A machine learning based multi-omics data hierarchical classification structure learning system

Abstract

The application discloses a kind of multi-omics data hierarchical classification structure learning systems based on machine learning, comprising: data import module, for loading multi-omics data, and the data is preprocessed;Hierarchical structure learning module, using the adaptive machine learning method driven by data carries out supervised learning, constructs category similarity matrix, uses bottom-up, top-down unsupervised clustering algorithm to preliminarily construct class label hierarchy, and integrates iterative algorithm into training process, finally obtains optimal class label hierarchy;Hierarchical classification verification module, using optimal class label hierarchy combines multi-omics data to carry out hierarchical classification, and provide result explanation.The application combines supervised, unsupervised machine learning and iterative algorithm, accurately infers the class label hierarchy in multi-omics data, improves the efficient classification and organization of complex multi-omics data, and provides classification result interpretability.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and biomedicine, and in particular to a machine learning-based multi-omics data hierarchical classification structure learning system. Background Technology

[0002] With the development of bioinformatics, multi-omics technologies have generated massive amounts of data, which are still accumulating rapidly. However, their significant value has only been initially utilized. In-depth analysis of various biological systems, especially achieving accurate classification, is a major challenge in multi-omics data research. However, biological systems generally possess unknown and complex hierarchical structures due to evolution, development, and differentiation. This makes it difficult for commonly used non-hierarchical and pre-defined hierarchical classification models to accurately reflect their internal structure and achieve precise sample classification. Therefore, how to infer the class label hierarchy based on multi-omics data for hierarchical classification is an urgent issue to be addressed.

[0003] Currently, hierarchical classification structure learning systems mainly fall into two categories: 1) Overall optimization of sample classification accuracy and hierarchical structure: This involves modeling and predicting data using methods such as combining multiple one-to-many classifiers, constructing a confusion matrix on the validation set, and normalizing the balance of the hierarchical structure using loss functions. 2) Optimization of class similarity metrics and class clustering methods: This involves modeling and predicting data using methods such as balanced class partitioning, K-means clustering based on similarity in the original or high-dimensional space, and nearest neighbor propagation. On the one hand, existing hierarchical structure learning systems have been proposed in image and text classification, but their development in multi-omics fields remains limited. This means that current systems cannot fully adapt to the unique characteristics of high throughput, high dimensionality, and high sparsity in multi-omics data, resulting in insufficient accuracy and reliability of the learned hierarchical classification structure. On the other hand, most current systems have low interpretability and cannot provide evidence to support the decisions made by the model. This means that current systems struggle to infer biologically meaningful class label hierarchies, thus failing to reveal biological evolutionary patterns and apply them to important tasks in biomedical fields such as accurate subtype discrimination and identification of potential targets. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings and deficiencies of existing technologies and propose a machine learning-based hierarchical classification structure learning system for multi-omics data. This system breaks through the limitations of traditional multi-omics classification methods in choosing classification structures, fully adapts to the unique properties of high throughput, high dimensionality, and high sparsity of multi-omics data, solves the problem of class label hierarchical structure inference in multi-omics data, thereby improving the efficient classification and organization of complex multi-omics data and providing interpretability of diagnostic results.

[0005] To achieve the above objectives, the technical solution provided by this invention is: a multi-omics data hierarchical classification structure learning system based on machine learning, comprising:

[0006] The data import module is used to load multi-omics data and preprocess the data to obtain a dataset that can be used to build machine learning models.

[0007] The hierarchical structure learning module uses data-driven adaptive machine learning methods for supervised learning, constructs a category similarity matrix, uses bottom-up and top-down unsupervised clustering algorithms to initially construct a class label hierarchy, and integrates iterative algorithms into the training process to finally obtain the optimal class label hierarchy.

[0008] The hierarchical classification verification module uses the optimal class label hierarchy structure combined with multi-omics data to perform hierarchical classification, and compares and analyzes it with non-optimal class label hierarchy structures, providing an interpretation of the results.

[0009] Furthermore, the data import module includes a data loading module and a data preprocessing module, wherein:

[0010] The data loading module reads multi-omics data and clinical data from the local machine, including CSV, RDF, h5ad, RData and xlsx formats. The multi-omics data includes genomics, epigenomics, transcriptomics and proteomics, and the data sources include complex disease genomes, microbial K-strings and single-cell sequencing. The clinical data includes survival status, survival time and relapse time.

