action and potential interaction of new drugs can be inferred, which is very helpful for the prediction of new drugs.

[0205] In order to further study the impact of each module on model performance, Figure 5 As shown, the present invention designs several different variants to compare with our model:

[0206] (1) MCF-DDI-CRLI: In this model, the original criss-cross attention module is retained, but the KAN module in our model is replaced with a traditional linear layer.

[0207] (2) MCF-DDI-CAKA: In this model, the original prediction module KAN is retained, but the cross attention module is replaced by a direct splicing method.

[0208] (3) MCF-DDI-CALI: In this model, the traditional linear layer is used to replace the KAN module, and the direct splicing method is used to replace the cross attention module.

[0209] (4) MCF-DDI-KCL: In this model, KAN and cross-attention mechanisms are retained, but batch normalization in the model is completely replaced by layer normalization.

[0210] By comparing these variant models with the original MCF-DDI model, we can analyze the impact of each module on the overall performance of the model, thereby gaining insight into the contribution of the model's structural design to prediction accuracy.

[0211] from Figure 5It can be clearly observed that the performance of all variant models is lower than the original MCF-DDI model. Specifically, when the traditional splicing strategy is used instead of the cross-attention feature fusion strategy (MCF-DDI-CAKA), the model performance decreases slightly, indicating that the cross-attention module can more effectively capture the relationship between features. When the traditional linear layer is used instead of KAN (MCF-DDI-CRLI), the performance indicators of the model also decrease slightly, showing the effectiveness of the learnable activation function in the KAN module. In contrast, the performance of the MCF-DDI model is better than that of MCF-DDI-KCL, which further proves the effectiveness of batch normalization in the MCF-DDI model because it helps to deal with the problem of uneven data distribution. These results deepen the understanding of the impact of each module on the overall model performance and provide guidance for future model optimization and design.

[0212] In this embodiment, an interpretable experiment of the MCF-DDI model is also provided, and the effectiveness and interpretability of the MCF-DDI model in substructure feature extraction are explored through visual analysis. Figure 6 Figure 1 shows a heat map of the importance of drug substructures in the setting of Experiment 1. By analyzing the importance score of each substructure, we can infer the degree of attention paid by the model to different substructures during the prediction process, and highlight the molecular substructures that the model pays the most attention to in red. Figure 6 (A) The model focuses on the fluorine and chlorine atoms in midazolam and the amide group in pentobarbital, respectively. Fluorine and chlorine are highly electronegative atoms that can form hydrogen bonds or coordination interactions with the nitrogen atom of the amide, leading to an increased risk or severity of sedation and central nervous system depression. Figure 6 (B) It can be observed that the model pays special attention to the hydroxyl functional groups in the aspirin and ibuprofen molecules. Both drugs are non-steroidal anti-inflammatory drugs and contain hydroxyl functional groups, which may make them compete at the same receptor. The reversible COX inhibition of ibuprofen may partially hinder the acetylation reaction of aspirin, thereby weakening the inhibitory effect of aspirin on platelets. Figure 6 (C) The model shows that the amino, amide and hydroxyl substructures of the two drugs are focused on. Since the amino or amide groups of carbamazepine form hydrogen bonds with the hydroxyl groups of dicoumarol, the anticoagulant effect may be enhanced, leading to an increased risk of bleeding. Figure 6 As shown in (D), the model pays special attention to the substructures of the pyrimidine ring, hydroxyl group and amino group of the two drugs. This is because the pyrimidine ring of dipyridamole contains nitrogen atoms, and the hydroxyl and amino groups of metoprolol can interact with these nitrogen atoms through hydrogen bonds, which may enhance or modify the binding ability of metoprolol to β receptors, thereby leading to excessive β receptor antagonism, which manifests as bradycardia.

