Microbe-Disease Association Prediction Method Based on Multi-View Multimodal and Cross-Attention Networks

Abstract

This invention discloses a microbiome-disease association prediction method based on multi-view multimodal and cross-attention networks. The method includes: first, constructing multiple similarity views of microorganisms and diseases; then, fusing these multi-view features together using similarity network fusion and view-level attention mechanisms; since the structural information encoded by nodes in homogeneous and heterogeneous networks differs, graph convolutional networks and multi-head cross-attention networks with shared weights are used to extract features of microorganisms and diseases respectively; a hierarchical attention mechanism adaptively assigns weights to the learned embedded features and initial features to generate the final feature representations of microorganisms and diseases; finally, an inner product decoder is used to reconstruct the microbiome-disease association matrix. This model comprehensively considers node information from homogeneous and heterogeneous networks, resulting in richer learned features and more accurate prediction results.

Description

Technical Field

[0001] This invention designs a microbial-disease association prediction method based on multi-view multimodal and cross-attention networks, which relates to the field of bioinformatics disease association prediction. Background Technology

[0002] The gut microbiome, comprising bacteria, archaea, fungi, viruses, and other tiny microorganisms, plays a vital role in various ecological, industrial, and medical fields. They inhabit various parts of the human body, such as the skin, mouth, digestive tract, and reproductive tract. The complex interactions between the gut microbiome and the human body have significant impacts on health and function, such as immune system regulation, defense against pathogens, and digestion and metabolism. Research indicates that the interaction between the gut microbiome and the nervous system may profoundly affect mood regulation, cognitive function, and behavioral performance. Certain diseases, such as depression, autism, asthma, and even some cancers, are closely related to disruptions in the composition and balance of the gut microbiome. There is also evidence that some microbial communities are beneficial to the human body. For example, lactic acid bacteria not only regulate the gut microbiome and alleviate polycystic ovary syndrome in women, but also provide certain benefits and effects in the prevention and treatment of obesity. Therefore, identifying the association between microbes and diseases can not only reveal the pathogenesis of diseases but also provide more precise and effective prevention, diagnosis, and treatment methods.

[0003] Because traditional biological experiments are time-consuming and labor-intensive, current research hotspots are using computational methods to identify potential associations between microorganisms and diseases.

[0004] Currently, there are four types of computational models used for predicting microbial disease associations: path-based methods, random walk-based methods, matrix factorization-based methods, and machine learning and deep learning methods. Although existing methods have achieved excellent predictive performance, they still have the following limitations:

[0005] (1) Many methods tend to construct heterogeneous networks for prediction without considering the impact of different modal networks on performance.

[0006] (2) Existing microbial-disease association prediction methods have limited sample sizes and insufficient entity information, which limits the predictive performance of the models.

[0007] (3) Imbalanced existing data can also lead to prediction bias. Summary of the Invention

[0008] To address at least one of the technical problems in the background art described above, the present invention provides a microbial-disease association prediction method based on multi-view multimodal and cross-attention networks.

[0009] To achieve the above objectives, this invention mainly comprises four modules: a microbial and disease similarity data processing module, a similarity feature fusion module, a feature extraction module, and an association prediction module. The main steps include:

[0010] (1) First, in the data processing stage, three similarity views for microorganisms and diseases were constructed, namely functional similarity, Gaussian interaction profile kernel (GIP) similarity and cosine similarity.

[0011] (2) Based on the multiple similarities between microorganisms and diseases, similarity network fusion (SNF) and view-level attention mechanism are used to perform similarity fusion to obtain comprehensive similarity information between microorganisms and diseases.

[0012] (3) Based on the fused microbial similarity and disease similarity, construct a microbial and disease isomorphic graph and encode it using a graph convolutional network (GCN) to learn the embedding features of microorganisms and diseases; based on the microbial and disease heteromorphic graph, encode it using a cross attention network (CANet) based on shared weights to extract the interaction information between microorganisms and diseases.

[0013] (4) Based on the learned feature information, a hierarchical attention mechanism is used to assign weights to the embedded features and initial feature information of different layers to obtain the final feature information.

[0014] (5) The microbial and disease features from the two networks are fused together, and the microbial-disease association matrix is ​​reconstructed using the inner product decoder. Attached Figure Description

[0015] Figure 1 This is an overall flowchart of the present invention;

[0016] Figure 2 This is a flowchart illustrating the similarity feature fusion process of the present invention;

[0017] Figure 3 This is a schematic diagram of the graph convolutional network encoding process of the present invention;

[0018] Figure 4 This is a schematic diagram of the cross-attention network encoding process of the present invention; Detailed Implementation

[0019] This invention is a microbial-disease association prediction method based on multi-view multimodal and cross-attention networks. The specific implementation process of this invention will be described in detail below with reference to the accompanying drawings.