[0011] The data preprocessing module is used to extract and organize different omics data, including: extracting sample-shared features for complex disease omics datasets, normalizing gene mutation and copy number variation data, removing features with more than 50% methylation gaps, filling gaps with K-nearest neighbors, normalizing methylation data, and integrating gene mutation, copy number variation, and methylation data; counting K-strings in FastQ files for microbiome datasets and retaining high-frequency K-strings with more than 10 occurrences; integrating data for single-cell omics datasets and using the Lasso method for feature selection; dividing each dataset into a training set X, a test set T, and a validation set V; and using a uniform format for all datasets, with rows representing samples and columns representing types and features.

[0012] Furthermore, the hierarchical learning module includes a category similarity module, a clustering module, and an iteration module, wherein:

[0013] The category similarity module constructs a category similarity matrix using an omics-driven adaptive machine learning method, including the following steps:

[0014] 1) For omics data containing K categories, construct a one-vs-Rest classification model f1, f2, ..., f based on its adaptive machine learning method. K , where f K For the Kth classification model, we obtain K prediction probability matrices:

[0015]

[0016] In the formula, T k Let x be the test set corresponding to the k-th class in the classification model. i For the samples in the test set, f k (x i P represents the classification model corresponding to the k-th category. k For the predicted probability matrix of the test set samples corresponding to the k-th category, its first column P k (x i ,1) represents sample x under the classification model i The probability of being predicted as the k-th class, in the second column P. k (x i ,2) represents the sample x under the classification model. i The probability of being predicted as a class other than the k-th class;

[0017] 2) Based on the prediction probability matrix P1, P2, ..., P K Construct a category confusion matrix and then transform it into a category similarity matrix:

[0018] ① First, construct a class confusion matrix C, where a represents the row number and b represents the column number. Each element in the matrix C... ab Represented as:

[0019] ② Standardize the class confusion matrix, that is, for the a-th row C of the class confusion matrix C... a All of them have:

[0020] ③ Further symmetrization of matrix A yields the category similarity matrix:

[0021] In the formula, T is the test set selected in the classification model, (x i ,y i Let P be the sample in T and its corresponding true class label. k (x i ,1) is the sample x under the classification model i The probability of being predicted as the k-th category, where A is the category similarity matrix. a Let A' be the a-th row of the matrix;

[0022] The clustering module is used to initially construct a hierarchical structure of class labels, including a bottom-up clustering module and a top-down clustering module. It constructs a clustering model based on the category similarity matrix to achieve the construction of the hierarchical structure of class labels, wherein:

[0023] The bottom-up clustering module includes the following steps:

[0024] 1) Initialize the hierarchical structure to be constructed as follows: Where H represents the hierarchical structure, N represents the node set, and E represents the edge set; initialize the category label list B = ({1},{2},...,{K}), and initialize the number of elements in the category label list B to m = K;

[0025] 2) When m < 2K, repeat the following steps:

[0026] ① Find the two most similar categories p and q from the category similarity matrix A, and merge them:

[0027] In the formula, A ab This represents the element in row a and column b of A, where n is the number of elements in A. m ={1,2,...,m} is the set of active nodes, which changes as m changes;

[0028] ② Add a new node to the node set N:

[0029] N = N∪{k+1}

[0030] ③ Merge the two most similar categories p and q into a new category and update the active node set n. k+1 :

[0031] n k+1 =(n k \{p,q})∪{k+1}

[0032] ④ Update category tag list B and add the newly created class:

[0033] B = (B, B[p] ∪ B[q]), where B[p] and B[q] represent the p-th and q-th elements in the category label list B, respectively;

[0034] ⑤ Add new edges to edge set E:

[0035] E=E∪(p,k+1)∪(q,k+1)

[0036] ⑥ Update the category confusion matrix among active nodes:

[0037] In the formula, A k+1,k A pk Aqk This represents the element in row (k+1)th column (k), row (p)th column (k), and row (q)th column (k) of A;

[0038] ⑦ Update m = m + 1, and continue to execute steps ①-⑥ above, moving upwards layer by layer until all classes are merged into one category;

[0039] 3) Output the hierarchical structure H(N,E) and the list of category labels B;

[0040] The top-down clustering module includes the following steps:

