Multilayer bayesian model based method for constructing a three-dimensional regulatory network of multi-omics data

Abstract

The application discloses a multi-omics data three-dimensional regulation network construction method based on a hierarchical Bayesian model, which comprises the following steps: S1, obtaining multi-omics data of at least two different molecular levels of samples; S2, constructing a hierarchical Bayesian model for integrating the multi-omics data and inferring the regulation relationship between molecules, wherein the hierarchical Bayesian model realizes cross-level information coupling by making the prior probability of the existence of a next-level regulation relationship dependent on the signal of a previous-level regulation relationship; S3, solving and performing posterior inference on the hierarchical Bayesian model to obtain the posterior probability of the regulation relationship between molecules; and S4, constructing a three-dimensional regulation network containing cross-molecular-level regulation directed edges based on the posterior probability. The application can systematically integrate multi-dimensional omics data, quantify and infer the cross-level regulation relationship between molecules, and construct a three-dimensional regulation network which comprehensively and accurately reflects the mechanism of tumor drug resistance.

Description

Technical Field

[0001] This invention relates to the field of network construction technology, specifically to a method for constructing a multi-omics data-based three-dimensional regulation network based on a hierarchical Bayesian model. Background Technology

[0002] Human epidermal growth factor receptor 2 (HER2) positive subtype accounts for approximately 20% to 30% of all breast cancer cases. This subtype is characterized by high invasiveness, short disease-free survival, and poor prognosis. Trastuzumab, a monoclonal antibody drug targeting the HER2 receptor, has become a cornerstone of clinical treatment for HER2-positive breast cancer.

[0003] However, the efficacy of trastuzumab is often severely limited due to primary or secondary resistance. A significant proportion of patients are unresponsive to initial treatment or develop resistance within one year of starting treatment, leading to disease progression. Current research has revealed several molecular mechanisms leading to drug resistance, including: (1) target binding barriers, such as HER2 receptor extracellular domain truncation (forming p95HER2) or cell membrane surface mucin 4 (MUC4) masking HER2 epitopes, which hinder the effective binding of drugs to targets; (2) abnormal activation of downstream signaling pathways, especially the continuous activation of the PI3K / AKT / mTOR signaling pathway driven by PTEN loss of function or PIK3CA activation mutation; (3) signal bypass activation, when HER2 signaling is inhibited, tumor cells maintain survival and proliferation by activating alternative pathways (such as the insulin-like growth factor-1 receptor (IGF-1R) pathway); (4) epigenetic and non-coding RNA regulation, DNA methylation, histone modification and microRNA (such as miR-375) participate in drug resistance formation by regulating the expression of key genes (such as IGF1R); (5) transcription factor regulation, the abnormal activity of transcription factors (such as SP1, FoxM1) has also been confirmed to be related to drug resistance.

[0004] While the above findings reveal multiple potential mechanisms of drug resistance, these understandings remain fragmented and incomplete. The fundamental reason is that cellular life processes are governed by a complex regulatory network comprised of molecules at multiple levels, including the genome, epigenome, transcriptome, and proteome. The emergence of drug resistance phenotypes is not the result of a single molecular event, but rather the ultimate manifestation of a systemic, multi-level dysregulation of the entire molecular network under drug stress.

[0005] Existing analytical methods have significant limitations: current data analysis strategies are mostly limited to a single omics level (e.g., analyzing only the transcriptome) or use simple statistical correlations to superficially integrate data from different omics. These methods struggle to capture directional causal regulatory relationships between different molecular levels. For example, they cannot effectively infer how epigenetic modifications (such as DNA methylation) regulate transcription factor expression, thereby affecting the transcription of downstream target genes and ultimately driving drug resistance phenotypes at the protein function level. Therefore, existing technologies cannot construct a comprehensive "three-dimensional" molecular network model that fully reflects cross-level regulatory logic. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, the present invention aims to provide a method for constructing a multi-omics data-based three-dimensional regulatory network based on a hierarchical Bayesian model. This method can systematically integrate multi-dimensional omics data, quantify and infer cross-level molecular regulatory relationships, and construct a comprehensive and accurate three-dimensional regulatory network that reflects the mechanism of tumor drug resistance. This provides a theoretical basis and key targets for overcoming drug resistance and developing novel treatment strategies.

