Microorganism-disease association prediction method and device based on multi-level perceptual aggregation and deep hierarchical graph mixed learner

Abstract

The invention belongs to the technical field of biomedical engineering, and particularly relates to a microorganism-disease association prediction method and device based on a multi-level perceptual aggregation and deep hierarchical graph mixed learner, and the method comprises the following specific steps: S1, obtaining known microorganism-disease association information, and constructing a microorganism-disease association matrix A1; meanwhile, constructing a microorganism comprehensive similarity matrix CM, a microorganism characterization mapping matrix MM, a disease comprehensive similarity matrix CD and a disease characterization mapping matrix DD; and S2, based on CM, MM, CD and DD, applying a multi-level perception aggregation strategy MLPA to obtain a microbial spring layer linkage feature matrix LM and a disease spring layer linkage feature matrix LD. According to the method, feature expression is enhanced by adopting a multi-level perception aggregation strategy, potential association is mined in combination with a double-drive association enhancement strategy, meanwhile, efficient learning and dynamic feature aggregation of a complex heterogeneous network structure are realized through a deep layered graph mixed learning device, and the accuracy and reliability of association prediction are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biomedical engineering, in particular to a microbe-disease association prediction method and device based on a multi-level perception aggregation and deep hierarchical graph hybrid learner. BACKGROUND

[0002] With the development of artificial intelligence technology, the research on microbe-disease association has become a hot spot in the cross-research of bioinformatics and artificial intelligence. Studies have shown that the imbalance of human intestinal flora, skin microbiome and other ecosystems is closely related to metabolic diseases, autoimmune diseases, nervous system diseases and even cancer and other complex diseases. Under this background, it is of great significance to develop an efficient and accurate microbe-disease association prediction method and device for understanding disease mechanisms and clinical diagnosis and treatment.

[0003] In the existing microbe-disease association prediction methods, the network propagation method constructs a microbe similarity network and a disease similarity network, and uses random walk, label propagation and other algorithms to diffuse and propagate the known association information in the network, so as to predict potential associations. This kind of method can effectively utilize the network topology, but it is strongly dependent on the quality of network construction, and it is difficult to capture nonlinear association patterns; the matrix decomposition method decomposes the microbe-disease association matrix into a low-dimensional feature matrix, and reconstructs the association matrix by linear combination to predict unknown associations. This kind of method has high computational efficiency, but its processing capacity for sparse data is limited, and it cannot effectively integrate multi-source heterogeneous information.

[0004] In recent years, graph neural networks (GNN) have shown great potential in the field of microbe-disease association prediction, among which graph convolutional networks (GCN) and graph attention networks (GAT) have become the most representative two methods. GCN can automatically learn the low-dimensional feature representation of microorganisms and diseases through an efficient neighborhood information aggregation mechanism. Its advantage lies in low computational complexity and easy implementation, especially suitable for processing large-scale biological network data. This method can capture network structure features of different scales through multi-layer convolution operation, providing an end-to-end solution for microbe-disease association prediction. GAT introduces an attention mechanism based on GCN, which can more finely describe the differences in association strength between microorganisms and diseases by calculating the dynamic weights between nodes. The core advantage of GAT is that it can automatically identify important association relationships in the network and assign them higher attention weights, thereby improving the prediction accuracy of the model. These two methods break through the limitations of traditional methods in automatic feature extraction through deep learning technology, providing a new technical path for microbe-disease association research.

[0005] Although the existing microbe-disease association prediction methods have made some progress, there are still several key problems: first, the similarity information is not fully utilized, and there is a lack of multi-level feature fusion mechanism for similarity information, resulting in incomplete and inaccurate prediction results; second, the microbe-disease association matrix has the problem of data sparsity, but the existing methods for dealing with the problem of data sparsity of the association matrix have certain limitations; third, the existing graph neural network architecture has the problems of insufficient high-order interaction modeling, lack of feature dynamic adjustment mechanism, and feature over-smoothing in microbe-disease association prediction, which seriously restricts the prediction performance of the model.

[0006] Based on the above, therefore, the application provides a microbe-disease association prediction method and device based on a multi-level perception aggregation and deep hierarchical graph hybrid learner. SUMMARY

[0007] To solve the above technical problems, according to one aspect of the application, the application provides the following technical solutions:

[0008] A microbe-disease association prediction method based on a multi-level perception aggregation and deep hierarchical graph hybrid learner includes the following specific steps:

[0009] S1: Obtain known microbe-disease association information and construct a microbe-disease association matrix A1; at the same time, construct a microbe comprehensive similarity matrix CM, a microbe representation mapping matrix MM, a disease comprehensive similarity matrix CD, and a disease representation mapping matrix DD;

[0010] S2: Based on CM, MM, CD, and DD, apply a multi-level perception aggregation strategy MLPA to obtain a microbe leap layer linkage feature matrix LM and a disease leap layer linkage feature matrix LD;

[0011] S3: Based on A1, apply a double-drive association enhancement strategy DDAES to alleviate the problem of data sparsity of the association matrix, and obtain an enhanced microbe-disease association matrix A2;

[0012] S4: Use A2, CM, and CD to construct an adjacency matrix B and an initial feature matrix X0 of the microbe-disease heterogeneous graph, and perform deep learning and dynamic aggregation on the features of microorganisms and diseases through a deep hierarchical graph hybrid learner DHGHL to obtain a microbe aggregated linkage feature matrix HM and a disease aggregated linkage feature matrix HD;

[0013] S5: Based on LM, HM, LD, and HD, reconstruct and predict the association between microorganisms and diseases to obtain a reconstruction prediction matrix A' and output the final association probability between microorganisms and diseases.

