Microbe-drug association prediction method and system based on structure-aware transformer

Abstract

The application provides a microorganism-drug association prediction method based on a structure perception converter, and comprises the following steps: S1, obtaining the similarity of microorganisms and drugs according to a microorganism-drug association network and a microorganism-disease-drug association network; S2, based on the similarity of the microorganisms and the drugs, extracting feature information between nodes by using a restart random walk algorithm; S3, constructing a heterogeneous network through the microorganism-drug association network, the microorganism similarity and the drug similarity; S4, inputting the heterogeneous network into a structure perception converter for training to obtain a trained structure perception converter; and S5, predicting potential microorganism-drug associations according to the trained structure perception converter. The structure perception converter can capture network structure information, capture structure and attribute features between nodes, solve the problems of over-smoothing and over-squeezing caused by a graph neural network, and improve the prediction accuracy of microorganism-drug associations.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, and in particular relates to a method and system for predicting microbial-drug associations based on a structure-aware converter. Background Technology

[0002] Numerous microorganisms reside within the human body, typically acting as dynamic, diverse, and complex genetic hosts that colonize human cells and play vital roles. Microorganisms are involved in drug therapy and disease development. On one hand, they promote metabolism, produce essential vitamins and genetic products, and prevent pathogen invasion. On the other hand, research indicates that abnormal growth or decay of microorganisms can seriously impact human health, leading to diseases such as obesity, diabetes, and even cancer.

[0003] Mounting clinical evidence suggests a close relationship between the gut microbiome and human health, with countless diseases caused by microorganisms requiring therapeutic drugs. Therefore, predicting the correlation between microorganisms and drugs plays a crucial role in new drug discovery, clinical drug application, and drug quality control.

[0004] However, due to the increasing variety of drugs being researched in the medical field, microbial resistance is also becoming stronger, and traditional experimental methods are limited by expensive equipment and heavy workloads. In recent years, rapid advancements in genomics, proteomics, life sciences, and pharmaceutical research have led to the rapid development of bioinformatics, resulting in the accumulation of a vast amount of biomedical data. Based on this, researchers have proposed many computational methods to infer potential microbe-drug associations.

[0005] Currently, methods for predicting microbe-drug associations mainly fall into two categories: similarity-based methods and graph representation learning-based methods. Similarity-based methods may fail to accurately measure the distance between nodes, leading to biased classification results. Graph representation learning-based methods primarily include graph neural network models such as Graph Convolutional Neural Networks (GCN), Graph Attention Networks (GAT), and Graph Variational Autoencoders (VGAE). Graph neural networks are highly effective in improving model predictive capabilities, but during computation, they often introduce problems such as oversmoothing and over-compression.

[0006] In summary, while the aforementioned models demonstrate reliable performance in some aspects, they still have certain limitations. Therefore, it is crucial to research new algorithms for calculating microbe-drug associations. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for predicting microbial-drug associations based on structure-aware converters, in order to solve the technical problems of expensive equipment and heavy workload in predicting microbial-drug associations by traditional experimental methods, as well as the problems of oversmoothing and over-compression that are prone to occur when using graph neural network models.

[0008] The technical solution of this invention is as follows:

[0009] S1: Obtain the similarity between microorganisms and drugs based on the microorganism-drug association network and the microorganism-disease-drug association network;

[0010] S2: Based on the similarity between the microorganisms and the drugs, a restarted random walk algorithm is used to extract feature information between nodes;

[0011] S3: Construct heterogeneous networks through microbial-drug association networks, microbial similarity, and drug similarity;

[0012] S4: Input the heterogeneous network into the structure-aware converter for training to obtain a trained structure-aware converter.

[0013] S5: Predict potential microbe-drug associations based on trained structure-aware converters.

[0014] Furthermore, step S1 includes:

[0015] Calculate microbial Gaussian kernel similarity and drug Gaussian kernel similarity using a microbial-drug association network;

[0016] By performing deduplication on the microorganism-disease-drug association network, we obtain the microorganism-drug association network, the microorganism-disease association network, and the drug-disease association network. We then calculate the semantic similarity of diseases and further calculate the functional similarity of microorganisms and drugs based on the semantic similarity of diseases.

