Method for predicting microorganism-drug interaction through double hypergraph contrast learning framework based on hierarchical attention

Abstract

The invention provides a method for predicting microorganism-drug interaction through a double hypergraph contrast learning framework based on hierarchical attention, and solves the problems that the graph structure is simple, the interpretable analysis is increased, and the prediction precision is improved. The method consists of three modules: S1, data preprocessing: carrying out nonlinear fusion on similarity data of microorganisms and drugs to obtain comprehensive similarity of the microorganisms and the drugs, and forming a double hypergraph by combining an original incidence matrix and the comprehensive similarity with KNN and K-means algorithms; s2, feature extraction: sending the double-hypergraph obtained in the step S1 into a hierarchical attention and hypergraph convolutional network to obtain feature embedding, updating features based on comparison learning of the double-hypergraph, and fusing the double-hypergraph features through an integration network. And S3, result prediction: mapping the microorganism and drug characteristics obtained in the step S2 by using a full connection layer to obtain final characteristics, and performing dot product to obtain a prediction score. And outputting the obtained prediction score in combination with specific literature verification and a network pharmacology verification model.

Description

Technical Field

[0001] The present invention relates to the field of predicting microorganism-drug interactions, and particularly to a method for predicting microorganism-drug interactions based on a hierarchical attention-based dual hypergraph contrast learning framework. Background Art

[0002] A variety of drugs can cause significant perturbations to the gut microbial community, leading to a decrease in the abundance of beneficial bacteria and triggering a series of adverse clinical consequences. Currently, traditional experiments have significant limitations in analyzing the mechanism of microorganism-drug interactions. Laboratory research requires expensive equipment and professional techniques, and it is difficult to fully simulate the complex physiological conditions in vivo in vitro, resulting in possible deviations between experimental results and actual situations. Although clinical observations have practical guiding significance, they rely on a large number of case accumulations and long-term follow-up observations, making it difficult to quickly summarize general laws. Moreover, for rare cases or newly emerging microorganism-drug combinations, there is often a lack of sufficient empirical data, and new computational methods need to be established to provide a theoretical basis for personalized and precise drug use. Early computational methods have made preliminary explorations in predicting microorganism-drug associations, but there are certain limitations. For example, existing methods such as HMDAKATZ (computing network nodes) are difficult to comprehensively and accurately reflect complex biological relationships. Currently, most computational methods are based on graph structures, but they cannot simulate the complex, diverse, and heterogeneous relationships between multiple drugs and microorganisms. For example, existing methods such as GCNMDA (graph convolutional network), EGATMDA (graph attention network), etc. may lead to incomplete or biased information learned by the model, thus affecting the accuracy of prediction. In recent years, hypergraph structures and graph contrast learning methods have brought new ideas and methods for predicting microorganism-drug associations. Hypergraphs generalize graphs by introducing hyperedges to represent higher-order relationships between node groups. This representation is highly relevant in many practical applications. By modeling higher-order relationships, hypergraphs can provide better performance than graph-based models in many data mining tasks. Interpretable analysis of model prediction results is necessary for microorganism-drug interactions. Existing methods use literature verification methods in interpretable analysis and can only verify based on the results of existing papers, while network pharmacology can explain the interactions between microorganisms and drugs from the perspectives of genes and targets. Summary of the Invention

[0003] Aiming at the deficiencies in the prior art, the present invention provides a method for predicting microorganism-drug interactions based on a hierarchical attention-based dual hypergraph contrast learning framework, which can effectively solve the problems of simple traditional model structure, increasing the interpretable analysis of microorganism-drug interactions, and improving the prediction accuracy.

[0004] According to the above problems to be solved, the present invention adopts the following technical solutions:

[0005] A method for predicting microorganism-drug interactions based on a hierarchical attention-based dual hypergraph contrast learning framework, comprising the following steps:

[0006] S1. Nonlinearly fuse the microorganism functional similarity and microorganism Gaussian kernel similarity of microorganisms to obtain the comprehensive microorganism similarity, and nonlinearly fuse the drug structure similarity and drug Gaussian kernel similarity of drugs to obtain the comprehensive drug similarity. Combine the original microorganism-drug association matrix and the comprehensive similarity, and use the KNN and K-means algorithms to construct a dual hypergraph.

[0007] S2. Feed the dual hypergraph obtained in step S1 into the hierarchical attention and hypergraph convolutional network to obtain feature embeddings. Use the two hypergraphs as positive and negative samples for each other, and update the features using the strategy of minimizing positive sample pairs and maximizing negative sample pairs in contrast learning. Integrate the obtained features through an integration network to dynamically fuse the dual hypergraph features.

[0008] S3. Map the microorganism and drug features obtained in step S2 using a fully connected layer to obtain the final microorganism and drug features, and calculate the dot product of the two to obtain the final prediction score.

[0009] S4. Combine the prediction score obtained in step S3 with specific literature verification and network pharmacology verification model outputs.

[0010] Further, in step S1, the data preprocessing steps include obtaining microorganism functional similarity, drug structure similarity, microorganism Gaussian kernel similarity, and drug Gaussian kernel similarity. Nonlinear fusion includes local information acquisition and normalization.

[0011] Further, the calculation of the microorganism Gaussian kernel similarity and the drug Gaussian kernel similarity in step S1 is as follows:

[0012]

[0013] Where MS2 represents the microorganism Gaussian kernel similarity, m i and m j represent the i-th and j-th microorganisms in the original association matrix, exp represents the exponential function, η m represents the normalized kernel bandwidth of microorganisms, A represents the original association matrix, ‖ ‖ 2 represents the square of the 2-norm. η m ' represents the original bandwidth, which is always set to 1, and N m represents the number of microorganisms.

[0014] Further, the comprehensive similarity is obtained by nonlinearly fusing the similarity data, as follows:

[0015]

[0016] where \(t = 1, 2, \ldots, M\), and \(M\) is the total number of similarity data. For each similarity data, the KNN algorithm is used to measure the local relationship as follows:

[0017]

[0018] where \(N\) i is a set of \(k\) nearest neighbors of node \(m\) in the microbial similarity data. \(K\) i represents the local affinity kernel of the \(t\)-th data type, and the neighbor parameter of KNN takes \(N\) t / 10. Finally, the model iteratively updates the similarity matrix of each type of data using the program as follows: m where \(t = 1, 2, \ldots, M\), and \(M\) is the total number of data types.

