Method for predicting microbe and drug associations based on dual attention graph convolution

Abstract

The application discloses a method and device for predicting microorganism and drug association based on double attention graph convolution, wherein the method for predicting microorganism and drug association based on double attention graph convolution comprises the following steps: constructing a microorganism and drug association network subgraph based on obtained known microorganism-drug association relations; learning an association prediction graph representation of the microorganism and drug association network subgraph by using an attention graph convolutional neural network; and inputting the association prediction graph representation into an attention pooling layer, thereby extracting information under maximum node representation and obtaining a microorganism-drug prediction model. The method for predicting microorganism and drug association based on double attention graph convolution automatically captures context information of a specific subgraph by using an attention graph convolutional network and an attention pooling layer to learn expression features, thereby constructing a prediction model, and the purpose of improving prediction accuracy and reliability is achieved.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics prediction, and more specifically to a method and apparatus for predicting microbial and drug associations based on dual attention map convolution. Background Technology

[0002] The human microbiome is a complex ecosystem composed of various microorganisms, including bacteria, fungi, archaea, viruses, and protozoa. These microorganisms are widely distributed in different ecological niches within the human body, such as the skin, oral cavity, respiratory tract, genitourinary tract, and digestive tract. This microbiome plays a vital role in the health and well-being of the human host. In recent years, with in-depth research into the human microbiome, it has become increasingly clear that these microorganisms play crucial roles in maintaining homeostasis, protecting the body from pathogens, assisting in nutrient metabolism, and regulating the immune response. Further research into the microbiome helps to better understand the internal micro-ecosystem of the human body, providing new ideas and methods for disease prevention and treatment. Microbiome imbalances often lead to various diseases, such as diabetes, obesity, and cancer. Furthermore, microorganisms play an important role in drug metabolism and absorption, regulating drug efficacy and toxicity. For example, the gut microbe *Actinobacterium eggerthella lenta* is responsible for the inactivation of the cardiac drug digoxin. Simultaneously, drugs can also affect the diversity and function of the microbiome. New drugs targeting microorganisms have been developed to help maintain the adaptability of microbial communities; however, interactions between microorganisms and drugs can also lead to the spread of drug-resistant bacteria, posing another serious threat to human health. Therefore, a deeper understanding of the relationship between microorganisms and drugs is needed.

[0003] Multiple databases have published substantial experimental evidence exploring the associations between microbes and drugs. These databases include the drug-microbe association databases MADA, Abiofilm, and Drugvirus, as well as the COVID-19 database HDVD. These datasets are freely accessible, facilitating the development of deep learning-based methods for predicting microbe-drug associations. Several computational methods for predicting drug-microbe associations have also been proposed. For example, Chen et al. proposed HMDAKATZ, a computational method based on the KATZ model, to predict potential microbe-drug associations; this was the first tool to predict potential associations between microbes and drugs. Long et al. proposed GCNMDA, a method based on GCN and conditional random fields to predict human microbe-drug associations. Long et al. also proposed HNERMDA, an embedding representation based on heterogeneous networks for predicting human microbe-drug associations. Subsequently, Long et al. proposed a novel graph attention network deep learning framework GATMDA with inductive matrix completion for predicting human microbe-disease associations. To focus on more important neighbors and filter out noise, a bidirectional interactive aggregator was designed to force the aggregation of representations of similar neighbors.

[0004] Currently, research on drug-microbe association prediction methods still has the following shortcomings:

[0005] 1. Limited sample size: Existing drug-microbe association prediction methods still face the problem of insufficient sample size, especially the lack of large-scale clinical data, which limits the robustness and accuracy of prediction models.

[0006] 2. Unclear mechanisms: Although existing research has revealed some mechanisms of microbial-drug interactions, the exact mechanisms between certain microbial-drug pairs are still not clear enough, which also affects the accuracy and reliability of prediction methods.

[0007] 3. High cost: Traditional culture medium-based microbial drug association validation methods are often time-consuming, labor-intensive, and costly.