[0017] Furthermore, the microbial Gaussian kernel similarity and microbial functional similarity, as well as the drug Gaussian kernel similarity and drug functional similarity, are calculated through microbial-drug association networks and microbial-disease-drug association networks, including the following steps:

[0018] An adjacency matrix was constructed based on the microbial-drug association network. Define the rows of the adjacency matrix A as n r n r Representing the quantity of drugs, the matrix has n columns. m n m Represents the number of microorganisms;

[0019] Let S(i) and S(j) represent the i-th and j-th rows of the adjacency matrix A, then the drug r iand drugs r j The Gaussian kernel similarity (GIR) between the drugs is calculated as follows:

[0020] GIR(r i ,r j )=exp(-μ||S(i)-S(j)|| 2 )

[0021]

[0022] Where μ is the standard kernel bandwidth, obtained by normalizing the parameters;

[0023] Similarly, let R(i) and R(j) represent the i-th and j-th columns of the adjacency matrix A, and let m be the number of microorganisms. i and microorganisms m j The Gaussian kernel similarity (GIM) between the microorganisms is calculated as follows:

[0024] GIM(m i ,m j )=exp(-μ||R(i)-R(j)|| 2 )

[0025]

[0026] Based on the microbiome-drug association network, the microbiome-disease association network, and the drug-disease association network, common diseases related to microorganisms and drugs are obtained. Semantic information about these diseases is then derived from disease ontology, and disease d is calculated. i and disease d j The semantic similarity DSE between them is calculated as follows:

[0027]

[0028] Let D1 be the subset containing disease d. i and the collection of all its related diseases, The semantic contribution value of all items in D1 is represented by the following formula:

[0029]

[0030] Same as above, in addition DV(d) represents the sum of the semantic contributions of all sub-items of D1. j Same as above;

[0031] Based on the above disease semantic similarity DSE, the drug functional similarity (FCR) is calculated as follows:

[0032]

[0033] Among them, for drug r i and drugs r j There are b1 diseases and r i Related, there are also B2 diseases and r j Related, the disease is represented as {d it (1≤t≤b1)} and {b js (1≤s≤b2)};

[0034] For microorganisms m i and microorganisms m j There are n1 diseases and m i Related, n2 diseases and m j Related, the disease is represented as {d ip (1≤p≤n1)} and {d jk (1≤k≤n2)}, the functional similarity of microorganisms (FCM) is calculated as follows:

[0035]

[0036] Drug similarity is represented as IRS(r i ,r j ):

[0037]

[0038] Microbial similarity is represented as IMS(m i ,m j ):

[0039]

[0040] Furthermore, by restarting the random walk algorithm, new drug similarity (IRS) and microbial similarity (IMS) are obtained by calculating the drug similarity (Sr) and microbial similarity (Sm). The formula for restarting the random walk algorithm is as follows:

[0041]

[0042]

[0043] Where λ is the restart probability and is set to 0.1, M is the transition probability matrix, and e i ∈R (1*m) It is the initial probability vector of node i. This represents the probability that node i reaches other nodes at time l.

[0044] Furthermore, a heterogeneous network is constructed based on the new drug similarity Sr, microbial similarity Sm, and adjacency matrix A. as follows:

[0045]

[0046] Among them, A T Let A be the transpose of the adjacency matrix A.

[0047] Furthermore, based on the aforementioned heterogeneous network, the model is trained using it as input to the structure-aware converter model, including:

[0048] The structure-aware converter includes a structure-aware attention mechanism and an encoder portion of the converter;

[0049] The calculation formula for the structure-aware attention mechanism is as follows:

[0050]

[0051] Among them, X v S represents the feature vector of node v. G (v) represents the subgraph of graph G centered at node v, f(x) μ ) = W v X is a linear function, W v It is a trainable parameter, and k graph It can be any kernel that compares a pair of subgraphs, k in the structure-aware converter graph The calculation method is as follows:

[0052]