[0213] In order to further verify the effectiveness of substructure extraction, a masking experiment of key nodes of substructure was conducted. In order to avoid the influence of other features on the experimental results, a channel ablation strategy was adopted, that is, only the substructure extraction channel was retained. Nodes with high weight scores were regarded as key nodes of substructure, and then two masking strategies were used for verification: the first one was to mask out the top 30% of nodes with the highest scores, and the second one was to randomly mask nodes. Figure 7 The predicted probabilities and visualization results under different masking strategies are shown. Under the first strategy, Figure 7 (A), a significant drop in prediction scores was observed. This indicates that key substructures play a decisive role in compound reactions. When this key information is missing, the model can only make predictions based on limited and unimportant data, resulting in a sharp drop in performance. Under the second strategy, Figure 7 (B) Random masking of nodes does not specifically affect those particularly important nodes, so the overall substructure function is relatively complete and the predicted score decreases slightly. This illustrates the model's sensitivity to key substructures and verifies the rationality of the designed substructure importance score. Figure 7 In (C), no operation is performed on the drug nodes, and the model can use all important substructures for prediction. It can be observed that the model predicts that the probability of interaction between the two drugs is close to 1, which proves that the substructures paid attention to by the model do play a good role in promoting the prediction of DDI.

[0214] In general, we have conducted an in-depth discussion on the performance of the MCF-DDI model in substructure feature extraction and its interpretability. Through visual analysis, we can clearly see the ability of the model in identifying key substructures in drug molecules. For example, in the heat map of drug substructure importance under the setting of Experiment 1, the model's focus on the hydroxyl functional groups in aspirin and ibuprofen molecules reveals the potential interaction mechanism between non-steroidal anti-inflammatory drugs. This discovery not only deepens our understanding of the mechanism of drug action, but also provides a new perspective for the study of drug interactions. Further masking experiments verify the effectiveness of the MCF-DDI model in substructure extraction from another perspective. Through the channel ablation strategy, we focus on the substructure extraction channel and use two masking strategies to test the sensitivity of the model to key substructures. The experimental results show that when the key substructures are masked, the prediction performance of the model decreases significantly, which fully demonstrates the decisive role of these substructures in drug interaction prediction. The strategy of randomly masking nodes shows the relative insensitivity of the model to non-key substructures, thus confirming the rationality of the substructure importance score. In summary, the MCF-DDI model has demonstrated excellent performance in interpretability experiments. It can not only accurately identify key substructures in drug molecules, but also verify the importance of these substructures in drug interaction prediction through experiments. These results not only provide strong theoretical support for drug development and clinical application, but also point out the direction for further optimizing the model structure and improving prediction accuracy. In the future, we expect the MCF-DDI model to play a greater role in the field of drug interaction research and make more contributions to human health.

[0215] In the embodiments of the present invention, a case analysis example of applying the above model to a real-life scenario is also provided.

[0216] In this embodiment, the ZhangDDI and ChCh-miner benchmark data sets are randomly divided into a ratio of 9:1 for training and testing. First, the model is trained using the training set, and then the drug pairs in the remaining test set are predicted. Finally, the drug pairs are sorted according to the probability scores predicted by the model. The higher the prediction score between the two drugs, the higher the possibility that they will interact. We selected the top ten drug pairs with the highest predicted probability scores. For these ten pairs of drugs, we used the drug reaction check tools Drug.com (https: / / www.drugs.com / ) and DrugBank (https: / / go.drugbank.com / interax / multi_search) to find evidence to support them and collect descriptions of their interactions. For these ten events, as shown in Table 6, there are 8 pairs of DDI events that can be determined and two pairs that are uncertain. For example, torsemide is a loop diuretic mainly used to treat hypertension and remove excess water from the body. Indacaterol is a long-acting β2 receptor agonist mainly used to treat chronic obstructive pulmonary disease. Since indacaterol may cause hypokalemia and torsemide may also cause potassium loss, the risk of hypokalemia or low blood potassium may be increased when using these two drugs. The application of the MCF-DDI model in the field of real-life DDI prediction has shown significant effects and value. The model can not only effectively identify potential drug interactions, but also serve as a powerful tool for clinical practice and drug safety monitoring, helping to reduce the occurrence of adverse drug reactions and improve patients' medication safety and treatment effects.