[0007] To achieve the objectives of this invention, the following technical solution is adopted:

[0008] A method for constructing a multi-omics data stereoregulation network based on a hierarchical Bayesian model includes the following steps:

[0009] S1. Obtain multi-omics data at at least two different molecular levels from the sample;

[0010] S2. Construct a hierarchical Bayesian model to integrate the multi-omics data and infer the regulatory relationships between molecules. The hierarchical Bayesian model achieves cross-level information coupling by depending the prior probability of the existence of the next level of regulatory relationship on the signal of the previous level of regulatory relationship.

[0011] S3. Solve the hierarchical Bayesian model and perform posterior inference to obtain the posterior probability of the intermolecular regulatory relationship;

[0012] S4. Based on the posterior probability, construct a three-dimensional regulatory network containing directed edges for cross-molecular level regulation.

[0013] Furthermore, after step S4, the method further includes: S5, performing biological analysis on the stereoregulatory network to identify key regulatory pathways, core regulatory factors, or key biological functions; wherein the core regulatory factors are identified through network topology analysis, and the key biological functions are identified through functional enrichment analysis.

[0014] Furthermore, before constructing the hierarchical Bayesian model, the multi-omics data is preprocessed, including standardization, logarithmic transformation, Z-score standardization, and missing value imputation of various types of omics data.

[0015] Furthermore, the hierarchical Bayesian model specifically includes a phenotypic prediction model, a cross-omics regulatory model, and a hierarchical prior distribution of parameters.

[0016] Furthermore, the phenotypic prediction model is a generalized linear model, and its connection function is the Logit function, which is used to associate phenotypic data with molecular features of downstream functional layers.

[0017] Furthermore, the cross-omics regulatory model includes:

[0018] The protein abundance model models protein abundance as a linear function of its corresponding mRNA expression level.

[0019] The mRNA expression model models the mRNA expression level as a linear function of upstream epigenetic data.

[0020] Furthermore, the multi-omics data also includes at least one of genome copy number variation data, non-coding RNA expression profile data, and phosphorylated proteomics data;

[0021] And / or, when constructing the stereo modulation network, the threshold for the posterior inclusion probability is set to 0.5 or higher, and directed edges are drawn between corresponding molecules only when the posterior inclusion probability exceeds the threshold.

[0022] And / or, the three-dimensional adjustment network uses visualization tools to arrange different types of nodes on different virtual levels to form a three-dimensional network diagram with cross-level control relationships;

[0023] And / or, the edges or nodes in the stereoregulation network are statistically evaluated using permutation tests and error detection rate correction methods;

[0024] And / or, the method for constructing a multi-omics data stereoregulatory network based on a hierarchical Bayesian model is executed in an online visualization platform; the platform receives the multi-omics data through its data interface and automatically executes steps S1 to S4.

[0025] Compared to existing technologies, the advantages of this invention are as follows: This invention possesses systematicity and integrativeness. It is the first to propose a computational framework capable of truly integrating multi-level omics data and inferring cross-level causal relationships, overcoming the limitations of traditional methods that can only perform single-level analysis or simple correlation integration. This invention exhibits profound mechanistic depth: by constructing a "three-dimensional" regulatory network, it can reveal the complete regulatory chain from upstream epigenetic modification to downstream protein function execution, thereby more profoundly and comprehensively elucidating the molecular mechanisms of complex diseases, such as the "systemic" dysregulation of tumor drug resistance. This invention possesses model flexibility and scalability; based on a hierarchical Bayesian framework, this model… It possesses high flexibility and scalability, allowing for the easy incorporation of more types of omics data (such as chromatin accessibility ATAC-seq, metabolomics data, etc.), simply by adding the corresponding hierarchy and regulatory relationships to the model. This invention ensures the probabilistic interpretation of the results, providing probabilistic measures for all inferred results (such as posterior inclusion probability PIP), enabling researchers to screen key molecules and regulatory relationships based on confidence levels, resulting in more reliable and robust results. This invention has broad application value, not only for studying tumor drug resistance but also for extensive application in precision medicine fields such as mechanism research of other complex diseases, biomarker discovery, and drug target screening. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method for constructing a multi-omics data stereoregulation network based on a hierarchical Bayesian model in an embodiment of the present invention;

