A multi-order similarity fusion learning method for predicting microbe-disease associations
By employing multi-level similarity fusion learning and label propagation methods, a high-quality microorganism-disease similarity network is constructed, which solves the problem of insufficient fusion of multi-level similarity networks in existing technologies and achieves more accurate microorganism-disease association prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGDONG UNIV OF TECH
- Filing Date
- 2023-09-07
- Publication Date
- 2026-05-29
AI Technical Summary
Existing methods for predicting microbial-disease associations have failed to effectively integrate multi-level similarity networks, resulting in limited prediction accuracy and difficulty in resisting noise.
A multi-level similarity fusion learning method is adopted. By constructing functional similarity matrices, linear neighborhood similarity matrices, and cosine similarity matrices between microorganisms and diseases, and combining them with label propagation methods, a high-quality similarity network is generated. Information is propagated from both microorganism and disease directions to improve prediction accuracy.
The generated similarity network can delve deeper into neighborhood similarity information, reduce the impact of noise, and improve the accuracy and precision of microbial-disease association prediction.
Smart Images
Figure CN117219173B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of combining machine learning with biological genes, and more specifically, to a method for predicting microbial-disease associations using multi-level similarity fusion learning. Background Technology
[0002] The human microbiome plays a crucial role in our health and disease; for example, the interaction between the gut microbiome and the immune system is considered a key factor in the pathogenesis of type 1 diabetes (T1D). Identifying disease-associated microbes is essential for disease prevention, diagnosis, and treatment, as well as advancing research in related fields. Given the exponential growth of data today, computer-based predictive methods can help medical professionals prioritize reliable disease-associated microbes for further validation. Based on previous research, methods for predicting microbe-disease associations can be broadly categorized into three types.
[0003] The first category is path-based methods. These methods typically focus on the weight scores of different paths between two nodes and treat the weights of potential paths as predicted scores for unknown associations. However, path-based methods rely on similarity networks to augment node paths and compute the weights of potential paths, and their predictive performance is limited by the reliability of the similarity network.
[0004] The second type of method is matrix factorization. This type of method continuously optimizes two latent information matrices by adding constraint terms until they approximate the input matrix, and then multiplies these two matrices to obtain the final predicted score. Matrix factorization uses similarity to approximate the constraints on the two decomposed matrices, which is beneficial for predicting unknown associations using similarity information, but it is also limited by the reliability of the similarity network. Furthermore, many matrix factorization methods do not address the problem of predicted values exceeding the [0,1] range, posing difficulties for biological interpretation.
[0005] The third category is network-based methods. These methods primarily focus on building a reliable similarity network and utilizing its topological information combined with graph methods such as random walks, network projection, and label propagation to obtain the final prediction score. Many methods rely on known associations, such as the association between microorganisms and drugs, and then construct a similarity network for microorganisms. However, these methods are limited by the richness of known information about drugs and may fail to predict associations with some new microorganisms and diseases. There are also attempts to fuse the similarity information of multiple microorganisms and diseases to construct a holistic similarity network and combine it with graph methods or machine learning algorithms to infer potential associations between microorganisms and diseases. These methods overcome the limitations of other initial data association information, which is beneficial for improving the predictive performance of new microorganism and disease associations. However, they also suffer from information bias, information loss, and noise during the fusion process, ultimately affecting the model's predictive performance.
[0006] Currently, among network-based methods for predicting the association between microorganisms and diseases, there are relatively few methods that construct similarity networks by fusing similarity information. Based on the assumption that similar microorganisms are usually associated with related diseases and vice versa, most prediction methods require the use of similarity information. Therefore, constructing a good similarity fusion method to efficiently obtain rich and reliable prior information is crucial. However, existing similarity fusion methods do not achieve multi-level fusion of similarity networks, failing to effectively fuse different similarity networks. The resulting similarity networks are not accurate enough and are not robust to noise, thus limiting the accuracy of the final prediction performance. Summary of the Invention
[0007] This invention provides a method for predicting microbial-disease associations using multi-level similarity fusion learning, which can better predict the association between microorganisms and diseases.
