Multi-omics data driven disease target discovery method, system, and computer storage medium

Abstract

The application discloses a multi-omics data driven disease target discovery method, system and computer storage medium, which combines causal analysis and multi-omics co-localization, greatly improves the biological credibility and verifiability of candidate genes, reduces noise and false positive results, integrates genetic, expression regulation and cell state information, systematically screens pathogenic factors, improves the comprehensiveness of target discovery, models and reasons based on a pure calculation process, greatly reduces experimental cost and time, is suitable for early exploration and rapid iteration of drug targets, and can flexibly configure omics input according to different disease types and sample data, and realizes adaptive modeling for various diseases.

Description

Technical Field

[0001] This invention relates to the field of disease target discovery technology, and more specifically, to a multi-omics data-driven method, system, and computer storage medium for disease target discovery. Background Technology

[0002] Current methods for disease target discovery rely too heavily on external knowledge bases, including databases of known drugs, proteins, or literature, making it difficult to discover new mechanisms. Furthermore, the algorithms used in these methods often fail to cover potential causal pathways. Simultaneously, cellular state modeling primarily relies on correlation analysis, making it difficult to determine the causal relationship between genes and disease states, thus limiting a deeper understanding of disease-driving mechanisms. Current technologies largely depend on single-genome datasets, resulting in a lack of systematic support for the discovery of key disease targets and hindering the formation of credible and verifiable candidate lists of pathogenic genes. Summary of the Invention

[0003] To overcome at least one of the defects described in the prior art, the present invention provides a multi-omics data-driven method, system, and computer storage medium for disease target discovery.

[0004] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: A first aspect of the present invention provides a multi-omics data-driven method for disease target discovery, comprising the following steps: Acquire single-cell transcriptome data and whole-genome sequencing data related to the disease under study; The initial input data is obtained based on the data of all healthy cells in the single-cell transcriptome data. The initial input data is perturbed to gradually generate a state evolution path from health to disease, and the perturbation amplitude and contribution value in the direction of Lie derivative of each gene are accumulated. Based on the perturbation amplitude and contribution value in the direction of Lie derivative for each gene, a list of first candidate genes is obtained; Based on the whole genome sequencing data, a first list of intermediate candidate genes was obtained through analysis and screening. Co-localization analysis was performed between the first intermediate candidate gene list and the expression quantitative trait sites to obtain the second intermediate candidate gene list. Based on the second intermediate candidate gene list and the single-cell transcriptome data, a sparse gene expression matrix is ​​constructed. Based on the sparse gene expression matrix, a second candidate gene list is obtained; Based on the first candidate gene list and the second candidate gene list, a third candidate gene list is obtained; Based on the third candidate gene list, protein interaction network score, matrix factorization score, transcriptional causality inference score and gene expression score were calculated respectively. Based on the protein interaction network score, matrix factorization score, transcriptional causal inference score, and gene expression score, a set of candidate genes to be verified was obtained.

[0005] Furthermore, the initial input data is perturbed to gradually generate a state evolution path from health to disease, accumulating the perturbation amplitude and contribution value in the Lie derivative direction for each gene, including: Initialize a perturbation vector with the same shape as the initial input data; Based on the perturbation vector and the initial input data, the perturbed input data is obtained; Based on the perturbed input data and the pre-trained cell state recognition model, the response gradient of the cell state recognition model to the current perturbation direction is obtained, which is used as the contribution value of the current perturbation in the Lie derivative direction. Based on the gradient backpropagation strategy and the preset target expectation, the perturbation vector is iterated step by step to generate the state evolution path from health to disease, and the perturbation vector corresponding to each step is recorded. The perturbation amplitude for each gene is obtained based on the recorded perturbation vector.

[0006] Furthermore, the target expectation includes the state expectation of the perturbed input data, the perturbation diversity, and the sparsity.

[0007] Furthermore, based on the second intermediate candidate gene list and the single-cell transcriptome data, a sparse gene expression matrix is ​​constructed, including: The single-cell transcriptome data includes two groups of data: disease and health. The single-cell transcriptome data is quality controlled to obtain quality-controlled single-cell transcriptome data. The second intermediate candidate gene list is used as the feature gene set of the quality-controlled single-cell transcriptome data. The expression values ​​on the feature gene set are retained, while the expression values ​​on other non-feature gene sets are set to zero, resulting in a sparse gene expression matrix.