[0027] Figure 2 This is an overall flowchart of the construction of the multi-omics three-dimensional regulatory network in an embodiment of the present invention;

[0028] Figure 3 for Figure 2 A magnified view of the fusion of multi-level prior information in Spike-and-Slab;

[0029] Figure 4 for Figure 2 A magnified view of network analysis and key driver identification in China;

[0030] Figure 5 for Figure 2 Enlarged diagram of the biological explanation and experimental verification in Chinese medicine;

[0031] Figure 6 This is a conceptual schematic diagram illustrating a multi-omics data three-dimensional regulatory network with different omics levels and their cross-level regulatory relationships in an embodiment of the present invention;

[0032] Figure 7 for Figure 6 Magnified view of mesomethylation;

[0033] Figure 8 for Figure 6 Enlarged view of the transcriptome;

[0034] Figure 9 for Figure 6 Magnified view of the proteome;

[0035] Figure 10 This is a schematic diagram of the probabilistic graphical model (DAG) of the hierarchical Bayesian model in an embodiment of the present invention. Detailed Implementation

[0036] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0037] like Figure 1 As shown, this embodiment of the invention provides a method for constructing a multi-omics data stereoregulation network based on a hierarchical Bayesian model, including the following steps:

[0038] S1. Obtain multi-omics data at at least two different molecular levels from the sample.

[0039] Acquisition of multi-omics data:

[0040] Data Acquisition: Collect multi-omics data from paired sensitive and resistant samples (e.g., cell lines or clinical tissues) for the biological question being studied (e.g., trastuzumab resistance). Data types should include at least two layers, with a typical combination as follows:

[0041] Upstream regulatory layer: Epigenomic data (such as DNA methylation chip / sequencing data), genome copy number variation (CNV) data.

[0042] Intermediate transcription layer: gene expression profile data (mRNA-seq), non-coding RNA expression profile data (miRNA-seq).

[0043] Downstream functional layer: proteomics data (mass spectrometry data), phosphorylated proteomics data.

[0044] Phenotypic data: The samples are clearly labeled as binary (e.g., sensitive / resistant) or continuous (e.g., drug IC50 value).

[0045] S2. Construct a hierarchical Bayesian model to integrate the multi-omics data and infer the regulatory relationships between molecules. The hierarchical Bayesian model achieves cross-level information coupling by depending on the prior probability of the existence of the next level of regulatory relationship on the signal of the previous level of regulatory relationship.

[0046] Constructing a multi-level integrated hierarchical Bayesian model (HBM):

[0047] A hierarchical Bayesian model was constructed to integrate multi-omics data and infer regulatory relationships. This model can simultaneously perform variable selection and parameter estimation, thereby identifying key molecules and their regulatory pathways related to phenotypes.

[0048] Top-level model: Phenotypic prediction model

[0049]

[0050] set up Phenotypic observations of the sample (e.g.: Indicates drug resistance. (This indicates sensitivity). A generalized linear model (GLM) is used to link phenotypes with molecular features of downstream functional layers (e.g., the proteome). It is a connection function (e.g., for a binary classification model, use the Logit function). It is a sample Expected value of the phenotype; It is a sample The Middle The elegance of a single protein; This is the coefficient of influence of the protein on the phenotype; It is the intercept; It refers to the total amount of protein.