[0053] Where, k exp This represents an asymmetric exponential kernel. It is a structure extractor based on a k-layer graph convolutional neural network (GCN). It extracts vector representations of some subgraphs G centered at node v, as follows:

[0054]

[0055] The encoder part of the converter consists of N identical stacked layers, each containing two sub-layers: a self-attention mechanism layer and a feedforward network layer. After defining the structure-aware attention mechanism, the encoder of the converter constructs the feedforward network layer as follows:

[0056] X' = ​​LayerNorm(X + SubLayer(X))

[0057] X″=FFN(X′)=ReLU(X′W1)W2

[0058] Here, X represents the input node features, W1 and W2 represent two trainable parameters, ReLU is used as the activation function, there is a residual connection in each sublayer of the structure-aware converter, and then a layer normalization operation is performed.

[0059] Furthermore, the reconstructed heterogeneous network is obtained by performing an aggregation operation on the flattened output of the self-attention layer through a fully connected layer. The computation steps are as follows:

[0060] F = ReLU(X″W) + b

[0061] Where W is a trainable parameter, b represents the bias, and ReLU is an activation function;

[0062] The reconstructed heterogeneous network is input into a deep neural network classifier to obtain the microbe-drug association score y. score The binary cross-entropy loss function is used, as shown below:

[0063] Loss=-∑[y pred log(y score )+(1-y pred log(1-y) score )]

[0064] Among them, y pred It is represented as a vector after the adjacency matrix A is flattened, and log represents the natural logarithm;

[0065] Through the above steps, a trained structure-aware converter model is obtained.

[0066] Furthermore, the accuracy of the trained structure-aware converter model is verified through the following steps:

[0067] Five-fold cross-validation was used to randomly divide the microbial-drug pair into five subsets of equal number. Each subset was used as a non-repeating test sample set, and the other four subsets were used as training sample sets. The structure-aware converter model was trained using the training set, and the test set was input into the trained structure-aware converter to obtain the classification results.

[0068] A positive classification result indicates that there is an association between the microorganism and the drug, while a negative classification result indicates that there is no association between the microorganism and the drug.

[0069] The AUC and AUPR values ​​of the trained structure-aware converter model are obtained based on the classification results. The accuracy of the structure-aware converter is verified based on these two evaluation metrics.

[0070] The present invention also provides a microbial-drug association prediction system based on a structure-aware converter, comprising: a processor that executes various instructions; and a storage device that stores multiple instructions, which are loaded by the processor and executed according to the above method.

[0071] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0072] It adds microbe-disease-drug associations and recalculates microbe functional similarity and drug functional similarity. Based on multiple similarity views, it learns the embedding features of microbes and drugs, fully exploring the potential features between microbes and drugs.

[0073] The structure-aware converter can capture network structure information as well as the structural and attribute features between nodes, and solves the problems of oversmoothing and over-compression caused by graph neural networks, thus improving the accuracy of predicting microbe-drug associations. Attached Figure Description

[0074] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0075] Figure 1 This is a framework diagram of a microbial-drug association prediction method based on a structure-aware converter according to the present invention.

[0076] Figure 2 This is an example diagram of the structure extractor in a structure-aware converter.

[0077] Figure 3 This is a flowchart of the present invention. Detailed Implementation

[0078] The inventive concept of this invention is as follows: First, the Gaussian kernel similarity between microorganisms and drugs is calculated based on a microorganism-drug association network. To obtain richer feature information, the functional similarity of microorganisms and drugs is recalculated using a microorganism-disease-drug association network. Next, the two similarity networks of microorganisms and drugs are integrated, and a restart random walk algorithm is applied to each. Then, a heterogeneous network is constructed based on the aforementioned microorganism-drug association network, microorganism similarity, and drug similarity. A structure-aware converter is used to learn the node features and network structure information in the heterogeneous network. Finally, a deep neural network classifier is used to obtain the association probability between microorganism-drug pairs. This invention learns the embedding features of microorganisms and drugs based on multiple similarity views, fully exploring the potential features between microorganisms and drugs. Furthermore, it solves the problems of oversmoothing and over-compression caused by prediction using graph neural network models in existing methods, improving the accuracy of predicting microorganism-drug associations.