[0014] As a preferred scheme of the microbe-disease association prediction method based on the multi-level perception aggregation and deep layered graph hybrid learner, wherein the specific steps of S1 are as follows:

[0015] S11: Obtain known microbe-disease association information, delete redundant and ambiguous entries in the obtained association information, and ensure the accuracy and consistency of the association information;

[0016] S12: Based on the collected microbe-disease association information, let nm and nd represent the number of microorganisms and diseases respectively, obtain the microbe-disease association matrix

[0017] S13: Construct a comprehensive similarity matrix of microorganisms based on A1 and a microorganism representation mapping matrix

[0018] S14: Construct a comprehensive similarity matrix of diseases based on A1 and a disease representation mapping matrix

[0019] As a preferred scheme of the microbe-disease association prediction method based on the multi-level perception aggregation and deep layered graph hybrid learner, wherein the specific steps of S13 are as follows:

[0020] S131: Based on A1, for any two microorganisms m i and m j , calculate the Gaussian kernel similarity SM i (m j ,m Gip ) between m i and m j ;

[0021] S132: For any two microorganisms m i and m j , calculate the cosine similarity SM i (m j ,m Cos ) between m i and m j ;

[0022] S133: For any two microorganisms m i and m j , calculate the Jaccard similarity SM i (m j ,m Jac ) between m i and m j ;

[0023] S134: For any two microorganisms m i and m j , compute the Sigmoid kernel similarity SM i (m j , m Sig ) between m i and m j ;

[0024] S135: Based on protein function association, obtain the function similarity SM i (m j , m Fun ) between any given two microorganisms m i and m j by the method of computing microorganism-microorganism function association index;

[0025] S136: Based on the Gaussian kernel similarity, cosine similarity, Jaccard similarity, Sigmoid kernel similarity and function similarity between any two microorganisms m i and m j , construct the microorganism comprehensive similarity matrix and the microorganism representation mapping matrix

[0026] The specific steps of the S14 are as follows:

[0027] S141: Based on A1, for any two diseases d i and d j , compute the Gaussian kernel similarity SD i (d j , d Gip ) between d i and d j ;

[0028] S142: For any two diseases d i and d j , compute the cosine similarity SD i (d j , d Cos ) between d i and d j ;

[0029] S143: For any two diseases d i and d j , compute the Jaccard similarity SD i (d j , d Jac ) between d i and d j ;

[0030] S144: For any two diseases d i and d j Calculate d i With d j Sigmoid kernel similarity between SD Sig (d i ,d j );

[0031] S145: Based on gene interaction information, calculate d for any given two diseases. i and d j Functional similarity between SD Fun (d i ,d j );

[0032] S146: Based on any two diseases d i and d j A comprehensive disease similarity matrix is ​​constructed by using Gaussian kernel similarity, cosine similarity, Jaccard similarity, Sigmoid kernel similarity, and functional similarity among the samples. Disease characterization mapping matrix

[0033] As a preferred embodiment of the microbial-disease association prediction method based on multi-level perceptual aggregation and deep hierarchical graph hybrid learner described in this invention, the specific steps of S2 are as follows:

[0034] S21: Based on CM and MM, the multi-level perception aggregation strategy MLPA is applied to obtain the microbial hierarchical linkage feature matrix LM;

[0035] S22: Based on CD and DD, the multi-level perception aggregation strategy MLPA is applied to obtain the disease hierarchical linkage feature matrix LD.

[0036] As a preferred embodiment of the microbial-disease association prediction method based on multi-level perceptual aggregation and deep hierarchical graph hybrid learner described in this invention, the specific steps of S21 are as follows:

[0037] S211: Based on CM and MM, construct a three-level sensory feature matrix: microbial single-layer sensory feature matrix Microbial dual-layer sensing feature matrix and the three-layer sensory feature matrix of microorganisms

[0038] S212: Based on and Constructing a microbial composite sensing feature matrix and microbial equilibrium residual characteristic matrix

[0039] S213: based on and constructing the microorganism-stratum linkage feature matrix through matrix splicing operation

[0040] The specific steps of S22 are as follows:

[0041] S221: based on CD and DD, constructing three levels of perception feature matrix: disease single-layer perception feature matrix disease double-layer perception feature matrix and disease three-layer perception feature matrix

[0042] S222: based on and constructing disease composite perception feature matrix and disease balanced residual feature matrix

[0043] S223: based on and constructing the disease-stratum linkage feature matrix through matrix splicing operation

[0044] As a preferred scheme of the microorganism-disease association prediction method based on multi-level perception aggregation and deep hierarchical graph hybrid learner, the specific steps of S3 are as follows:

[0045] S31: constructing the microorganism partner matrix based on A1

[0046] S32: constructing the disease partner matrix based on A1

[0047] S33: normalizing M1 and D1 to obtain normalized microorganism partner matrix and normalized disease partner matrix

[0048] S34: based on A1, M norm , D norm , constructing the association potential score matrix

[0049] S35: based on A pot , calculating the average potential score score i of the associated microorganism for each disease d i ;

[0050] S36: based on A1, Apot and score i , constructing an enhanced microbe-disease association matrix

[0051] As a preferred scheme of the microbe-disease association prediction method based on multi-level perception aggregation and deep hierarchical graph hybrid learner, wherein the specific steps of S4 are as follows:

[0052] S41: based on A2, CM and CD, constructing an adjacency matrix of a microbe-disease heterogeneous graph and an initial feature matrix

[0053] S42: based on B and X0, applying a deep hierarchical graph hybrid learner DHGHL to obtain a microbe aggregation linkage feature matrix HM and a disease aggregation linkage feature matrix HD.

[0054] As a preferred scheme of the microbe-disease association prediction method based on multi-level perception aggregation and deep hierarchical graph hybrid learner, wherein the specific steps of S42 are as follows:

[0055] S421: based on B and X0, using a graph convolution network GCN in the first layer of DHGHL to obtain a first layer collaborative representation matrix

[0056] S422: based on B and H (1) , using a graph convolution network GCN in the second layer of DHGHL to obtain a second layer collaborative representation matrix

[0057] S423: based on H (1) and H (2) , applying a double-layer feature dynamic fusion strategy DL-FDFS in the third layer of DHGHL to obtain a third layer collaborative representation matrix

[0058] S424: based on B and H (3) , applying an enhanced double-path graph attention network E-DPGAT in the fourth layer of DHGHL to further retain original features and explore high-order structure information, to obtain a microbe aggregation linkage feature matrix HM and a disease aggregation linkage feature matrix HD.

[0059] The specific steps of S424 are as follows:

[0060] S4241: based on B, constructing a direct neighbor node set of each node i and an indirect neighbor node set

[0061] S4242: based on the set and H (3) , calculate the attention coefficient between each node i and its direct neighbor node j Then normalize to get

[0062] S4243: according to the normalized attention coefficient Weighted aggregation of direct neighbor features, get direct neighbor aggregation feature vector

[0063] S4244: based on the node's own feature vector h i And the direct neighbor aggregation feature vector Construct the first path feature vector of the node

[0064] S4245: based on the set And H (3) , calculate the attention coefficient between each node i and its indirect neighbor node j Then normalize to get

[0065] S4246: according to the normalized attention coefficient Weighted aggregation of indirect neighbor features, get indirect neighbor aggregation feature vector

[0066] S4247: based on the node's own feature vector h i And the indirect neighbor aggregation feature vector Construct the second path feature vector of the node

[0067] S4248: based on And Construct the fourth layer collaborative representation feature vector for any node i

[0068] S4249: based on Construct the microorganism aggregation linkage feature matrix And the disease aggregation linkage feature matrix

[0069] As a preferred scheme of the microorganism-disease association prediction method based on the multi-level perception aggregation and deep hierarchical graph hybrid learner, wherein: the specific steps of S5 are as follows:

[0070] S51: based on LM, HM, LD, HD, multi-modal feature splicing is performed to construct a reconstructed microorganism feature matrix and reconstructing the disease feature matrix

[0071] S52: Reconstructing the association between microorganisms and diseases based on the FM and the FD to obtain a reconstructed prediction matrix

[0072] S53: Defining a loss function to measure the difference between the reconstructed prediction matrix A' and the real association matrix A1, and training the model parameters by optimizing the loss function;

[0073] S54: Based on A', outputting the final association probability between microorganisms and diseases.