[0041] 1) Initialize the hierarchical structure Where r represents the root node, N represents the node set, and E represents the edge set; initialize the category label set L. r = {1,2,...,K}, training set X and test set T, and push the root node r into the empty stack S;

[0042] 2) When the stack is not empty, repeat the following steps:

[0043] ① Pop the last pushed node t;

[0044] ②When |L t When | > 1, train |L t One-vs-Rest classifier Each category has its own binary classifier, where... For the |L t A classification model is used, and the category confusion matrix C and category similarity matrix A are calculated:

[0045]

[0046]

[0047]

[0048] In the formula, L t f represents the set of category labels for the corresponding node t. k (x i ) represents sample x i The type predicted by the k-th classification model;

[0049] ③ Construct a diagonal matrix D, and then construct the Laplace matrix L using the diagonal matrix:

[0050] D[i,i]=sum(A[i,:])

[0051] L=DA

[0052] In the formula, D[i,i] represents the diagonal element of the i-th row and i-th column, and A[i,:] represents all elements of the i-th row;

[0053] ④ Use the elbow method to find the optimal number of clusters k';

[0054] ⑤ Perform spectral clustering, decompose the Laplacian matrix L into eigenvalues ​​to obtain eigenvalues ​​x' and eigenvectors λ, sort the eigenvalues ​​x in ascending order, select the k'-1 smallest non-zero eigenvalues ​​x", extract the corresponding eigenvectors λ', and form a matrix Z with λ' as the column. Normalize the rows of Z and use them as the k'-dimensional embedding of the class.

[0055] ⑥ Perform K-means clustering, setting the number of clusters to k', and apply it to the rows of the embedding matrix Z to obtain the corresponding cluster labels;

[0056] ⑦ Based on the clustering labels obtained in step ⑥, partition the set of child nodes N of node t. t Create or update category tag sets {L r :c∈N t}, where c represents N t The child nodes in the array satisfy the condition that they do not overlap and

[0057] ⑧ Update edge set E:

[0058] E = E∪{(t,c):c∈N} t}

[0059] ⑨ Update node set N:

[0060] N = N∪N t

[0061] ⑩ Set the child node set N t Push all nodes onto stack S;

[0062] 3) Output the hierarchical structure H(N,E) and the classification label {L} corresponding to each node. e :e∈N}, where e represents a node in the node set;

[0063] The iterative module retrains the model based on the initial class label hierarchy to achieve iterative updates, including the following steps:

[0064] 1) Input the training set X with class labels, the number of classes K, and the initial hierarchical structure H0 obtained by the bottom-up clustering module or the top-down clustering module; initialize H = H0; initialize the number of iterations d = 1;

[0065] 2) Repeat the following steps until the hierarchy H is reached. d =H d-1 Or d > 100:

[0066] ① A classification model is constructed using a hierarchical structure H, resulting in a new class confusion matrix C. d :C d = f(X,H), where f represents the classifier used;

[0067] ②Standardization:

[0068] ③Symmetricalization:

[0069] ④ By using a bottom-up or top-down clustering module, update the hierarchical structure based on the new category similarity matrix to obtain a new hierarchical structure H. d ;

[0070] ⑤ Update H = H d d = d + 1, return to step ①;

[0071] 3) Output the optimal class label hierarchy H d .

[0072] Furthermore, the hierarchical classification verification module includes the following steps:

[0073] 1) Use the data loading module in the data import module to load multi-omics data, use the data preprocessing module to process the multi-omics data, and obtain the validation set V;

[0074] 2) Input the validation set V processed in step 1) into the optimal class label hierarchy learned by the hierarchical structure learning module for hierarchical classification;

[0075] 3) Extract the classification results from step 2) above, including accuracy, precision, recall, AUC value and F1 value, as the precision measure O1 under the current prediction result;

[0076] 4) Input the validation set V processed in step 1) into the non-optimal class label hierarchy for hierarchical classification, and compare their accuracy measures O2 and O1 to give interpretability to the classification effect.

[0077] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0078] 1. Solve the problem of class label hierarchical structure inference in multi-omics data by using machine learning methods, fill the gap in building hierarchical classification structure learning systems in the field of multi-omics, overcome the limitations of traditional classification methods, and improve classification accuracy and efficiency.

[0079] 2. Imparting a certain degree of interpretability to the classification results of multi-omics data and exploring the deeper biological relationships behind the data helps to further verify the classification results. Attached Figure Description

[0080] Figure 1 This is a schematic diagram showing the relationship between the various modules of the system of the present invention.

