Biomarker prediction system based on multi-omics data integration, storage medium and method
The biomarker prediction system, which integrates multi-omics data, utilizes the XGBoost and SHAP algorithms to optimize model parameters and combines them with pivot gene screening. This solves the problem of insufficient biomarker discovery caused by single-omics analysis, and achieves efficient and reliable panel screening for biomarker monitoring, thereby improving prediction accuracy and stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING MENTAL HEALTH CENT
- Filing Date
- 2026-01-14
- Publication Date
- 2026-04-21
Smart Images
Figure CN121905281A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of data processing and bioinformatics, and in particular to a biomarker prediction system, storage medium, and method based on multi-omics data integration. Background Technology
[0002] With the deepening of precision medicine and molecular biology research, the efficient and accurate discovery of clinically or biologically significant biomarkers from massive biological data has become crucial for disease mechanism research, drug target development, and personalized medicine. Biomarkers generally refer to molecular indicators that can objectively measure and evaluate normal physiological processes, pathological processes, or responses to therapeutic interventions. Traditional biomarker discovery often relies on single-type omics data, such as analyzing only gene expression profiles or gene mutation profiles. However, this existing approach, which relies on single-omics data, has inherent flaws. These flaws are not simply due to insufficient performance, but rather stem from fundamental methodological limitations. First, complex biological states or disease processes are often driven by multi-level molecular changes, including genomic, transcriptomic, and epigenomic changes. Analyzing from only one dimension makes it difficult to comprehensively and systematically reveal the complete biological mechanisms behind them, resulting in narrow coverage of candidate biomarkers, one-sided biological interpretations, and potential omission of key pathways or network information. In other words, the lack of comprehensiveness in biomarker discovery due to the singularity of the analytical dimension leads to low reliability and stability of the results. Summary of the Invention
[0003] The main technical problem solved by this invention is to provide a biomarker prediction system, storage medium and method based on multi-omics data integration, so as to solve the problem of insufficient comprehensiveness of biomarker discovery caused by the single analysis dimension in the prior art.
[0004] To achieve the above objectives, the technical solution adopted in this application is as follows: Firstly, a biomarker prediction system based on multi-omics data integration, comprising: a data acquisition and preprocessing module: used to acquire multi-omics data samples, classify and label the multi-omics data samples, and output feature data;
[0005] Multi-omics integrated analysis module: used to perform gene mutation, CNV difference, and transcriptome expression profile analysis on the feature data, identify differentially expressed genes related to biomarker status, and obtain a set of differentially expressed genes;
[0006] Monitoring panel selection and model building module: Based on the selected differentially expressed gene set, the XGBoost algorithm is used for model training, and the parameters are optimized by combining grid search and K-fold cross-validation, where K is a positive integer. The SHAP algorithm is used to interpret the feature importance, and the optimal feature combination is selected as the monitoring panel through multiple random repeated training and validation.
[0007] Optionally, the data acquisition and preprocessing module acquires multi-omics data samples including clinical data samples, mutation data samples, CNV data samples, and / or transcriptome data samples; performs state annotation on the clinical data samples, mutation data samples, CNV data samples, and / or transcriptome data samples, and outputs a feature dataset containing annotation information.
[0008] Optionally, the preprocessing operations performed by the data acquisition and preprocessing module include:
[0009] Perform correlation analysis on the clinical information data of the samples;
[0010] Mutation load calculation of genomic mutation data;
[0011] Differential analysis was performed on CNV difference data to screen out CNV genes that showed significant differences between different state groups;
[0012] Transcriptome expression data were standardized and differentially expressed genes were screened.
[0013] Optionally, the XGBoost algorithm is used for model training, and its optimization process includes:
[0014] The hyperparameters of the XGBoost algorithm are optimized by using grid search combined with K-fold cross-validation.
[0015] During model training, multiple iterations with different random seeds are used to evaluate model stability.
[0016] Optionally, the feature selection step using the SHAP algorithm in the monitoring panel screening and model building module includes:
[0017] Calculate the SHAP value of each gene feature for the model prediction result;
[0018] Genes were sorted according to the mean of their SHAP values, and the top N genes in terms of importance were selected.