[0008] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:
[0009] A method for predicting microbial-disease associations using multi-level similarity fusion learning, comprising the following steps:
[0010] S1: Obtain the microbial functional similarity matrix Disease functional similarity matrix ;
[0011] S2: Obtain the microbiome-disease association matrix and utilize Calculate the linear neighborhood similarity matrix of microorganisms Similarity matrix with cosine Simultaneously, the linear neighborhood similarity matrix of the disease is calculated. Similarity matrix with cosine ;
[0012] S3: Construct a multi-level similarity fusion learning method to fuse the three types of microbial similarities and three types of disease similarities described in steps S1 and S2 respectively to generate a microbial similarity network. Disease Similarity Network ;
[0013] S4: Finally, the label propagation method is used to obtain the final prediction result.
[0014] Preferably, the microbial functional similarity matrix in step S1 Specifically:
[0015] Protein-protein functional association networks were retrieved from the STRING v11 database, and then the gene neighbor method was used to predict functional associations between proteins. Microbial functional similarity was calculated by capturing the potential for interactions between proteins encoded in the genomes of two microorganisms. ( The microbial functional similarity matrix is represented by the number of microorganisms, where... Representative microorganisms and microorganisms Similar ratings between them.
[0016] Preferably, the disease functional similarity matrix in step S1 Specifically:
[0017] Based on the assumption that similar diseases tend to interact with similar genes, functional similarity of diseases is calculated using functional associations between diseases and related genes. The newly released HumanNet v2.0 database provides efficient access to gene interactions, where each interaction has an associated log-likelihood score (LLS) to assess the probability of functional linkage between genes. (Regarding genes...) With gene set Functional relationships between As shown below:
[0018]
[0019] in, The functional similarity between genes is defined as follows:
[0020]
[0021] in Derived from LLS standardization, it is defined as follows:
[0022]
[0023] in and These represent the largest values in HumanNet. and the smallest .
[0024] For a disease and We first derived their respective gene sets. and ,in and They are respectively and The number of genes in the genome. Therefore, we can calculate the functional similarity of diseases:
[0025] .
[0026] Preferably, in step S2, the microorganism-disease association matrix is obtained. Specifically:
[0027] We obtained known microbe-disease associations from the Human Microbe-Disease Association Database (HMDAD). This included 483 experimentally validated associations between 39 diseases and 292 microbes. After removing redundant associations, 450 distinct association pairs were obtained. A microbe-disease association matrix was then used. ,in and These represent the number of microorganisms and the number of diseases, respectively; microorganism-disease association matrix. The definition is as follows:
[0028]
[0029] Preferably, in step S2, the microbial linear neighborhood similarity matrix Specifically:
[0030] Assumption Indicates microorganisms Given the feature vectors, our goal is to minimize:
[0031]
[0032] (5)
[0033] in ( )express of (A free parameter) The set of nearest neighbors (by Euclidean distance); yes The One neighbor; measure right The reconstruction contribution, i.e., the weights of a data point's neighbors, can be used as a measure of similarity. Therefore, we can obtain a linear neighborhood similarity matrix between microorganisms. .
[0034] Preferably, in step S2, the disease linear neighborhood similarity matrix The calculation method is similar to the microbial linear neighborhood similarity matrix. The calculation method is the same.
[0035] Preferably, the microbial cosine similarity matrix in step S2 Specifically:
[0036] If microorganisms and microorganisms If they are similar, then in the adjacency matrix... In the context of binary vectors and They should also be similar. Based on the known adjacency matrix... cosine similarity matrix of microorganisms It is expressed as follows:
[0037] (6)
[0038] in, Indicates microorganisms and Cosine similarity between two vectors and Represent the adjacency matrix respectively The first in row and number OK.
[0039] Preferably, in step S2, the disease cosine similarity matrix The calculation method and the microbial cosine similarity matrix The calculation method is the same.
[0040] Preferably, in step S3, a multi-level similarity fusion learning method is constructed to fuse the three similarities of microorganisms into a single similarity network. The specific steps are as follows:
[0041] Before performing multi-similarity fusion, we need to analyze the three similarity matrices of the microorganisms. , and Selective column normalization is performed to obtain three initial similarity kernels, and three neighbor constraint kernels are created. Selective row normalization is then performed on the three kernels.