[0020] like Figure 1As shown, this invention provides a microbial-disease association prediction method based on multi-view multimodal and cross-attention networks, mainly comprising four modules: microbial and disease similarity data processing, similarity feature fusion, feature extraction, and association prediction. Each step is described in detail below.

[0021] 1. Processing of microbial and disease similarity data:

[0022] 1.1: This invention uses two datasets, HMDAD and Disbiome. HMDAD includes 39 diseases, 292 microorganisms and 450 associations; Disbiome includes 218 diseases, 1052 microorganisms and 4351 associations.

[0023] An association matrix was constructed. To represent the association between microbial diseases, among which and These represent the number of microorganisms and the number of diseases, respectively. For a microorganism-disease pair, if the number of microorganisms is... With disease The relationship between them is known, then ,otherwise .

[0024] 1.2: In the HMDAD and Disbiome datasets, there are 450 and 4351 positive samples, respectively, accounting for approximately 3.9% and 1.8% of the total number of samples. Due to the imbalance between positive and negative samples, this invention uses the k-means clustering method to select negative samples that are balanced with the number of positive samples.

[0025] First, the negative samples were divided into k clusters. Then, samples were randomly selected from each cluster as negative samples, while the positive samples remained unchanged. To maintain data balance, 460 negative samples were selected from the HMDAD dataset and 4374 negative samples were selected from the Disbiome dataset.

[0026] 1.3: Calculating Microbial Functional Similarity Microbial GIP nuclear similarity cosine similarity with microorganisms .

[0027] Calculating disease functional similarity Disease GIP core similarity cosine similarity to disease .

[0028] Ultimately, this yields three similarity views of microorganisms and diseases.

[0029] 2. Similarity feature fusion:

[0030] like Figure 2 As shown, this invention employs two similarity fusion methods: SNF and view-level attention mechanism, to generate comprehensive similarity information between microorganisms and diseases.

[0031] 2.1: Based on multiple similarity views of microorganisms and diseases, SNF is used to generate comprehensive microbial similarity. Similarity to disease First, calculate the normalized weight matrix for microorganisms and diseases:

[0032]

[0033]

[0034] Wherein, NMF represents the normalized weight matrix of the microbial functional similarity network; similarly, the normalized weight matrices of the microbial GIP similarity and cosine similarity networks are NMG and NMC, respectively. NDF represents the normalized weight matrix of the disease functional similarity network; similarly, the normalized weight matrices of the disease GIP similarity and cosine similarity networks are NDG and NDC, respectively. Then, the KNN method is used to calculate the local relationships between microorganisms and diseases.

[0035]

[0036]

[0037] in The matrix represents the neighbors of a microorganism / disease. KMF represents the local relation matrix of microbial functional similarity. Similarly, KMG and KMC are the relation matrices of microbial GIP similarity and cosine similarity, respectively. KDF represents the local relation matrix of disease functional similarity. Similarly, KDG and KDC are the relation matrices of disease GIP similarity and cosine similarity, respectively.

[0038] Finally, the fusion process for each similarity is as follows:

[0039]

[0040]

[0041] By integrating all similarities, the final comprehensive similarity matrix of microorganisms and diseases is as follows:

[0042]

[0043]

[0044] 2.2: Based on multiple similarity views of microorganisms and diseases, a view-level attention mechanism is used to adaptively assign weights to different similarity views.

[0045] Let the original feature matrices of the microbial nodes and disease nodes in the p-th view be denoted as follows: and Since microorganisms and disease nodes have different dimensions, feature alignment is performed first. This involves projecting microbial nodes and disease nodes onto the same vector space. The projected feature matrix of the microorganisms is shown below. and the projection feature matrix of the disease The calculation is as follows:

[0046]

[0047]

[0048] in and Represents a linear transformation matrix that can project nodes from different vectors into the same D-dimensional vector space.

[0049] X M = [ X M ( 1 ) , X M ( 2 ) ,..., X M ( P ) ] and X D = [ X D ( 1 ) , X D ( 2 ) ,..., X D ( P ) ] These represent the aligned multi-view features for microorganisms and diseases, respectively. The final embedding features of the nodes can be represented as follows:

[0050] X = [ X M X D ] ∈ R ( Nm + Nd ) × P × D

[0051] Then, a view-level attention mechanism is used to distinguish key information during the multi-view feature fusion process. For each feature with multiple views... f i = [ f i ( 1 ) , f i ( 2 ) ,..., f i ( P ) ] nodes View-level attention is used to distinguish the importance of different views and assign weights to them. The calculation process is as follows:

[0052]

[0053] First, the attention coefficient for each view is calculated as follows:

[0054]

[0055] Among them is Attention vector It is a trainable weight matrix It is the bias vector.