[0008] Furthermore, the protein interaction network score includes: Based on the preset protein interaction database and the third candidate gene list, a protein interaction network graph is constructed, with the protein corresponding to each gene as a node in the protein interaction network graph. If there is an interaction between the proteins corresponding to two genes, an edge is connected between the nodes, and the weight of the edge is provided by the protein interaction database. Based on the protein interaction network graph, the connection number score, path centrality score, and interaction strength score are calculated respectively. The connection number score represents the number of interactions between the current gene and other candidate genes. The path centrality score represents the frequency with which the node corresponding to the gene is the intermediate node of the shortest path in the protein interaction network graph. The interaction strength score is represented by the weight of the corresponding edge. The protein interaction network score is obtained by weighted summing of the connection number score, path centrality score, and interaction strength score.

[0009] Furthermore, the matrix factorization scoring includes: Based on a pre-defined disease-gene relationship database and the third candidate gene list, a two-dimensional sparse matrix of relationships between different genes and different diseases is constructed. The two-dimensional relational sparse matrix is ​​decomposed using nonnegative matrix factorization to obtain the embedding vector of each gene and the embedding vector of the target disease. The matrix factorization score is obtained based on the Euclidean distance or cosine similarity between each gene embedding vector and the target disease embedding vector.

[0010] Furthermore, the transcriptional causal inference score includes: Based on a pre-defined transcription factor regulation database and the third candidate gene list, a transcription regulation network is constructed. The nodes of the transcription regulation network consist of transcription factors and regulated target genes, and the edges represent regulatory relationships. Based on the transcriptional regulatory network, the path length penalty term, the number of downstream regulatory genes, and whether the target is a disease-related gene are calculated respectively. The path length penalty term is defined as the reciprocal of the regulatory path length from the candidate gene to a certain downstream gene. The number of downstream regulatory genes represents the number of all downstream targets regulated by the candidate gene. The value of whether the target is a disease-related gene is assigned different values ​​depending on whether the target is a disease-related gene. The causal impact score of a single path is obtained by summing the path length penalty term, the number of downstream regulatory genes, and whether the target is a disease-related gene. A transcriptional causal inference score is obtained based on the causal influence scores of all pathways of the candidate gene.

[0011] Furthermore, the gene expression score includes: Using single-cell transcriptome data including both disease and health data, a cell-gene-state gene expression matrix is ​​constructed. The disease and health labels of cells are used as classification targets, and a random forest classifier is trained. The feature importance coefficients output by the classifier are used to evaluate the contribution of each gene to disease prediction, which is then used as a gene expression score.

[0012] A second aspect of the present invention provides a multi-omics data-driven disease target discovery system, comprising: The acquisition module acquires single-cell transcriptome data and whole-genome sequencing data related to the disease under study. An initialization module obtains initial input data based on data from all healthy cells in the single-cell transcriptome data. The perturbation module perturbs the initial input data, gradually generating a state evolution path from health to disease, and accumulating the perturbation amplitude and contribution value in the direction of Lie derivative for each gene; The first calculation module obtains a first candidate gene list based on the perturbation amplitude and the contribution value in the direction of the Lie derivative of each gene. The first analysis module analyzes and screens the whole genome sequencing data to obtain a first list of intermediate candidate genes; The second analysis module performs co-localization analysis between the first intermediate candidate gene list and the expression quantitative trait sites to obtain the second intermediate candidate gene list. The construction module constructs a sparse gene expression matrix based on the second intermediate candidate gene list and the single-cell transcriptome data; The second calculation module obtains a second candidate gene list based on the sparse gene expression matrix; The third calculation module obtains a third candidate gene list based on the first candidate gene list and the second candidate gene list; The first scoring module calculates protein interaction network score, matrix factorization score, transcriptional causal inference score and gene expression score based on the third candidate gene list. The first screening module, based on the protein interaction network score, matrix factorization score, transcriptional causal inference score, and gene expression score, screens a set of candidate genes to be verified.

[0013] A third aspect of the present invention provides a computer storage medium storing a computer program, which, when executed by a processor, performs the aforementioned multi-omics data-driven disease target discovery method.