[0042] The operation for selective column normalization is as follows:
[0043]
[0044]
[0045] Represents the microbial similarity matrix. Represents the number of microorganisms. Represents similarity categories, The initial kernel representing the microbial similarity after normalization satisfies .
[0046] In constructing a higher-order pathfinding algorithm and fusing more similarity subspace information, to avoid amplifying weak similarities and introducing more noise, we employ K-Nearest Neighbor (KNN) preprocessing to constrain the initial similarity, before searching for each microorganism. Reliable neighbor similarity information is incorporated into higher-order fusion to ensure the reliability of the enhanced local similarity. The selective row normalization process for constructing the neighbor constraint kernel is as follows:
[0047]
[0048]
[0049] in The neighbor constraint kernel represents the normalized microbial similarity matrix. Represents the corresponding microorganism The set of neighbors also includes itself, satisfying .
[0050] In the process of multi-level similarity fusion learning, we take microorganisms as an example. Assume the size of the microbial similarity matrix is... By constructing ( The matrix is used for low-order global learning of each network and high-order feature learning across intermediate nodes of multiple networks. Finally, the final microbial similarity fusion matrix is obtained through the following formula. :
[0051]
[0052] First, there is low-order global learning without intermediate nodes. This involves learning the common latent features of all similarity networks. The learning scope covers the entire similarity network, which can be achieved by minimizing the average loss, as shown below:
[0053]
[0054] To learn higher-order similarity features, enable interaction between networks, and further explore the structural information of different network subspaces, we rely on neighboring nodes of similar networks as intermediate nodes to bridge the gap between two networks. This approach enables cross-node higher-order feature learning. The specific formula is shown below:
[0055]
[0056] in for The transpose of the matrix. However, relying solely on a single neighbor node is insufficient to describe the complex structure of the entire network when merging multiple similar networks. To increase the number of node paths and achieve deeper interaction between different network subspaces, we use two neighbor nodes to graft the networks together, which better restores the complex structure of the similarity network. The specific formula is as follows:
[0057]
[0058]
[0059] in for The mean matrix after KNN processing. Finally, combining the above formulas, we propose a new formula as shown below, and minimize it using an optimization algorithm. Finally, a fused similarity network adjacency matrix is generated. :
[0060]
[0061]
[0062] Utilizing the decomposability of the Scaled Form ADMM algorithm, we can minimize The problem is solved by sub-problems and average updates. For example, in solving the nth sub-problem... The augmented Lagrangian function for the subproblem is:
[0063]
[0064] in, It is a penalty parameter. Lagrange multipliers. By definition The above formula can be simplified to the following expression:
[0065]
[0066]
[0067] The following update steps (update at t+1) are then used until convergence:
[0068]
[0069]
[0070] Using the gradient descent algorithm to optimize and solve the above formula, we can first reconstruct it as follows:
[0071]
[0072]
[0073]
[0074]
[0075] in The trace of the matrix is given by the formula above. Taking the derivative and setting it to zero yields the following result: The iterative update formula is:
[0076]
[0077] in It is a matrix consisting entirely of 1s. Therefore, we can update each... Until convergence, ultimately by taking all The average value is used as the adjacency matrix of the fused disease similarity network, as shown below:
[0078]
[0079] To ensure that the similarity data is scaled to the interval [0,1] and that the constructed similarity network is an undirected graph, we optimize the converged adjacency matrix. Perform normalization and symmetry operations as follows:
[0080]
[0081]
[0082] From this, we can obtain the final fused adjacency matrix of the microbial similarity network. A fused similarity network is constructed from this adjacency matrix. .
[0083] Preferably, the disease similarity network in step S3 The computational method and microbial similarity network The calculation method is the same.
[0084] Preferably, the label propagation method is used in step S4 to obtain the final prediction result; the specific process of propagation from the microbial direction is as follows: the initial known microbial-disease associations are regarded as labels and propagated through the similarity network constructed above. In the microbial similarity network, the label of node i in each iteration is probabilistically... From similarity networks It obtains information from the neighborhood and maintains its initial label information. The speed is updated until convergence (the predicted label of the node does not change in several consecutive iterations).
