Cancer driver gene identification method based on conflict perception Markov blanket discovery
By using a conflict-aware Markov blanket-based approach, the problems of missing causal features and noise interference in high-dimensional cancer gene data were solved, resulting in more accurate driver gene identification and improved biological interpretability.
Patent Information
- Application Number
- CN202511640047.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-11
- Publication Date
- 2026-02-03
AI Technical Summary
Existing feature selection methods suffer from problems such as omission of causal features, noise interference, and insufficient biological consistency on cancer datasets, making it difficult to accurately identify driver genes in high-dimensional cancer gene expression data.
A conflict-aware Markov blanket-based approach was adopted to screen stable causal genes through preliminary conditional independence tests, delayed pruning, and consistency verification. These genes were then compared with known cancer gene databases to improve identification accuracy.
It significantly reduced the omission rate of causal features, improved the accuracy of driver gene identification and the matching rate with known cancer genes, and enhanced biological consistency.
Smart Images
Figure CN121459938A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bioinformatics, and more particularly to a method for identifying cancer-driving genes based on conflict-aware Markov blanket discovery. Background Technology
[0002] With the deepening research into the molecular mechanisms of cancer, a large amount of high-dimensional data at the genomic and transcriptomic levels has accumulated. How to identify the driver genes that play a causal role in the occurrence and development of cancer from thousands or tens of thousands of candidate genes is a key issue for precision medicine and targeted therapy for cancer.
[0003] Existing feature selection methods have the following problems on cancer datasets:
[0004] 1. Omission of causal features: Under different sets of conditions, candidate genes may exhibit a conflict between conditional independence and dependence, leading to the premature elimination of true cancer driver genes.
[0005] 2. High-dimensional noise interference: Cancer gene expression data has extremely high dimensionality and is accompanied by redundancy and noise. Existing methods are unable to maintain the integrity of the causal structure, resulting in a low detection rate.
[0006] 3. Insufficient biological consistency: The candidate genes identified by traditional methods do not match well with known cancer-related gene databases, affecting their practicality.
[0007] Therefore, how to avoid the omission of causal features due to conditional independence conflicts and noise interference in high-dimensional cancer gene expression datasets, and improve the identification accuracy of driver genes and the matching rate with known cancer genes, is a technical problem that urgently needs to be solved. Summary of the Invention
[0008] The purpose of this invention is to address the above-mentioned problems by providing a cancer driver gene identification method based on conflict-aware Markov blanket discovery that can resolve conditional independence conflicts and improve the accuracy of causal feature recognition.
[0009] To achieve the above objectives, the technical solution of the present invention is as follows:
[0010] A method for identifying cancer-driving genes based on conflict-aware Markov blanket discovery includes the following steps:
[0011] S1. Perform preliminary conditional independence tests on cancer miRNA or mRNA expression data, screen for genes related to cancer phenotype, and obtain the initial candidate parent-child node set ICPCT.
[0012] S2. Detect the independent or dependent behavior of genes in the initial candidate parent-child node set ICPCT under different condition sets, and obtain the candidate parent-child node set CPCT and the candidate spouse node set CSPT according to the delayed pruning strategy.
[0013] S3. Perform multi-condition pruning on the genes in the candidate mate node set CSPT, and obtain the mate set SPT after screening.
[0014] S4. Perform conflict-aware parent-child node filtering on the genes in the candidate parent-child node set CPCT to obtain the parent-child set PCT.
[0015] S5. After integrating the spouse set SPT and the parent-child set PCT, a Markov blanket of the target gene T is obtained. This is then output as the cancer driver gene set, and the output cancer driver gene set is compared with a known cancer gene database to evaluate the matching accuracy.
[0016] Furthermore, step S2 specifically includes the following steps:
[0017] S21. Sort the candidate variable genes in the initial candidate parent-child node set ICPCT according to their dependency strength;
[0018] S22. Select candidate variable gene Y in order of dependency strength, and determine whether candidate variable gene Y and target gene T are independent under condition set Z; if independent, remove candidate variable gene Y from the initial candidate parent-child node set ICPCT; if not independent, keep candidate variable gene Y in the candidate parent-child node set ICPCT.