[0074] A microorganism-disease association prediction device based on a multi-level perception aggregation and deep hierarchical graph hybrid learner, comprising:

[0075] A data acquisition and preprocessing module for automatically collecting microorganism-disease association information from authoritative biomedical databases, deleting redundant and ambiguous entries in the obtained association information, and then constructing a microorganism-disease association matrix A1 to provide a standardized data basis for subsequent feature calculation;

[0076] A feature calculation engine for extracting various similarity information and enhancing the association information; first, based on the microorganism-disease association matrix A1, the comprehensive similarity matrix and the representation mapping matrix of microorganisms and diseases are constructed by integrating Gaussian kernel similarity, cosine similarity, Jaccard similarity, Sigmoid kernel similarity and functional similarity; then, an enhanced microorganism-disease association matrix A2 is generated through a double-drive association enhancement strategy to provide high information density feature input for the subsequent intelligent analysis and learning module;

[0077] An intelligent analysis and learning module for applying a multi-level perception aggregation strategy and a deep hierarchical graph hybrid learner to realize deep modeling of microorganism features and disease features, and to reconstruct and predict the association between microorganisms and diseases to obtain the association probability between microorganisms and diseases;

[0078] An interaction and output module for converting complex prediction results into intuitive and usable information.

[0079] Compared with the prior art:

[0080] 1. The multi-level perception aggregation strategy (MLPA) is based on the systematic integration of multiple similarities, and uses progressive matrix multiplication operations to extract feature information layer by layer. Meanwhile, two new types of feature expression forms, namely composite perception feature matrix and balanced residual feature matrix, are designed. Through a nonlinear fusion mechanism, dynamic optimization and fusion of multi-level features are realized. The limitations of traditional methods in the utilization of similarity features are effectively overcome, and the depth mining and full utilization of microbial and disease similarity information are realized, which significantly improves the comprehensiveness of feature expression and the robustness of the model;

[0081] 2. The double-drive association enhancement strategy (DDAES) solves the problem of data sparsity in the microbial-disease association matrix, effectively solving the problem of large fluctuations and poor stability of the prediction results caused by data sparsity in existing sparse matrix processing methods, and significantly improving the recognition ability of the model for potential microbial-disease associations;

[0082] 3. The deep hierarchical graph hybrid learner (DHGHL) adopts a unique four-layer architecture design, which systematically solves the key problems of existing graph neural networks in microbial-disease association prediction. The first and second layers of DHGHL use graph convolutional networks (GCN) to capture direct neighborhood features and indirect neighborhood features in the microbial-disease heterogeneous graph, respectively. The third layer realizes adaptive weighted fusion of direct neighborhood features and indirect neighborhood features through a double-layer feature dynamic fusion strategy (DL-FDFS), balancing the contributions of local neighborhood information and indirect neighborhood information. The fourth layer innovatively introduces an enhanced double-path graph attention network (E-DPGAT) to accurately model the interaction between one-hop direct neighbors and two-hop indirect neighbors, while combining residual connection and layer normalization techniques to realize efficient learning and dynamic feature aggregation of complex heterogeneous network structures. This learner not only effectively solves the long-range dependency modeling problem of traditional graph neural networks, but also avoids the problem of excessive smoothing of node features, ensuring the complete preservation of key feature information. BRIEF DESCRIPTION OF DRAWINGS

[0083] Fig. 1 The figure is a schematic diagram of the overall process of the present application;

[0084] Fig. 2 The figure is a schematic diagram of the overall process of the present application;

[0085] Fig. 3 The figure is a schematic diagram of the overall process of the present application; DETAILED DESCRIPTION

[0086] In order to make the purpose, technical scheme and advantages of the present application clearer, the embodiments of the present application will be described in detail below with reference to the drawings.

[0087] The application provides a microbe-disease association prediction method based on a multi-level perception aggregation and deep hierarchical graph hybrid learner, please refer to Figs. 1-3 , including the specific steps as follows:

[0088] S1: Obtain known microbe-disease association information, and construct a microbe-disease association matrix A1; meanwhile, construct a microbe comprehensive similarity matrix CM, a microbe representation mapping matrix MM, a disease comprehensive similarity matrix CD, and a disease representation mapping matrix DD.

[0089] The specific steps of S1 are as follows:

[0090] S11: Obtain known microbe-disease association information, delete redundant and fuzzy entries in the obtained association information, and ensure the accuracy and consistency of the association information.

[0091] Embodiment: The known microbe-disease association information is downloaded from the Disbiome database (https: / / disbiome.ugent). The database provides verified microbe-disease association information. The database initially contains 5573 associations, and 4351 non-redundant and non-fuzzy associations are retained in the experiment, involving 218 diseases and 1052 microorganisms.

[0092] S12: Based on the collected microbe-disease association information, let nm and nd represent the number of microorganisms and diseases respectively, and obtain the microbe-disease association matrix The specific construction method of A1 is as follows: for any given microorganism m i and disease d j , if there is a known association between them, then A1(i,j) = 1, otherwise A1(i,j) = 0. Where A1(i,j) represents the value of the element in the i-th row and the j-th column of A1.

[0093] S13: Based on A1, construct the microbe comprehensive similarity matrix and the microbe representation mapping matrix

[0094] The specific steps of S13 are as follows:

[0095] S131: Based on A1, for any two microorganisms m i and m j , calculate the Gaussian kernel similarity SM Gip (m i ,m j ) between m i and m j ; the specific calculation process is as follows:

[0096] SM Gip (mi ,m j )=exp(-λ1||IP(m i )-IP(m j )|| 2 )

[0097]

[0098] Where λ1 represents the normalized bandwidth parameter; ||.|| represents the Euclidean distance, ||.|| 2 IP(m) represents the square of the Euclidean distance. i ) represents a microorganism m constructed based on A1 i A binary associative vector used to record m i Information on associations with all diseases; IP(m) j ) represents a microorganism m constructed based on A1 j A binary associative vector used to record m j Information on associations with all diseases.