[0051] Intermediate layer of the model: cross-omics regulatory model

[0052] It is assumed that the expression level of downstream molecules (proteins) is regulated by upstream molecules (mRNA), and the expression of mRNA is in turn regulated by even more upstream epigenetic factors such as DNA methylation. This biological prior knowledge is encoded into the prior distribution of model parameters, forming a hierarchical structure.

[0053] Protein abundance model: This model calculates the abundance of each protein. Modeling its corresponding mRNA expression level The function.

[0054]

[0055] in, This represents the translation efficiency or regulatory strength from mRNA to protein.

[0056] mRNA expression model: Similarly, the expression level of each mRNA is... Modeling as multiple upstream regulatory factors (e.g., DNA methylation level) The function of ).

[0057]

[0058] in, Representing the The methylation site corresponding to the th methylation site The regulatory effect coefficient of gene expression; It is related to genes Number of related methylation sites

[0059] Model Underlying Structure: Hierarchical Prior Distribution of Parameters and Variable Selection

[0060] To achieve variable selection (i.e., identifying which regulatory relationships are significant) and integrate cross-omics information, key effect coefficients in the model ( Set a hierarchical "Spike-and-Slab" prior distribution.

[0061] right , and Priors: A similar Spike-and-Slab prior is adopted.

[0062]

[0063] in, , and It is a binary indicator variable, when hour, From a "thick plate" distribution with large variance ( The fact that the protein was extracted from the sample indicates that the protein has a significant effect on the phenotype. and hour, and Similarly, this indicates that mRNA and methylation have significant effects on phenotype. When hour, From a distribution with extremely small variance (a "peaked" distribution) Extraction from () indicates that the protein has no significant effect on the phenotype; and hour, and As above, this indicates that mRNA and methylation have no significant effect on phenotype. It is a protein The prior probability of being selected into the model. It is mRNA The prior probability of being selected into the model. It is methylation The prior probability of being selected into the model.

[0064] Core Innovation: Prior Integration of Cross-Level Information

[0065] The key innovation of this invention lies in using regulatory information at the molecular level as prior evidence for the existence of a regulatory relationship at another level. Specifically, it uses the prior probability of the indicator variable ( Model it so that it depends on whether there is a regulatory signal in the previous level.

[0066] For example, a protein It has a significant effect on phenotype (i.e. The probability of ) It should depend on whether the translation / regulation of its upstream mRNA to protein is significant (i.e., This dependency relationship can be described using a logistic regression model:

[0067]

[0068] Similarly, the probability of significant mRNA-to-protein translation / regulation can depend on the presence of significant methylation regulation upstream:

[0069]

[0070] This structure couples regulatory relationships at different levels at a probabilistic level, enabling the model to preferentially discover regulatory pathways with continuous chains of biological evidence (e.g., from methylation to mRNA to protein, ultimately affecting the model).

[0071] Probabilistic graphical model representation:

[0072] The entire hierarchical Bayesian model can be intuitively represented by a directed acyclic graph (DAG), clearly showing the conditional dependencies between data, parameters, and hyperparameters. The structure of this graph embodies the core idea of ​​this invention: the flow and integration of information at different molecular levels.

[0073] S3. Solve the hierarchical Bayesian model and perform posterior inference to obtain the posterior probability of the intermolecular regulatory relationship.

[0074] Model Solving and Posterior Inference:

[0075] Due to the complexity of the model, its posterior distribution lacks an analytical solution. Markov Chain Monte Carlo (MCMC) methods, such as Gibbs Sampling or the Metropolis-Hastings algorithm, are used to sample the model's posterior distribution. Through analysis of the MCMC chain, each parameter (e.g., ...) is obtained. The posterior distribution of ), and the posterior inclusion probability (PIP) of each indicator variable, i.e. .

[0076] S4. Based on the posterior probability, construct a three-dimensional regulatory network containing directed edges for cross-molecular level regulation.

[0077] Construction and visualization of three-dimensional regulation networks:

[0078] Definitions of network nodes and edges:

[0079] Nodes: Nodes in a network represent different biomolecules, such as methylation sites, mRNA, and proteins. Node types can be distinguished by their shape or color.