[0019] S23. Generate a candidate spouse node set CSPT from the candidate parent-child node set ICPCT; generate a candidate parent-child node set CPCT from the candidate variable genes removed from the candidate parent-child node set ICPCT.
[0020] Furthermore, step S3 specifically includes the following steps:
[0021] S31. Traverse the candidate spouse gene X in the candidate spouse node set CSPT, and determine whether the candidate spouse gene X and the target gene T are independent under the required conditions. If they are independent, remove the candidate spouse gene X from the candidate spouse node set CSPT. If they are not independent, proceed to the next step.
[0022] S32. Determine whether candidate spouse gene X is independent of candidate variable gene Y in the initial candidate parent-child node set ICPCT; if independent, remove candidate spouse gene X from the candidate spouse node set CSPT; if not independent, proceed to the next step.
[0023] S33. Determine whether candidate spouse gene X and target gene T are independent under the expansion condition; if they are independent, remove candidate spouse gene X from the candidate spouse node set CSPT; if they are not independent, retain candidate spouse gene X.
[0024] S34. Generate a set of spouses SPT from the remaining candidate spouse genes X.
[0025] Furthermore, step S4 specifically includes the following steps:
[0026] S41. Traverse the parent-child node gene X in the candidate parent-child node set CPCT and determine whether there is a conflicting variable Z. If it exists, correct it and record it to obtain the conflict-corrected candidate set SX. If it does not exist, directly obtain the candidate set SX.
[0027] S42. Select parent-child node genes X in the candidate set SX in turn, and determine whether they depend on the target gene T under a certain condition; if yes, proceed to the next step; if not, remove the parent-child node gene X from the candidate set SX.
[0028] S43. Determine whether there is a dependency or independence contradiction between the parent and child node genes X; if so, remove the parent and child node genes X from the candidate set SX; if not, retain the parent and child node genes X.
[0029] S44. Generate a parent-child set PCT from the retained parent-child node gene X.
[0030] Compared with the prior art, the advantages and positive effects of this invention are:
[0031] 1. This invention uses a "conflict detection + delayed pruning" mechanism to retain genes that exhibit conflict dependence under certain conditions, thereby significantly reducing the omission rate of causal features.
[0032] 2. This invention uses a consistency verification mechanism to retain only genes that exhibit stable causal dependence under multiple conditions, effectively filtering out noise and improving the robustness of high-dimensional data.
[0033] 3. On various cancer datasets of TCGA, the gene set identified by this invention has the highest matching rate with the known cancer driver gene library (NCG 7.2), which is 2%–4% higher than traditional algorithms such as BAMB, effectively improving biological consistency. Attached Figure Description
[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0035] Figure 1 This is the overall framework logic diagram of the present invention;
[0036] Figure 2 A flowchart for identifying the candidate parent-child node set and the candidate spouse node set;
[0037] Figure 3 A flowchart of the pruning process for the set of candidate spouse nodes;
[0038] Figure 4 A flowchart for conflict-aware filtering of candidate parent-child node sets. Detailed Implementation
[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art to all other embodiments obtained without creative effort should be included within the protection scope of the present invention.
[0040] This invention proposes a cancer driver gene identification method based on Conflict-Aware Markov Blanket (CAMB) to identify key genes with driving effects in high-dimensional cancer gene expression datasets. The core idea is to introduce conflict detection and delayed pruning mechanisms into the traditional causal feature selection framework to address the problem of causal feature omission caused by conflicts in conditional independence determination, thereby improving the completeness and accuracy of causal structure identification. It mainly includes the following screening and verification operations:
[0041] 1. Candidate Feature Generation: Perform preliminary conditional independence (CI) tests on cancer miRNA or mRNA expression data to screen for genes significantly associated with cancer phenotypes (tumor / normal state). This yields a candidate parent-child node set (CPCT) and a candidate spouse node set (CSPT).