[0099] in:

[0100] The core idea of ​​the Gaussian Interaction Profile (GIP) is to use a Gaussian kernel function to transform the distance between the association vectors of entities into a similarity score. The advantages of GIP are its simplicity and efficiency; by mapping discrete interaction information to a continuous similarity space, it effectively captures latent patterns.

[0101] Euclidean distance is used to calculate the straight-line distance between two points in a multidimensional space.

[0102] S132: For any two microorganisms m i and m j Calculate m i With m j Cosine similarity between SM Cos (m i ,m j The specific calculation formula is as follows:

[0103]

[0104] Where · represents the dot product operation of vectors.

[0105] S133: For any two microorganisms m i and m j Calculate m i With m j Jaccard similarity between SM Jac (m i ,mj ) ; the specific calculation formula is as follows:

[0106]

[0107] wherein, |IP(m i )∩IP(m j )| represents the number of diseases associated with m i and m j ; |IP(m i )∪IP(m j )| represents the total number of all diseases associated with m i and m j after deduplication.

[0108] wherein: the Jaccard similarity is an index for measuring the similarity of two sets, defined as the ratio of the intersection size to the union size of two sets, which can intuitively reflect the proportion of common features and ignore the interference of unique features.

[0109] S134: for any two microorganisms m i and m j , calculate the Sigmoid kernel similarity SM Sig (m i ,m j ) between m i and m j ; the specific calculation formula is as follows:

[0110]

[0111] wherein, tanh represents the hyperbolic tangent function; θ m represents the bias term;

[0112] wherein: the Sigmoid kernel similarity is a nonlinear similarity measurement method based on the hyperbolic tangent function (tanh), which evaluates the correlation strength by calculating the nonlinear transformation of the dot product of two vectors.

[0113] S135: based on protein function association, the function similarity SM Fun (m i ,m j ) between any given two microorganisms m i and m j is obtained by calculating the microorganism-microorganism function association index;

[0114] wherein:

[0115] The calculation of the functional similarity of microorganisms is based on the existing proposed microbial functional index (MFI) method, which quantifies the functional similarity between different microorganisms by analyzing the protein-protein functional association network in the microbial genome. Specifically, this method uses database retrieval to obtain the protein-protein functional association network, and counts the information of microbial pairs in the association network, including the presence of network nodes (gene families) in at least one genome and the connections (i.e. edges in the network) connecting these gene families. Gene families are labeled to represent the presence or absence of proteins of a specific gene family in genome A, genome B, or both, thereby forming three types of gene families: belonging to genome A, belonging to genome B, and common. Such a network contains three types of undirected edges, connecting three different types of starting nodes and ending nodes (A connected to A, A interconnected with B, B connected to B). The microbial functional association index between two microorganisms is defined as the total number of edges that must cross the gene families encoded exclusively in one genome to connect to the gene families in the other genome. The functional similarity of microorganisms can be represented by the following formula:

[0116]

[0117] where MFI represents the functional association index between microorganisms, i.e. the functional similarity between microorganisms, AB represents the number of edges interconnected between A and B, AA represents the number of edges connected to A, and BB represents the number of edges connected to B.

[0118] S136: Construct a comprehensive similarity matrix of microorganisms based on the Gaussian kernel similarity, cosine similarity, Jaccard similarity, Sigmoid kernel similarity and functional similarity between any two microorganisms m i and m j and the microorganism representation mapping matrix

[0119] where the specific construction process of CM is as follows:

[0120] First, calculate the comprehensive similarity SM i (m j ,m Com ) between m i and m j , and the calculation formula is as follows:

[0121]

[0122] Then, integrate the comprehensive similarity between all microorganisms to construct CM. The integration process is as follows:

[0123] ​

[0124] In addition, the specific construction process of MM is as follows:

[0125] First, calculate the feature mapping similarity SM i between m j , the calculation formula is as follows: Map (m i ,m j )

[0126]

[0127] Then, the feature mapping similarity between all microorganisms is integrated, and MM is constructed. The integration process is as follows:

[0128]

[0129] Among them: comprehensive similarity is a concept for measuring the similarity between two or more objects, which is not dependent on a single similarity to judge, but uses multiple similarity calculation methods to evaluate the similarity between objects more comprehensively and accurately.

[0130] S14: Constructing disease comprehensive similarity matrix based on A1 and disease feature mapping matrix

[0131] The specific steps of S14 are as follows:

[0132] S141: Based on A1, for any two diseases d i and d j , calculate the Gaussian kernel similarity SD i (d j ,d Gip ) between d i and d j , the specific calculation process is as follows:

[0133] SD Gip (d i ,d j )=exp(-λ2||IP(d i )-IP(d j )|| 2 )

[0134]

[0135] Among them, λ2 represents the normalized bandwidth parameter; IP(d i ) represents the binary association vector of disease d i based on A1, which records d iInformation on associations with all microorganisms; IP(d) j ) represents the disease d constructed based on A1. j The binary associative vector, recording d j Information on associations with all microorganisms.

[0136] S142: For any two diseases d i and d j Calculate d i With d j Cosine similarity between SD Cos (d i ,d j The specific calculation formula is as follows:

[0137]

[0138] S143: For any two diseases d i and d j Calculate d i With d j Jaccard similarity between SD Jac (d i ,d j The specific calculation formula is as follows:

[0139]

[0140] Among them, |IP(d i )∩IP(d j )| represents d i and d j The number of co-related microorganisms; |IP(d i )∪IP(d j )| represents d i and d j The total number of all related, deduplicated microorganisms.

[0141] S144: For any two diseases d i and d j Calculate d i With d j Sigmoid kernel similarity between SD Sig (d i ,d j The specific calculation formula is as follows:

[0142]

[0143] Where, θ d This indicates the bias term.

[0144] S145: Based on the gene interaction information, calculate the functional similarity SD i and d j for any given two diseases d Fun (d i ,d j );

[0145] The method for calculating the functional similarity between diseases based on gene interaction information: first, derive the relevant gene set for each disease; then measure the similarity between diseases according to the functional similarity between genes (evaluated by log-likelihood score). Specifically, the similarity between disease pairs is calculated by calculating the maximum functional similarity between the relevant gene sets and considering the functional association of all genes, and finally obtaining the functional similarity between diseases.