[0085]
[0086] This indicates that similarity networks are accessed through tags. The association prediction score matrix of microorganisms and diseases obtained during the spread of the disease For the initial incidence matrix Ultimately, a microbiome-disease association prediction score matrix can be obtained. As shown below:
[0087]
[0088] The predicted rating matrix output after normalization is more accurate, so we further refine the propagated rating matrix. Normalization is performed:
[0089]
[0090] Preferably, the scoring matrix obtained in step S4 from the direction of disease transmission The calculation method and the scoring matrix obtained from the propagation from the microbial direction The calculation method is the same.
[0091] In step S4, the final comprehensive score matrix is obtained by propagating the linear neighborhood label propagation method across the microbial and disease networks. for:
[0092] .
[0093] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:
[0094] This invention constructs a similarity fusion method to mine deeper neighborhood similarity information and generate a high-quality, low-noise similarity network. Combined with the linear neighborhood label propagation method, it can effectively utilize the constructed similarity network. Furthermore, propagation from both microorganism and disease directions can make better use of existing similarity information, resulting in more accurate predictions of microorganism-disease associations. Attached Figure Description
[0095] Figure 1 This is a flowchart of the method of the present invention.
[0096] Figure 2 The diagram below illustrates the AUCs of MOSFL-LNP, MNNMDA, NTSHMDA, LRLSHMDA, KATZHMDA, BiRWHMDA, and BRWMDA implemented based on LOOCV in this embodiment.
[0097] Figure 3 The diagram illustrates AUCs implemented in this embodiment based on 5-fold-CV using MOSFL-LNP, MNNMDA, NTSHDDA, LRLSHMDA, KATZHMDA, BiRWHMDA, and BRWMDA. Detailed Implementation
[0098] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent.
[0099] To better illustrate this embodiment, some parts in the accompanying drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions;
[0100] It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.
[0101] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.
[0102] Example 1
[0103] This embodiment provides a method for predicting microbial-disease associations using multi-level similarity fusion learning, such as... Figure 1 As shown, it includes the following steps:
[0104] S1: Obtain the microbial functional similarity matrix Disease functional similarity matrix ;
[0105] S2: Obtain the microbiome-disease association matrix and utilize Calculate the linear neighborhood similarity matrix of microorganisms Similarity matrix with cosine Simultaneously, the linear neighborhood similarity matrix of the disease is calculated. Similarity matrix with cosine ;
[0106] S3: Construct a multi-level similarity fusion learning method to fuse the three types of microbial similarities and three types of disease similarities described in steps S1 and S2 respectively to generate a microbial similarity network. Disease Similarity Network ;
[0107] S4: Finally, the label propagation method is used to obtain the final prediction result.
[0108] Microbial functional similarity matrix in step S1 Specifically:
[0109] Protein-protein functional association networks were retrieved from the STRING v11 database, and then the gene neighbor method was used to predict functional associations between proteins. Microbial functional similarity was calculated by capturing the potential for interactions between proteins encoded in the genomes of two microorganisms. ( The microbial functional similarity matrix is represented by the number of microorganisms, where... Representative microorganisms and microorganisms Similar ratings between them.
[0110] Disease functional similarity matrix in step S1 Specifically:
[0111] Based on the assumption that similar diseases tend to interact with similar genes, functional similarity of diseases is calculated using functional associations between diseases and related genes. The newly released HumanNet v2.0 database provides efficient access to gene interactions, where each interaction has an associated log-likelihood score (LLS) to assess the probability of functional linkage between genes. (Regarding genes...) With gene set Functional relationships between As shown below:
[0112]
[0113] in, The functional similarity between genes is defined as follows:
[0114]
[0115] in Derived from LLS standardization, it is defined as follows:
[0116]
[0117] in and These represent the largest values in HumanNet. and the smallest .
[0118] For a disease and We first derived their respective gene sets. and ,in and They are respectively and The number of genes in the genome. Therefore, we can calculate the functional similarity of diseases:
[0119] .