[0019]

[0020] where \(t = 1, 2, \ldots, M\), and \(M\) is the total number of data types. is the state matrix of the \(t\)-th data type after the \(r\)-th iteration.

[0021] where is the state matrix of the \(t\)-th data type after the \(r\)-th iteration, and \(T\) represents matrix inversion.

[0022] Furthermore, in step S2, hierarchical attention is a process of aggregating information on points and hyperedges in the hypergraph structure. First, information aggregation from points to edges, hyperedge-level attention, is performed, and then information aggregation from edges to points, node-level attention, is carried out. Here, the set of all nodes connected to hyperedge \(e\) i is defined as \(y\) i , and the set of hyperedges connected to node \(n\) i is located as \(\rho\) i . Hyperedge-level attention captures the different importance of all nodes \(n\) j connected to the given hyperedge \(e\) p \(\in y\) i . The attention score \(a\) i of node \(n\) j to hyperedge \(e\) ji is defined as follows:

[0023]

[0024] where \(w1\) is a learnable parameter matrix, \(u\) is a trainable weight vector, and \(S(\cdot)\) is a similarity function, defined as scaled dot-product attention where \(D\) is the dimension.

[0025] Use the updated hyperedge representation to obtain the node representation. Node attention determines the different importance of different hyperedges, and the hyperedge e j for the node n i the attention score b ij is defined as follows:

[0026]

[0027] where w2 and w3 are learnable parameter matrices, S(·) is a similarity function, and the Scaled Dot-Product Attention is defined as

[0028] Finally, the hyperedge representation e j and the updated node representation Z i are derived by aggregating their neighbor information and are defined as follows:

[0029]

[0030] where φ(·) consists of two layers of MLP and an ELU activation function.

[0031] Furthermore, dual hypergraph contrast learning is adopted to use two hypergraphs as positive and negative samples for each other, and the features are updated by minimizing the positive sample pairs and maximizing the negative sample pairs in contrast learning. The obtained features are dynamically fused with the dual hypergraph features through an integration network. For the selection of positive and negative samples, for each node v, the embedding v i generated in one view is defined as the anchor point, and the embedding generated in the other view is denoted as u i , so that the different embeddings v i and u i of the same node in two different views form a positive sample pair. The embeddings v k and u k (k≠i) of other nodes are regarded as negative sample pairs, where v k and the anchor point v i form an intra-view negative sample pair, while u k and the anchor point v i form a cross-view negative sample pair. At this time, the contrast loss is defined as follows:

[0032]

[0033] where sim(·) is the cosine similarity function, g(·) is a two-layer neural network projection head used to enhance the information ability of nodes. τ is the temperature control parameter.

[0034] Furthermore, in step S2, the integrated network is used to dynamically fuse the features of the dual hypergraphs. First, the attention mechanism between the two views is used to update the embeddings of the two views, and then the multi-head attention mechanism is used to fuse the features of the two hypergraphs. The final microbial embedding representation and drug embedding representation are obtained.

[0035] Furthermore, in step S3, the final microbial embedding representation and drug embedding representation are used to obtain the embedding matrix through a fully connected neural network, and then the reconstructed correlation matrix is obtained through the matrix multiplication of the two features.

[0036] Furthermore, in step S4, two cases, namely the literature verification method and network pharmacology, are designed to verify the interpretability.

[0037] The present invention has the following beneficial effects:

[0038] 1. The method provided by the present invention extracts microbial-drug association information, calculates the functional similarity and Gaussian kernel similarity of microorganisms, and the structural similarity and Gaussian kernel similarity of drugs. The comprehensive similarity matrix is generated through non-linear fusion. The fused comprehensive similarity matrix and the original association information are combined to construct a dual hypergraph using the KNN and K-means algorithms, which can better simulate the association information between microorganisms and drugs. A hierarchical attention mechanism is designed to calculate the hyperedge attention and node attention of the hypergraph respectively, and then the topological features are extracted through hypergraph convolution, capturing the global information of the hypergraph more comprehensively. The embedding quality across hypergraph views is improved using contrastive learning. Visual attention dynamically fuses multi-perspective features, and multi-head attention adaptively fuses dual-channel embeddings. Finally, the embedding is optimized through a fully connected layer, and the correlation probability is reconstructed through matrix multiplication.

[0039] 2. The AUC and AUPR scores of the method provided by the present invention are better than those of other benchmark methods.

[0040] 3. The method provided by the present invention uses the gene and target information of network pharmacology to analyze the microbial-drug interaction in the interpretability analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is the overall flowchart of DHCLHAM in the embodiment of the present invention. Among them, in Figure A, the functional similarity and Gaussian kernel similarity of microorganisms, and the structural similarity and Gaussian kernel similarity of drugs are calculated. The comprehensive similarity matrix is generated through non-linear fusion. The hypergraph is constructed using the KNN and K-means algorithms.; In Figure B, the topological features are extracted through hierarchical attention and hypergraph convolution. The embedding quality across hypergraph views is improved using contrastive learning. The dual hypergraph embeddings are fused through the integrated network; In Figure C, the correlation probability is reconstructed through matrix multiplication. The output of the model is verified using network pharmacology.

[0042] Figure 2Scatter plot of the data set for the embodiments of the present invention.

[0043] Figure 3 Nonlinear fusion flow chart for the embodiments of the present invention.

[0044] Figure 4 Hierarchical attention flow chart for the embodiments of the present invention.

[0045] Figure 5 Visualization graph of the comparative experiment for the embodiments of the present invention, where Figure 5 a and b are the AUC and AUPR values of the MDAD data set, c and d are the AUC and AUPR values of the aBiofilm data set, and e and f are the AUC and AUPR values of the DrugVirus data set.

[0046] Figure 6 Ablation experiment graph for the embodiments of the present invention, where Figure 6 a, b, and c are the MDAD, aBiofilm, and DrugVirus data sets.

[0047] Figure 7 Effect of nonlinear fusion in the MDAD data set for the embodiments of the present invention.

[0048] Figure 8 Analysis of the number of multi-head attention heads parameter for the embodiments of the present invention, where Figure 8 a, b, and c are the MDAD, aBiofilm, and DrugVirus data sets.

[0049] Figure 9 Effect of predicting performance under different K values and C values on DrugVirus for the embodiments of the present invention, where Figure 9 a and b are the parameter analyses of the K value and C value.

[0050] Figure 10 Graph of the top 20 relevant microorganisms predicted for ciprofloxacin and moxifloxacin for the embodiments of the present invention.