[0008] 4. Technological imperfections: Although deep learning is very effective in predictive performance, the methods for predicting drug-microbe associations are not yet comprehensive enough and require further development and improvement. Summary of the Invention

[0009] To address at least some of the technical problems in the prior art, the present invention provides a method and apparatus for predicting microbial and drug associations based on dual attention map convolution. Specifically, the present invention includes the following.

[0010] A first aspect of the present invention provides a method for predicting microbial-drug associations based on dual attention map convolution, comprising:

[0011] Based on the known microbial-drug associations obtained, a subgraph of the microbial-drug association network is constructed;

[0012] The association prediction graph representation of the microorganism and drug association network subgraph is learned using an attention graph convolutional neural network;

[0013] The association prediction graph representation is fed into the attention pooling layer to extract information from the maximized node representation, thus obtaining the microbiome-drug prediction model.

[0014] Optionally, the construction of a microbial-drug association network subgraph based on the acquired known microbial-drug associations includes:

[0015] Based on known microbe-drug associations, a dichotomous microbe-drug association network diagram is constructed.

[0016] The microbiome-drug association network graph is divided into several subgraphs, thereby transforming the prediction of the association between microorganisms and drugs into predicting the association by inducing the representation of each midpoint by learning the subgraph embeddings around a given midpoint in the association network graph.

[0017] Optionally, the step of learning the association prediction graph representation of the microbial and drug association network subgraph using an attention graph convolutional neural network includes:

[0018] The convolution results of attention-map convolutional neural networks are presented in a hierarchical manner, forming a hierarchical node representation δv. n ,

[0019]

[0020] Where α is the attention weight, Representative node v n Local features in k-hops, node v n These are nodes in the subgraph of the microbial and drug association network.

[0021] Optionally, obtain the optimal hierarchical node representation δv n This involves adding the output of the previous layer of the attention map convolutional neural network and the initial node features as the input to each attention convolutional layer.

[0022] Optionally, the step of learning the association prediction graph representation of the microbial and drug association network subgraph using an attention graph convolutional neural network includes:

[0023] The subgraph is represented as an n×t matrix, where t is the number of attention convolutional layers stacked using residual learning techniques.

[0024] Optionally, the step of feeding the association prediction graph representation into the attention pooling layer to extract information under the maximized node representation to obtain the microbe-drug prediction model includes:

[0025] The weight vector B is obtained by representing the correlation prediction graph as input.

[0026] By performing a weighted summation on B, we obtain the graph representation matrix of the subgraph.

[0027] Will The input is used to complete the classification of the graph, thereby obtaining the final microbial and drug association prediction results.

[0028]

[0029] Where W is the trainable weight and C is the bias.

[0030] Optionally, the step of obtaining the microbiome-drug prediction model further includes:

[0031] The microbiome-drug prediction model was optimized.

[0032] Optionally, optimizing the microbiome-drug prediction model includes:

[0033] The model is trained using the cross-entropy loss function. Known microbial-drug associations in the dataset are considered positive samples, while others are considered negative samples. That is, the classification label of the subgraph around the extracted positive sample is 1, and otherwise it is 0.

[0034] Optionally, the formula for the classification labels of the sub-images surrounding the extracted positive samples is:

[0035]

[0036] Where (i, j) represents the microorganism-drug pair, y ij For truth labels, Let be the predicted association probability, and Loss be the classification label.

[0037] A second aspect of the present invention provides an apparatus for predicting microbial and drug associations based on dual attention map convolution, comprising:

[0038] The module is used to construct a subgraph of the microbial-drug association network based on the acquired known microbial-drug associations;

[0039] The representation module is used to learn the association prediction graph representation of the microbial and drug association network subgraph using an attention graph convolutional neural network;

[0040] The extraction module is used to input the association prediction graph representation into the attention pooling layer, thereby extracting the information under the maximized node representation to obtain the microbial-drug prediction model.