[0014] Compared with the prior art, the beneficial effects of the technical solution of the present invention are: This invention combines causal analysis with multi-omics co-localization, greatly improving the biological credibility and verifiability of candidate genes and reducing noise and false positive results; by integrating genetic, expression regulation, and cellular state information, it systematically screens pathogenic factors, improving the comprehensiveness of target discovery; modeling and inference are based on a pure computational process, greatly reducing experimental costs and time, making it suitable for early exploration and rapid iteration of drug targets; and omics inputs can be flexibly configured according to different disease types and sample data to achieve adaptive modeling for multiple diseases. Attached Figure Description

[0015] Figure 1 A flowchart illustrating a multi-omics data-driven disease target discovery method provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the overall process of a multi-omics data-driven disease target discovery method provided in an embodiment of the present invention. Figure 3 This is an example diagram of the state evolution process based on a cell model and Lie group theory provided in an embodiment of the present invention. Figure 4 This is an example diagram of a differential gene screening process combined with Mendelian randomized causal analysis provided in an embodiment of the present invention. Figure 5 This is an example diagram of a multi-dimensional candidate gene comprehensive scoring process provided in an embodiment of the present invention; Figure 6 A visualization example of the ranking of genes in different groups based on the multi-dimensional scoring provided in this embodiment of the invention; Figure 7 This is a schematic diagram of a multi-omics data-driven disease target discovery system provided in an embodiment of the present invention. Detailed Implementation

[0016] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. 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; 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.

[0017] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] Example 1 This embodiment provides a multi-omics data-driven method for disease target discovery, such as... Figure 1 and Figure 2 As shown, it includes the following steps: Acquire single-cell transcriptome data and whole-genome sequencing data related to the disease under study; The initial input data is obtained based on the data of all healthy cells in the single-cell transcriptome data. The initial input data is perturbed to gradually generate a state evolution path from health to disease, and the perturbation amplitude and contribution value in the direction of Lie derivative of each gene are accumulated. Based on the perturbation amplitude and contribution value in the direction of Lie derivative for each gene, a list of first candidate genes is obtained; Based on the whole genome sequencing data, a first list of intermediate candidate genes was obtained through analysis and screening. Co-localization analysis was performed between the first intermediate candidate gene list and the expression quantitative trait sites to obtain the second intermediate candidate gene list. Based on the second intermediate candidate gene list and the single-cell transcriptome data, a sparse gene expression matrix is ​​constructed. Based on the sparse gene expression matrix, a second candidate gene list is obtained; Based on the first candidate gene list and the second candidate gene list, a third candidate gene list is obtained; Based on the third candidate gene list, protein interaction network score, matrix factorization score, transcriptional causality inference score and gene expression score were calculated respectively. Based on the protein interaction network score, matrix factorization score, transcriptional causal inference score, and gene expression score, a set of candidate genes to be verified was obtained.

[0019] This invention automatically mines new regulatory relationships and unsupervised discovery mechanisms from data to discover novel drug targets, breaking through the limitations of existing databases. It introduces a causal reasoning framework and combines the ideas of Lie group theory to simulate the impact of genes on the state after intervention, more accurately identifying pathogenic driver genes and improving biological explanatory power.

[0020] The method in this embodiment can be divided into three key steps: 1. Combining Lie group theory with cellular evolutionary states and causal analysis; 2. Causal differential gene analysis combining genomics and single-cell transcriptomics; Three: A gene scoring and ranking process that combines multiple omics data and multiple biological dimensions; Among them, an example diagram combining Lie group theory with causal analysis of cell evolution states is shown in the figure below. Figure 3 As shown, specifically: In an optional embodiment, the initial input data, obtained from all healthy cell data in the single-cell transcriptome data, can be: The data of all healthy cells in the single-cell transcriptome data were read, and the expression value of each gene was averaged as the initial input data.