[0051] Figure 11 Graph related to gene pathways for the embodiments of the present invention.

[0052] Figure 12 Enrichment analysis bubble chart for the embodiments of the present invention. Detailed implementation manners

[0053] The technical solutions in the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0054] The present invention provides a method for predicting microorganism-drug interactions based on a hierarchical attention-based dual hypergraph contrast learning framework, namely DHCLHAM. DHCLHAM mainly includes three modules: a data preprocessing module, a feature extraction module, and a result prediction module.

[0055] The overall flowchart of DHCLHAM is as Figure 1 shown. Among them, the data preprocessing module mainly includes calculating the functional similarity and Gaussian kernel similarity of microorganisms, the structural similarity and Gaussian kernel similarity of drugs. A comprehensive similarity matrix is generated through non-linear fusion. The hypergraph is constructed using the KNN and K-means algorithms. The feature extraction module takes the hypergraph of the previous module as input, calculates the hyperedge attention and node attention of the hypergraph through hierarchical attention respectively, and then extracts topological features through hypergraph convolution to more comprehensively capture the global information of the hypergraph. Contrast learning is used to improve the embedding quality across hypergraph views. Visual attention dynamically fuses multi-view features, and multi-head attention adaptively fuses dual-channel embeddings. The result prediction module obtains the final feature matrix by using a fully connected layer for the output of the previous layer, and reconstructs the association probability through matrix multiplication of the feature matrix. Network pharmacology is used to verify the model output.

[0056] 1. Method content

[0057] 1.1 Data preprocessing module

[0058] Regarding the microorganism-drug interaction dataset, including three datasets: MDAD, aBiofilm, and DrugVirus, as shown in Table 1 below. At the same time, we calculated the densities of each dataset to be 1.04%, 1.19%, and 5.61% respectively. The scatter plots of the three datasets are as Figure 2 shown.

[0059] Table 1 Statistical information of microorganism and drug datasets

[0060]

[0061] 1.1.1 Microorganism similarity network

[0062] The similarity of microorganisms is measured from two aspects. The first microorganism similarity is the functional similarity of microorganisms, which is calculated using the Kamneva algorithm. Suppose there are two microorganisms, m i and m j , then their microorganism functional similarity can be expressed as MS1(m i , m j) is represented by. However, many microorganisms have no similarity scores in MS1, and it is obvious that MS1 is sparse. To discover more valuable microbial information, more similarity information needs to be found. The second microbial similarity is the Gaussian interaction profile kernel similarity. Its core idea is that similar microorganisms have similar functions, thus producing similar interaction profiles. Gaussian kernel similarity can make full use of the node interaction information in the network. Therefore, it provides an effective method to quantify the similarity between nodes. Specifically, in the original matrix A, microorganism m i and m j can be represented as the i-th row and the j-th row in A. Then, the Gaussian kernel similarity MS2(m i and m j ) of microorganism m i and m j is defined as:

[0063]

[0064] where MS2 represents the Gaussian kernel similarity of microorganisms, m i and m j represent the i-th and j-th microorganisms in the original association matrix, exp represents the exponential function, η m represents the normalized kernel bandwidth of microorganisms, A represents the original association matrix, || || 2 represents the square of the 2-norm. η m ' represents the original bandwidth, which is always set to 1, and N m represents the number of microorganisms.

[0065] 1.1.2 Nonlinear Fusion

[0066] After calculating the functional similarity and Gaussian interaction profile kernel similarity of microorganisms, the two are fused. Integrating different similarity information can not only avoid the one-sidedness of data, but also obtain a more accurate and reasonable comprehensive similarity between microorganisms and drugs. However, simple linear similarity combination methods are difficult to apply to the fusion of multiple biological similarities. When traditional linear fusion methods integrate multi-view similarity information, the strategies adopted are often too simple

[30] , such as filling the missing part of one similarity with another similarity or directly averaging different types of similarities. This way is difficult to capture the complex nonlinear relationships between different types of similarities and may lead to information loss or poor fusion effects. Nonlinear fusion methods can, through complex iterative calculation processes, adaptively learn the nonlinear interaction relationships between different similarity networks, thus more fully mining the deep information hidden in multi-source data and providing the possibility for constructing a more accurate and comprehensive comprehensive similarity network. Therefore, this model adopts nonlinear fusion. The fused microbial similarity is represented as MS. First, the normalized weight is calculated and defined as:

[0067]

[0068] where \(t = 1, 2, \ldots, M\), and \(M\) is the total number of similarity data. For each similarity data, the KNN algorithm is used to measure the local relationship as follows:

[0069]

[0070] where \(N\) i is a set of \(k\) nearest neighbors of node \(m\) in the microbial similarity data. \(K\) i represents the local affinity kernel of the \(t\)-th data type, and the neighbor parameter of KNN is taken as \(N\) t / 10. Finally, the model iteratively updates the similarity matrix of each type of data as follows: m where \(t = 1, 2, \ldots, M\), and \(M\) is the total number of data types. Where

[0071]

[0072] is the state matrix of the \(t\)-th data type after the \(r\)-th iteration, and \(T\) represents matrix inversion. In this model, when \(MS\)

[0073] reaches the convergence criterion, the iteration stops. The convergence criterion is defined as the relative change t '(r+1) being less than \(10\) -5 . After iterative update, the final comprehensive similarity matrix \(MS\) can be obtained:

[0074]

[0075] However, the obtained similarity matrix is not a symmetric matrix. Therefore, \(MS=(MS+(MS) T ) / 2\) is the final microbial similarity matrix. The process is as Figure 3 shown.

[0076] 1.1.3 Drug Similarity Network

[0077] Similarly, the model measures the similarity of drugs from two aspects. The first drug similarity is the drug structure similarity, which is calculated using the SIMCOMP2 algorithm. Suppose there are two drugs \(d\) i and \(d\) j , then their drug structure similarity can be expressed as \(DS1(d\) i ,d j ​) It is represented by []. However, many drugs do not have similarity scores in DS1, and it is obvious that DS1 is sparse. In order to discover more valuable similarity information, other similarity information needs to be searched. The second drug similarity is the Gaussian interaction profile kernel similarity, and the calculation method is the same as that of the microbial Gaussian interaction profile kernel similarity. The drug comprehensive similarity matrix DS is the same as the above-mentioned non-linear fusion for calculating the microbial comprehensive similarity matrix.