[0041] The present invention relates to a method and apparatus for predicting microbial and drug associations based on dual attention graph convolution. The method for predicting microbial and drug associations based on dual attention graph convolution applies attention graph convolutional networks and attention pooling layers to automatically capture contextual information of specific subgraphs for expressive feature learning, thereby constructing a prediction model to improve prediction accuracy and reliability. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method for predicting microbial-drug associations based on dual attention map convolution disclosed in this embodiment;

[0043] Figure 2 This is a schematic diagram showing the AUC values ​​of MDPAGCN on LOOCV and other existing methods on the MDAD dataset disclosed in this embodiment;

[0044] Figure 3 This is a schematic diagram of all parameter combinations with a jump count of 2 disclosed in this embodiment. In each group of columns, the left column corresponds to AUC, and the right column corresponds to AUPR. Detailed Implementation

[0045] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as a limitation of the present invention, but rather as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0046] It should be understood that the terminology used in this invention is merely for describing particular embodiments and is not intended to limit the invention. Furthermore, with respect to numerical ranges in this invention, it should be understood that the upper and lower limits of the range and each intermediate value between them are specifically disclosed. Any stated value or intermediate value within a stated range, as well as each smaller range between any other stated value or intermediate value within said range, are also included in this invention. The upper and lower limits of these smaller ranges may be independently included or excluded from the range.

[0047] Unless otherwise stated, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. While only preferred methods and materials have been described herein, any methods and materials similar or equivalent to those described herein may be used in the implementation or testing of this invention. All references to this specification are incorporated by way of citation to disclose and describe methods and / or materials associated with those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

[0048] Terminology Explanation:

[0049] KATZ: A classic language smoothing model.

[0050] GCN: Graph Convolutional Neural Network, is a feature extractor, except that its object is graph data.

[0051] Adam optimizer: an algorithm that performs first-order gradient optimization on a stochastic objective function, based on adaptive low-order moment estimation.

[0052] Dropout: It is a relatively effective way to alleviate overfitting when training neural network models, and achieves a regularization effect to a certain extent.

[0053] like Figure 1 As shown, a method for predicting microbial-drug associations based on dual attention map convolution includes:

[0054] Based on the known microbial-drug associations obtained, a subgraph of the microbial-drug association network is constructed;

[0055] The association prediction graph representation of the microorganism and drug association network subgraph is learned using an attention graph convolutional neural network;

[0056] The association prediction graph representation is fed into the attention pooling layer to extract information from the maximized node representation, thus obtaining the microbiome-drug prediction model.

[0057] In a specific application scenario, predicting microbe-drug associations based on dual attention map convolution includes the following steps:

[0058] Step 1: Construct a subgraph of the microbial-drug association network based on known microbial-drug associations in the database. Further detailed steps include:

[0059] Based on known microbe-drug associations, a binary microbe-drug association network G = {O, P, E} containing N types of microbes and M types of drugs was constructed, where O = {o1, o2, ..., o3}. n Let} be the set of microbial nodes, P = {p1, p2, ..., p...} n Let} be the set of drug nodes. E = {(oi, pj) | oi ∈ O, pj ∈ P} is the set of edges in graph G, representing known associations between known microorganisms and drugs. The adjacency matrix of graph G is A = R. N×M If the microorganisms and drugs are related, then a ij =1, otherwise a ij =0.

[0060] The microbiome-drug association network graph is divided into several subgraphs. The prediction of microbiome-drug associations is transformed into predicting the association by inducing the representation of each midpoint by learning the subgraph embeddings around a given midpoint in the graph. Here, a simple graph without self-loops is considered, where there is at most one edge between two nodes; the existence of one edge is called a hop. For the target microbiome-drug pair to be predicted (o... i p j Treat it as a center point and only construct a subgraph from its surrounding k-hop neighbors. For each subgraph GV ij Here, node marking is performed, and the microbial nodes in the central node are marked. i Mark it as 0, and set the drug node p in the central node to 0. j Mark it as 1. Similarly, mark the microbial node of the k-hop node as 2k, and the drug node of the k-hop node as 2k+1.