[0019] The selected Top N genes are combined to construct a feature gene combination that can distinguish different sample states.
[0020] Optionally, the monitoring panel screening and model building module is further used for:
[0021] Multiple prediction models constructed based on different gene combinations were validated, and the model whose performance index reached a predetermined threshold was selected as the optimized model.
[0022] The model performance metrics include the area under the receiver operating characteristic (AUC).
[0023] Optionally, it also includes a hub gene screening module, which is used for:
[0024] Based on the differentially expressed genes, a gene interaction network was constructed using a protein interaction network database;
[0025] Key subnetworks are extracted from the gene interaction network, and hub genes are screened based on a preset algorithm.
[0026] Optionally, the specific methods by which the hub gene screening module screens hub genes include:
[0027] Based on the connectivity of genes in the network, genes with connectivity greater than a preset threshold are selected to form key subnetworks;
[0028] The Boruta feature selection algorithm was used, and after multiple random training sessions, genes that were confirmed a number of times reached a preset threshold were selected as hub genes.
[0029] Secondly, this application provides a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the functions of each module of the biomarker prediction system based on multi-omics data integration.
[0030] Thirdly, a biomarker prediction method based on multi-omics data integration includes the following steps:
[0031] Acquire multi-omics data samples, classify and label the multi-omics data samples, and output feature data;
[0032] Multi-omics integrated analysis was performed on the feature data to identify differentially expressed genes associated with biomarker status;
[0033] Based on the differentially expressed genes, the XGBoost algorithm was used for model training. The model parameters were optimized by combining grid search and cross-validation. The SHAP algorithm was used to interpret the importance of features. Through multiple random repetitions of training and validation, the optimal feature combination was selected as the prediction panel.
[0034] The beneficial effects of this application are: Data acquisition and preprocessing module: used to acquire multi-omics data samples, classify and label the multi-omics data samples, and output feature data;
[0035] Multi-omics integrated analysis module: used to perform gene mutation, CNV difference, and transcriptome expression profile analysis on the feature data, identify differentially expressed genes related to biomarker status, and obtain a set of differentially expressed genes;
[0036] Monitoring panel selection and model building module: Based on the selected differentially expressed gene set, the XGBoost algorithm is used for model training, and the parameters are optimized by combining grid search and K-fold cross-validation, where K is a positive integer. The SHAP algorithm is used to interpret the feature importance, and the optimal feature combination is selected as the monitoring panel through multiple random repeated training and validation.
[0037] By employing the XGBoost algorithm combined with grid search and K-fold cross-validation for model training and optimization, and utilizing the SHAP algorithm to interpret feature importance, overfitting is effectively prevented, and the problem of large fluctuations in model results is solved. The monitoring panel constructed in this invention not only has excellent prediction accuracy but also ensures the model's generalization ability and stability. Furthermore, this method can clearly indicate the key gene features that contribute the most to the prediction results. Overall, it achieves the technical effect of automatically, stably, and interpretably discovering small but precise monitoring panels of key biomarkers from complex multi-omics data. Attached Figure Description
[0038] Figure 1 This is a schematic diagram of the data flow in a biomarker prediction system based on multi-omics data integration.
[0039] Figure 2 This is a schematic diagram of the data flow in a biomarker prediction system based on multi-omics data integration, which includes a hub gene screening module.
[0040] Figure 3 A schematic diagram of the functional modules for monitoring panel filtering and model building;
[0041] Figure 4 A schematic diagram of the functional modules of the hub gene screening module;
[0042] Figure 5 This is a flowchart illustrating the steps of a biomarker pre-method based on multi-omics data integration.
[0043] Figure 6 This is a schematic diagram of the structure of an embodiment of the electronic device of this application. Detailed Implementation
[0044] The present invention will now be described in further detail with reference to specific embodiments and accompanying drawings. Similar elements in different embodiments are referred to by associated similar element reference numerals. In the following embodiments, many details are described to facilitate a better understanding of this application. However, those skilled in the art will readily recognize that some features may be omitted in different situations, or may be replaced by other elements, materials, or methods. In some cases, certain operations related to this application are not shown or described in the specification. This is to avoid obscuring the core parts of this application with excessive description. For those skilled in the art, detailed description of these related operations is not necessary; they can fully understand the related operations based on the description in the specification and general technical knowledge in the art.