[0056] Then the attention coefficient is adjusted using the softmax function. Perform normalization to obtain Weights:

[0057]

[0058] A higher value indicates greater importance of the corresponding feature view. Subsequently, based on the learned weight values, the node's... The comprehensive feature vector is shown below:

[0059]

[0060] Finally, the fusion features can be obtained. .

[0061] 3. Feature extraction:

[0062] 3.1 Based on the comprehensive microbial similarity and disease similarity fused by SNF, microbial isomorphic networks and disease isomorphic networks are constructed respectively, and encoded using GCN to extract the embedding features of microorganisms and diseases. The specific process is as follows: Figure 3 As shown.

[0063] For an adjacency matrix of Microbial similarity network and the adjacency matrix is Disease similarity network The propagation rules of GCN are as follows:

[0064]

[0065]

[0066] in and Indicates the microorganisms and disease nodes at the 1st Layer embedding features . express The degree matrix, express The degree matrix, and A weight matrix representing microorganisms and diseases. This represents a non-linear activation function.

[0067] 3.1 Based on the fusion of comprehensive microbial and disease similarities using a view-level attention mechanism, a microbial-disease heterogeneous network is constructed and encoded using CANet to learn the interaction characteristics between microorganisms and diseases. The specific process is as follows: Figure 4 As shown.

[0068] The attention module consists of a query vector, a key vector, and a value vector. It calculates the correlation coefficient between the query vector and the key vector, which serves as the weight of the value vector. This invention employs a multi-head cross-attention network based on shared weights to extract interaction information between microorganisms and diseases, capturing their interrelationships to enhance feature representation.

[0069] CANet consists of microbial and disease attention, which have equal weights. Microbial attention comprises three parts: using microbial features... Microbial query vector calculated through linear layers Use of disease characteristics Microbial value vectors calculated through two additional linear layers and microbial bond vector The formulas for calculating the query, key, and value of microorganisms are as follows:

[0070]

[0071] in , , , , and These are the weight matrices for three linear layers. It is the channel dimension. It refers to the number of heads to focus on.

[0072] Disease attention includes three parts: using disease features Disease query vector calculated through linear layer Use of microbial characteristics The disease value vector is calculated through two additional linear layers. and disease key vector The formulas for querying diseases, calculating keys, and values ​​are as follows:

[0073]

[0074] in , and It is a weight matrix shared with microbial attention.

[0075] Based on the query, key, and value vectors of microorganisms and diseases, the attention matrix for microorganisms and diseases is calculated using the softmax function:

[0076]

[0077] in It is the dimension of the query and key vector.

[0078] The attention matrix of each head is multiplied by the corresponding value matrix to obtain the features of each head. Subsequently, these specific features of each head are concatenated through a linear layer to generate the final microbial attention matrix and disease attention matrix.

[0079]

[0080]

[0081] in and These are the final microbial attention matrix and the disease attention matrix, respectively.

[0082] 4. Correlation prediction:

[0083] Based on the encoded features, multi-layer embedded features are fused through a hierarchical attention mechanism to obtain the features of microorganisms and diseases, and an inner product decoder is used to reconstruct the microorganism-disease association matrix.

[0084] 4.1: Based on the features of GCN and CANet encoding, considering the insufficient information of single-layer embedding aggregation, multiple layers are needed to learn the node representations of microorganisms and diseases. Node features in different layers contain different structural information, and combining neighborhood information from different layers can yield richer node representations. In this invention, a hierarchical attention mechanism is employed to combine embedding information from different layers.

[0085] Specifically, for nodes Its multi-layer embedding features are The importance weights of each feature layer are calculated as follows:

[0086]

[0087] in and These represent the initial features. and the Layer Embedded Features The node characteristics.

[0088] Similar to the computation process of the view-level attention mechanism, the computation process of the hierarchical attention mechanism for aggregating node features is as follows:

[0089]

[0090]

[0091]

[0092] final, and This represents the microbial and disease characteristics obtained through GCN encoding; and This represents the microbial and disease characteristics obtained through CANet encoding.