[0120] In step S2, the microbiome-disease association matrix is obtained. Specifically:
[0121] We obtained known microbe-disease associations from the Human Microbe-Disease Association Database (HMDAD). This included 483 experimentally validated associations between 39 diseases and 292 microbes. After removing redundant associations, 450 distinct association pairs were obtained. A microbe-disease association matrix was then used. ,in and These represent the number of microorganisms and the number of diseases, respectively; microorganism-disease association matrix. The definition is as follows:
[0122]
[0123] Microbial linear neighborhood similarity matrix in step S2 Specifically:
[0124] Assumption Indicates microorganisms Given the feature vectors, our goal is to minimize:
[0125]
[0126] (5)
[0127] in ( )express of (A free parameter) The set of nearest neighbors (by Euclidean distance); yes The One neighbor; measure right The reconstruction contribution, i.e., the weights of a data point's neighbors, can be used as a measure of similarity. Therefore, we can obtain the linear neighborhood similarity matrix between microorganisms. .
[0128] Disease linear neighborhood similarity matrix in step S2 The calculation method is similar to the microbial linear neighborhood similarity matrix. The calculation method is the same.
[0129] Microbial cosine similarity matrix in step S2 Specifically:
[0130] If microorganisms and microorganisms If they are similar, then in the adjacency matrix... In the context of binary vectors and They should also be similar. Based on the known adjacency matrix... cosine similarity matrix of microorganisms It is expressed as follows:
[0131] (6)
[0132] in, Indicates microorganisms and Cosine similarity between two vectors and Represent the adjacency matrix respectively The first in row and number OK.
[0133] Disease cosine similarity matrix in step S2 The calculation method and the microbial cosine similarity matrix The calculation method is the same.
[0134] Step S3 involves constructing a multi-level similarity fusion learning method to fuse the three similarities of microorganisms into a single similarity network. The specific steps are as follows:
[0135] Before performing multi-similarity fusion, we need to analyze the three similarity matrices of the microorganisms. , and Selective column normalization is performed to obtain three initial similarity kernels, and three neighbor constraint kernels are created. Selective row normalization is then performed on the three kernels.
[0136] The operation for selective column normalization is as follows:
[0137]
[0138]
[0139] Represents the microbial similarity matrix. Represents the number of microorganisms. Represents similarity categories, The initial kernel representing the normalized microbial similarity of the column satisfies .
[0140] In constructing a higher-order pathfinding algorithm and fusing more similarity subspace information, to avoid amplifying weak similarities and introducing more noise, we employ K-Nearest Neighbor (KNN) preprocessing to constrain the initial similarity, before searching for each microorganism. Reliable neighbor similarity information is incorporated into higher-order fusion to ensure the reliability of the enhanced local similarity. The selective row normalization process for constructing the neighbor constraint kernel is as follows:
[0141]
[0142]
[0143] in The neighbor constraint kernel represents the normalized microbial similarity matrix. Represents the corresponding microorganism The set of neighbors also includes itself, satisfying .
[0144] In the process of multi-level similarity fusion learning, we take microorganisms as an example. Assume the size of the microbial similarity matrix is... By constructing ( The matrix is used for low-order global learning of each network and high-order feature learning across intermediate nodes of multiple networks. Finally, the final microbial similarity fusion matrix is obtained through the following formula. :
[0145]
[0146] First, there is low-order global learning without intermediate nodes. This involves learning the common latent features of all similarity networks. The learning scope covers the entire similarity network, which can be achieved by minimizing the average loss, as shown below:
[0147]
[0148] To learn higher-order similarity features, enable interaction between networks, and further explore the structural information of different network subspaces, we rely on neighboring nodes of similar networks as intermediate nodes to bridge the gap between two networks. This approach enables cross-node higher-order feature learning. The specific formula is shown below:
[0149]
[0150] in for The transpose of the matrix. However, relying solely on a single neighbor node is insufficient to describe the complex structure of the entire network when merging multiple similar networks. To increase the number of node paths and achieve deeper interaction between different network subspaces, we use two neighbor nodes to graft the networks together, which better restores the complex structure of the similarity network. The specific formula is as follows:
[0151]
[0152]
[0153] in for The mean matrix after KNN processing. Finally, combining the above formulas, we propose a new formula as shown below, and minimize it using an optimization algorithm. Finally, a fused similarity network adjacency matrix is generated. :
[0154]
[0155]
[0156] Utilizing the decomposability of the Scaled Form ADMM algorithm, we can minimize The problem is solved by sub-problems and average updates. For example, in solving the nth sub-problem... The augmented Lagrangian function for the subproblem is:
[0157]
[0158] in, It is a penalty parameter. Lagrange multipliers. By definition The above formula can be simplified to the following expression:
[0159]
[0160]
[0161] The following update steps (update at t+1) are then used until convergence:
[0162]
[0163]
[0164] Using the gradient descent algorithm to optimize and solve the above formula, we can first reconstruct it as follows:
[0165]
[0166]
[0167]
[0168]
[0169] in The trace of the matrix is given by the formula above. Taking the derivative and setting it to zero yields the following result: The iterative update formula is:
[0170]
[0171] in It is a matrix consisting entirely of 1s. Therefore, we can update each... Until convergence, ultimately by taking all The average value is used as the adjacency matrix of the fused disease similarity network, as shown below:
[0172]
[0173] To ensure that the similarity data is scaled to the interval [0,1] and that the constructed similarity network is an undirected graph, we optimize the converged adjacency matrix. Perform normalization and symmetry operations as follows:
[0174]
[0175]
[0176] From this, we can obtain the final fused adjacency matrix of the microbial similarity network. A fused similarity network is constructed from this adjacency matrix. .