[0021] In a further embodiment, the initial input data is perturbed to gradually generate a state evolution path from health to disease, accumulating the perturbation amplitude and contribution value in the Lie derivative direction for each gene, including: Initialize a perturbation vector with the same shape as the initial input data; In this embodiment, the initial perturbation vector can be a vector [1,1,1,1,1,1,1…,1] consisting entirely of 1s, which has the same shape as the initial input data, and is used to simulate the perturbation of the expression level of each gene. Based on the perturbation vector and the initial input data, the perturbed input data is obtained; In this embodiment, the perturbation vector is applied to the initial input data to obtain the perturbed input data. The specific method for applying the perturbation vector to the initial input data is to perform a dot product between the perturbation vector and the initial input data, and to multiply each value in the perturbation vector by the corresponding expression value at the same position in the initial input data to obtain the perturbed input data. Based on the perturbed input data and the pre-trained cell state recognition model, the response gradient of the cell state recognition model to the current perturbation direction is obtained, which is used as the contribution value of the current perturbation in the Lie derivative direction. In this embodiment, the pre-trained cell state recognition model can be trained using single-cell transcriptome data from both diseased and healthy groups. The subsequent cell state recognition model can predict the cell state corresponding to the input data based on the input data. The specific method for obtaining the response gradient of the current perturbation direction using the cell state recognition model is as follows: Let the current cell state recognition model function be... The input data is The perturbated input data is The perturbation direction vector is defined as the difference between the current perturbation vector and the perturbation vector of the previous round:

[0022] The Lie group derivative estimate calculated along the perturbation direction vector is:

[0023] The chain rule is approximately as follows:

[0024] Right now

[0025] in, It is the gradient of the cell state recognition model output with respect to the input, which can be obtained through backpropagation. It is the perturbation direction vector; By calculating the mapping relationship between the direction of gene perturbation and changes in cell state using Lie group derivatives, this falls under the category of "directional response" modeling in causal reasoning. Compared to conventional gradient analysis (which only considers the first-order response at a point), Lie group derivatives focus on the rate of change of the perturbation trajectory during continuous evolution, providing a quantitative basis for path sensitivity. The contribution value in the direction of the Lie derivative is specifically the response gradient, representing the sensitivity of the gene in the direction of the Lie derivative, reflecting the strength of the response to the predicted state brought about by the current perturbation direction. Based on the gradient backpropagation strategy and the preset target expectation, the perturbation vector is iterated step by step to generate the state evolution path from health to disease, and the perturbation vector corresponding to each step is recorded. In an optional embodiment, the corresponding state recognition results and the expected values ​​of multiple targets are also recorded. Based on the recorded perturbation vector, the perturbation amplitude of each gene is obtained; In this embodiment, the perturbation amplitude refers to the absolute difference between the value in the perturbation vector and 1 for the i-th position (gene) under the current perturbation operation, which represents the current regulatory intensity on the i-th gene. In a further embodiment, the target expectation includes the state expectation of the perturbed input data, the perturbation diversity, and the sparsity.

[0026] In this embodiment, the state recognition expectation specifically requires that the perturbation-induced expression pattern should be predicted as a disease state by the cell state recognition model; the perturbation sparsity constraint specifically encourages perturbations to concentrate on a small number of key genes; and the perturbation diversity constraint specifically avoids the perturbation process from getting stuck in local or repetitive paths. A total loss function is constructed by combining the above objectives, and the perturbation vector is iteratively optimized step-by-step through a backpropagation strategy.

[0027] In a further embodiment, a first candidate gene list is obtained based on the perturbation amplitude and the contribution value in the Lie derivative direction of each gene, specifically as follows: The perturbation amplitude and the contribution value in the direction of Lie derivative of each gene are weighted and summed as a sensitivity index to sort each position of the perturbation variable. Each position represents a different gene, thus obtaining the first candidate gene list.

[0028] The second step combines genomics and single-cell transcriptomics for causal differential gene analysis. (Example diagram follows) Figure 4 As shown, specifically: In a further embodiment, based on the whole-genome sequencing data, the first intermediate candidate gene list is analyzed and screened, which may be: Using SNPs from whole-genome sequencing data as instrumental variables, a causal model of exposure (phenotypic information) on disease outcome (disease state) was constructed and Mendelian randomization analysis (MR) was performed. Based on Egger and other methods, a candidate gene list (WGS-list, also known as the first intermediate candidate gene list) that is significantly associated with the disease was screened.

[0029] In a further embodiment, the first intermediate candidate gene list is co-localized with expression quantitative trait loci to obtain a second intermediate candidate gene list, which may be: Co-localization analysis was performed on the first intermediate candidate gene list selected and eQTL sites (expression quantitative trait sites) obtained from public databases or proprietary data to identify sites with significant signal overlap between the first intermediate candidate gene list and eQTL sites, and to determine the more functional pathogenic regulatory gene list COL-list (i.e., the second intermediate candidate gene list).