[0078] 1.1.4 Dual-view hypergraph construction

[0079] Most existing studies rely on simple graphs, where the adjacency matrix reflects node relationships. However, for microbe-drug association prediction, simple graph structures cannot clearly capture the complex interactions between entities. Hypergraphs, as an extension of simple graphs, can contain more complex graph information. The relationship between microbes and drugs is not simply between two nodes. One microbe can metabolize multiple drugs, and one drug can affect multiple microbes. Drugs that can affect the same microbe have some similar properties. Similarly, there are also similarities among microbes. If drugs with similar properties are connected by hyperedges, the entire network can be represented by a high-order graph. Then, hypergraph learning can be used to capture this high-order association information, using hyperedges to represent the potential similarities between microbe-drug nodes in the heterogeneous graph, and achieving high-precision prediction of microbes and drugs. By constructing hyperedges to explore the high-order relationships between nodes, the complex interactions within the biological system can be analyzed more comprehensively. Therefore, we adopt the hypergraph structure as an intermediate node for transmitting microbe and drug information to achieve global high-order information propagation between microbe and drug nodes. To collect high-dimensional relationships beyond pairwise relationships from the original heterogeneous association network, this model uses a weighted hypergraph G=(V, E, W) to represent the hyperedges related to microbes and drugs. Here, V is the node set representing microbe nodes and drug nodes. E is the hyperedge set representing the association relationships between microbes and drugs. W is the weight set, which is a diagonal matrix. In this model, we connect microbe-drug associations and integrate the comprehensive similarity between microbes and drugs into the features of microbe nodes. Similarly, we connect drug-microbe associations and integrate drug similarities into the features of drug nodes. Based on the concatenated features, this model uses the KNN algorithm and the K-means algorithm to construct hypergraphs of microbes and drugs. Briefly, in the KNN algorithm, we first calculate the nearest k neighbors of each microbe according to the Euclidean distance, and determine a subset, i.e., the hyperedge, from the k neighbors. This hypergraph structure is called View1 in this paper. Additionally, we use the K-means method to randomly select cluster centers and determine the distance between each microbe and the cluster centers using the Euclidean distance, thus grouping those with closer distances into one class to form a subset, i.e., the hyperedge. This hypergraph structure is called View2 in this paper. After multiple iterations until the cluster centers no longer change. At this time, the relationship matrix of the hypergraph hypernodes and hyperedges obtained can be expressed as H∈R V×E . Particularly, in the View1 hypergraph, the number of hyperedges constructed by KNN for the hypergraph is the number of nodes. In the View2 hypergraph, the number of hyperedges constructed by K-means for the hypergraph is the number of cluster centers. The hypergraph incidence matrix H is defined as follows:

[0080]

[0081] where v represents nodes and e represents hyperedges.

[0082] 1.2 Feature Extraction Module

[0083] 1.2.1 Hierarchical Attention

[0084] Whether it is graph attention or hypergraph attention, their calculation methods are to capture the relationships between nodes and edges (hyperedges) in the graph (hypergraph), so as to determine the contribution degree of each edge (hyperedge) to the nodes. Since the adjacency matrix of the hypergraph consists of nodes and hyperedges, and a hyperedge can contain multiple nodes. In View1, theoretically, it can include all nodes. Hypergraph attention takes the initial node features of the hypergraph as input, then updates the representation of the hyperedges, and aggregates the information from the updated hyperedges to update the node representation. Some nodes may be informative, while others may not. Therefore, we designed hierarchical attention to consider the attention of hyperedges to nodes. When calculating the attention of nodes, the attention of hyperedges is calculated in advance. The importance degrees of hyperedges may also be different, which can better retain the features of the global structure of the hypergraph. To solve this problem, this solution designs hyperedge-level attention and node-level attention. It realizes the dynamic information transfer and feature enhancement between nodes and hyperedges in the hypergraph as Figure 4 shown. Here, the set of all nodes connected to the hyperedge e i is defined as y i , and the set of hyperedges connected to the node n i is located as ρ i .

[0085] The hyperedge attention in this solution aggregates the information of all nodes connected to each hyperedge. Here, the hyperedge-level attention captures the different importance of all nodes n j connected to the given hyperedge e p ∈y i . The attention score a i of the node n j to the hyperedge e ji is defined as follows:

[0086]

[0087] where w1 is a learnable parameter matrix, u is a trainable weight vector, S(·) is a similarity function, and is defined using scaled dot-product attention as where D is the dimension.

[0088] The updated hyperedge representation is used to obtain the node representation. The node attention determines the different importance of different hyperedges. The attention score b j of the hyperedge e i to the node n ij is defined as follows:

[0089]

[0090] Among them, w2 and w3 are learnable parameter matrices, S(·) is a similarity function, and is defined using Scaled Dot-Product Attention as

[0091] Finally, the hyperedge representation e is derived by aggregating their neighbor information j and the updated node representation Z i , which are defined as follows:

[0092]

[0093] where φ(·) consists of two layers of MLP and an ELU activation function.

[0094] 1.2.2 Hypergraph Convolutional Network

[0095] The hypergraph convolutional network HGCN that uses spectral convolution on the updated nodes can well encode the high-order relationships in the hypergraph structure. Through feature transformation and hypergraph aggregation, deeper features can be captured, the expressive power can be enhanced, and the generalization ability can be improved. According to the matrix after updating the nodes and the weight W of the hyperedges, the following update formula is defined:

[0096]

[0097] where X l is the aggregated information at layer l, X 0 = X.θ l is a learnable weight matrix, σ() is a non-linear activation function. D e is the degree matrix of the hyperedges, defined as d(e)=∑ v∈V H(v,e). D v is the degree matrix of the nodes, defined as d(v)=∑ e∈E w(e)H(v,e).