[0079] The present invention will be further described below with reference to the embodiments.

[0080] Example 1

[0081] A method and system for predicting microbial-drug associations based on structure-aware converters includes the following steps:

[0082] S1: Obtain the similarity between microorganisms and drugs based on the microorganism-drug association network and the microorganism-disease-drug association network;

[0083] S2: Based on the similarity between the microorganisms and the drugs, a restarted random walk algorithm is used to extract feature information between nodes;

[0084] S3: Construct heterogeneous networks through microbial-drug association networks, microbial similarity, and drug similarity;

[0085] S4: Input the heterogeneous network into the structure-aware converter for training to obtain a trained structure-aware converter.

[0086] S5: Predict potential microbe-drug associations based on trained structure-aware converters.

[0087] Example 2

[0088] The specific steps of step S1 include:

[0089] Microbe-drug association network was established, and microbe Gaussian kernel similarity and drug Gaussian kernel similarity were calculated.

[0090] By performing deduplication on the microorganism-disease-drug association network, we obtain the microorganism-drug association network, the microorganism-disease association network, and the drug-disease association network. We then calculate the semantic similarity of diseases and further calculate the functional similarity of microorganisms and drugs based on the semantic similarity of diseases.

[0091] In one specific embodiment, the calculation of drug Gaussian kernel similarity and microbial Gaussian kernel similarity through a microbial-drug association network specifically includes the following steps:

[0092] An adjacency matrix was constructed based on the microbial-drug association network. Define the rows of the adjacency matrix A as n r , representing the quantity of drugs, the matrix has n columns. m , representing the number of microorganisms;

[0093] Let S(i) and S(j) represent the i-th and j-th rows of the adjacency matrix A, then the drug r i and drugs r jThe Gaussian kernel similarity (GIR) between the drugs is calculated as follows:

[0094] GIR(r i ,r j )=exp(-μ||S(i)-S(j)|| 2 )

[0095]

[0096] Where μ represents the standard kernel bandwidth, which is obtained by normalizing the parameters;

[0097] Similarly, let R(i) and R(j) represent the i-th and j-th columns of the adjacency matrix A, then the number of microorganisms m i and microorganisms m j The Gaussian kernel similarity (GIM) between the microorganisms is calculated as follows:

[0098] GIM(m i ,m j )=exp(-μ||R(i)-R(j)|| 2 )

[0099]

[0100] In one specific embodiment, the calculation of drug functional similarity and microbial functional similarity through a microbe-disease-drug association network specifically includes the following steps:

[0101] Based on the microbiome-drug association network, the microbiome-disease association network, and the drug-disease association network, common diseases related to microorganisms and drugs are obtained. Semantic information about these diseases is then derived from disease ontology, and disease d is calculated. i and disease d j The semantic similarity DSE between them is calculated as follows:

[0102]

[0103] Here, it is assumed that D1 is a subset containing disease d. i And the collection of all its related diseases, then The semantic contribution value of all items in D1 is represented by the following formula:

[0104]

[0105] Same as above, in addition, DV(d) represents the sum of the semantic contributions of all sub-items of D1. j Same as above;

[0106] Based on the above disease semantic similarity DSE, the drug functional similarity (FCR) is calculated as follows:

[0107]

[0108] Among them, for drug r i and drugs r j Suppose there are b1 diseases and r i Related, there are also B2 diseases and r j Relatedly, the disease can be represented as {d} it (1≤t≤b1)} and {d js (1≤s≤b2)};

[0109] Similarly, for microorganisms m i and microorganisms m j Suppose there are n1 kinds of diseases and m i Related, n2 diseases and m j Relatedly, the disease can be represented as {d} ip (1≤p≤n1)} and {d jk (1≤k≤n2)}, then the microbial functional similarity FCM is calculated as follows:

[0110]

[0111] To efficiently utilize the model for node feature extraction and node classification prediction, drug Gaussian kernel similarity (GIR) will be combined with drug functional similarity (FCR) to obtain drug similarity IRS.