[0030] In a further embodiment, based on the second intermediate candidate gene list and the single-cell transcriptome data, a sparse gene expression matrix is ​​constructed, including: The single-cell transcriptome data includes two groups of data: disease and health. The single-cell transcriptome data is quality controlled to obtain quality-controlled single-cell transcriptome data. The second intermediate candidate gene list is used as the feature gene set of the quality-controlled single-cell transcriptome data. The expression values ​​on the feature gene set are retained, while the expression values ​​on other non-feature gene sets are set to zero, resulting in a sparse gene expression matrix.

[0031] In a further embodiment, based on the sparse gene expression matrix, a second candidate gene list is obtained, which may be: Cells were grouped according to their grouping labels (healthy / disease), and differential expression analysis was performed based on a sparse expression matrix. The Wilcoxon rank-sum test was used to extract a list of second candidate genes that were statistically significant and had large effect sizes.

[0032] The third step combines multi-omics data and a multi-biological dimension gene scoring and ranking process, illustrated in the diagram below. Figure 5 As shown, specifically: In a further embodiment, a third candidate gene list is obtained based on the first candidate gene list and the second candidate gene list, specifically as follows: The intersection of the first and second candidate gene lists yields a broader candidate gene list, resulting in the third candidate gene list.

[0033] In a further embodiment, the protein interaction network scoring includes: Based on a pre-defined protein interaction database (such as the STRING or BioGRID database) and the third candidate gene list, a protein interaction network graph is constructed. The protein corresponding to each gene is used as a node in the protein interaction network graph. If there is an interaction between the proteins corresponding to two genes, an edge is connected between the nodes. The weight of the edge is provided by the protein interaction database. In this embodiment, the number of interactions between each gene and other genes, whether the gene is located on the interaction path, and the comprehensive interaction score between two genes directly provided by the public database can also be counted in advance based on the protein interaction network diagram for protein interaction network scoring. Based on the protein interaction network graph, the connection number score, path centrality score, and interaction strength score are calculated respectively. The connection number score represents the number of interactions between the current gene and other candidate genes. The path centrality score represents the frequency with which the node corresponding to the gene is the intermediate node of the shortest path in the protein interaction network graph. The interaction strength score is represented by the weight of the corresponding edge. In this embodiment, a public database (STRING or BioGRID) is used to extract whether there are interaction records between the coding proteins corresponding to the third candidate gene list, including direct interactions (direct binding) and indirect interactions (shortest path formed by 1-2 intermediate proteins). Then, the protein corresponding to each gene is used as a node in the graph. If there is an interaction (direct or indirect) between the proteins corresponding to two genes, an edge is connected between the nodes; the weight of the edge depends on the interaction score provided by the database, such as STRING's combined_score (range 0~1000, normalizable). The protein interaction network score has three dimensions: first, the connection number score, which represents the number of interactions between this gene and other candidate genes; let the i-th node... The neighbor set is The score would then be:

[0034] Next is the path centrality score, which represents the frequency with which the node is an intermediate node in the shortest path in the network, for all gene pairs. ,judge Whether it is a relay point in its shortest interaction path is used to obtain a centrality score:

[0035] In the formula, It means arrive The number of shortest paths, It was through The number of paths; Finally, there is the interaction strength score, based on the combined score provided by the STRING database, i.e. ; The protein interaction network score is obtained by weighted summing of the connection number score, path centrality score, and interaction strength score. In this embodiment, the protein interaction network score Specifically:

[0036] In the formula, , , The empirical weights can be set to 1:1:1 or adjusted based on biological validation. norm() indicates that each score is normalized to 0-1.

[0037] In a further embodiment, the matrix factorization scoring includes: Based on a pre-defined disease-gene relationship database (such as the PrimeKG database) and the third candidate gene list, a two-dimensional sparse matrix of relationships between different genes and different diseases is constructed. The two-dimensional relational sparse matrix is ​​decomposed using nonnegative matrix factorization to obtain the embedding vector of each gene and the embedding vector of the target disease. In this embodiment, the two-dimensional relational sparse matrix is ​​assumed to be:

[0038] Where D represents the total number of diseases and G represents the total number of genes. Indicates gene Is it related to disease? Given co-occurrence records (represented by 0 and 1), the two-dimensional relational sparse matrix is ​​decomposed using nonnegative matrix decomposition to obtain:

[0039] In the formula, The embedding matrix represents the disease (each row is an embedding vector for a disease). This represents the gene embedding matrix (each row is the embedding vector of a gene), where k is the set embedding dimension, typically set to 64.