[0098] 1.2.3 Dual-View Contrastive Learning

[0099] In view of the data sparsity and complex relationships in real hypergraph structures, contrast learning has gradually become popular in recent years to improve and enhance the embedding quality. At the same time, many studies have started to combine graph-structured data with contrast learning to strengthen the embedding representation. Existing feature extraction methods based on GCL can be divided into two categories: structure enhancement and feature enhancement. Specifically, structure augmentation randomly discards nodes or edges of the graph to obtain an enhanced graph structure, and then feeds the augmented graph into the encoder for contrastive representation. Feature enhancement adds random noise as a contrastive view to the node embeddings. First, structure enhancement randomly deletes nodes or edges, which easily destroys the inherent properties of the input graph. Second, feature enhancement adds the same proportion of noise to each node, which ignores the unique features of the nodes on the graph. Therefore, this project adopts dual-view contrast learning. Dual hypergraphs can just solve these two problems, one is the problem of destroying the inherent properties of the input graph, and the other is the problem of feature neglect due to adding noise. The dual-view hypergraph contrast learning proposed by this model is the consistency of the same node in different views and the difference of different nodes. That is, according to the two hypergraph structures constructed previously, a contrastive objective function is used to ensure that the encoded embeddings of each node in two different views are not only consistent with each other, but also can be clearly distinguished from the embeddings of other nodes. That is, for each node v, the embedding v generated by it in one view i is defined as the anchor point, and the embedding representation generated in the other view is denoted as u i , so that the different embeddings v i and u i of the same node in two different views form a positive sample pair. The embeddings v k and u k of other nodes (k≠i) are regarded as negative sample pairs, where v k and the anchor point v i form an intra-view negative sample pair, while u k and the anchor point v i form a cross-view negative sample pair. Inspired by InfoNCE

[18] , the sample pair of each positive example in this model is defined as follows:

[0100]

[0101] where sim() is the cosine similarity function, and g(·) is a two-layer neural network projection head used to enhance the information ability of the nodes. τ is the temperature control parameter.

[0102] Taking the hypergraph constructed by View1 as an example, the hypergraph contrast loss between microorganisms and drugs is defined as follows:

[0103]

[0104] where N m and N dis the number of nodes of microorganisms and drugs. At the same time, when targeting the nodes of two different hypergraph views of View1 and View2, the views of microorganisms and drugs are symmetric. Therefore, we can obtain the hypergraph comparison view of View2 as and So the overall comparison loss of microorganisms and drugs obtained finally is defined as follows:

[0105]

[0106] 1.2.4 Integrating Networks

[0107] After contrastive learning, we performed contrastive loss training from two different views. Next, we need to integrate the two views to form a more perfect embedded feature vector. First of all, since the View1 and View2 views are essentially different, the differences in hyperedges will affect the embedded feature vectors of microorganisms and drugs. Therefore, the preferences of the two views are inconsistent, and thus the impact on the final microorganism-drug association prediction is also different. So a basic global average pooling followed by a fully connected neural network is used to calculate the weights of each view. Finally, the embedded representation and the attention weights are combined. Taking microorganisms as an example, it is defined as follows:

[0108]

[0109] where GAP(·) is the global average pooling layer, and FNN(·) is a two-layer fully connected neural network. The nonlinear activation functions of the two layers are ReLU() and Sigmoid() functions. represent the embeddings of the View1 and View2 views of microorganisms. Finally, the embedded representation of microorganisms with attention weights is obtained Similarly, the attention-weighted embedding of drugs can be obtained

[0110] Through the attention embedding, we obtained the embedded information of the two hypergraph structures. Inspired by Graph-Transformer

[25] , we introduced a multi-head attention mechanism to integrate the different views of microorganisms and drugs. Briefly speaking, taking microorganisms as an example, in the multi-head attention mechanism, the features extracted by each self-attention layer are obtained from different subspaces, and then these features from different subspaces are concatenated together to obtain the features of microorganisms. First, the calculated and are connected to obtain the comprehensive representation of microorganisms Finally Based on the basic architecture of Transformer, we use three projection weight matrices W q , W k and Wv Project the finally represented data of the input microorganisms onto three basic components, namely the microorganism query matrix Microorganism key matrix and the microorganism value matrix Here we continue to use the scaled dot - product function as the attention function. Finally, the inter - view attention matrix is defined as follows:

[0111]

[0112] where j represents View1 and View2, and A m (i, j) represents the attention degree of the i - th view to the j - th view of the current microorganism, and d f is the dimension of the microorganism embedding representation. Among them, A m (i, j) ∈ R 2×2 represents the inter - view attention matrix between the two views of the microorganism. It should be noted that the attention matrix here is the same as the number of microorganism nodes. According to the attention scores of the two views, the interaction between different views can be highlighted. At the same time, in order to learn to capture richer feature representation information and improve the robustness and generalization ability of the model learning process, the self - attention mechanism is constructed into multi - head attention here. The definition is as follows:

[0113]

[0114] where, N represents the number of heads, which is specified through parameter analysis in the experiment. Finally, we encode the feature vector embedding obtained by multi - head attention through a two - layer feed - forward network to obtain the final embedding representation, and the definition is as follows:

[0115] h m = W h ·Vec(V_ave m ) (3)

[0116] where, W h represents the parameter in the feed - forward network, and Vec(·) represents the vectorization operation of row connection, that is, connecting multiple vectors into a long vector by row. Then, for N m microorganisms, the embedding matrix can be expressed as Similarly, the embedding matrix representation of the drug can be obtained as

[0117] 1.3 Result prediction module

[0118] 1.3.1 Prediction score

[0119] In the final prediction score stage, we use the final embedding representations H m and Hd , the microorganism Y is obtained through a fully connected neural network (FNN) m = FNN(H m ), and the embedding matrix of drug Y d = FNN(H d ). Then, the reconstructed correlation matrix is obtained through the matrix multiplication of the two features, and the definition is as follows:

[0120]

[0121] 2. Experimental content

[0122] 2.1 Evaluation metrics

[0123] This scheme adopts a five-fold cross-validation strategy to evaluate the performance of DHCLHAM and baseline methods on the MDAD, aBiofilm, and DrugVirus datasets respectively. Briefly, for the above three datasets, all known microorganism-drug association pairs are regarded as positive samples and form a positive sample set, while all the remaining unknown microorganism-drug association pairs are regarded as a negative sample set. First, all the experimental data are divided into five sub-parts with the same data. Then, each sub-part is used as the test set in turn, and the other four sub-parts are used as the training set. Ensure that each part serves as both the training set and the test set. The evaluation of the model is based on the area under the ROC curve (AUC), the area under the precision-recall curve (AUPR), and the F1 score. The definitions are as follows.

[0124]

[0125] Among them, TP represents true positive, FP represents false positive, TN represents true negative, and FN represents false negative.

[0126] 2.2 Comparative experiments

[0127] To illustrate the superiority of this scheme, the proposed DHCLHAM of this scheme is compared with six leading methods in this field as Figure 5 shown, including respectively:

[0128] GCNMDA: Feature processing is performed using Random Walk with Restart (RWR). Then, a Conditional Random Field (CRF) layer is embedded in GCN and an attention mechanism is designed to update node embeddings. See Long Y, Wu M, Kwoh CK, et al. Predicting human microbe-drug associations via graph convolutional network with conditional random field. Bioinformatics. 2020;36(19):4918-4927. DOI: 10.1093 / bioinformatics / btaa598.