[0042] Input data: miRNA / mRNA expression matrix, where each row represents a sample and each column represents a gene. The target variable T represents the cancer status.
[0043] Conditional independence test: The chi-square G² statistic is used to test the dependence between genes and cancer status. Essentially, it is a statistical significance test of the probability distribution, representing the causal correlation between gene expression level and cancer phenotype.
[0044] 2. Conflict Detection Mechanism: This mechanism detects contradictory independent / dependent behaviors exhibited by genes under different sets of conditions. A "conflict record set" is created for these conflicting genes to avoid direct rejection during the initial testing, thereby reducing the risk of incorrectly excluding the true driver genes.
[0045] Conflict detection: If a gene is independent of cancer under the condition set S, but is associated with cancer under the extended condition S∪Z, it indicates a statistical dependency conflict. Physically, this means the gene may be an upstream or downstream node of a cancer driver, appearing "falsely independent" after being masked.
[0046] 3. Delayed pruning and consistency verification: A delayed pruning mechanism is used to re-confirm genes with dependency conflicts. Global consistency verification (repetition testing under multiple condition sets) is then used to confirm the robustness of the causal relationship between genes and cancer phenotypes.
[0047] Delayed pruning: to avoid the loss of driver genes due to a single misjudgment, and in physical terms, to verify a stable causal relationship between genes and cancer through multiple combinations of conditions.
[0048] 4. Spouse gene identification and spurious correlation elimination: Multi-stage conditional tests are performed on candidate spouse genes to eliminate noise or spurious correlation genes and retain the real genes related to cancer driving mechanisms.
[0049] Spouse detection: Spouse genes represent causal paths that share child nodes with the target gene T. Physically, this means discovering hidden driving factors through triangular relationships (X→T←Y).
[0050] 5. Driver Gene Output and Validation: Output the final set of cancer driver genes and compare it with a cancer gene database to validate the biological interpretability of the results.
[0051] The overall process is described in detail below, such as Figure 1 As shown, the steps are as follows:
[0052] 1. Cancer data input and preprocessing;
[0053] 2. Identification of candidate parent-child nodes and spouse nodes (e.g.) Figure 2 (as shown)
[0054] 2.1 Initial candidate set generation;
[0055] 2.2 Input the cancer gene expression matrix V and the target variable T;
[0056] 2.3 Perform zero-order conditional independence test to obtain the initial candidate parent-child node set ICPCT;
[0057] 2.4 Sort candidate genes by dependence intensity;
[0058] 2.5 Iterate through each candidate gene Y. If it is independent of T under a certain condition set, delay pruning; otherwise, retain it.
[0059] 2.6 Construct the candidate parent-child node set CPCT and the candidate spouse node set CSPT.
[0060] 3. Multi-condition pruning of spouse nodes (e.g.) Figure 3 (as shown)
[0061] 3.1 Traverse each parent-child node Y;
[0062] 3.2 Perform a three-stage conditional test on its candidate spouse X:
[0063] ① X and T are independent under certain conditions → Remove them;
[0064] ② X and Y are independent → Remove;
[0065] ③ X and T are independent under the extended condition → Remove;
[0066] 3.3 Through the above tests, the simplified spouse set SPT is obtained.
[0067] 4. Conflict-aware parent-child node filtering (e.g.) Figure 4 (as shown)
[0068] 4.1 Initialize the conflict set C;
[0069] 4.2 Traverse each candidate parent-child node X. If there is a conflicting variable Z that causes the conditional independence to be reversed, record it in set C.
[0070] 4.3 Construct the conflict-corrected candidate set SX;
[0071] 4.4 Perform dependency verification on X: If X does not depend on T under any conditions, then remove it;
[0072] 4.5 Perform consistency checks: If X exhibits dependency / independence contradictions under different conditions, it is removed; otherwise, it is retained.
[0073] 4.6 Output the final parent-child set PCT.
[0074] 5. Finally, the parent-child nodes and spouse nodes are integrated to obtain the Markov blanket (MB) of the target variable T. This set is the set of cancer driver genes and is output. This set of genes is compared with the known cancer gene database (NCG 7.2), which can significantly improve the identification rate and biological interpretability of driver genes.