[0081] Figure 2 This is a flowchart of the system training and prediction process of the present invention. Detailed Implementation

[0082] The present invention will be further described below with reference to specific embodiments.

[0083] This embodiment discloses a multi-omics data hierarchical classification structure learning system based on machine learning. It is a multi-omics data hierarchical classification structure learning system developed using the R language and capable of running on Windows devices. The relationships between the system modules are as follows: Figure 1 As shown, the system training and prediction process is as follows: Figure 2 As shown. It includes:

[0084] The data import module is used to load multi-omics data and preprocess the data to obtain a dataset that can be used to build machine learning models.

[0085] The hierarchical structure learning module uses data-driven adaptive machine learning methods for supervised learning, constructs a category similarity matrix, uses bottom-up and top-down unsupervised clustering algorithms to initially construct a class label hierarchy, and integrates iterative algorithms into the training process to finally obtain the optimal class label hierarchy.

[0086] The hierarchical classification verification module uses the optimal class label hierarchy structure combined with multi-omics data to perform hierarchical classification, and compares and analyzes it with non-optimal class label hierarchy structures, providing an interpretation of the results.

[0087] Specifically, the data import module includes a data loading module and a data preprocessing module, wherein:

[0088] The data loading module reads breast cancer data in formats such as CSV, RDF, h5ad, RData, and xlsx from the local machine. This includes multi-omics data and clinical data. The multi-omics data of breast cancer includes genomics and epigenomics, while the clinical data of breast cancer includes survival status, survival time, and recurrence time.

[0089] The data preprocessing module is used to extract multi-omics shared samples in breast cancer, extract sample-shared features for breast cancer data, normalize gene mutation and copy number variation data, remove features with more than 50% of methylation gaps, fill gaps with K-nearest neighbors, normalize methylation data, integrate gene mutation, copy number variation and methylation data, and divide them into training set X, test set T and validation set V.

[0090] Specifically, the hierarchical learning module includes a category similarity module, a clustering module, and an iteration module, wherein:

[0091] The category similarity module constructs a breast cancer category similarity matrix using an omics-driven adaptive machine learning method, including the following steps:

[0092] 1) For omics data containing K categories, construct a one-vs-Rest classification model f1, f2, ..., f based on the Lasso-Logistic machine learning method. K , where f K For the Kth classification model, we obtain K prediction probability matrices:

[0093]

[0094] In the formula, T k Let x be the test set corresponding to the k-th class in the classification model. i For the samples in the test set, f k (x i P represents the classification model corresponding to the k-th category. k For the predicted probability matrix of the test set samples corresponding to the k-th category, its first column P k (x i ,1) represents sample x under the classification model i The probability of being predicted as the k-th class, in the second column P. k (x i ,2) represents the sample x under the classification model. i The probability of being predicted as a class other than the k-th class;

[0095] 2) Based on the prediction probability matrix P1, P2, ..., P K Construct a breast cancer category confusion matrix and transform it into a breast cancer category similarity matrix:

[0096] ① First, construct a class confusion matrix C, where a represents the row number and b represents the column number. Each element in the matrix C... ab Represented as:

[0097] ② Standardize the class confusion matrix, that is, for the a-th row C of the class confusion matrix C...a All of them have:

[0098] ③ Further symmetrization of matrix A yields the category similarity matrix:

[0099] In the formula, T is the test set selected in the classification model, (x i ,y i Let P be the sample in T and its corresponding true class label. k (x i ,1) is the sample x under the classification model i The probability of being predicted as the k-th category, where A is the breast cancer category similarity matrix, A a Let A' be the a-th row of the matrix;

[0100] The clustering module is used to initially construct a hierarchical structure of class labels among breast cancer categories. It includes a bottom-up clustering module and a top-down clustering module, and constructs a clustering model based on a category similarity matrix to achieve the construction of the breast cancer class label hierarchy.