[0129] Graph2MDA: Based on Variational Graph Autoencoder (VGAE) and Deep Neural Network (DNN). See Deng L, Huang Y, Liu X, et al. Graph2mda: a multi-modal variational graph embedding model for predicting microbe-drug associations[J]. Bioinformatics 2022;38(4):1118-1125. DOI: 10.1093 / bioinformatics / btab792.

[0130] LCASPMDA: Extract node features using a Learnable Graph Convolutional Attention Network (LCAT). See Yang Z, Wang L, Zhang X, et al. LCASPMDA: a computational model for predicting potential microbe-drug associations based on learnable graph convolutional attention networks and self-paced iterative sampling ensemble. Front Microbiol. 2024;15:1366272. DOI: 10.3389 / fmicb.2024.1366272.

[0131] SCSMDA: Construct a similarity network and a meta-path induced network of microbes and drugs, and then use a structure-enhanced contrastive learning strategy to enhance node embeddings. See Tian Z, Yu Y, Fang H, et al. Predicting microbe-drug associations with structure-enhanced contrastive learning and self-paced negative sampling strategy. Brief Bioinform 2023; 24(2): bbac634. DOI: 10.1093 / bib / bbac634.

[0132] NRGCNMDA: Extract shallow features through Node2vec, then use a residual graph convolutional network (REGCN) to capture long-range dependencies through skip connections, introduce context constraints to smooth feature embeddings with a conditional random field (CRF) layer, and finally calculate the association score through a bilinear decoder. See Du X, Li J, Wang B, et al. NRGCNMDA: Microbe-Drug Association Prediction Based on Residual Graph Convolutional Networks and Conditional Random Fields. Interdiscip Sci. Published online January 7, 2025. DOI: 10.1007 / s12539-024-00678-z.

[0133] MCHAN: Extract key information with a graph convolutional network and an attention mechanism. See Li G, Cao Z, Liang C, et al. MCHAN: Prediction of human microbe-drug associations based on multiview contrastive hypergraph attention network. Current Bioinformatics 2024. 20: 70-86. DOI: 10.2174 / 0115748936288616240212073805.

[0134] The comparison results show that the proposed DHCLHAM in this scheme is compared with other baseline methods in terms of AUC and AUPR metrics, and the corresponding results on the MDAD, aBiofilm, and DrugVirus datasets. Although the AUC value of 0.9223 on the DrugVirus dataset is slightly lower than the AUC value of 0.9267 of the NRGCNMDA method, the AUPR value of 0.9214 is significantly higher than the AUPR value of 0.9024 of the NRGCNMDA method, which may be due to the relatively small amount of data in the DrugVirus dataset. Additionally, on the other two datasets, MDAD and aBiofilm, the AUC values of our method DHCLHAM are 0.9827 and 0.9861, and the AUPR values are 0.9787 and 0.9833, both significantly higher than other methods. This may be because we utilized the hypergraph structure and simultaneously adopted a two-level hypergraph attention mechanism to extract high-dimensional information, and conducted contrastive learning on two hypergraph structures. Generally speaking, DHCLHAM can more accurately reflect the complex relationship between microorganisms and drugs, thereby improving the prediction performance.

[0135] 2.3 Ablation Experiments

[0136] To verify the importance of different modules in the proposed DHCLHAM model of this scheme, we conducted ablation analysis on three important modules in three datasets to determine the impact of these modules on the performance of the final model.

[0137] 2.3.1. Impact of Hierarchical Attention and Contrastive Learning

[0138] NoDHA: For hierarchical attention, the NoDHA module of hierarchical attention is omitted and replaced with ordinary graph attention.

[0139] NoCL: For contrastive learning, the NoCL module of hypergraph contrastive learning is omitted, and the features are directly fused through the integration network after hypergraph convolution.

[0140] The performance of NoDHA in terms of the four evaluation metrics of AUC, AUPR, F1, and ACC on the three datasets is lower than that of DHCLHAM. This may be attributed to the sparse interaction data between microorganisms and drugs, while the two-level attention mechanism provides potential high-order information and enriches the network. By introducing attention, the model can initially update the nodes of the hyperedges and then feedback the update of node information, which further explains the importance of the hierarchical attention module. Additionally, the performance of the NoCL model also decreases in the four evaluation metrics. Hypergraph contrastive learning helps to identify and enhance the node information of different views, thereby improving the accuracy. These findings verify the importance of each component of DHCLHAM and the overall efficacy of the model. As Figure 6 shown.

[0141] 2.3.2. Influence of Nonlinear Fusion

[0142] Nonlinear fusion can naturally incorporate multiple similarity information from the microbial and drug feature spaces, thereby effectively improving the prediction ability of the model. The effectiveness of nonlinear fusion is compared by comparing the model with single similarity and two-similarity linear fusion. Specifically,

[0143] MS1&DS1: In the data preprocessing stage, only microbial functional similarity and drug structural similarity are used.

[0144] MS2&DS2: In the data preprocessing stage, only microbial Gaussian kernel similarity and drug Gaussian kernel similarity are used.

[0145] Linear: In the data preprocessing stage, linear average fusion is adopted for microbial functional similarity and microbial Gaussian kernel similarity, and linear average fusion is adopted for drug structural similarity and drug Gaussian kernel similarity.

[0146] The model using the above similarity combinations proposed by us conducted experiments with the same parameter settings on the MDAD dataset. As Figure 7 shown, the prediction results of MS1&DS1, MS2&DS2, and Linear all decreased on the MDAD dataset. This indicates that nonlinear fusion helps to accurately model the nonlinear relationship between microbial similarity views and drug similarity, thereby improving the overall performance of the model.

[0147] 2.4 Parameter Analysis

[0148] For the DHCLHAM model proposed in this scheme, some key parameters will affect its performance. Here we mainly focus on four parameters: the number of multi-head attention heads head, the number of HGCL layers, the value of K in the KNN method in hypergraph construction, and the number of cluster centers C in the K-means method. Corresponding experiments were carried out, and the results were evaluated using AUC, AUPR, F1, and ACC.