[0040] The matrix factorization score is based on the Euclidean distance or cosine similarity between each gene embedding vector and the target disease embedding vector. A higher similarity or a shorter distance indicates that the gene co-occurs more frequently with the target disease.

[0041] The significance of matrix factorization scoring lies in the fact that the original data on the relationship between genes and target diseases is explicit and sparse and discrete. The vector representation after embedding is a compression and generalization of these sparse relationships. If the embedding vector of a gene and the embedding vector of the target disease are very close in the feature space, it means that they often co-occur in the co-occurrence space. Even if there is no direct association between the gene and the disease in the original matrix, the potential association between the gene and the target disease can be inferred through the approximation of the embedding space.

[0042] In a further embodiment, the transcriptional causal inference score includes: Based on a pre-defined transcription factor regulation database (such as the TKLink database) and the third candidate gene list, a transcription regulation network is constructed. The nodes of the transcription regulation network consist of transcription factors and regulated target genes, and the edges represent regulatory relationships. Based on the transcriptional regulatory network, the path length penalty term, the number of downstream regulatory genes, and whether the target is a disease-related gene are calculated respectively. The path length penalty term is defined as the reciprocal of the regulatory path length from the candidate gene to a certain downstream gene. The number of downstream regulatory genes represents the number of all downstream targets regulated by the candidate gene. The value of whether the target is a disease-related gene is assigned different values ​​depending on whether the target is a disease-related gene. In a specific embodiment, assuming a candidate gene There is at least one regulatory pathway in the transcriptional regulatory network. ,in These are downstream genes that it directly or indirectly regulates. If the endpoint of this pathway... If a gene is a known pathogenic gene for the target disease, it can be considered to potentially participate in disease regulation. To quantify this causal relationship, the following weighting factors are defined: w1: Path length penalty term, defined as w1 = 1 / path length. The shorter the path, the more direct the control relationship. w2: Number of downstream regulatory genes, i.e. The number of all downstream targets controlled indicates its network influence; w3: Is the target gene a disease-related gene? If the target is known for the disease, then w3=1; otherwise, it is 0. Based on the path length penalty term, the number of downstream regulatory genes, and whether the target gene is a disease-related gene, the causal impact score of a single path is obtained by summing the factors. Based on the above weights, the causal impact score of a single path is defined as follows:

[0043] A transcriptional causal inference score is obtained based on the causal influence scores of all pathways of the candidate gene. like If multiple pathways satisfy the conditions, the causal scores of all pathways can be summed to obtain a cumulative score, or the highest score among the pathways can be taken as a representative causal score. Finally, the "transcriptional causal inference score" of a gene is defined as its aggregated score across all target disease pathways in the transcriptional regulatory network. If a gene does not have any downstream regulatory pathways related to the target disease, its causal score is 0.

[0044] In a further embodiment, the gene expression score includes: Using single-cell transcriptome data including both disease and health data, a cell-gene-state gene expression matrix is ​​constructed. The disease and health labels of cells are used as classification targets, and a random forest classifier is trained. The feature importance coefficients output by the classifier are used to evaluate the contribution of each gene to disease prediction, which is then used as a gene expression score.

[0045] In a further embodiment, based on the protein interaction network score, matrix factorization score, transcriptional causal inference score, and gene expression score, a set of candidate genes to be validated is obtained, including: The protein interaction network score, matrix factorization score, transcriptional causal inference score, and gene expression score were fused together in a weighted sum manner. All candidate genes were ranked according to the comprehensive score, and the top 30 genes with the highest total score were selected as the final set of candidate genes to be validated.

[0046] The final visualization of gene sorting for different groupings is shown below. Figure 6 As shown, light gray represents known disease-causing genes, dark gray represents novel disease-related genes given in the embodiments of the present invention, and black represents genes that are clearly unrelated to the disease.