[0101] The bottom-up clustering module includes the following steps:

[0102] 1) Initialize the hierarchical structure to be constructed as follows: Where H represents the hierarchical structure, N represents the node set, and E represents the edge set; initialize the category label list B = ({1},{2},...,{K}), and initialize the number of elements in the category label list B to m = K;

[0103] 2) When m < 2K, repeat the following steps:

[0104] ① Find the two most similar categories p and q in the breast cancer category similarity matrix A, and merge them: In the formula A ab This represents the element in row a and column b of A, where n is the number of elements in A. m ={1,2,...,m} is the set of active nodes, which changes as m changes;

[0105] ② Add a new node to the node set N:

[0106] N = N∪{i+1}

[0107] ③ Merge the two most similar categories p and q into a new category and update the active node set:

[0108] n k+1 =(n k \{p,q})∪{k+1}

[0109] ④ Update category tag list B and add the newly created class:

[0110] B = (B, B[p] ∪ B[q]), where B[p] and B[q] represent the p-th and q-th elements in the category label list B, respectively;

[0111] ⑤ Add new edges to edge set E:

[0112] E=E∪(p,k+1)∪(q,k+1)

[0113] ⑥ Update the category confusion matrix among active nodes:

[0114] In the formula A k+1,k A pk A qk This represents the element in row (k+1)th column (k), row (p)th column (k), and row (q)th column (k) of A;

[0115] ⑦ Update m = m + 1, and continue to execute steps ①-⑥ above, moving upwards layer by layer until all classes are merged into one category;

[0116] 3) Output the hierarchical structure H(N,E) and the list of category labels B.

[0117] The top-down clustering module includes the following steps:

[0118] 1) Initialize the hierarchical structure Where r represents the root node, N represents the node set, and E represents the edge set; initialize the category label set L. r = {1,2,...,K}, training set X and test set T, and push the root node r into the empty stack S;

[0119] 2) When the stack is not empty, repeat the following steps:

[0120] ① Pop the last pushed node t;

[0121] ②When |L t When | > 1, train |L t One-vs-Rest classifier Each category has its own binary classifier, where... For the |L t A classification model is used, and the category confusion matrix C and category similarity matrix A are calculated:

[0122]

[0123]

[0124]

[0125] In the formula, L t f represents the set of category labels for the corresponding node t. k (x i ) represents sample x i The type predicted by the k-th classification model;

[0126] ③ Construct a diagonal matrix D, and then construct the Laplace matrix L using the diagonal matrix:

[0127] D[i,i]=sum(A[i,:])

[0128] L=DA

[0129] Where D[i,i] represents the diagonal element of the i-th row and i-th column, and A[i,:] represents all elements of the i-th row;

[0130] ④ Use the elbow method to find the optimal number of clusters k';

[0131] ⑤ Perform spectral clustering, decompose the Laplacian matrix L into eigenvalues ​​to obtain eigenvalues ​​x' and eigenvectors λ, sort the eigenvalues ​​x in ascending order, select the k'-1 smallest non-zero eigenvalues ​​x", extract the corresponding eigenvectors λ', and form a matrix Z with λ' as the column. Normalize the rows of Z and use them as the k'-dimensional embedding of the class.

[0132] ⑥ Perform K-means clustering, setting the number of clusters to k', and apply it to the rows of the embedding matrix Z to obtain the corresponding cluster labels;

[0133] ⑦ Based on the clustering labels obtained in step ⑥, partition the set of child nodes N of node t. t Create or update category tag sets {L r :c∈N t}, where c represents N t The child nodes in the array satisfy the condition that they do not overlap and

[0134] ⑧ Update edge set E:

[0135] E = E∪{(t,c):c∈N} t}

[0136] ⑨ Update node set N:

[0137] N = N∪N t

[0138] ⑩ Set the child node set N t All nodes are pushed onto stack S.

[0139] 3) Output the hierarchical structure H(N,E) and the classification label {L} corresponding to each node.e :e∈N}, where e represents a node in the node set;

[0140] The iterative module retrains the model based on the initial breast cancer class label hierarchy to achieve iterative updates, including the following steps:

[0141] 1) Input the training set X with class labels, the number of classes K, and the initial hierarchical structure H0 obtained by the bottom-up clustering module or the top-down clustering module; initialize H = H0; initialize the number of iterations d = 1;

[0142] 2) Repeat the following steps until the hierarchy H is reached. d =H d-1 Or d > 100:

[0143] ① A classification model is constructed using a hierarchical structure H, resulting in a new class confusion matrix C. d :C d = f(X,H), where f represents the classifier used;

[0144] ②Standardization:

[0145] ③Symmetricalization:

[0146] ④ By using a bottom-up or top-down clustering module, update the hierarchical structure based on the new category similarity matrix to obtain a new hierarchical structure H. d ;

[0147] ⑤ Update H = H d d = d + 1, return to step ①.