[0149] (1) Number of multi-head attention heads, as Figure 8 shown, we change the value of head in the set {2, 3, 4, 5, 6}. It can be observed that for the MDAD and aBiofilm datasets, the optimal results in various metrics are achieved when the number of heads is 5, while for the DrugVirus dataset, the highest evaluation metrics are achieved when the number of heads is 4.

[0150] (2) Number of HGCN layers, as shown in Table 2, the settings of HGCN are {1, 2, 3, 4}. The ones in bold in the table indicate the best performance. Obviously, when the number of layers is 2, the evaluation metrics of the model reach the peak for all three datasets.

[0151] (3) In the KNN method for hypergraph construction, the value of K and the number of cluster centers C in the K-means method can determine the length and quantity of hyperedges. Selecting appropriate quantities is crucial. Here, the DrugVirus dataset is taken as an example. As Figure 9 shown, when K is 13 and C is 9, the performance of DHCLHAM is optimal.

[0152] Table 2 Influence of DHCLHAM on performance under different HGCN layers predicted on three datasets

[0153]

[0154] 2.5 Case Analysis

[0155] To comprehensively verify the DHCLHAM proposed in this scheme for finding new microbial-drug interactions and the analysis of the interpretability of prediction results, we designed two cases, namely the literature verification method and network pharmacology, for verification.

[0156] 2.5.1. Case 1

[0157] We conducted a case study on two commonly used antibacterial drugs, ciprofloxacin and moxifloxacin, on the MDAD dataset, which is the same as previous studies. Specifically, for the selected target drugs, all known microbial-drug associations will be set as unknown, and then all candidate microorganisms will be sorted in descending order according to the scores predicted by DHCLHAM. Finally, we screened out the top 20 microorganisms and verified them through published literature. Figure 10 Visually shows the prediction results of ciprofloxacin and moxifloxacin for microorganisms.

[0158] Table 3 The top 20 predicted ciprofloxacin-related microorganisms

[0159]

[0160] Ciprofloxacin is a second-generation fluoroquinolone drug used to treat respiratory infections and sepsis. It can treat a variety of bacterial infections, especially those caused by Gram-negative pathogens. The top 20 microorganisms predicted for ciprofloxacin are shown in Table 3, among which 18 microorganisms are verified by literature records, and the accuracy rate of the top 20 reaches 90%.

[0161] Moxifloxacin is a broad-spectrum antibacterial drug belonging to the fourth-generation quinolone antibiotics. It is usually used to treat upper and lower respiratory tract infections, such as acute sinusitis, pneumonia, and skin and soft tissue infections. The top 20 microorganisms predicted for moxifloxacin are shown in Table 4, among which 16 microorganisms are verified by literature records, and the accuracy rate of the top 20 reaches 80%.

[0162] Table 4 The top 20 predicted moxifloxacin-related microorganisms

[0163]

[0164] Table 5 The top 20 predicted SARS-COV-2-related drugs

[0165]

[0166] 2.5.2. Case 2

[0167] To further verify the interpretability of the model for predicting microorganism-drug associations, taking the SARS-COV-2 virus as an example, we used the method of network pharmacology for analysis. First, we obtained the top twenty drugs most relevant to the SARS-COV-2 virus using the prediction scores of the model. Fourteen drugs were verified to be related to the SARS-COV-2 virus in the literature search, as shown in Table 5. Secondly, we obtained the gene information of the SARS-COV-2 virus and the target information of 20 verified and unverified drugs respectively. For the genes of the SARS-COV-2 virus, in order to obtain more comprehensive gene information, we obtained the gene information of the SARS-COV-2 virus from four databases: Genecards, NCBI, Uniport, and OMIM. In the Genecards database, in order to obtain more accurate gene information, we took the median of the "Score" column to obtain the specific gene information. In the NCBI, Uniport, and OMIM databases, we selected the "Human" label to obtain the specific gene information. We removed duplicates from the gene information of the four databases to obtain the final gene information of the SARS-COV-2 virus. For the target gene information of the drugs, we obtained the target protein information of the drugs from ChEMBL, selected "Human" using Uniport to convert it into target genes, and for drugs that did not find target information in ChEMBL, we used the Swiss database for target prediction. When obtaining the target information of all verified and unverified drugs, only two target genes were different from the verified drugs. Therefore, for the target genes of all drugs, we removed the two different genes of the unverified drugs, and then removed duplicates from all the genes finally. Finally, we performed Go enrichment analysis using the gene information of the SARS-COV-2 virus and the target information of the drugs. Figure 12The specific parameters that can be enriched by the bubble chart show that the best-enriched pathway is transcription factor binding, and the worst is IkappaB kinase activity. Obviously, for the pathway protein kinase bindinggo, after SARS-CoV-2 infects host cells, it will utilize the host's protein kinases to promote its own replication and transcription. Regarding transcription factor binding, SARS-CoV-2 infection can activate or interfere with host transcription factors to trigger immune responses, etc. The virus's own replication and transcription rely on host transcription factors and affect their recruitment, and this pathway is of great significance in both immune cell functions and the pathological processes caused by infection. Figure 11 The enrichment information of specific genes in these 20 biological processes can be seen. For example, for the pathway ubiquitin-like protein ligase binding, after SARS-CoV-2 infects host cells, it will utilize or interfere with the binding process of ubiquitin-like protein ligases in host cells. Obviously, through enrichment analysis, it can well verify and explain the prediction of the model for the interaction between microorganisms and drugs.