[0075] This invention is mainly implemented using a computer program, which consists of the following functional modules:
[0076] Data input module: Read the cancer gene expression matrix (LIHC, LUAD, COAD, STAD from TCGA) and use cancer status (tumor / normal) as the target variable T.
[0077] Candidate set generation module: Generates an initial set of candidate genes that are significantly correlated with the target variable T by passing the zero-order conditional independence test.
[0078] Conflict detection module: Identifies genes that exhibit independent / dependent contradictions under different sets of conditions and records them in the conflict set C.
[0079] Delayed pruning module: Performs multi-condition verification on genes with conflict dependencies to avoid premature removal due to local errors.
[0080] Spouse identification and correction module: Through a three-stage conditional test, redundant or spurious genes are eliminated, and true causal spouses are retained.
[0081] Results output module: Outputs the set of driver genes and compares them with a cancer database to assess the accuracy of the match.
[0082] Through the combined action of the above modules, a stable restoration of the causal structure is ensured in cancer data analysis.
[0083] The judgment model and formula for computer programs are as follows:
[0084] 1. Definition of conditional independence:
[0085] ;
[0086] Where X and Y are any two gene variables; S is the condition set (which is also a subset of other genes); if the equation holds, then X and Y are considered independent under condition S.
[0087] 2. Conflict condition determination formula:
[0088] ;
[0089] Meaning: If adding a variable Z reverses the causal dependency between T and X, then Z causes a conflict.
[0090] 3. Dependency intensity measure; using the G² statistic:
[0091] ;
[0092] Where: n ijk Under condition S, the observed frequencies of gene X taking value i and T taking value j; the larger the G² value, the more significant the dependency.
[0093] This invention is primarily implemented through a computer program and can run on conventional hardware systems, including:
[0094] Central Processing Unit (CPU): Executes conditional independence checking and conflict detection logic;
[0095] Memory (RAM / hard disk): Stores the cancer gene expression matrix and calculation results;
[0096] Input / output interface: Input gene data file, output driver gene set.
[0097] This invention focuses on protecting the identification method and its application in cancer driver gene identification. The hardware part is not innovative and will not be described in detail.
[0098] This invention is specifically applied to the field of cancer gene data analysis. The technical problem it solves is how to identify truly driving genes in high-dimensional cancer gene data. By introducing a conflict-sensing mechanism, this method can improve the accuracy of driver gene identification and enhance biological interpretability, which can be used to support precision diagnosis of tumors and the development of targeted drugs.
[0099] This invention differs from traditional CI tests that rely solely on a single condition set. By detecting "dependency conflicts" and delaying decisions, it significantly reduces the omission of causal features. Furthermore, through a multi-condition verification mechanism, it ensures that the causal relationships of the retained genes are robust under different conditions, rather than being accidental dependencies. Experiments have shown that this invention can significantly improve the matching rate with known cancer driver genes in various TCGA cancer datasets (LIHC, LUAD, COAD, STAD).
[0100] The present invention has the following beneficial effects:
[0101] 1. Technical Performance: In experiments on five benchmark Bayesian networks and ten real-world datasets, the F1 score, accuracy, and recall of this invention are generally superior to the nine comparative algorithms. In the cancer-related gene identification task on four TCGA datasets (LIHC, LUAD, COAD, STAD), CAMB's performance in matching accuracy with the known oncogene database (NCG 7.2) is as follows: LIHC: 31.43% (3.81 percentage points higher than the second-best algorithm BAMB); LUAD: 13.01% (4.92 percentage points higher than the second-best algorithm CFSMI); COAD: 14.77% (lower than the optimal algorithms IAMB / FBED / EAMB's 33.33%, a difference of 18.56 percentage points); STAD: 20.00% (1.82 percentage points higher than the second-best algorithm BAMB).
[0102] 2. Social Impact: Enhancing the ability to identify cancer driver genes contributes to cancer mechanism research and targeted drug development. It also supports precision medicine and the formulation of personalized treatment plans.