[0177] Disease similarity network in step S3 The computational method and microbial similarity network The calculation method is the same.
[0178] Step S4 employs a label propagation method to obtain the final prediction result. The specific process of propagation from the microbial direction is as follows: Initial known microbial-disease associations are treated as labels and propagated through the similarity network constructed above. In the microbial similarity network, the label of node i in each iteration is determined by probability. From similarity networks It obtains information from the neighborhood and maintains its initial label information. The speed is updated until convergence (the predicted label of the node does not change in several consecutive iterations).
[0179]
[0180] This indicates that similarity networks are accessed through tags. The association prediction score matrix of microorganisms and diseases obtained during the spread of the disease For the initial incidence matrix Ultimately, a microbiome-disease association prediction score matrix can be obtained. As shown below:
[0181]
[0182] The predicted rating matrix output after normalization is more accurate, so we further refine the propagated rating matrix. Normalization is performed:
[0183]
[0184] The scoring matrix obtained from the spread of the disease in step S4 The calculation method and the scoring matrix obtained from the propagation from the microbial direction The calculation method is the same.
[0185] In step S4, the final comprehensive score matrix is obtained by propagating the linear neighborhood label propagation method across the microbial and disease networks. for:
[0186]
[0187] To verify the prediction accuracy of the method (MOSFL-LNP) in this embodiment, MOSFL-LNP is compared with six advanced methods: MNNMDA, NTSHDDA, LRLSHMDA, KATZHMDA, BiRWHMDA, and BRWMDA. Figure 2 As can be seen, under the LOOCV framework, the area under the AUC of the MOSFL-LNP curve is 0.9718, which is larger than other calculation methods (MNNMDA 0.9701, NTSHANMDA 0.907, LRLSHMDA 0.9056, KATZHMDA 0.88, BiRWHMDA 0.8952, BRWMDA 0.9379), indicating that MOSFL-LNP outperforms other calculation methods. To further verify the predictive performance of MOSFL-LNP, a 5x CV framework was used for validation. From... Figure 3 As can be seen, MOSFL-LNP achieves a reliable AUC of 0.9701±0.0028, significantly exceeding the AUC values of 0.9534±0.0027, 0.8843±0.0042, 0.8708±0.0029, 0.8357±0.0033, 0.8769±0.0032, and 0.9037±0.004. This indicates that MOSFL-LNP is more effective than other methods within the 5-fold CV framework. These results fully demonstrate that the MOSFL-LNP method is superior to comparative methods and is more conducive to predicting the association between microorganisms and diseases.
[0188] We also used MOSFL-LNP to predict the application of known microorganisms in practical microbe-disease prediction. In predicting new microbe-disease associations, we used known microbe-disease associations as the training dataset for MOSFL-LNP, then calculated and ranked the prediction scores for each unknown microbe-disease pair. We selected type 1 diabetes and asthma as case studies. The top 10 microorganisms were validated in the PubMed database. The results, shown in Tables 1 and 2, indicate that 100% and 90% of the predicted microorganisms, respectively, were associated with the diseases.