[0047] Example 2 This embodiment provides a multi-omics data-driven disease target discovery system, such as... Figure 7 As shown, the discovery system applies the multi-omics data-driven disease target discovery method described in Example 1, including: The acquisition module acquires single-cell transcriptome data and whole-genome sequencing data related to the disease under study. An initialization module obtains initial input data based on data from all healthy cells in the single-cell transcriptome data. The perturbation module perturbs the initial input data, gradually generating a state evolution path from health to disease, and accumulating the perturbation amplitude and contribution value in the direction of Lie derivative for each gene; The first calculation module obtains a first candidate gene list based on the perturbation amplitude and the contribution value in the direction of the Lie derivative of each gene. The first analysis module analyzes and screens the whole genome sequencing data to obtain a first list of intermediate candidate genes; The second analysis module performs co-localization analysis between the first intermediate candidate gene list and the expression quantitative trait sites to obtain the second intermediate candidate gene list. The construction module constructs a sparse gene expression matrix based on the second intermediate candidate gene list and the single-cell transcriptome data; The second calculation module obtains a second candidate gene list based on the sparse gene expression matrix; The third calculation module obtains a third candidate gene list based on the first candidate gene list and the second candidate gene list; The first scoring module calculates protein interaction network score, matrix factorization score, transcriptional causal inference score and gene expression score based on the third candidate gene list. The first screening module, based on the protein interaction network score, matrix factorization score, transcriptional causal inference score, and gene expression score, screens a set of candidate genes to be verified.

[0048] Example 3 This embodiment provides a computer storage medium storing a computer program. When the computer program is executed by a processor, it completes the multi-omics data-driven disease target discovery method.

[0049] The same or similar labels correspond to the same or similar parts; 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. 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 multi-omics data-driven method for disease target discovery, characterized in that, Includes the following steps: Acquire single-cell transcriptome data and whole-genome sequencing data related to the disease under study; The initial input data is obtained based on the data of all healthy cells in the single-cell transcriptome data. The initial input data is perturbed to gradually generate a state evolution path from health to disease, and the perturbation amplitude and contribution value in the direction of Lie derivative of each gene are accumulated. Based on the perturbation amplitude and contribution value in the direction of Lie derivative for each gene, a list of first candidate genes is obtained; Based on the whole genome sequencing data, a first list of intermediate candidate genes was obtained through analysis and screening. Co-localization analysis was performed between the first intermediate candidate gene list and the expression quantitative trait sites to obtain the second intermediate candidate gene list. Based on the second intermediate candidate gene list and the single-cell transcriptome data, a sparse gene expression matrix is ​​constructed. Based on the sparse gene expression matrix, a second candidate gene list is obtained; Based on the first candidate gene list and the second candidate gene list, a third candidate gene list is obtained; Based on the third candidate gene list, protein interaction network score, matrix factorization score, transcriptional causality inference score and gene expression score were calculated respectively. Based on the protein interaction network score, matrix factorization score, transcriptional causal inference score, and gene expression score, a set of candidate genes to be verified was obtained.

2. The multi-omics data-driven disease target discovery method according to claim 1, characterized in that, The initial input data is perturbed to gradually generate a state evolution path from health to disease, accumulating the perturbation amplitude and contribution value in the Lie derivative direction for each gene, including: Initialize a perturbation vector with the same shape as the initial input data; Based on the perturbation vector and the initial input data, the perturbed input data is obtained; Based on the perturbed input data and the pre-trained cell state recognition model, the response gradient of the cell state recognition model to the current perturbation direction is obtained, which is used as the contribution value of the current perturbation in the Lie derivative direction. Based on the gradient backpropagation strategy and the preset target expectation, the perturbation vector is iterated step by step to generate the state evolution path from health to disease, and the perturbation vector corresponding to each step is recorded. The perturbation amplitude for each gene is obtained based on the recorded perturbation vector.

3. The multi-omics data-driven disease target discovery method according to claim 2, characterized in that, The target expectation includes the state expectation of the perturbed input data, the perturbation diversity, and the sparsity.

4. The multi-omics data-driven disease target discovery method according to claim 1, characterized in that, Based on the second intermediate candidate gene list and the single-cell transcriptome data, a sparse gene expression matrix is ​​constructed, including: The single-cell transcriptome data includes two groups of data: disease and health. The single-cell transcriptome data is quality controlled to obtain quality-controlled single-cell transcriptome data. The second intermediate candidate gene list is used as the feature gene set of the quality-controlled single-cell transcriptome data. The expression values ​​on the feature gene set are retained, while the expression values ​​on other non-feature gene sets are set to zero, resulting in a sparse gene expression matrix.