[0080] Edges: Directed edges in a network represent regulatory relationships between molecules. The existence of an edge is determined by the posterior inclusion probability (PIP) of the corresponding indicator variable.

[0081] Building a network:

[0082] Set a PIP threshold (such as 0.5 or higher, for example 0.9).

[0083] if The threshold is at the methylation site. and genes Draw a regulatory edge between mRNAs ( ).

[0084] if The threshold is in the gene mRNA and protein Draw a control edge between them ( ).

[0085] if The threshold, then, is in the protein Draw a control edge between the phenotype and the control edge ( ).

[0086] The thickness or color of the edge can be used to represent the magnitude of the moderating effect (the post-validation mean of the corresponding effect coefficient).

[0087] 3D visualization:

[0088] Different types of nodes (methylation, mRNA, protein, phenotype) are arranged on different virtual planes or levels and connected by the cross-layer regulatory edges obtained in step 2 to form an intuitive "three-dimensional" network diagram.

[0089] Furthermore, after step S4, the method further includes: S5, performing biological analysis on the stereoregulatory network to identify key regulatory pathways, core regulatory factors, or key biological functions; wherein the core regulatory factors are identified through network topology analysis, and the key biological functions are identified through functional enrichment analysis.

[0090] Network analysis and biological applications:

[0091] Key regulatory pathway identification: In the constructed three-dimensional network, we seek long pathways that start from top-level regulatory molecules (such as methylation), pass through the middle layer (mRNA), and finally reach downstream functional molecules (proteins) to influence phenotype. These pathways represent key regulatory axes that lead to phenotypic changes.

[0092] Core regulatory factor (Hub node) identification: Through network topology analysis (such as calculating node degree, betweenness centrality, etc.), hub molecules that play a core role in the network are identified. These molecules may be potential biomarkers or novel targets for drug intervention.

[0093] Functional enrichment analysis: Pathway enrichment analysis (such as KEGG, GO analysis) is performed on key molecules identified in the network to reveal their underlying biological functions and signaling pathways.

[0094] Furthermore, before constructing the hierarchical Bayesian model, the multi-omics data is preprocessed, including standardization, logarithmic transformation, Z-score standardization, and missing value imputation of various types of omics data.

[0095] Preprocessing of multi-omics data:

[0096] Data preprocessing: Standardize the data for each omics dataset. For example, perform logarithmic transformation and Z-score standardization on gene expression data; calculate β values ​​for methylation data; and standardize protein abundance data. Ensure that all features (genes, proteins, etc.) are comparable across different samples. Impute missing values ​​in the data.

[0097] Furthermore, the hierarchical Bayesian model specifically includes a phenotypic prediction model, a cross-omics regulatory model, and a hierarchical prior distribution of parameters.

[0098] Furthermore, the phenotypic prediction model is a generalized linear model, and its connection function is the Logit function, which is used to associate phenotypic data with molecular features of downstream functional layers.

[0099] Furthermore, the cross-omics regulatory model includes:

[0100] The protein abundance model models protein abundance as a linear function of its corresponding mRNA expression level.

[0101] The mRNA expression model models the mRNA expression level as a linear function of upstream epigenetic data.

[0102] Furthermore, the multi-omics data also includes at least one of genome copy number variation data, non-coding RNA expression profile data, and phosphorylated proteomics data.

[0103] Furthermore, when constructing the stereoregulation network, the threshold for the posterior inclusion probability is set to 0.5 or higher, and directed edges are drawn between corresponding molecules only when the posterior inclusion probability exceeds the threshold.

[0104] Furthermore, the three-dimensional adjustment network uses visualization tools to arrange different types of nodes on different virtual levels, forming a three-dimensional network diagram with cross-level control relationships.

[0105] Furthermore, the permutation test and error detection rate correction method are used to evaluate the statistical significance of the edges or nodes in the three-dimensional adjustment network.