[0148] 3) Output the optimal class label hierarchy H d .

[0149] Specifically, the hierarchical classification verification module includes the following steps:

[0150] 1) Use the data loading module in the data import module to load new breast cancer multi-omics data, use the data preprocessing module to process the new breast cancer multi-omics data, and obtain the validation set V;

[0151] 2) Input the breast cancer validation set V processed in step 1) into the optimal class label hierarchy of breast cancer learned by the hierarchical structure learning module for hierarchical classification;

[0152] 3) Extract the classification results from step 2) above, including accuracy, precision, recall, AUC value and F1 value, as the precision measure O1 under the current prediction result;

[0153] 4) Input the breast cancer validation set V processed in step 1) into the non-optimal class label hierarchy for hierarchical classification, and compare their accuracy measures O2 and O1 to give interpretability to the classification effect.

[0154] The above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Therefore, any changes made in accordance with the shape and principle of the present invention should be covered within the protection scope of the present invention.

Claims

1. A multi-omics data hierarchical classification structure learning system based on machine learning, characterized in that, include: The data import module is used to load multi-omics data and preprocess the data to obtain a dataset that can be used to build machine learning models. The hierarchical structure learning module uses data-driven adaptive machine learning methods for supervised learning, constructs a category similarity matrix, uses bottom-up and top-down unsupervised clustering algorithms to initially construct a class label hierarchy, and integrates iterative algorithms into the training process to finally obtain the optimal class label hierarchy. The hierarchical classification verification module uses the optimal class label hierarchy structure combined with multi-omics data to perform hierarchical classification, and compares and analyzes it with the non-optimal class label hierarchy structure, providing an interpretation of the results. The hierarchical learning module includes a category similarity module, a clustering module, and an iterative module, wherein: The category similarity module constructs a category similarity matrix using an omics-driven adaptive machine learning method, including the following steps: 1) Targeting those containing Using omics data from various categories, a one-vs-Rest classification model is constructed based on adaptive machine learning methods. ,in For the first A classification model was obtained. One prediction probability matrix: ； In the formula, For the corresponding number in the classification model Test sets for each category, For the samples in the test set, For the corresponding number A classification model for each category. For the corresponding number The predicted probability matrix of test set samples of each category, its first column Indicates the sample under the classification model Predicted as the first The probability of each category, in the second column. Indicates the sample under the classification model Predicted as non-first The probability of each category; 2) Based on the prediction probability matrix Construct a category confusion matrix and then transform it into a category similarity matrix: ① First, construct the category confusion matrix. ,make Indicates the line number. Indicates the column number, each element in the matrix Represented as: ; ② Standardize the class confusion matrix, that is, for the class confusion matrix The first in OK All of them have: ; ③ Further analysis of the matrix Symmetricization yields the category similarity matrix: ; In the formula, The test set selected for the classification model. for The samples and their corresponding true class labels For samples under the classification model Predicted as the first The probability of each category This is a category similarity matrix. For the first matrix OK, Representation matrix transpose; The clustering module is used to initially construct a hierarchical structure of class labels, including a bottom-up clustering module and a top-down clustering module. It constructs a clustering model based on the category similarity matrix to achieve the construction of the hierarchical structure of class labels, wherein: The bottom-up clustering module includes the following steps: 1) Initialize the hierarchical structure to be constructed as follows: ,in Representing a hierarchical structure, Represents a set of nodes. Represent the edge set; initialize the category label list. Initialize the category tag list Number of elements ; 2) When Repeat the following steps: ① From the category similarity matrix Find the two most similar categories Merge them: In the formula, express The Middle Line 1 Column elements, As the active node set, with It changes with the changes; ② In the node set Add a new node: ； ③ Select the two most similar categories Merge into a new class and update the active node set. : ； ④ Update the category tag list Add the newly created class: In the formula, and These represent the list of category tags. The Middle The and the first One element; ⑤ In the edge set Add a new edge: ； ⑥ Update the category confusion matrix among active nodes: In the formula, express The Middle Line 1 Column, No. Line 1 Column and the Line 1 Column elements; ⑦ Update Continue executing steps ①-⑥ above, moving upwards layer by layer, until all classes are merged into one category; 3) Output hierarchy and category tag list ; The top-down clustering module includes the following steps: 1) Initialize the hierarchical structure ,in Represents the root node. Represents a set of nodes. Represent the edge set; initialize the category label set. Training set and test set and the root node Push an empty stack middle; 2) When the stack is not empty, repeat the following steps: ① Pop the last pushed node ; ② When During training One-vs-Rest classifier , respectively, correspond to binary classifiers for each category, where For the first A classification model was developed, and the class confusion matrix was calculated. Similarity matrix with category : ； ； ； In the formula, Indicates the corresponding node Category tag set, Indicates sample In the The type predicted by a classification model; ③ Construct a diagonal matrix And construct the Laplace matrix using the diagonal matrix. : ； ； In the formula, Indicates the first Line 1 The diagonal elements of the column, Indicates the first All elements of the row; ④ Use the elbow method to find the optimal number of clusters. ; ⑤ Perform spectral clustering on the Laplace matrix. Perform eigenvalue decomposition to obtain eigenvalues. and eigenvectors eigenvalues ​​in ascending order Sort and select The smallest non-zero eigenvalues Extract the corresponding feature vectors and with Columns form a matrix ,Will The rows are normalized and used as categories. Dimensional embedding; ⑥ Perform K-means clustering, setting the number of clusters to [number]. Applied to embedding matrices The rows are used to obtain the corresponding cluster labels; ⑦ Based on the clustering labels obtained in step ⑥, divide the nodes. child node set Create or update category tag sets ,in express The child nodes in the array satisfy the condition that they do not overlap and ; ⑧ Update edge sets : ； 9. Update the node set : ； ⑩ Set of child nodes Push all nodes onto the stack middle; 3) Output hierarchy and the category label corresponding to each node ,in Represents the nodes in the node set; The iterative module retrains the model based on the initial class label hierarchy to achieve iterative updates, including the following steps: 1) Input a training set with class labels Number of categories The initial hierarchical structure obtained by bottom-up clustering modules or top-down clustering modules ;initialization Initialize the number of iterations. ; 2) Repeat the following steps until the hierarchy is complete. or : ① Use hierarchical structure A classification model is constructed to obtain a new class confusion matrix. : ,in Indicates the classifier used; ② Standardization: ; ③ Symmetry: ; ④ Update the hierarchical structure based on the new category similarity matrix using either a bottom-up or top-down clustering module to obtain a new hierarchical structure. ; ⑤ Update , Return to step ①; 3) Output the optimal class tag hierarchy structure .