5. The multi-omics data-driven disease target discovery method according to claim 1, characterized in that, The protein-protein interaction network score includes: Based on the preset protein interaction database and the third candidate gene list, a protein interaction network graph is constructed, with the protein corresponding to each gene as a node in the protein interaction network graph. If there is an interaction between the proteins corresponding to two genes, an edge is connected between the nodes, and the weight of the edge is provided by the protein interaction database. Based on the protein interaction network graph, the connection number score, path centrality score, and interaction strength score are calculated respectively. The connection number score represents the number of interactions between the current gene and other candidate genes. The path centrality score represents the frequency with which the node corresponding to the gene is the intermediate node of the shortest path in the protein interaction network graph. The interaction strength score is represented by the weight of the corresponding edge. The protein interaction network score is obtained by weighted summing of the connection number score, path centrality score, and interaction strength score.

6. The multi-omics data-driven disease target discovery method according to claim 1, characterized in that, The matrix factorization scoring includes: Based on a pre-defined disease-gene relationship database and the third candidate gene list, a two-dimensional sparse matrix of relationships between different genes and different diseases is constructed. The two-dimensional relational sparse matrix is ​​decomposed using nonnegative matrix factorization to obtain the embedding vector of each gene and the embedding vector of the target disease. The matrix factorization score is obtained based on the Euclidean distance or cosine similarity between each gene embedding vector and the target disease embedding vector.

7. The multi-omics data-driven disease target discovery method according to claim 1, characterized in that, The transcriptional causal inference score includes: Based on a pre-defined transcription factor regulation database and the third candidate gene list, a transcription regulation network is constructed. The nodes of the transcription regulation network consist of transcription factors and regulated target genes, and the edges represent regulatory relationships. Based on the transcriptional regulatory network, the path length penalty term, the number of downstream regulatory genes, and whether the target is a disease-related gene are calculated respectively. The path length penalty term is defined as the reciprocal of the regulatory path length from the candidate gene to a certain downstream gene. The number of downstream regulatory genes represents the number of all downstream targets regulated by the candidate gene. The value of whether the target is a disease-related gene is assigned different values ​​depending on whether the target is a disease-related gene. The causal impact score of a single path is obtained by summing the path length penalty term, the number of downstream regulatory genes, and whether the target is a disease-related gene. A transcriptional causal inference score is obtained based on the causal influence scores of all pathways of the candidate gene.

8. The multi-omics data-driven disease target discovery method according to claim 1, characterized in that, The gene expression score includes: Using single-cell transcriptome data including both disease and health data, a cell-gene-state gene expression matrix is ​​constructed. The disease and health labels of cells are used as classification targets, and a random forest classifier is trained. The feature importance coefficients output by the classifier are used to evaluate the contribution of each gene to disease prediction, which is then used as a gene expression score.

9. A multi-omics data-driven disease target discovery system, characterized in that, include: The acquisition module acquires single-cell transcriptome data and whole-genome sequencing data related to the disease under study. An initialization module obtains initial input data based on data from all healthy cells in the single-cell transcriptome data. The perturbation module perturbs the initial input data, gradually generating a state evolution path from health to disease, and accumulating the perturbation amplitude and contribution value in the direction of Lie derivative for each gene; The first calculation module obtains a first candidate gene list based on the perturbation amplitude and the contribution value in the direction of the Lie derivative of each gene. The first analysis module analyzes and screens the whole genome sequencing data to obtain a first list of intermediate candidate genes; The second analysis module performs co-localization analysis on the first intermediate candidate gene list and the expression quantitative trait sites to obtain the second intermediate candidate gene list. The construction module constructs a sparse gene expression matrix based on the second intermediate candidate gene list and the single-cell transcriptome data; The second calculation module obtains a second candidate gene list based on the sparse gene expression matrix; The third calculation module obtains a third candidate gene list based on the first candidate gene list and the second candidate gene list; The first scoring module calculates protein interaction network score, matrix factorization score, transcriptional causal inference score and gene expression score based on the third candidate gene list. The first screening module, based on the protein interaction network score, matrix factorization score, transcriptional causal inference score, and gene expression score, screens a set of candidate genes to be verified.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, performs the multi-omics data-driven disease target discovery method as described in any one of claims 1 to 8.

Images