[0112] Specifically, for drug r i and r j If there is drug functional similarity between them, the drug similarity is defined as the average of GIR and FCR; otherwise, it is defined as GIR. The calculation formula is as follows:

[0113]

[0114] Similarly, combining microbial Gaussian kernel similarity (GIM) with microbial functional similarity (FCM) yields microbial similarity index (IMS). The calculation formula is as follows:

[0115]

[0116] Example 3

[0117] The specific steps of S2 include:

[0118] Based on the drug similarity IRS and microbial similarity IMS obtained in step S1, the restarted random walk algorithm is used to calculate the IRS and IMS to obtain new drug similarity Sr and microbial similarity Sm. The formula for the restarted random walk algorithm is:

[0119]

[0120]

[0121] Where λ is the restart probability, which is set to 0.1, and M is the transition probability matrix. i ∈R (1*m) It is the initial probability vector of node i. This represents the probability that node i reaches other nodes at time l.

[0122] Example 4

[0123] The specific steps of S3 include:

[0124] Based on the new drug similarity Sr and microbial similarity Sm obtained in step S2, a heterogeneous network is constructed by combining it with the adjacency matrix A. as follows:

[0125]

[0126] Among them, A T Let A be the transpose of the adjacency matrix A.

[0127] Example 5

[0128] The specific steps of S4 are as follows:

[0129] The heterogeneous network obtained in step S3 is used as input to train the structure-aware converter model.

[0130] The structure-aware converter mainly includes a structure-aware self-attention mechanism and the encoder part of the converter;

[0131] The attention mechanism in the transformer can be rewritten as a kernel smoother, where the kernel is a trainable exponential kernel defined on the node features that can capture the attribute features of the node pairs.

[0132] To further obtain the structural similarity of node pairs, the structure-aware converter constructs an attention mechanism that can interpret the local substructure around each node, namely the structure-aware attention mechanism, whose calculation formula is as follows:

[0133]

[0134] Among them, X v S represents the feature vector of node v.G (v) represents the subgraph of graph G centered at node v, f(x) μ ) = W v X is a linear function, W v It is a trainable parameter, and k graph It can be any kernel that compares a pair of subgraphs, k in the structure-aware converter graph The calculation method is as follows:

[0135]

[0136] Where, k exp This represents an asymmetric exponential kernel. It is a structure extractor based on a k-layer graph convolutional neural network (GCN). It extracts vector representations of some subgraphs G centered at node v, as follows:

[0137]

[0138] The encoder of the converter consists of N identical stacked layers, each containing two sub-layers: a self-attention mechanism layer and a feedforward network layer. After defining the structure-aware attention mechanism, the encoder part of the converter constructs the feedforward network layer as follows:

[0139] X' = ​​LayerNorm(X + SubLayer(X))

[0140] X″=FFN(X′)=ReLU(X′W1)W2

[0141] Here, X represents the input node features, W1 and W2 represent two trainable parameters, ReLU is used as the activation function, there is a residual connection in each sublayer of the structure-aware converter, and then a layer normalization operation is performed.

[0142] The reconstructed heterogeneous network is obtained by performing an aggregation operation on the flattened output of the self-attention layer using a fully connected layer. The computation steps are as follows:

[0143] F = ReLU(X″W) + b

[0144] Where W is a trainable parameter, b represents the bias, and ReLU is the activation function;

[0145] The AdamW optimizer is used to optimize the entire structure-aware converter. As a variant of Adam, it can not only automatically adjust the learning rate but also automatically adjust the weight coefficients, making the model more stable and thus avoiding overfitting.

[0146] The reconstructed heterogeneous network is input into a deep neural network classifier to obtain the microbe-drug association score y. score The binary cross-entropy loss function is used, as shown below:

[0147] Loss=-∑[y pred log(y score )+(1-y pred log(1-y) score )]

[0148] Where y pred It is represented as a vector after the adjacency matrix A has been flattened, and log represents the natural logarithm;

[0149] Through the above steps, a trained structure-aware converter model is obtained.