[0045] Furthermore, the features, operations, or characteristics described in the specification can be combined in any suitable manner to form various embodiments. At the same time, the steps or actions in the method description can be rearranged or adjusted in a manner obvious to those skilled in the art. Therefore, the various orders in the specification and drawings are only for the clear description of a particular embodiment and do not imply a necessary order, unless otherwise stated that a particular order must be followed.
[0046] The serial numbers assigned to components in this article, such as "first" and "second", are used only to distinguish the objects being described and have no sequential or technical meaning.
[0047] Single-gene analysis often fails to fully reveal complex biological process relationships, and the resulting predictive models lack interpretability and stability. Researchers aimed to discover biomarkers associated with resistance to experimental compound 'Y' in specific cell lines. Single-gene analysis often fails to fully reveal complex biological process relationships, and the resulting predictive models lack interpretability and stability. They divided cell line samples into 'sensitive' and 'resistant' groups (status tags). Using the system of this invention, they integrated and analyzed the transcriptomic and genomic data of the two groups of cells to screen for differentially expressed genes. These genes were used to construct predictive models, ultimately obtaining a 'compound Y resistance prediction panel' consisting of several genes and a list of potential pivot genes. This result is used to guide subsequent molecular biology experiments (such as gene knockout / overexpression) to verify the function of these genes and explore drug resistance mechanisms, rather than for the diagnosis or prognosis of any human or animal disease.
[0048] The biomarker prediction system based on multi-omics data integration described in this application is described in the appendix. Figure 1As shown, it includes: a data acquisition and preprocessing module 10, which is used to acquire multi-omics data samples, classify and label the multi-omics data samples, and output feature data; the multi-omics data samples include clinical data samples, mutation data samples, CNV data samples and transcriptome data samples, and the clinical data samples, the mutation data samples, the CNV data samples and / or the transcriptome data samples are labeled with status, and a feature dataset containing labeling information is output.
[0049] In processing clinical data samples, the data acquisition and preprocessing module first performs factor transformation on categorical variables, then statistically analyzes the frequency distribution of each category, and subsequently uses the Pearson chi-square test and the likelihood ratio chi-square test to analyze the significance of the association between each categorical variable and the recurrence outcome. Simultaneously, the contingency coefficient and Cramer's V coefficient are used to assess the strength of the association. By introducing factor transformation and frequency statistics for categorical variables, combined with the Pearson chi-square test and the likelihood ratio chi-square test, the system can effectively identify clinical features significantly associated with the recurrence outcome and quantify the strength of the association using the contingency coefficient and the Cramer's V coefficient, thereby enhancing the interpretability and reliability of the model. By performing factor transformation on categorical variables and jointly applying the Pearson chi-square test, the likelihood ratio chi-square test, and the contingency coefficient, the system can objectively and quantitatively identify clinical features significantly associated with the sample status, enhancing the reliability and interpretability of clinical dimension features in subsequent model construction.
[0050] In processing mutation data samples, the data acquisition and preprocessing module first uses the tcgaCompare function to statistically compare the overall mutation load of each group of samples. Second, combining BSgenome.Hsapiens.UCSC.hg38, the NMP package, and the COSMIC database, it analyzes the mutation feature profiles of each group based on cosine similarity. Finally, the pathways function systematically identifies the mutation status of key oncogenic pathways in each group. This processing comprehensively reveals the differences in mutation load, feature profiles, and pathway activation patterns, thereby extracting mutation dimensional features closely related to biomarkers and providing high-value mutation biomarker inputs for multi-omics integrated prediction. By comparing the overall mutation load of each group of samples, analyzing the cosine similarity of mutation feature profiles, and identifying key oncogenic pathway mutations, it is possible to systematically extract differentiated mutation tags from three levels: mutation quantity, pattern, and functional pathway, providing rich genomic variation information for integrated analysis.