[0146] S146: Based on the Gaussian kernel similarity, cosine similarity, Jaccard similarity, Sigmoid kernel similarity and functional similarity between any two diseases d i and d j , construct the disease comprehensive similarity matrix and disease representation mapping matrix

[0147] Wherein, the specific construction process of CD is as follows:

[0148] First, calculate the comprehensive similarity SD i (d j ,d Com ) between d i and d j , the calculation formula is as follows:

[0149]

[0150] Then, integrate the comprehensive similarity between all diseases, and construct CD. The integration process is as follows:

[0151]

[0152] In addition, the specific construction process of DD is as follows:

[0153] First, calculate the representation mapping similarity SD i (d j ,d Map ) between d i and d j , the calculation formula is as follows:

[0154]

[0155] Then, the representation mapping similarities between all diseases are integrated two by two to construct DD. The integration process is as follows:

[0156]

[0157] S2: Based on CM, MM, CD and DD, a multi-level perceptive aggregation strategy MLPA (Multi-Level Perceptive Aggregation) is applied to obtain a microorganism stratum linkage feature matrix LM and a disease stratum linkage feature matrix LD.

[0158] The specific steps of S2 are as follows:

[0159] S21: Based on CM and MM, a multi-level perceptive aggregation strategy MLPA is applied to obtain a microorganism stratum linkage feature matrix LM.

[0160] The specific steps of S21 are as follows:

[0161] S211: Based on CM and MM, three levels of perceptive feature matrices are constructed: a microorganism single-layer perceptive feature matrix a microorganism double-layer perceptive feature matrix and a microorganism three-layer perceptive feature matrix The construction process of is as follows:

[0162]

[0163] Wherein, * represents matrix multiplication.

[0164] S212: Based on and a microorganism composite perceptive feature matrix and a microorganism balanced residual feature matrix are constructed. The construction process of and

[0165] is as follows:

[0166] Wherein, Sigmoid() is an activation function.

[0167] S213: Based on and a microorganism stratum linkage feature matrix LM is constructed through matrix splicing operation. The specific construction process of LM is as follows:

[0168]

[0169] where concat() represents the horizontal concatenation operation of matrix, which completes the expansion of feature dimension.

[0170] S22: Based on CD and DD, the multi-level perception aggregation strategy MLPA is applied to obtain the disease leap layer linkage feature matrix LD.

[0171] The specific steps of S22 are as follows:

[0172] S221: Based on CD and DD, three levels of perception feature matrices are constructed: disease single layer perception feature matrix disease double layer perception feature matrix and disease three layer perception feature matrix The construction process of is as follows:

[0173]

[0174] S222: Based on and disease composite perception feature matrix and disease balanced residual feature matrix The construction process of is as follows:

[0175]

[0176] S223: Based on and disease leap layer linkage feature matrix LD is constructed through matrix concatenation operation. The specific construction process of LD is as follows:

[0177]

[0178] S3: Based on A1, the dual-driven association enhancement strategy DDAES is applied to alleviate the data sparsity problem of the association matrix, and the enhanced microorganism-disease association matrix A2 is obtained.

[0179] The specific steps of S3 are as follows:

[0180] S31: Based on A1, the microorganism partner matrix M1 is constructed. The specific construction process of M1 is as follows: for any given microorganism m i and microorganism m j , if there is at least one same associated disease between them, that is, there is at least one disease d kIf A1(i,k) = 1 and A1(j,k) = 1, then M1(i,j) = 1; otherwise, M1(i,j) = 0.

[0181] S32: Constructing a Disease Partner Matrix Based on A1 The specific construction process of D1 is as follows: For any given disease d i and disease d j If there is at least one common associated microorganism among them, that is, if there is at least one microorganism m k If A1(k,i)=1 and A1(k,j)=1, then D1(i,j)=1; otherwise, D1(i,j)=0.

[0182] S33: Normalize M1 and D1 to obtain the normalized microbial partner matrix. and normalized disease partner matrix

[0183] Among them, M norm The construction process is as follows:

[0184] For M norm For any element in M1: if M1(i,j)=0, then M norm (i,j)=0; if M1(i,j)=1, then M norm (i,j) is calculated according to the following formula:

[0185]

[0186] Where |S(m i )| indicates that in A1, there is a relationship with microorganism m i The number of associated diseases represents the number of elements in row i of A1 with a value of 1; |S(m j )| indicates that in A1, there is a relationship with microorganism m j Number of associated diseases; |S(m i )∩S(m j )| indicates that it is simultaneously with m i and m j The number of associated diseases.

[0187] In addition, D norm The construction process is as follows:

[0188] For D norm Any element in the array: If D1(i,j)=0, then D norm (i,j)=0; if D1(i,j)=1, then D norm (i,j) is calculated according to the following formula:

[0189]

[0190] where |S(d i )| represents the number of microorganisms associated with disease d i in A1, i.e., the number of elements with value 1 in the i-th column of A1; |S(d j )| represents the number of microorganisms associated with disease d j in A1; |S(d i )∩S(d j )| represents the number of microorganisms associated with both d i and d j .

[0191] S34: Based on A1, M norm , D norm , construct the associated potential score matrix The specific construction process is as follows:

[0192] A pot = M norm *A1+A1*(D norm ) T

[0193] S35: Based on A pot , for each disease d i , calculate the average potential score score i of its associated microorganisms; the calculation formula of score i is as follows:

[0194]

[0195] where N(d i ) represents the set of microorganisms with non-zero associated potential score with d pot in A i ; |N(d i )| represents the number of microorganisms with non-zero associated potential score with d pot in A i ; A pot (k,i) represents the value of the element in the k-th row and i-th column of A pot , recording the associated potential score between m k and d i .

[0196] S36: Based on A1, A pot and score i , construct the enhanced microorganism-disease association matrix The specific construction process is as follows: if A1(k,i) = 1, or the microorganism m kdisease d i In A pot the association potential score A pot (k, i) > score i , it is considered that m k and d i are associated, and A2(k, i) = 1 is set. The formula is constructed as follows:

[0197]

[0198] S4: using A2, CM, CD, constructing the adjacency matrix B and the initial feature matrix X0 of the microorganism-disease heterogeneous graph, and using the deep hierarchical graph hybrid learner DHGHL to perform deep learning and dynamic aggregation on the features of microorganisms and diseases, to obtain the microorganism aggregated linkage feature matrix HM and the disease aggregated linkage feature matrix HD.