[0150] Example 6

[0151] The specific steps of S5 are as follows:

[0152] Based on the trained structure-aware converter model obtained in step S4, the accuracy of the model is verified. The steps include:

[0153] Five-fold cross-validation was used to randomly divide the microbial-drug pair into five subsets of equal number. Each subset was used as a non-repeating test sample set, and the other four subsets were used as training sample sets. The structure-aware converter model was trained using the training set, and the test set was input into the trained structure-aware converter to obtain the classification results.

[0154] A positive classification result indicates that there is an association between the microorganism and the drug, while a negative classification result indicates that there is no association between the microorganism and the drug.

[0155] The AUC and AUPR values ​​of the trained structure-aware converter model are obtained based on the classification results. The accuracy of the structure-aware converter is verified based on these two evaluation metrics.

[0156] Example 7

[0157] The present invention also designs a microbial-drug association prediction system based on a structure-aware converter, comprising: a processor that executes various instructions; and a storage device that stores multiple instructions, which are loaded by the processor and executed according to the above method.

[0158] Clearly, the examples of the present invention provided above are intended only to further illustrate the present invention and not to limit the implementation of the present invention. Any modifications, equivalent substitutions, or improvements made that fall within the spirit and principles of the present invention should be considered to be within the protection scope of the claims of the present invention.

Claims

1. A method for predicting microbial-drug associations based on structure-aware converters, characterized in that, Includes the following steps: S1: Obtain the similarity between microorganisms and drugs based on the microorganism-drug association network and the microorganism-disease-drug association network; S2: Based on the similarity between the microorganisms and the drugs, a restarted random walk algorithm is used to extract feature information between nodes; S3: Construct heterogeneous networks through microbial-drug association networks, microbial similarity, and drug similarity; S4: Input the heterogeneous network into the structure-aware converter for training to obtain a trained structure-aware converter. S5: Predict potential microbe-drug associations based on trained structure-aware converters; Step S1 includes: Calculate microbial Gaussian kernel similarity and drug Gaussian kernel similarity using a microbial-drug association network; By performing deduplication on the microbiome-disease-drug association network, we obtain the microbiome-drug association network, the microbiome-disease association network, and the drug-disease association network. We then calculate the semantic similarity of diseases and further calculate the functional similarity of microorganisms and drugs based on the semantic similarity of diseases. The calculation of microbial Gaussian kernel similarity and microbial functional similarity, as well as drug Gaussian kernel similarity and drug functional similarity, using microbial-drug association networks and microbial-disease-drug association networks includes the following steps: An adjacency matrix was constructed based on the microbial-drug association network. Define the rows of the adjacency matrix A as n r n r Representing the quantity of drugs, the matrix has n columns. m n m Represents the number of microorganisms; Let S(i) and S(j) represent the i-th and j-th rows of the adjacency matrix A, then the drug r i and drugs r j The Gaussian kernel similarity (GIR) between the drugs is calculated as follows: GIR(r i ,r j )=exp(-μ||S(i)-S(j)|| 2 ) Where μ is the standard kernel bandwidth, obtained by normalizing the parameters; Similarly, let R(i) and R(j) represent the i-th and j-th columns of the adjacency matrix A, and let m be the number of microorganisms. i and microorganisms m j The Gaussian kernel similarity (GIM) between the microorganisms is calculated as follows: GIM(m i ,m j )=exp(-μ||R(i)-R(j)|| 2 ) Based on the microbiome-drug association network, the microbiome-disease association network, and the drug-disease association network, common diseases related to microorganisms and drugs are obtained. Semantic information about these diseases is then derived from disease ontology, and disease d is calculated. i and disease d j The semantic similarity DSE between them is calculated as follows: Let D1 be the subset containing disease d. i and the collection of all its related diseases, The semantic contribution value of all items in D1 is represented by the following formula: Same as above, in addition DV(d) represents the sum of the semantic contributions of all sub-items of D1. j Same as above; Based on the above disease semantic similarity DSE, the drug functional similarity (FCR) is calculated as follows: Among them, for drug r i and drugs r j There are b1 diseases and r i Related, there are also B2 diseases and r j Related, the disease is represented as {d it (1≤t≤b1)} and {d js (1≤s≤b2)}; For microorganisms m i and microorganisms m j There are n1 diseases and m i Related, n2 diseases and m j Related, the disease is represented as {d ip (1≤p≤n1)} and {d jk (1≤k≤n2)}, the functional similarity of microorganisms (FCM) is calculated as follows: Drug similarity is represented as IRS(r i ,r j ): Microbial similarity is represented as IMS(m i ,m j ): New drug similarity scores Sr and microbial similarity scores Sm are obtained by calculating drug similarity IRS and microbial similarity IMS using the restarted random walk algorithm. The formula for the restarted random walk algorithm is as follows: Where λ is the restart probability and is set to 0.1, M is the transition probability matrix, and e i ∈R (1*m) It is the initial probability vector of node i. This represents the probability that node i reaches other nodes at time l; Heterogeneous networks are constructed based on the new drug similarity Sr, microbial similarity Sm, and adjacency matrix A. as follows: Among them, A T This represents the transpose of the adjacency matrix A; Based on the aforementioned heterogeneous network, the model is trained using it as input to the structure-aware converter model, including: The structure-aware converter includes a structure-aware attention mechanism and an encoder portion of the converter; The calculation formula for the structure-aware attention mechanism is as follows: Among them, X v S represents the feature vector of node v. G (v) represents the subgraph of graph G centered at node v, f(x) μ ) = w v X is a linear function, W v It is a trainable parameter, and k graph It can be any kernel that compares a pair of subgraphs, k in the structure-aware converter graph The calculation method is as follows: Where, k exp This represents an asymmetric exponential kernel. It is a structure extractor based on a k-layer graph convolutional neural network (GCN). It extracts vector representations of some subgraphs G centered at node v, as follows: The encoder part of the converter consists of N identical stacked layers, each containing two sub-layers: a self-attention mechanism layer and a feedforward network layer. After defining the structure-aware attention mechanism, the encoder of the converter constructs the feedforward network layer as follows: X' = ​​LayerNorm(X + SubLayer(X)) X″=FFN(X′)=ReLU(X′W1)W2 Here, X represents the input node features, W1 and W2 represent two trainable parameters, ReLU is used as the activation function, there is a residual connection in each sublayer of the structure-aware converter, and then a layer normalization operation is performed.