[0093] Through the above steps, the final embedding features of microorganisms and diseases can be represented as follows:

[0094] [ X M X D ] = [ X MGCN || X MCAN X DGCN || X DCAN ]

[0095] in The final characteristic representation of a microorganism, It is the final embedded feature of the disease. Used to connect vectors.

[0096] 4.2: Reconstruct the association matrix using an inner product decoder, where the association score is represented as:

[0097]

[0098] in This represents the prediction score between the microbe-disease pair.

[0099] Known and unknown microbial-disease associations in the dataset are used as positive and negative samples, respectively. A weighting factor is introduced to balance the dataset, and a focus loss function is used as the objective function.

Claims

1. A method for predicting microbial-disease associations based on multi-view multimodal and cross-attention networks, characterized in that: step: Step 1: Preprocess the microbial and disease data, construct multiple similarity views of microorganisms and diseases, and select reliable negative samples through k-means clustering; Step 2: Based on multiple similarity views of microorganisms and diseases, similarity network fusion (SNF) is used to fuse them and construct a comprehensive similarity between microorganisms and diseases; Microbial isomorphism maps are constructed based on comprehensive microbial similarity, and disease isomorphism maps are constructed based on disease similarity. Graph convolutional networks are used to encode the microbial and disease isomorphism maps to extract feature information of microorganisms and diseases. Step 2.1: Use SNF to fuse multiple similarities between microorganisms and diseases. By calculating the normalized weight matrix of the similarity between microorganisms and diseases, the KNN algorithm is used to obtain the local relationship between microorganisms and diseases, and then the similarity matrix between microorganisms and diseases is iteratively updated. Step 2.2: Based on the fused similarity matrix, construct an isomorphic graph of microorganisms and diseases, use a graph convolutional network for feature extraction, aggregate neighbor node information in the network, and update the node features; Step 3: Based on multiple similarity views of microorganisms and diseases, a view-level attention mechanism is used for fusion to obtain initial feature information of microorganisms and diseases; For heterogeneous networks of microbial diseases, the interaction information between microorganisms and diseases is learned by encoding through a cross-attention network with shared weights. Step 3.1: The similarities between microorganisms and diseases are fused through a view-level attention mechanism. First, the microorganism nodes and disease nodes are projected into the same vector space. Step 3.2: Calculate the attention coefficient for each view of microorganisms and diseases; Step 3.3: Normalize the attention coefficients using the softmax function; Step 3.4: Obtain initial feature information of microorganisms and diseases based on the learned attention coefficients; Step 3.5: Encode the initial features of microorganisms and diseases using a cross-attention network based on shared weights; Step 3.6: Microbial attention includes a microbial query vector calculated from microbial features through a linear layer, and microbial key vectors and value vectors calculated from disease features through two other linear layers; Disease attention includes a disease query vector calculated from disease features through a linear layer, and disease key vectors and value vectors calculated from microbial features through two other linear layers. Step 3.7: Calculate the attention matrix for each head of microorganisms and diseases using the softmax function; Step 3.8: Multiply the attention matrix of each head by the corresponding value matrix to obtain the features of each head; The features of each head are concatenated and the final attention matrix is ​​generated through a linear layer; Step 4: Fuse the encoded features using a hierarchical attention mechanism; Step 5: Reconstruct the microbiome-disease association matrix based on the inner product decoder.

2. The microbial-disease association prediction method based on multi-view multimodal and cross-attention networks according to claim 1, characterized in that, In step 1: Step 1.1: Construct multiple similarity views of microorganisms and diseases, including functional similarity, Gaussian interaction profile kernel similarity, and cosine similarity; Step 1.2: Use k-means clustering to select negative samples that are balanced with the number of positive samples. Divide the negative samples in the dataset into k clusters, and randomly select samples from each cluster as negative samples, while keeping the positive samples unchanged.

3. The microbial-disease association prediction method based on multi-view multimodal and cross-attention networks according to claim 1, characterized in that, In step 4: Step 4.1: Based on the features encoded by the graph convolutional network and the cross-attention network, a layer attention mechanism is adopted to fuse the initial features from the two networks and the features from each layer of the network respectively; Step 4.2: Calculate the features of each layer and the attention coefficients of the initial features; Step 4.3: Based on the weighted features, obtain the microbial embedding features and the disease embedding features.

4. The microbial-disease association prediction method based on multi-view multimodal and cross-attention networks according to claim 1, characterized in that, In step 5: Step 5.1: Connect the microbial embedding features and disease embedding features from the two networks respectively to obtain the final microbial features and disease features; Step 5.2: Reconstruct the microbe-disease association matrix using the inner product decoder.

Images