[0199] The specific steps of S4 are as follows:

[0200] S41: based on A2, CM, CD, constructing the adjacency matrix B and the initial feature matrix X0 of the microorganism-disease heterogeneous graph and initial feature matrix The specific construction process is as follows:

[0201]

[0202] S42: based on B and X0, applying the deep hierarchical graph hybrid learner DHGHL (Deep Hierarchical Graph Hybrid Learner) to obtain the microorganism aggregated linkage feature matrix HM and the disease aggregated linkage feature matrix HD; the overall architecture diagram of DHGHL is shown in Fig. 2 .

[0203] The specific steps of S42 are as follows:

[0204] S421: based on B and X0, using the graph convolution network GCN in the first layer of DHGHL to obtain the first layer collaborative representation matrix The specific process of obtaining H (1) is as follows:

[0205]

[0206] wherein, is the adjacency matrix after normalization based on B, and the construction process is as follows:

[0207]

[0208] D is the degree matrix of B, is the inverse of the square root of each element of D.

[0209] is the first layer weight matrix obtained by training, d represents the feature mapping dimension; ELU is an activation function, and the complete formula is as follows:

[0210]

[0211] Among them, the graph convolution network GCN is a deep learning model specially processing graph structure data, which realizes efficient learning and propagation of node features by extending the traditional convolution operation to the graph data in non-Euclidean space.

[0212] S422: based on B and H (1) , in the second layer of DHGHL, the second layer collaborative representation matrix H is obtained by using the graph convolution network GCN. (2) The specific process of obtaining H (1) is as follows:

[0213]

[0214] Among them, is the second layer weight matrix obtained by training.

[0215] S423: based on H (2) , in the third layer of DHGHL, the double-layer feature dynamic fusion strategy DL-FDFS (Double-Layer Feature Dynamic Fusion Strategy) is applied to obtain the third layer collaborative representation matrix H The specific process of obtaining H (3) is as follows:

[0216] H (3) = α1H (1) + α2H (2)

[0217] Among them, α1 and α2 represent the weight contribution values obtained by training.

[0218] Among them: in DHGHL: direct neighborhood information is obtained by the first layer GCN, which effectively models the direct neighborhood features in the heterogeneous graph to obtain H (1) ; the second layer GCN is used to obtain deeper and more complex indirect neighborhood information, which effectively models the indirect neighborhood features in the heterogeneous graph to obtain H (2) ; the third layer adopts the DL-FDFS strategy to dynamically balance the contribution of direct neighborhood features and indirect neighborhood features obtained from the double-layer GCN, and finally obtains H (3) .

[0219] S424: based on B and H (3) In the fourth layer of the DHGHL, an enhanced dual-path graph attention network (E-DPGAT) is applied to further retain original features and explore high-order structural information, obtaining a microbial aggregated linkage feature matrix HM and a disease aggregated linkage feature matrix HD. The overall design diagram of the E-DPGAT is shown in FIG. 4. Fig. 3

[0220] The specific steps of S424 are as follows:

[0221] S4241: based on B, constructing a direct neighbor node set of each node i and an indirect neighbor node set The direct neighbor node set of node i is constructed as follows: based on B, if there is a node j satisfying B(i, j) = 1, then node j belongs to The indirect neighbor node set of node i is constructed as follows: based on B 2 , if there is a node j satisfying B(i, j) = 1, then node j belongs to 2

[0222] wherein: the nodes in the microbe-disease heterogeneous graph represent any microbe or disease. In the following description, the nodes include microbe nodes and disease nodes.

[0223] S4242: based on the sets and H (3) , calculating the attention coefficient between each node i and its direct neighbor node j Then, normalization is performed to obtain The specific calculation process is shown as follows:

[0224]

[0225] wherein: LeakyReLU is an activation function for alleviating negative interval gradient disappearance; is an attention parameter vector; is a learned weight matrix; is a feature vector of node i learned from H (3) , i.e., a column vector composed of all elements in the i-th row of H (3) ; and is a feature vector of node j learned from H (3) , i.e., a column vector composed of all elements in the i-th row of H (3) ​​The column vector formed by all elements in the jth row of the matrix; || represents the vector splicing operation, and the purpose is to splice two feature vectors into a long vector.

[0226] S4243: According to the normalized attention coefficient Weighted aggregation of direct neighbor features to obtain a direct neighbor aggregated feature vector The specific aggregation process is as follows:

[0227]

[0228] Where ReLU is an activation function.

[0229] S4244: Based on the node itself feature vector h i and the direct neighbor aggregated feature vector Construct the first path feature vector of the node The specific construction formula is as follows:

[0230]

[0231] Where LayerNorm represents the layer normalization operation; MLP is a multi-layer perceptron, which is used to process data and learn the complex relationship between input and output. The purpose of using MLP here is to reduce the dimension of the spliced result of W1h i and So that the dimension of the reduced vector is the same as h i , which is convenient for subsequent operation and calculation.

[0232] S4245: Based on the set and H (3) , calculate the attention coefficient between each node i and its indirect neighbor node j Then normalize to get The specific calculation process is as follows:

[0233]

[0234] Where, is the attention parameter vector; is the learned weight matrix; h i is the feature vector of node i learned from H (3) ; h j is the feature vector of node j learned from H (3) .

[0235] S4246: According to the normalized attention coefficient Weighted aggregation of indirect neighbor features to obtain an indirect neighbor aggregated feature vector The specific polymerization process is as follows:

[0236]

[0237] S4247: based on the node itself feature vector h i And the indirect neighbor aggregation feature vector Construct the second path feature vector of the node The specific construction formula is as follows:

[0238]

[0239] Where, the purpose of using MLP here is to reduce the dimension of the spliced result of W2h i And Make the dimension of the reduced vector the same as h i , Facilitating subsequent operations and calculations.

[0240] S4248: based on And Construct the fourth layer collaborative representation feature vector The construction process is as follows:

[0241]

[0242] S4249: based on Construct the microorganism aggregation linkage feature matrix And the disease aggregation linkage feature matrix Where, HM is composed of the fourth layer collaborative representation feature vectors of all microorganism nodes; HD is composed of the fourth layer collaborative representation feature vectors of all disease nodes. The specific construction formula of HM and HD is as follows:

[0243]

[0244] Where, Indicates the fourth layer collaborative representation feature vector corresponding to the microorganism node m k ; Indicates the fourth layer collaborative representation feature vector corresponding to the disease node d k .