[0061] Step 2: After extracting the microbial-drug association network subgraph, an attention graph convolutional neural network is used to learn the association prediction graph representation of the subgraph. Further detailed procedures include:

[0062] After extracting the association graphs of the subgraphs, a graph neural network is used to learn the association prediction graph representation of the subgraphs. For traditional graph convolutional networks, the most general form of a graph convolution of depth k can be recursively represented by a widely followed convolutional structure, as follows:

[0063]

[0064] H 0 =X (1),

[0065] in, Each subgraph has a self-connected adjacency matrix. for The angle matrix, The graph structure is normalized, ω represents the model parameters to be trained, and φ is the non-linear activation function. After this operation is performed k times, the node attribute vector H... k It becomes a node attribute vector that contains k-hop local structure information. This represents the output node embedding of layer k, where N is the number of nodes in the subgraph, and d k This is the number of output channels in layer k. For k=1, the initial node characteristic H... 0 It is a one-time encoding of the node label in the subgraph.

[0066] Currently, most graph neural networks are driven by Equation (1), which employs a k-hop message aggregation mechanism. This allows the node representation to capture the local structural information of its k-hop neighbors. However, as the number of layers increases, a large amount of early information is lost in each convolution step, which severely affects the final prediction output and limits the model's capacity. This embodiment aims to ensure that the graph convolutional network not only relies on the k-hop convolution result but also captures valuable information from each hop. Therefore, the convolution result will be a hierarchical representation containing the most valuable information from different hop convolution processes. The attention behavior is expressed in Equation (2), forming a hierarchical node representation δv. n As shown below:

[0067]

[0068] For simplicity, layer attention is used to determine the importance of the aggregation result at each hop, where α is the attention weight. Represents node v n The local structure within k hops. The last node carries hierarchical data.

[0069] To extract richer features and fully leverage the advantages of neural networks, a residual learning method is employed to stack t attention convolutional layers to obtain a better final node representation δv. n The output of the previous layer and the initial node features are added together and used as the input to each attention convolutional layer, as shown in Equation (3).

[0070]

[0071]

[0072]

[0073] Here, Dense() is a fully connected layer that aggregates the outputs of all attention graph convolutional layers. Now, given the node representation δ for all vertices v∈GV, for simplicity, the subgraph is represented as an n×t matrix. Each row represents a node, as shown in formula (4).

[0074]

[0075] Step 3: The output of the attention graph convolutional layer is encoded into a fixed-size embedding matrix and fed into the attention pooling layer. Simultaneously, information from the maximized node representation is extracted to obtain the microbiome-drug prediction result. Further detailed steps include:

[0076] Using an attention mechanism, the graph node representations learned from the convolution module are used as input and output weight vectors B, as shown in Equation (5).

[0077]

[0078] Here, u is the weight matrix, and its size is a hyperparameter set for the number of subgraphs. Each row of B represents the weight of a node in different subspaces. A weighted summation of B yields the graph representation matrix of the subgraph GV. Each row represents a graph in a subspace, and the entire matrix provides a comprehensive representation of the graph. Finally, fully connected layers and softmax layers take the BGV as input to perform graph classification, thus obtaining the final microbial o. i and drug p j Association prediction results W is the trainable weight, and C is the bias. As shown in formula (6) below.

[0079]

[0080] Step 4: Optimize the designed model. Further detailed steps include:

[0081] Since the task is to predict microbe-drug associations, this is a binary classification problem. The main objective is to minimize the difference between the predicted label and the ground truth. The microbe-drug association prediction problem is transformed into a graph classification task, and the model is trained using a loss function based on cross-entropy. Known microbe-drug associations in the dataset are considered positive samples, and others are considered negative samples. Therefore, the classification label of the subgraph surrounding the extracted positive samples is 1, and otherwise 0. As shown in Equation (7):

[0082]

[0083] Where (i, j) represents microorganisms o i and drug p j Yes. ij For truth labels, For (o) i p j The predicted association probability is calculated. To optimize the model, the Adam optimizer is used, and the model is trained in a setting that denoises by randomly removing all outgoing messages of a specific edge with a fixed probability. Regular dropout is applied to the prediction layer.