[0103] 3. Economic Benefits: Improves the accuracy and stability of the algorithm, reducing the waste of experimental resources caused by screening errors. It has high application value for pharmaceutical companies and medical research institutions.
[0104] 4. Overall advantages: This invention balances algorithm accuracy and running efficiency, and its running time on some datasets is even better than existing algorithms, demonstrating good scalability and practicality.
Claims
1. A method for identifying cancer driver genes based on conflict-aware Markov blanket discovery, characterized in that: Includes the following steps: S1. Perform preliminary conditional independence tests on cancer miRNA or mRNA expression data, screen for genes related to cancer phenotype, and obtain the initial candidate parent-child node set ICPCT. S2. Detect the independent or dependent behavior of genes in the initial candidate parent-child node set ICPCT under different condition sets, and obtain the candidate parent-child node set CPCT and the candidate spouse node set CSPT according to the delayed pruning strategy. S3. Perform multi-condition pruning on the genes in the candidate mate node set CSPT, and obtain the mate set SPT after screening. S4. Perform conflict-aware parent-child node filtering on the genes in the candidate parent-child node set CPCT to obtain the parent-child set PCT. S5. After integrating the spouse set SPT and the parent-child set PCT, a Markov blanket of the target gene T is obtained. This is then output as the cancer driver gene set, and the output cancer driver gene set is compared with a known cancer gene database to evaluate the matching accuracy.
2. The cancer driver gene identification method based on conflict-aware Markov blanket discovery as described in claim 1, characterized in that: Step S2 specifically includes the following steps: S21. Sort the candidate variable genes in the initial candidate parent-child node set ICPCT according to their dependency strength; S22. Select candidate variable gene Y in order of dependency strength, and determine whether candidate variable gene Y and target gene T are independent under condition set Z; if independent, remove candidate variable gene Y from the initial candidate parent-child node set ICPCT; if not independent, keep candidate variable gene Y in the candidate parent-child node set ICPCT. S23. Generate a candidate spouse node set CSPT from the candidate parent-child node set ICPCT; generate a candidate parent-child node set CPCT from the candidate variable genes removed from the candidate parent-child node set ICPCT.
3. The cancer driver gene identification method based on conflict-aware Markov blanket discovery as described in claim 2, characterized in that: Step S3 specifically includes the following steps: S31. Traverse the candidate spouse gene X in the candidate spouse node set CSPT, and determine whether the candidate spouse gene X and the target gene T are independent under the required conditions. If they are independent, remove the candidate spouse gene X from the candidate spouse node set CSPT. If they are not independent, proceed to the next step. S32. Determine whether candidate spouse gene X is independent of candidate variable gene Y in the initial candidate parent-child node set ICPCT; if independent, remove candidate spouse gene X from the candidate spouse node set CSPT; if not independent, proceed to the next step. S33. Determine whether candidate spouse gene X and target gene T are independent under the expansion condition; if they are independent, remove candidate spouse gene X from the candidate spouse node set CSPT; if they are not independent, retain candidate spouse gene X. S34. Generate a set of spouses SPT from the remaining candidate spouse genes X.
4. The cancer driver gene identification method based on conflict-aware Markov blanket discovery as described in claim 3, characterized in that: Step S4 specifically includes the following steps: S41. Traverse the parent-child node gene X in the candidate parent-child node set CPCT and determine whether there is a conflicting variable Z. If it exists, correct it and record it to obtain the conflict-corrected candidate set SX. If it does not exist, directly obtain the candidate set SX. S42. Select parent-child node genes X in the candidate set SX in turn, and determine whether they depend on the target gene T under a certain condition; if yes, proceed to the next step; if not, remove the parent-child node gene X from the candidate set SX. S43. Determine whether there is a dependency or independence contradiction between the parent and child node genes X; if so, remove the parent and child node genes X from the candidate set SX; if not, retain the parent and child node genes X. S44. Generate a parent-child set PCT from the retained parent-child node gene X.