[0245] S5: based on LM, HM, LD, HD, reconstruct the association between microorganisms and diseases to obtain the reconstruction prediction matrix A' and output the final association probability between microorganisms and diseases;

[0246] The specific steps of S5 are as follows:

[0247] S51: Based on LM, HM, LD, HD, multi-modal feature splicing is performed to construct a reconstructed microorganism feature matrix and a reconstructed disease feature matrix The specific process is as follows:

[0248] FM = concat(LM, HM)

[0249] FD = concat(LD, HD)

[0250] S52: Based on FM and FD, the association between microorganisms and diseases is reconstructed to obtain a reconstructed prediction matrix The specific process design of reconstruction is as follows:

[0251] A' = g(FM * W' * FD T )

[0252] Wherein, g() represents an activation function, and a Sigmoid function is used as the activation function. is a learned training matrix.

[0253] S53: Define a loss function to measure the difference between the reconstructed prediction matrix A' and the real association matrix A1, and train the model parameters by optimizing the loss function; the formula of the loss function is as follows:

[0254]

[0255] Wherein, represents the Frobenius norm, that is, the mean square error of the matrix elements.

[0256] S54: Based on A', the final association probability between microorganisms and diseases is output; the element A'(i,j) in A' represents the association probability between the microorganism m i and the disease d j .

[0257] A microorganism-disease association prediction device based on a multi-level perception aggregation and deep hierarchical graph hybrid learner, comprising:

[0258] A data acquisition and preprocessing module is used to automatically collect microorganism-disease association information from authoritative biomedical databases, delete redundant and ambiguous entries in the obtained association information, and then construct a microorganism-disease association matrix A1 to provide a standardized data basis for subsequent feature calculation;

[0259] The feature calculation engine is used to extract multiple similarity information and enhance the association information; first, based on the microbe-disease association matrix A1, the comprehensive similarity matrix and the representation mapping matrix of microbe and disease are constructed by integrating Gaussian kernel similarity, cosine similarity, Jaccard similarity, Sigmoid kernel similarity and functional similarity; then, the enhanced microbe-disease association matrix A2 is generated by the double-driven association enhancement strategy (DDAES), which provides high information density feature input for the subsequent intelligent analysis and learning module;

[0260] The intelligent analysis and learning module is used to realize the deep modeling of microbe features and disease features by applying the multi-level perception aggregation strategy (MLPA) and the deep hierarchical graph hybrid learner (DHGHL), and to reconstruct and predict the association between microbe and disease, so as to obtain the association probability between microbe and disease.

[0261] The interaction and output module is used to convert the complex prediction results into intuitive and usable information. As the user interface layer of the microbe-disease association prediction device, the module provides a visual interactive interface, automatically matches the information in the database and calls the reconstructed prediction results of the association between microbe and disease, and finally outputs the microbe-disease association prediction results in a visual form, which is convenient for medical professionals to quickly interpret and make decisions.

[0262] Although the present application has been described with reference to the embodiments above, various improvements can be made thereto and components thereof can be substituted with equivalents without departing from the scope of the present application. In particular, each feature disclosed in the described embodiments of the present application can be used alone or in combination with other features unless explicitly stated otherwise, and the description herein is not exhaustive of all possible combinations made of the features described herein. Therefore, the present application is not limited to the specific embodiments disclosed herein but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for predicting microbe-disease associations based on a multi-level perception aggregation and deep hierarchical graph hybrid learner, characterized in that, The specific steps include the following: S1: Obtain known microorganism-disease association information, and construct a microorganism-disease association matrix A1; meanwhile, construct a microorganism comprehensive similarity matrix CM, a microorganism representation mapping matrix MM, a disease comprehensive similarity matrix CD, and a disease representation mapping matrix DD; S2: Based on CM, MM, CD, and DD, apply a multi-level perception aggregation strategy MLPA to obtain a microorganism leap layer linkage feature matrix LM and a disease leap layer linkage feature matrix LD; S3: Based on A1, apply a double-drive association enhancement strategy DDAES to alleviate the data sparsity problem of the association matrix, and obtain an enhanced microorganism-disease association matrix A2; S4: Use A2, CM, and CD to construct an adjacency matrix B and an initial feature matrix X0 of a microorganism-disease heterogeneous graph, and use a deep hierarchical graph hybrid learner DHGHL to perform deep learning and dynamic aggregation on the features of microorganisms and diseases, to obtain a microorganism aggregated linkage feature matrix HM and a disease aggregated linkage feature matrix HD; S5: Based on LM, HM, LD, HD, the correlation between microorganisms and diseases is reconstructed and predicted to obtain a reconstructed prediction matrix A ′ And output the final correlation probability between microorganisms and diseases.

2. The method of claim 1, wherein the method is based on a multi-level perception aggregation and deep hierarchical graph hybrid learner. The specific steps of S1 are as follows: S11: Obtain known microorganism-disease association information, delete redundant and fuzzy entries in the obtained association information, and ensure the accuracy and consistency of the association information; S12: Based on the collected microbe-disease association information, let nm and nd represent the number of microbes and diseases respectively, and obtain the microbe-disease association matrix S13: Constructing a comprehensive similarity matrix of microorganisms based on A1 and a microorganism characterization mapping matrix S14: Constructing a disease comprehensive similarity matrix based on A1 and disease characterization mapping matrix 3. The method of claim 2, wherein the method is characterized by, The specific steps of S13 are as follows: S131: Based on A1, for any two microorganisms m i and m j Calculate m i and m j Gaussian kernel similarity between SM Gip (m i ,m j ); S132: For any two microorganisms m i and m j Calculate m i With m j Cosine similarity between SM Cos (m i ,m j ); S133: For any two microorganisms m i and m j Calculate m i With m j Jaccard similarity between SM Jac (m i ,m j ); S134: For any two microorganisms m i and m j , compute a Sigmoid kernel similarity SM i (m j , m Sig ) between m i and m j ; S135: Based on the protein function association, the function similarity SM between any given two microorganisms m i and m j is obtained by calculating the microorganism-microorganism function association index, i.e. Fun (m i ,m j ). S136: Constructing a comprehensive similarity matrix of microorganisms based on the Gaussian kernel similarity, cosine similarity, Jaccard similarity, Sigmoid kernel similarity and functional similarity between any two microorganisms m i and m j and microorganism characterization mapping matrix ​ The specific steps of S14 are as follows: S141: Based on A1, for any two diseases d i and d j Calculate d i With d j Gaussian kernel similarity between SD Gip (d i ,d j ); S142: For any two diseases d i and d j , compute the cosine similarity SD i between d j and d Cos (d i , d j ); S143: For any two diseases d i and d j Calculate d i With d j Jaccard similarity between SD Jac (d i ,d j ); S144: For any two diseases d i and d j , compute the Sigmoid kernel similarity SD i between d j and d Sig (d i , d j ); S145: Based on the gene interaction information, compute the functional similarity SD between any given two diseases d i and d j ; S146: Compute the similarity between any given two diseases d Fun (d i ,d j ); S146: Constructing disease comprehensive similarity matrix based on Gaussian kernel similarity, cosine similarity, Jaccard similarity, Sigmoid kernel similarity and functional similarity between any two diseases d i and d j and disease representation mapping matrix ​ 4. The method of claim 1, wherein the method is based on a multi-level perception aggregation and deep hierarchical graph hybrid learner. The specific steps of S2 are as follows: S21: Based on CM and MM, apply a multi-level perception aggregation strategy MLPA to obtain a microorganism leap layer linkage feature matrix LM; S22: Based on CD and DD, apply a multi-level perception aggregation strategy MLPA to obtain a disease leap layer linkage feature matrix LD.