[0051] In processing CNV data samples, the data acquisition and preprocessing modules label gene states based on CNV difference values. States with a CNV value equal to 2 are labeled "neutral," those not equal to 2 are labeled "change," with values greater than 2 further labeled "gain" and values less than 2 labeled "loss." Subsequently, based on these labeling results, a 2×2 frequency contingency table is constructed, with group as columns and gene CNV state as rows. This provides a clearly structured and directly statistically processed data foundation for subsequent inter-group difference analysis. By standardizing CNV data states and constructing contingency tables, this system can transform continuous CNV difference values into discrete state categories with clear biological significance, thereby significantly improving the data's processability and interpretability in subsequent difference analysis. Transforming complex copy number data into a structured format suitable for statistical hypothesis testing greatly improves the processing efficiency and clarity of results in inter-group difference analysis of CNV data.
[0052] In processing transcriptome data samples, the data acquisition and preprocessing module first divides the sample system into four comparison groups based on clinical outcomes: NR vs Other, BMR vs Other, CR vs Other, and BMCR vs Other. Then, the DESeq2 package is used to standardize the raw read count data and perform differential expression analysis between groups. Finally, genes that meet statistical significance and whose expression changes reach a certain threshold are identified as differentially expressed genes in each comparison group. By standardizing the comparison group division and conducting rigorous differential expression analysis of the transcriptome data, high-throughput and high-reliability screening of gene expression changes significantly associated with different clinical outcomes is possible. The resulting set of differentially expressed genes can serve as important transcriptome biomarkers. The data acquisition and preprocessing module, through standardization and status annotation of clinical, mutation, CNV, and transcriptome data, provides a high-quality, structured feature dataset for subsequent analysis, ensuring the reliability and consistency of the analytical foundation from the outset.
[0053] The multi-omics integration analysis module 20 receives the feature dataset from the data acquisition and preprocessing module. It cleans and standardizes the raw data, assigning a "state label" to each sample based on its source or experimental design, thus outputting a labeled feature dataset. Only samples with clear and consistent state labels are retained for subsequent analysis, providing a grouping basis for later differential analysis. The multi-omics integration analysis module 20 performs gene mutation analysis, CNV differential analysis, and transcriptome expression profiling in parallel or sequentially. For mutation data analysis, it uses the maftools software package to parse whole-exome sequencing data, calculates and compares tumor mutation burden between different state groups using the tmb function, and can draw mutation maps for visual visualization. For CNV differential analysis, it uses the "loss" or "gain" status of labeled genes, employs statistical methods such as the chi-square test to screen for CNV genes showing significant differences between different state groups, and calculates their frequency distribution. Transcriptome data analysis involves standardizing raw read counts using tools such as DESeq2 and screening differentially expressed genes (DEGs) based on set thresholds. Principal component analysis (PCA) dimensionality reduction and gene ontology (GO) and reactorome pathway enrichment analyses can then be performed on these DEGs. The multi-omics integrated analysis module 20 performs differential analysis on genomic mutations, copy number variations, and transcriptome expression profiles in parallel. This multi-dimensional integration strategy overcomes the limitations of single data sources, more comprehensively capturing key molecular features related to biological states, thereby improving the coverage and biological relevance of the candidate biomarker pool.
[0054] Please refer to the appendix. Figure 3As shown, the monitoring panel screening and model building module 30 uses the differentially expressed gene set output by the multi-omics integration analysis module 20 as the initial candidate feature pool. The monitoring panel screening and model building module 30 uses the XGBoost algorithm as the basic classifier architecture prediction model. The basic classifier architecture prediction model process divides the data into training and testing sets. On the training set, the monitoring panel screening and model building module 30 employs a grid search combined with K-fold cross-validation, such as a 10-fold cross-validation strategy, to systematically optimize key hyperparameters in the XGBoost algorithm, such as the learning rate, the maximum tree depth, and the subsample ratio, aiming to obtain the hyperparameter combination that performs best on the cross-validation set. Using grid search combined with K-fold cross-validation to optimize XGBoost hyperparameters can systematically find the parameter combination that may perform best on unknown data, effectively improving the model's generalization performance and preventing underfitting or overfitting due to improper parameter selection. After model optimization, the monitoring panel screening and model building module 30 applies the SHAP algorithm to interpret the contribution of each gene feature to the model prediction results, i.e., calculates its SHAP value. Subsequently, it sorts the genes based on the average absolute value of their SHAP values, selecting the top N genes by importance (e.g., genes with N=5, 8, 10). Based on this, the