[0217] Table 6. Top 10 DDIs with predicted probability.

[0218]

[0219] The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

[0220] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A drug-drug interaction prediction method based on multi-channel feature fusion, characterized in that: The following steps are involved: S1. Preprocessing benchmark datasets of different sizes containing different drug responses; S2. constructing a drug-drug reaction prediction model and training it using the benchmark dataset; The drug-drug reaction prediction model comprises a feature extraction module, a feature fusion module and a predictor connected in sequence; wherein the feature extraction module is used to extract drug substructure features, drug molecular fingerprint features and drug similarity matrix features, and the feature fusion module is used to fuse the features extracted by the feature extraction module through a cross attention mechanism to obtain fused drug features; and the predictor is used to predict the probability of drug-drug interaction for the fused drugs; S3, inputting any two target drugs into the trained drug-drug reaction prediction model, and outputting the predicted probability of drug-drug interaction; In the drug-drug response prediction model: In the feature extraction module, the drug molecular structure features are extracted through the graph neural network GNN, the drug molecular fingerprint features are extracted through the one-dimensional convolutional network, and the drug similarity feature matrix is ​​extracted based on the two properties of drug molecular structure and target through the similarity matrix feature extraction channel; In the feature fusion module, a multi-head attention mechanism is used to calculate the cross attention between the drug substructure features, drug molecular fingerprint features, and drug similarity matrix features, and feature fusion is performed based on them to obtain fused drug features; The predictor is a Kolmogorov-Arnold Network predictor.

2. The drug-drug interaction prediction method based on multi-channel feature fusion according to claim 1, characterized in that: In the step S2, in the drug-drug reaction prediction model, the drug substructure feature is extracted through a drug substructure feature extraction channel; In the drug substructure feature extraction channel, the drug molecular structure features are extracted by using the graph neural network GNN. The specific extraction method is as follows: S2-A1, represent the origin in the drug molecule as a node, and the chemical bonds as an edge, and construct a drug molecule graph; S2-A2, determine the bond-order characteristics of all edges in the drug molecule graph; S2-A3, calculate the importance score of each bond-level feature in the entire drug molecule graph; S2-A4, according to the importance score of the bond-level features, the drug molecule graph is pooled to obtain the drug molecule graph pooling features; S2-A5, calculate the attention weight score of each layer of substructure according to the pooling characteristics of the drug molecule graph; S2-A6, normalize the attention weight scores of all layer substructures, and combine the bond-level features of each layer of the graph neural network GNN to determine the final bond-level features of the edges in the drug molecule graph; S2-A7. According to the characteristics of each node in the drug molecule graph and the final bond-level characteristics of its adjacent edges, the substructure representation of each node is obtained as the drug substructure feature.

3. The drug-drug interaction prediction method based on multi-channel feature fusion according to claim 2, characterized in that: In step S2-A2, the bond-level features of the edge It is expressed as: ； In the formula, represents the characteristics of the starting node a, represents the characteristics of the terminal node b, Represents the characteristics of the bond between nodes a and b; Step S2-A4, drug molecule graph pooling features It is expressed as: ； In the formula, represents the pooling layer, Indicates the current number of GNN layers, Represents the matrix composed of all key-level features in the current layer t of GNN, Represents the adjacency matrix of the current layer graph; In step S2-A5, the attention weight score It is expressed as: ； In the formula, represents the weight vector of the tth layer, represents the tanh activation function, W represents the learnable matrix, represents the bias term; In step S2-A6, the final bond-level feature It is expressed as: ； In the formula, represents the key-level feature representation between nodes a and b after processing at the tth layer, express Activation function, T represents the number of network layers of the substructure; In step S2-A7, the substructure of each node is represented by It is expressed as: ； In the formula, represents the original feature representation of node a, represents a nonlinear function, Represents the set of all neighbor nodes of node a.