[0106] Furthermore, the method for constructing a multi-omics data stereoregulatory network based on a hierarchical Bayesian model is executed in an online visualization platform; the platform receives the multi-omics data through its data interface and automatically executes steps S1 to S4.

[0107] Experimental example:

[0108] like Figure 2-5 As shown, data acquisition and preprocessing:

[0109] Samples: HER2-positive breast cancer cell lines sensitive to trastuzumab (such as BT-474) and drug-resistant cell lines induced by escalation of drug concentration (BT-474-R) were collected, with 3 biological replicates for each.

[0110] Multi-omics data acquisition: The following tests were performed simultaneously on all 6 samples:

[0111] Whole-genome DNA methylation sequencing (WGBS);

[0112] mRNA expression profiling sequencing (RNA-seq);

[0113] Protein mass spectrometry (DIA-MS) that is independent of data acquisition.

[0114] Data preprocessing:

[0115] Methylation data: WGBS data were aligned to a reference genome, and the average methylation level (B value) of each gene promoter region (2kb upstream to 500bp downstream of TSS) was calculated.

[0116] RNA-seq data: alignment was performed using STAR, counting was performed using HTSeq, and finally converted to TPM (Transcripts Per Million) values, followed by log2 (TPM+1) transformation and Z-score normalization.

[0117] Proteomic data: Spectronaut was used to analyze the raw DIA-MS data to obtain the relative abundance of proteins, followed by log2 transformation and Z-score normalization.

[0118] Phenotypic data: Sensitive strain Y=0, resistant strain=1.

[0119] like Figure 6-10 As shown, construct and solve the hierarchical Bayesian model:

[0120] Model specification: Based on the mathematical formula described in this invention, a HBM consisting of three levels: methylation, mRNA, and protein is constructed.

[0121] Top floor:

[0122] Intermediate layer:

[0123] Bottom layer:

[0124] Prior distribution settings: for all effect coefficients ( Set the Spike-and-Slab prior. For the variance parameter ( ) and weight parameters ( Set no-information or weak-information priors, such as inverse gamma distribution and normal distribution.

[0125] Model Solving: Write model code using open-source software (such as Python or R) and perform posterior sampling using MCMC. Run four chains, each iterating 20,000 times, with 10,000 pre-burns and the last 10,000 iterations used for inference. Check the convergence of the chains (e.g., using Gelman-Rubin diagnostics).

[0126] This invention employs a Bayesian statistical Markov Chain Monte Carlo (MCMC) method to perform posterior estimation of the parameters of the hierarchical model, thereby inferring a robust biomolecular regulatory network. The specific estimation steps are as follows:

[0127] X1. Parameter Initialization

[0128] Set the total number of iterations to Assign initial values ​​to all parameters to be estimated in the model. Specifically:

[0129] The adjustment weight coefficients for each edge Its initial value From a value with zero mean and variance of It is generated by random sampling from a normal distribution.

[0130] The selection indicator variable for representing the existence of edges Its initial value Generated based on the Spike-and-Slab prior distribution.

[0131] X2. Iterative sampling process

[0132] Subsequently, proceed from arrive The iterative loop. In each iteration, for each control edge (including...) , , , Given the parameters, perform the following operations:

[0133] (a) Generating candidate parameters: based on the first The parameter values ​​of the next iteration generate a set of candidate values ​​for the parameters of the current edge.

[0134] To adjust the weighting coefficient Generate candidate values It originates from a value that is currently... The mean and the preset variance are... It was sampled from a normal distribution (proposal distribution).

[0135] To select indicator variables Generate candidate values It is sampled from a Bernoulli distribution (proposal distribution).

[0136] (b) Calculate the joint likelihood: Calculate the joint likelihood of the observed data under the candidate parameters. The joint likelihood probability. This joint likelihood is formed by multiplying the conditional probabilities and prior probabilities of each layer in the model, specifically:

[0137]

[0138] This step quantifies the interpretation of candidate parameters in actual observational data (methylation). mRNA ,protein Drug resistance The rationality of ).