2. The machine learning-based multi-omics data hierarchical classification structure learning system according to claim 1, characterized in that: The data import module includes a data loading module and a data preprocessing module, wherein: The data loading module reads multi-omics data and clinical data from the local machine, including CSV, RDF, h5ad, RData and xlsx formats. The multi-omics data includes genomics, epigenomics, transcriptomics and proteomics, and the data sources include complex disease genomes, microbial K-strings and single-cell sequencing. The clinical data includes survival status, survival time and relapse time. The data preprocessing module is used to extract and organize different omics data, including: extracting sample-shared features for complex disease omics datasets, normalizing gene mutation and copy number variation data, removing features with more than 50% methylation gaps, filling gaps with K-nearest neighbors, normalizing methylation data, and integrating gene mutation, copy number variation, and methylation data; counting K-strings in FastQ files for microbiome datasets and retaining high-frequency K-strings with more than 10 occurrences; integrating data and using the Lasso method for feature selection for single-cell omics datasets; and dividing each dataset into training sets. Test set With the validation set All datasets are formatted uniformly, with rows representing samples and columns representing types and features.

3. The machine learning-based multi-omics data hierarchical classification structure learning system according to claim 2, characterized in that: The hierarchical classification verification module includes the following steps: 1) Use the data loading module in the data import module to load multi-omics data, use the data preprocessing module to process the multi-omics data, and obtain the validation set. ; 2) The validation set processed in step 1) The input is the optimal class label hierarchy learned by the hierarchy learning module, which is used for hierarchical classification. 3) Extract the classification results from step 2) above, including accuracy, precision, recall, AUC, and F1 score, as a measure of accuracy for the current prediction result. ; 4) The validation set processed in step 1) Perform hierarchical classification on non-optimal class label hierarchies and measure their accuracy. and By making comparisons, the classification results can be made interpretable.

Images