[0084] Furthermore, in step four, after training the graph neural network model, the model was validated. The specific validation steps are as follows:

[0085] To comprehensively evaluate the effectiveness of the MDPAGCN model in predicting potential drug-associated microorganisms, two commonly used cross-validation methods were employed: Leave-one-out-of-the-world cross-validation (LOOCV) and n-fold cross-validation (n-fold CV). These two validation methods are widely used in association prediction studies because they avoid overfitting while making full use of the dataset for model evaluation.

[0086] Specifically, in 5-fold cross-validation (5-fold CV), all known microbe-drug associations are considered positive samples and randomly divided into five equal subsets. In each iteration, four subsets are combined to form a positive training set, while the remaining subsets are used as a positive test set. The advantage of this cross-validation method is that it avoids model bias and variance issues and makes full use of the dataset for model evaluation. Furthermore, in leave-one-out-of-the-world cross-validation (LOOCV), each known microbe-drug association is used as a test sample, while the remaining data is used as training samples. This cross-validation method ensures that every sample appears in the test set, thus more accurately evaluating model performance. Although this method has a higher computational cost, the model can withstand this cost because the dataset is not large.

[0087] Furthermore, AUC and AUPR were used to measure the performance of MDPAGCN. Specifically, AUC and AUPR represent the areas under the ROC and PR curves, respectively. The ROC curve is obtained by setting a set of thresholds from [0, 1], with each threshold resulting in a (TPR, FPR) pair. The vertical axis represents TPR, and the horizontal axis represents FPR; connecting all (TPR, FPR) pairs yields the ROC curve. Similarly, the PR curve is obtained by setting a set of thresholds from [0, 1], with each threshold resulting in a (Presession, Recall) pair. The vertical axis represents Presession, and the horizontal axis represents Recall; connecting all (Presession, Recall) pairs yields the PR curve. By using both ROC and PR curves, the classifier's performance can be evaluated more comprehensively. To improve the reliability and stability of the experiments, multiple experiments are typically conducted, and the average value is calculated. To provide a reliable performance estimate, the 5-fold cross-validation procedure was repeated 10 times, and the average results were reported.

[0088]

[0089]

[0090]

[0091]

[0092] Wherein, TP is the number of known microbe-drug association pairs classified as having an association, FN is the number of known microbe-drug association pairs classified as not having an association, FP is the number of unknown microbe-drug association pairs classified as having an association, TN is the number of unknown microbe-drug association pairs classified as not having an association, Precision is precision, and recall is recall value.

[0093] To further evaluate the effectiveness and robustness of the model, 5-fold cross-validation was performed on three datasets—MDAD, aBiofilm, and DrugVirus—using both MDPAGCN and all baseline methods. The results show that the model outperforms the baseline methods, significantly outperforming the other seven models in both AUC and AUPR.

[0094] To investigate known microbe-drug interactions, the model used three datasets: MDAD, aBiofilm, and DrugVirus. The MDAD dataset contained 5505 clinically or experimentally validated microbe-drug interactions involving 1388 drugs and 174 microorganisms. After filtering out redundant information, 2470 interactions between 1373 drugs and 173 microorganisms were obtained. The aBiofilm dataset included 1720 unique antibiofilm agents / drugs that act on 140 organisms / microorganisms, including bacteria and fungi. After removing duplicate data, 2884 microbe-drug interaction data points were obtained, involving 1720 drugs and 140 microorganisms. The DrugVirus dataset provided activity and development status information for 118 drugs / compounds targeting 83 human viruses. In addition, drug-virus interactions between 76 clinically or experimentally validated drugs and 12 viruses were collected from drug databases and related publications. A total of 933 drug-virus interactions involving 175 drugs and 95 viruses were collected. Overall, the statistical data for the three microbial-drug association datasets mentioned above are shown in Table 1.

[0095] Table 1. Statistical data of the three microbial-drug association datasets used in this embodiment.

[0096]

[0097] There are not enough existing prediction methods specifically designed to address the microbe-drug association prediction problem.

[0098] The first two methods do not have an attention mechanism, while the latter two methods add a single attention mechanism.