4. The drug-drug interaction prediction method based on multi-channel feature fusion according to claim 1, characterized in that: In the step S2, in the drug-drug reaction prediction model, the drug molecular fingerprint feature is extracted through a molecular fingerprint feature extraction channel; In the molecular fingerprint feature extraction channel, the drug molecular fingerprint feature is extracted by a one-dimensional convolutional network, and the extraction method is specifically as follows: S2-B1. Converting drug molecules into molecular fingerprints: S2-B2, input the molecular fingerprint into a one-dimensional convolutional network to generate a feature map; The one-dimensional convolutional network includes three sequentially connected convolutional layers with the same convolution kernel size and different numbers; S2-B3, the generated feature map is aggregated and information filtered through the pooling layer to obtain the molecular fingerprint feature.

5. The drug-drug interaction prediction method based on multi-channel feature fusion according to claim 4, characterized in that: In step S2-B2, in the one-dimensional convolutional network, the output sequence length of the l+1th convolutional layer It is expressed as: ； In the formula, represents the length of the l-th layer sequence, Indicates the length of the padding. Represents the length of the convolution kernel, and s represents the step size of each movement of the convolution kernel; For each element in the output sequence, it is represented as: ； ； In the formula, represents the input of the l+1th layer, represents the feature vector value output by the l+1th layer, represents the bias term of the convolutional layer, represents the weight corresponding to the kth channel in the l+1 layer, i represents the index of the output feature map on each element, Indicates an incrementing parameter, ; In the step S2-B3, the operation of feature aggregation and information filtering performed by the pooling layer is expressed as: ； In the formula, represents the output value of the pooling, Indicates setting parameter value.

6. The drug-drug interaction prediction method based on multi-channel feature fusion according to claim 1, characterized in that: In the step S2, in the drug-drug reaction prediction model, the drug similarity feature matrix is ​​extracted based on the two properties of drug molecular structure and target through the similarity matrix feature extraction channel, and the extraction method is specifically as follows: S2-C1, Morgan fingerprint is used to extract the structural features of drug molecules, and then the Jaccard coefficient is used to calculate the similarity scores of two molecular fingerprints, and then a similarity matrix based on molecular substructure is obtained; S2-C2, construct a drug-target heterogeneous graph network through the drug-target dataset, use the Node2Vec algorithm to learn the embedding of nodes in the network, and map each node in the drug-target heterogeneous graph network to a low-dimensional space to obtain the low-dimensional embedding of each node; S2-C3, extract the low-dimensional embedding representation of the drug based on the node-based low-dimensional embedding, and obtain the cosine similarity of the low-dimensional embedding representation of the two drugs, and then obtain the drug-target based similarity matrix; S2-C4, concatenate the similarity matrix based on molecular substructure and the similarity feature matrix based on drug-target to obtain the drug similarity feature matrix.

7. The drug-drug interaction prediction method based on multi-channel feature fusion according to claim 6, characterized in that: In step S2-C1, the similarity score I of the molecular fingerprint is expressed as: ； In the formula, and Represents the first The value of the element; In the step S2-C2, a random walk strategy is used in the Node2Vec algorithm to balance the sampling of BFS and DFS, which is expressed as: ； In the formula, represents the current sampling node, It means that the next sampling node is assumed to be , Indicates that the next sampling node is The probability of Representation Node Jump to the next node The probability of represents the normalization constant, Represents the set of edges in a graph.

8. The drug-drug interaction prediction method based on multi-channel feature fusion according to claim 1, characterized in that: In the step S2, in the drug-drug reaction prediction model, the predictor is a Kolmogorov-Arnold Network predictor, which is expressed as: ； In the formula, represents the function representation of the KAN predictor, … represents the number of nodes in the 0th to L-1th KAN layers, … represents the activation function connecting two adjacent KAN layers, … represents the activation value of neurons in the 0th to L-1th KAN layers, … represents the node index, and the subscript L represents the KAN layer index.

Images