[0168] In summary, this scheme proposes a new DHCLHAM based on hierarchical attention-based dual hypergraph contrast learning. By extracting microorganism-drug associations, then calculating the functional similarity and Gaussian kernel similarity of microorganisms, the structural similarity and Gaussian kernel similarity of drugs. A comprehensive similarity matrix is generated through non-linear fusion. Use the KNN and K-means algorithms to construct a hypergraph, combine the fused similarity matrix with the original associations to establish hyperedges. Design a hierarchical attention to calculate the hyperedge attention and node attention of the hypergraph respectively, and then extract topological features through hypergraph convolution. Then use contrast learning to improve the embedding quality across hypergraph views. Visual perception attention dynamically fuses multi-perspective features, and multi-head attention adaptively fuses dual-channel embeddings. Finally, optimize the embedding through a fully connected layer and reconstruct the association probability through matrix multiplication. Use network pharmacology to verify including target localization, pathway analysis, and functional enrichment to biologically explain the predicted drug-microorganism interactions. In addition, the research results of this study have important reference value for optimizing clinical treatment plans and lay a theoretical foundation for the development of precision drug strategies targeting the gut microbiota.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A method for predicting microbial-drug interactions based on a hierarchical attention-based dual hypergraph contrastive learning framework, characterized in that: It includes the following steps: S1. Nonlinearly fuse the microbial functional similarity and microbial Gaussian kernel similarity of microorganisms to obtain the comprehensive similarity of microorganisms. Nonlinearly fuse the drug structure similarity and drug Gaussian kernel similarity of drugs to obtain the comprehensive similarity of drugs. Combine the original microorganism-drug association matrix and the comprehensive similarity, and use the KNN and K-means algorithms to construct a bi-hypergraph. S2. Feed the bi-hypergraph obtained in step S1 into the hierarchical attention and hypergraph convolutional network to obtain feature embeddings. Use the two hypergraphs as positive and negative samples for each other, and update the features by minimizing the positive sample pairs and maximizing the negative sample pairs using contrastive learning. Integrate the obtained features through an integration network to dynamically fuse the bi-hypergraph features. S3. Map the microorganism and drug features obtained in step S2 using a fully connected layer to obtain the final microorganism and drug features, and calculate the dot product of the two to obtain the final prediction score. S4. Combine the prediction score obtained in step S3 with specific literature verification and network pharmacology verification model outputs.

2. A method for predicting microbial-drug interactions using a dual hypergraph contrastive learning framework based on hierarchical attention according to claim 1, characterized in that: In step S1, the data preprocessing steps include obtaining microbial functional similarity, drug structure similarity, microbial Gaussian kernel similarity, and drug Gaussian kernel similarity. The nonlinear fusion includes local information acquisition and normalization.

3. A method for predicting microbial-drug interactions based on a hierarchical attention-based bi-hypergraph contrastive learning framework according to claim 1, characterized in that: In step S1, the calculation of the microbial Gaussian kernel similarity and the drug Gaussian kernel similarity is as follows: MS2(m i ,m j ) = exp(-η m ||A(m i ,·) - A(m j ,·)|| 2 ) Among them, MS2 represents the microbial Gaussian kernel similarity, m i and m j represent the i-th and j-th microorganisms in the original correlation matrix, exp represents the exponential function, η m represents the normalized kernel bandwidth of microorganisms, A represents the original correlation matrix, || || 2 represents the square of the 2-norm. η m ' represents the original bandwidth, which is always set to 1, N m represents the number of microorganisms.

4. A method for predicting microorganism-drug interactions based on a hierarchical attention-based dual hypergraph contrastive learning framework according to claim 3, characterized in that: The comprehensive similarity is obtained by nonlinearly fusing the similarity data, as follows: where \(t = 1, 2, \ldots, M\), and \(M\) is the total number of similarity data. Use the KNN algorithm to measure the local relationship for each similarity data, as follows: Where N i is a set of k nearest neighbors of node m i in the microbial similarity data. K t represents the local affinity kernel of the t-th data type, and the neighbor parameter of KNN takes N m / 10. Finally, the model iteratively updates the similarity matrix of each type of data using the program as follows: Among them is the state matrix of the t-th data type after the r-th iteration, and T represents the matrix transpose.

5. A method for predicting microbial-drug interactions based on a hierarchical attention-based bi-hypergraph contrastive learning framework according to claim 4, characterized in that: In step S2, hierarchical attention is a process of aggregating information from points and hyperedges on the hypergraph structure. First, information aggregation from points to edges, i.e., hyperedge-level attention, is performed, and then information aggregation from edges to points, i.e., node-level attention. Here, the set of all nodes connected to hyperedge e i is defined as y i , and the set of hyperedges connected to node n i is located as ρ i . Hyperedge-level attention captures the different importance of all nodes n j connected to the given hyperedge e p ∈y i . The attention score a i of node n j to hyperedge e ji is defined as follows: where w1 is a learnable parameter matrix, u is a trainable weight vector, and S(·) is a similarity function, defined as Scaled Dot-Product Attention, where D is the dimension. Use the updated hyperedge representation to obtain the node representation. Node attention determines the different importance of different hyperedges, and hyperedge e j for node n i has an attention score b ij defined as follows: where w2 and w3 are learnable parameter matrices, and S(·) is a similarity function, defined as ScaledDot-Product Attention, Finally, the hyperedge representation e is derived by aggregating their neighbor information j and the updated node representation Z i , which are defined as follows: where \(\varphi(\cdot)\) consists of two layers of MLP and an ELU activation function.

6. A method for predicting microbial-drug interactions using a dual hypergraph contrastive learning framework based on hierarchical attention as claimed in claim 5, characterized in that: By using dual hypergraph contrastive learning, two hypergraphs are regarded as positive and negative samples for each other, and the features are updated by minimizing positive sample pairs and maximizing negative sample pairs in contrastive learning. The obtained features are dynamically fused with the dual hypergraph features through an integration network. For the selection of positive and negative samples, for each node v, the embedding v generated in one view i is defined as the anchor point, and the embedding generated in the other view is denoted as u i . In this way, different embeddings v i and u i of the same node in two different views form a positive sample pair. The embeddings v k and u k (k≠i) of other nodes are regarded as negative sample pairs, where v k and the anchor point v i form an intra-view negative sample pair, while u k and the anchor point v i form a cross-view negative sample pair. At this time, the contrastive loss is defined as follows: where \(sim(\cdot)\) is the cosine similarity function, \(g(\cdot)\) is a two-layer neural network projection head used to enhance the information ability of nodes. \(\tau\) is the temperature control parameter.

7. A method for predicting microorganism-drug interactions based on a hierarchical attention-based dual hypergraph contrastive learning framework according to claim 1, characterized in that: In step S2, use the integration network to dynamically fuse the bi-hypergraph features. First, use the attention mechanism between the two views to update the embeddings of the two views, and then use the multi-head attention mechanism to fuse the features of the two hypergraphs to obtain the final microorganism embedding representation and drug embedding representation.

8. A method for predicting microbial-drug interactions based on a hierarchical attention-based dual hypergraph contrastive learning framework according to claim 1, characterized in that: In step S3, obtain the embedding matrix for the final microorganism embedding representation and drug embedding representation through a fully connected neural network, and then obtain the reconstructed association matrix through the matrix multiplication of the two features.

9. A method for predicting microbial-drug interactions based on a hierarchical attention-based bi-hypergraph contrastive learning framework according to claim 1, characterized in that: In step S4, two cases of literature verification method and network pharmacology are designed to verify the interpretability.

Images