[0139] (c) Calculate the acceptance probability: Calculate the acceptance probability of accepting this set of candidate parameters. The calculation formula is as follows:

[0140]

[0141] in, It is the joint prior probability of the candidate parameters. This is the probability density function of the proposal distribution. This formula ensures that the sampling process tends to accept parameters that improve the model's fit to the data.

[0142] (d) Judgment and Update: Generate a uniform random number between 0 and 1 .like Then, accept candidate parameters and set... , Otherwise, reject the candidate parameter and retain the value from the previous round. , .

[0143] X3. Convergence and Result Extraction

[0144] Convergence diagnosis: After all iterations are completed, in order to ensure that the sampling results are stable and reliable, methods such as the Gelman-Rubin statistic are used to diagnose the convergence of the sampling sequence.

[0145] Calculate the final statistic: Discard the initial "burn-in" sample and calculate using the remaining stable sample:

[0146] Average control weight of each edge: ,in From the start of the firing period to .

[0147] The probability of each edge existing: ,in From the start of the firing period to Here A preset threshold is used to determine the existence of edges.

[0148] X4. Network Construction

[0149] Finally, the final biomolecular regulatory network is constructed using edges whose probability exceeds the threshold, and the nodes include... The weight of an edge is its corresponding average control weight. This network is inferred from multi-omics data D, capable of revealing underlying biological mechanisms and used to predict drug responses. The computational model.

[0150] Based on the above, the construction and visualization of the stereoregulation network are as follows:

[0151] 1. Calculate the posterior inclusion probability (PIP) of the indicator variable corresponding to each regulation relationship from the MCMC sampling results.

[0152] 2. Set the PIP threshold to 0.9.

[0153] 3. Use Cytoscape software for network visualization. Place methylation sites, mRNA, and protein nodes in three parallel virtual layers.

[0154] 3.1. If Then a regulatory edge is drawn between the methylation node and the mRNA. ).

[0155] 2.2. If Then, a regulatory edge is drawn between the mRNA node and the protein node. ).

[0156] 2.3. If Then, a regulatory edge is drawn between the protein node and the phenotype. ).

[0157] 2.4. The final result is the three-dimensional adjustment network shown in the figure.

[0158] Results analysis and verification:

[0159] Key regulatory pathways identified: In the network, a significant regulatory pathway is: "IGF1R gene promoter hypomethylation → IGF1R mRNA hyperexpression → IGF1R protein hyperexpression → promotion of drug resistance." This is consistent with the IGF-1R bypass activation mechanism mentioned in the background art, but this invention further reveals the upstream epigenetic regulatory drivers.

[0160] Hub node analysis: Network topology analysis revealed that, in addition to IGF1R, the transcription factor FOXM1 protein is also a highly connected hub node. It is simultaneously influenced by multiple upstream signals and regulates the expression of multiple downstream genes, ultimately affecting drug resistance. This provides data support for FOXM1 as a target for drug resistance intervention.

[0161] Performance Evaluation: To verify the superiority of the method of this invention, a simulated dataset containing known regulatory relationships was constructed. The HBM method of this invention was compared with traditional Lasso regression and methods that perform differential analysis on each omics data separately. As shown in the table below, the HBM method of this invention exhibits the highest accuracy, recall, and AUC in identifying true regulatory relationships.

[0162] List of performance comparison results of this invention with other multi-omics fusion methods on simulated datasets

[0163]

[0164] In other embodiments, machine learning alternatives can be used: random forests or deep learning can be used to integrate multi-omics data.

[0165] In other embodiments, the Bayesian model can be simplified by integrating only two types of regulatory factors (such as transcription factors + miRNAs).

[0166] In other embodiments, the invention may be based on a correlational network, such as WGCNA, but it cannot distinguish between direct and indirect regulatory relationships.

[0167] The present invention preferably uses a hierarchical Bayesian framework because it balances accuracy and interpretability.