2. The microbial-drug association prediction method based on a structure-aware converter according to claim 1, characterized in that, The reconstructed heterogeneous network is obtained by performing an aggregation operation on the flattened output of the self-attention layer using a fully connected layer. The computation steps are as follows: F′=ReLU(X″W)+b Where W is a trainable parameter, b represents the bias, and ReLU is an activation function; The reconstructed heterogeneous network is input into a deep neural network classifier to obtain the microbe-drug association score y. score The binary cross-entropy loss function is used, as shown below: Loss=-∑[y pred log(y score )+(1-y pred )log(1-y score )] Among them, y pred It is represented as a vector after the adjacency matrix A is flattened, and log represents the natural logarithm; Through the above steps, a trained structure-aware converter model is obtained.

3. A microbial-drug association prediction method based on a structure-aware converter according to claim 2, characterized in that, The steps to verify the accuracy of the trained structure-aware converter model include: Five-fold cross-validation was used to randomly divide the microbial-drug pair into five subsets of equal number. Each subset was used as a non-repeating test sample set, and the other four subsets were used as training sample sets. The structure-aware converter model was trained using the training set, and the test set was input into the trained structure-aware converter to obtain the classification results. A positive classification result indicates that there is an association between the microorganism and the drug, while a negative classification result indicates that there is no association between the microorganism and the drug. The AUC and AUPR values ​​of the trained structure-aware converter model are obtained based on the classification results. The accuracy of the structure-aware converter is verified based on these two evaluation metrics.

4. A microbial-drug association prediction system based on a structure-aware converter, characterized in that, The method includes a processor that executes instructions; and a storage device that stores multiple instructions, which are loaded by the processor and executed according to the method of claims 1-3.

Images