[0099] This embodiment selects the optimal parameters through multiple parameter combination experiments, rather than simply improving performance by fine-tuning the model. In the experiments, using the HMDAD dataset, the size of the fully connected layer was set to 64, and the hidden layer size of the convolutional layer was also set to 64. The initial learning rate (lr) was selected from the set {0.01, 0.001, 0.0001}, the dropout rate was selected from the set {0.1, 0.3, 0.5}, and the number of hops (k) was selected from the set {1, 2, 3, 4}.

[0100] First, the learning rate and dropout rate are set to 0.01 and 0.1 by default. Then, the choice of the number of hops is discussed. As shown in Table 2, without fine-tuning the model, extracting only hops k=1 results in the lowest performance, while hops k=2 provides the model with better representational capabilities. Further increasing the hop depth does not significantly improve performance. This indicates that 2 hops are sufficient to distinguish the specific context of the molecular graph. Therefore, this embodiment recommends choosing 2 hops to balance performance and computational efficiency. Based on hops k=2, the settings of other parameters are discussed. The optimal parameter combination for the MDPAGCN model is: k=2, lr=0.001, dropout=0.1. Figure 3 All parameter combinations with a hop count of 2 are listed to demonstrate that the parameters selected for this model make MDPAGCN perform optimally.

[0101] Table 2. Comparison of AUC and AUPR for different hop counts

[0102]

[0103] Therefore, in order to evaluate the performance of this model, the MDPAGCN method disclosed in this embodiment is compared with four advanced methods proposed to solve the co-prediction task in the field of computational microbial networks.

[0104] (1) HMDAKATZ is a computational method based on the KATZ model to predict potential microbe-drug associations. It is the first tool developed to predict potential associations between microbes and drugs.

[0105] (2) HNERMDA is an embedding representation based on heterogeneous networks designed to predict potential associations between human microbes and drugs.

[0106] (3) GCNMDA is a method for predicting human microbiome-drug associations based on GCN and conditional random fields, and it also uses an attention mechanism.

[0107] (4) GATMDA is a deep learning framework for predicting human microbiome-disease associations. It can also be used to predict potential associations between human microbiome and drugs. It mainly uses a graph attention network with inductive matrix completion.

[0108] All four benchmark methods used the default parameter values ​​from their original implementations and were compared on the same benchmark dataset. For better visual comparison, Figure 2 The table shows the AUC values ​​of MDPAGCN and the baseline methods HMDAKATZ, GCNMDA, HNERMDA, and GATMDA on the MDAD dataset within the global LOOCV evaluation framework. Clearly, among all methods, the MDPAGCN model achieves the best predictive performance, with an AUC value of 0.9504. This demonstrates that the MDPAGCN model can accurately identify drug-related microorganisms. To further evaluate the model's effectiveness and robustness, this embodiment also used 5-fold cross-validation to perform MDPAGCN and all baseline methods on three datasets: MDAD, aBiofilm, and DrugVirus. The results are shown in Table 3, where bold represents the best performance and horizontal lines represent the second-best performance. This indicates that the MDPAGCN model outperforms the baseline methods, significantly outperforming the other four models in both AUC and AUPR. MDPAGCN is the prediction method used in this embodiment.

[0109] Table 3. Performance of our model and the four baseline methods at 5-fold cross-validation on three datasets.

[0110]

[0111] To further validate the performance of the MDPAGCN model in predicting microbe-drug associations, the following experiments were conducted in this embodiment. Specifically, the MDPAGCN model was implemented using two popular drugs, ciprofloxacin and moxifloxacin, as a case study. First, all known association pairs were set as unknown association pairs. Then, the MDPAGCN model was used to predict all candidate microbes, and the candidates were ranked according to their prediction scores, prioritizing those with higher scores. During this process, the top 20 predicted microbes or drugs were selected, and the performance of the MDPAGCN model was measured by cross-checking synonyms in MeSH and DrugBank, and by checking whether predicted microbe-drug associations were reported in PubMed literature.