[0168] In other embodiments, the construction of the stereo conditioning network can also be achieved using machine learning-based ensemble methods, instead of methods based on hierarchical Bayesian models. For example, models such as random forests or deep neural networks can be used to perform end-to-end ensemble and pattern recognition of multi-omics data.

[0169] In other embodiments, the construction of the stereoregulatory network can also be achieved using a simplified Bayesian model instead of a full-level hierarchical Bayesian model. For example, a simplified two-layer network model can be constructed by integrating only two types of regulatory factors (such as transcription factors and microRNA).

[0170] In other embodiments, the construction of the stereoregulatory network can also be achieved using correlation-based network analysis methods, instead of methods based on hierarchical Bayesian models. For example, methods such as weighted gene co-expression network analysis (WGCNA) can be used to directly construct molecular association networks based on the statistical correlation of molecular expression levels in multi-omics data.

[0171] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for constructing a multi-omics data stereoregulation network based on a hierarchical Bayesian model, characterized in that, Includes the following steps: S1. Obtain multi-omics data at at least two different molecular levels from the sample; S2. Construct a hierarchical Bayesian model to integrate the multi-omics data and infer the regulatory relationships between molecules. The hierarchical Bayesian model achieves cross-level information coupling by depending the prior probability of the existence of the next level of regulatory relationship on the signal of the previous level of regulatory relationship. S3. Solve the hierarchical Bayesian model and perform posterior inference to obtain the posterior probability of the intermolecular regulatory relationship; S4. Based on the posterior probability, construct a three-dimensional regulatory network containing directed edges for cross-molecular level regulation.

2. The method for constructing a multi-omics data stereoregulation network based on a hierarchical Bayesian model according to claim 1, characterized in that, After step S4, the method further includes: S5, performing biological analysis on the three-dimensional regulatory network to identify key regulatory pathways, core regulatory factors, or key biological functions; wherein the core regulatory factors are identified through network topology analysis, and the key biological functions are identified through functional enrichment analysis.

3. The method for constructing a multi-omics data stereoregulation network based on a hierarchical Bayesian model according to claim 1, characterized in that, Before constructing the hierarchical Bayesian model, the multi-omics data is preprocessed, including standardization, logarithmic transformation, Z-score standardization, and missing value imputation of various types of omics data.

4. The method for constructing a multi-omics data stereoregulation network based on a hierarchical Bayesian model according to claim 1, characterized in that, The hierarchical Bayesian model specifically includes a phenotypic prediction model, a cross-omics regulatory model, and a hierarchical prior distribution of parameters.

5. The method for constructing a multi-omics data stereoregulation network based on a hierarchical Bayesian model according to claim 4, characterized in that, The phenotypic prediction model is a generalized linear model, and its connection function is the Logit function, which is used to associate phenotypic data with molecular features of downstream functional layers.

6. The method for constructing a multi-omics data stereoregulation network based on a hierarchical Bayesian model according to claim 4, characterized in that, The cross-omics regulatory model includes: The protein abundance model models protein abundance as a linear function of its corresponding mRNA expression level. The mRNA expression model models the mRNA expression level as a linear function of upstream epigenetic data.

7. The method for constructing a multi-omics data stereoregulation network based on a hierarchical Bayesian model according to claim 1, characterized in that, The multi-omics data also includes at least one of the following: genome copy number variation data, non-coding RNA expression profile data, and phosphorylated proteomics data; And / or, when constructing the stereo modulation network, the threshold for the posterior inclusion probability is set to 0.5 or higher, and directed edges are drawn between corresponding molecules only when the posterior inclusion probability exceeds the threshold. And / or, the three-dimensional adjustment network uses visualization tools to arrange different types of nodes on different virtual levels to form a three-dimensional network diagram with cross-level control relationships; And / or, the edges or nodes in the stereoregulation network are statistically evaluated using permutation tests and error detection rate correction methods; And / or, the method for constructing multi-omics data stereo conditioning networks based on hierarchical Bayesian models is executed in an online visualization platform; The platform receives the multi-omics data through its data interface and automatically executes steps S1 to S4.

Images