5. The method of claim 4, wherein the method is characterized by, The specific steps of S21 are as follows: S211: Based on CM and MM, three levels of perception feature matrix are constructed: Microbial single-layer perception feature matrix Microbial double-layer perception feature matrix and Microbial three-layer perception feature matrix S212: based on and constructing a microorganism complex perception feature matrix and a microorganism balanced residual feature matrix S213: based on and By matrix splicing operation, the microbial jump layer linkage feature matrix is constructed The specific steps of S22 are as follows: S221 : Based on CD and DD, three levels of perception feature matrix are constructed: disease single-layer perception feature matrix disease double-layer perception feature matrix and disease three-layer perception feature matrix S222: based on and constructing a disease composite perception feature matrix and a disease balanced residual feature matrix S223: based on and A disease jump layer linkage feature matrix is constructed through a matrix splicing operation 6. The method of claim 1, wherein the method is based on a multi-level perception aggregation and deep hierarchical graph hybrid learner. The specific steps of S3 are as follows: S31: Constructing a microbial partner matrix based on A1 S32: Constructing a disease partner matrix based on A1 S33: normalizing M1, D1 to obtain a normalized microorganism partner matrix and a normalized disease partner matrix S34: Based on A1, M norm , D norm , construct the correlation potential score matrix S35: Based on A pot , for each disease d i , compute the average potential score score i of its associated microorganisms; S36: Based on A1, A pot and score i , construct an enhanced microbe-disease association matrix 7. The method of claim 1, wherein the method is based on a multi-level perception aggregation and deep hierarchical graph hybrid learner. The specific steps of S4 are as follows: S41: Construct an adjacency matrix of the microorganism-disease heterogeneous graph based on A2, CM, CD and the initial feature matrix S42: Based on B and X0, apply a deep hierarchical graph hybrid learner DHGHL to obtain a microorganism aggregated linkage feature matrix HM and a disease aggregated linkage feature matrix HD.

8. The method of claim 7, wherein the method is characterized by, The specific steps of S42 are as follows: S421: Based on B and X0, using a graph convolution network GCN to obtain a first layer collaborative representation matrix at the first layer of the DHGHL S422: based on B and H (1) At the second layer of the DHGHL, a graph convolution network GCN is used to obtain a second layer collaborative representation matrix S423: based on H (1) and H (2) In the third layer of the DHGHL, a double-layer feature dynamic fusion strategy DL-FDFS is applied to obtain a third layer collaborative representation matrix S424: based on B and H (3) In the fourth layer of the DHGHL, the enhanced dual-path graph attention network (E-DPGAT) is applied to further retain the original features and explore high-order structure information, obtaining the microbial aggregated linkage feature matrix HM and the disease aggregated linkage feature matrix HD. The specific steps of S424 are as follows: S4241 : Based on B, construct the direct neighbor node set of each node i and the indirect neighbor node set S4242: based on the set and H (3) , compute the attention coefficient between each node i and its direct neighbor node j Then normalize to get S4243: obtaining the attention coefficient according to the normalization weighting and aggregating the direct neighbor features to obtain a direct neighbor aggregated feature vector S4244: based on the node's own eigenvector h i and direct neighbor aggregated eigenvector constructing a node first path eigenvector S4245: based on the set and H (3) , compute the attention coefficient between each node i and its indirect neighbor node j Then normalize to get S4246: obtain the attention coefficient according to the normalization weighting and aggregating the indirect neighbor features to obtain an indirect neighbor aggregated feature vector S4247: based on the node's own eigenvector h i and indirect neighbor aggregated eigenvector constructing a node's second path eigenvector S4248: Based on and Constructing the fourth layer of collaborative representation feature vector for any node i S4249: based on constructing a microorganism polymeric linkage feature matrix and a disease polymeric linkage feature matrix 9. The method of claim 1, wherein the method is based on a multi-level perception aggregation and deep hierarchical graph hybrid learner. The specific steps of S5 are as follows: S51: based on LM, HM, LD, HD, multi-modal feature splicing is performed to construct a reconstructed microorganism feature matrix and a reconstructed disease feature matrix S52: Reconstruct the association between microorganism and disease based on FM and FD, and obtain a reconstructed prediction matrix S53: define a loss function for measuring the difference between the reconstructed prediction matrix A ′ the difference between the real correlation matrix A1, and train the model parameters by optimizing the loss function; S54: Based on A ′ , output the final probability of association between the microbe and the disease.

10. The method of claim 1, wherein the method of predicting microbe-disease associations based on multi-level perception aggregation and deep layered graph hybrid learner comprises a prediction device for running the method of predicting microbe-disease associations based on multi-level perception aggregation and deep layered graph hybrid learner. The prediction device comprises: A data acquisition and preprocessing module, configured to automatically acquire microorganism-disease association information from authoritative biomedical databases, delete redundant and fuzzy entries in the obtained association information, and then construct a microorganism-disease association matrix A1 to provide a standardized data basis for subsequent feature calculation; A feature calculation engine, configured to extract multiple similarity information and enhance the association information; first, based on the microorganism-disease association matrix A1, construct a comprehensive similarity matrix and a representation mapping matrix of microorganisms and diseases by integrating Gaussian kernel similarity, cosine similarity, Jaccard similarity, Sigmoid kernel similarity, and functional similarity; then, generate an enhanced microorganism-disease association matrix A2 through a double-drive association enhancement strategy to provide high information density feature input for subsequent intelligent analysis and learning modules; An intelligent analysis and learning module is configured to apply a multi-level perception aggregation strategy and a deep hierarchical graph hybrid learner to implement deep modeling of microbial features and disease features, and to reconstruct and predict the correlation between the microbes and the diseases to obtain the correlation probability between the microbes and the diseases. An interaction and output module is configured to be responsible for converting complex prediction results into intuitive and usable information.

Images