[0112] Ciprofloxacin is a second-generation fluoroquinolone drug primarily used to treat infectious diseases caused by Gram-negative pathogens. Increasing research indicates that ciprofloxacin interacts closely with a wide range of human microorganisms. Common interactions include drug activity and toxicity to microorganisms, as well as microbial resistance. For example, Maheshwari et al. showed that Enteric bacteria exhibit high resistance to ciprofloxacin, and the MDPAGCN model predicted it as the most likely candidate target microorganism for ciprofloxacin. Kim and Woo demonstrated that Candida albicans is a highly resistant microorganism to ciprofloxacin. Cho et al. demonstrated that Mycobacterium avium is highly sensitive to ciprofloxacin. The final results showed that 10 and 17 of the top 10 and top 20 predicted ciprofloxacin-associated microorganisms, respectively, could be validated using previously published literature, with accuracies of 100% and 85%, respectively. Table 4 shows the top 20 candidate microorganisms for ciprofloxacin.

[0113] Table 4. Top 20 candidate microorganisms for Ciprofloxacin

[0114]

[0115] Moxifloxacin is an antibacterial drug belonging to the broad-spectrum 8-methoxyfluoroquinolone class, used to treat conditions such as chronic bronchitis, pelvic inflammatory disease, and skin diseases. For example, Aldulaimi et al. demonstrated that moxifloxacin can destroy Staphylococcus aureus, leading to leakage of cellular contents. Dosanj et al. demonstrated that olive oil is highly sensitive to moxifloxacin. Alharbi et al. found that more than 50% of E. coli isolates obtained from wound infections were resistant to moxifloxacin. As shown in Table 5, 9 and 17 of the top 10 and 20 predicted candidate moxifloxacin-associated microorganisms, respectively, have been validated by existing reports. Accuracy rates of 90% and 85%, respectively, indicate that MDPAGCN is a very promising tool for screening candidate compounds for drug development in practical applications.

[0116] Table 5. Top 20 candidate microorganisms for Moxifloxacin

[0117]

[0118] Compared with the prior art, this embodiment has the following beneficial effects:

[0119] 1. This method proposes a microbial-drug prediction method based on attention mechanism and graph convolutional neural network. Compared with previous microbial-drug association prediction methods, the MDPAGCN method applies attention graph convolutional network and attention pooling layer to automatically capture the contextual information of specific subgraphs for expressive feature learning.

[0120] 2. The model was evaluated using leave-one-out cross-validation (LOOCV) and 5-fold cross-validation (5-fold CV). In the prediction of microbiome-drug associations, the AUC values ​​of LOOCV and 5-fold CV were 0.9504 and 0.9599 + / - 0.0058, respectively. These results are better than existing methods for predicting microbiome-drug associations.

[0121] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.

[0122] In this disclosure, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Words such as "including," "comprising," "having," etc., are open-ended terms meaning "including but not limited to," and are used interchangeably with them. The terms "or" and "and" as used herein refer to the terms "and / or," and are used interchangeably with them unless the context explicitly indicates otherwise. The term "such as" as used herein refers to the phrase "such as but not limited to," and is used interchangeably with it.

[0123] Additionally, as used herein, the “or” used in a list of items beginning with “at least one” indicates a separate list, such that a list of, for example, “at least one of A, B, or C” means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word “exemplary” does not imply that the described example is preferred or better than other examples.

[0124] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.

[0125] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.

[0126] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.

[0127] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.