[0189] Furthermore, MOSFL-LNP predicts some unproven microbial-disease associations, including firmicutes and asthma. These predicted associations have not yet been reported in the current literature, but their existence is highly probable and awaits further investigation and validation by medical researchers.
[0190] Table 1. Top 10 Potential Microorganisms Predicted by MOSFL-LNP to be Associated with Type 1 Diabetes
[0191]
[0192] Table 2. Top 10 Potential Microbes Associated with Asthma Predicted by MOSFL-LNP
[0193]
[0194] The same or similar labels correspond to the same or similar parts;
[0195] The terms used to describe positional relationships in the accompanying drawings are for illustrative purposes only and should not be construed as limiting this patent.
[0196] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.
Claims
1. A method for predicting microbial-disease associations using multi-level similarity fusion learning, characterized in that, Includes the following steps: S1: Obtain the microbial functional similarity matrix Disease functional similarity matrix ; S2: Obtain the microbiome-disease association matrix and utilize Calculate the linear neighborhood similarity matrix of microorganisms Similarity matrix with cosine Simultaneously, the linear neighborhood similarity matrix of the disease is calculated. Similarity matrix with cosine ; S3: Construct a multi-level similarity fusion learning method to fuse the three microbial similarity matrices and the three disease similarity matrices from steps S1 and S2 respectively to generate a microbial similarity network. Disease Similarity Network ; S4: Finally, the label propagation method is used to obtain the final prediction result; Step S3 involves constructing a multi-level similarity fusion learning method to fuse the three similarities of microorganisms into a single similarity network. The specific steps are as follows: Before performing multi-similarity fusion, it is necessary to analyze the three similarity matrices of the microorganisms. , and Selective column normalization is performed to obtain three initial similarity kernels, and three neighbor constraint kernels are created. Selective row normalization is then performed on the three kernels. In constructing a higher-order pathfinding algorithm and fusing more similarity subspace information, to avoid amplifying weak similarities and introducing more noise, K-Nearest Neighbor preprocessing is used to constrain the initial similarity. This preprocessing is performed before finding each microorganism. Reliable neighbor similarity information is incorporated into higher-order fusion to ensure that the enhanced local similarity is reliable; In the process of multi-order similarity fusion learning, it is assumed that the size of the microbial similarity matrix is... By constructing The matrix is used for low-order global learning of each network and high-order feature learning across intermediate nodes of multiple networks. Finally, the final microbial similarity fusion matrix is obtained through the following formula. : First, there is low-order global learning without intermediate nodes. This involves learning the common latent features of all similarity networks. The learning scope covers the entire similarity network, and this is achieved by minimizing the average loss, as expressed in the following formula: Secondly, there is the learning of higher-order features across intermediate nodes in multiple networks, which enables interaction between networks and further explores the structural information of different network subspaces; relying on neighboring nodes of similar networks as intermediate nodes to achieve the grafting between two networks, this method is used to realize higher-order feature learning across nodes. The specific formula is shown below: in for The matrix transpose; in order to increase more node paths and realize the interaction between different network subspaces, the network is grafted by relying on two neighboring nodes to restore the complex structure of similar networks.
2. The method for predicting microbial-disease associations based on multi-level similarity fusion learning according to claim 1, characterized in that, Microbial functional similarity matrix in step S1 Specifically: Protein-protein functional association networks were retrieved from the STRING v11 database, and then the gene neighbor method was used to predict the functional associations between proteins. The functional similarity of the microorganisms was calculated by capturing the potential for interactions between proteins encoded in the genomes of two different microorganisms. To represent the microbial functional similarity matrix Represents the number of microorganisms, among which Representative microorganisms and microorganisms Similar ratings between them.
3. The method for predicting microbial-disease associations based on multi-level similarity fusion learning according to claim 1, characterized in that, In step S2, the microbiome-disease association matrix is obtained. Specifically: Known microbe-disease associations were obtained from the Human Microbe-Disease Association Database (HMDAD); this included 483 experimentally validated associations between 39 diseases and 292 microbes; after removing redundant associations, 450 distinct association pairs were obtained. This represents the microbiome-disease association matrix, where and These represent the number of microorganisms and the number of diseases, respectively; microorganism-disease association matrix. The definition is as follows: 。