Claims

1. A method for predicting microbial-drug associations based on dual attention map convolution, characterized in that, include: Based on the known microbial-drug associations obtained, a subgraph of the microbial-drug association network is constructed; The association prediction graph representation of the microorganism and drug association network subgraph is learned using an attention graph convolutional neural network; The association prediction graph representation is fed into the attention pooling layer to extract information from the maximized node representation, thus obtaining the microbiome-drug prediction model. The construction of a microbial-drug association network subgraph based on the acquired known microbial-drug associations includes: Based on known microbe-drug associations, a dichotomous microbe-drug association network diagram is constructed. The microbiome-drug association network graph is divided into several subgraphs, thereby transforming the prediction of the association between microorganisms and drugs into predicting the association by learning the subgraph embeddings around a given midpoint in the association network graph to induce the representation of each midpoint. The method of learning the association prediction graph representation of the microbial and drug association network subgraph using an attention graph convolutional neural network includes: The convolution results of attention-map convolutional neural networks are presented in a hierarchical manner, forming a hierarchical node representation. , ， in, For attention weights, Representative node Local features in k-hops, nodes These are nodes in a subgraph of the microbial and drug association network. Obtain the optimal hierarchical node representation This includes adding the output of the previous layer of the attention map convolutional neural network and the initial node features as the input to each attention convolutional layer; The method of learning the association prediction graph representation of the microbial and drug association network subgraph using an attention graph convolutional neural network includes: The subgraph is represented as an n × t matrix, where t is the number of attention convolutional layers stacked using residual learning techniques; The step of feeding the association prediction graph representation into the attention pooling layer to extract information under the maximized node representation and obtain the microbiome-drug prediction model includes: The weight vector B is obtained by representing the correlation prediction graph as input. By performing a weighted summation on B, we obtain the graph representation matrix of the subgraph. , Will The input is used to complete the classification of the graph, thereby obtaining the final microbial and drug association prediction results. , ， Where W is the trainable weight and C is the bias; The step of obtaining the microbial-drug prediction model further includes: Optimize the microbe-drug prediction model; The optimization of the microbiome-drug prediction model includes: The model is trained using the cross-entropy loss function. Known microbial-drug associations in the dataset are considered positive samples, while others are considered negative samples. That is, the classification label of the subgraph around the extracted positive sample is 1, and otherwise it is 0. The formula for the classification labels of the sub-images surrounding the extracted positive samples is: Where (i, j) represents the microorganism-drug pair, For truth labels, In order to predict the probability of association, For category tags.

2. A device for predicting microbial and drug associations based on dual attention map convolution, characterized in that, include: The module is used to construct a subgraph of the microbial-drug association network based on the acquired known microbial-drug associations; The representation module is used to learn the association prediction graph representation of the microbial and drug association network subgraph using an attention graph convolutional neural network; The extraction module is used to input the association prediction graph representation into the attention pooling layer, thereby extracting the information under the maximized node representation to obtain the microbiome-drug prediction model; The method of learning the association prediction graph representation of the microbial and drug association network subgraph using an attention graph convolutional neural network includes: The convolution results of attention-map convolutional neural networks are presented in a hierarchical manner, forming a hierarchical node representation. , ， in, For attention weights, Representative node Local features in k-hops, nodes These are nodes in a subgraph of the microbial and drug association network. Obtain the optimal hierarchical node representation This includes adding the output of the previous layer of the attention map convolutional neural network and the initial node features as the input to each attention convolutional layer; The method of learning the association prediction graph representation of the microbial and drug association network subgraph using an attention graph convolutional neural network includes: The subgraph is represented as an n × t matrix, where t is the number of attention convolutional layers stacked using residual learning techniques; The step of feeding the association prediction graph representation into the attention pooling layer to extract information under the maximized node representation and obtain the microbiome-drug prediction model includes: The weight vector B is obtained by representing the correlation prediction graph as input. By performing a weighted summation on B, we obtain the graph representation matrix of the subgraph. , Will The input is used to complete the classification of the graph, thereby obtaining the final microbial and drug association prediction results. , ， Where W is the trainable weight and C is the bias; The step of obtaining the microbial-drug prediction model further includes: Optimize the microbe-drug prediction model; The optimization of the microbiome-drug prediction model includes: The model is trained using the cross-entropy loss function. Known microbial-drug associations in the dataset are considered positive samples, while others are considered negative samples. That is, the classification label of the subgraph around the extracted positive sample is 1, and otherwise it is 0. The formula for the classification labels of the sub-images surrounding the extracted positive samples is: Where (i, j) represents the microorganism-drug pair, For truth labels, In order to predict the probability of association, For category tags.

Images