module freely combines the selected Top N genes to construct multiple candidate gene combinations, i.e., potential monitoring panels. Introducing the SHAP algorithm to calculate feature importance provides consistent and comparable global and local interpretations. This not only helps to select the Top N genes that contribute the most to the model predictions, but more importantly, it makes the model decision-making process transparent, providing biologists with quantitative evidence to understand 'why these genes,' significantly enhancing the biological credibility and acceptability of the selected monitoring panels. For each potential monitoring panel, the monitoring panel screening and model building module 30 retrains the model using the same model training framework, such as one including grid search and cross-validation, and performs performance validation on independent test sets or datasets generated through multiple random and repeated partitions. Finally, the monitoring panel selection and model building module 30 selects gene combinations whose model performance indicators reach a predetermined threshold and exhibit stable performance to determine the final monitoring panel. The model performance indicators include the area under the receiver operating characteristic (AUC). By using different random seeds for multiple iterations of training and validation, and evaluating the stability of the model performance indicators, the robustness of the model and the selected feature combinations can be effectively assessed. This ensures that the final monitoring panel maintains high performance across different data subsets, reducing the randomness of the results.The monitoring panel selection and model building module 30 employs the XGBoost algorithm combined with SHAP for feature importance interpretation. This design not only improves the accuracy of the prediction model but, more importantly, endows the model with interpretability: the SHAP algorithm can quantify the contribution of each gene feature to the prediction result, thereby transforming the "black box" output of machine learning into an understandable ranking of biomarker importance. Based on this, the optimal feature combination is selected through multiple randomized repeated training and validation processes. This process significantly enhances the stability and generalization ability of the final 'monitoring panel', effectively avoiding overfitting.
[0055] The biomarker prediction system based on multi-omics data integration may further include a hub gene screening module 40. Please refer to the appendix. Figure 2 and attached Figure 4 As shown, this module aims to deeply mine the differentially expressed genes (DEGs) obtained from the multi-omics integrated analysis module 20 from the perspective of biological functional associations, in order to identify key genes at the core of the regulatory network. First, the screened list of differentially expressed genes is imported into protein interaction network databases such as STRING to construct an interaction network between the proteins encoded by these genes, thereby revealing the functional associations between genes at the system level. Then, the topological properties of the network are analyzed, for example, genes with connectivity higher than a preset threshold are screened. These highly connected genes collectively constitute key subnetworks in the network, indicating their potential importance in biological processes. To further improve the robustness and reliability of the screening results, this module can also introduce a wrapper feature selection algorithm such as Boruta. This algorithm creates shadow features and repeatedly compares the importance of real features and shadow features in multiple rounds of random training, ultimately screening out genes that are "confirmed" in all training rounds. These "confirmed" genes are identified as highly reliable hub genes. The final output of the hub gene screening module 40 is a list of key hub genes closely related to the sample state. This list of hub genes can provide clear research targets and directions for subsequent in-depth biological experimental verification. Compared to traditional single-omics analysis, this invention provides a systematic framework for computational biomarker discovery by constructing a collaborative system that integrates data acquisition and preprocessing, multi-omics integrated analysis, monitoring panel screening, and model building. The hub gene screening module 40 provides supplementary validation from the perspective of biological functional networks. By constructing a protein-protein interaction network of differentially expressed genes and screening for highly connected nodes, genes at the core of the regulatory network can be identified; these genes often have more important biological functions. Further confirmation is achieved using robust feature selection algorithms such as Boruta. This process, through creating shadow features and performing multiple rounds of comparison, resists random fluctuations and screens out truly relevant and stable core genes, facilitating the derived biological mechanism explanation from the computational screening results.
[0056] The biomarker prediction system based on multi-omics data integration provided in this embodiment achieves automated and standardized screening and validation of monitoring panels and pivot genes for high-value biomarkers from multi-omics data through the collaborative work of multiple modules, thereby improving the overall efficiency and reliability of biomarker discovery.
[0057] Example a: This example aims to demonstrate the discovery of biomarkers associated with differences in the response of cell models to specific experimental stimuli. For example, the method for constructing a specific serum monitoring panel using a biomarker prediction system based on multi-omics data integration is as follows: The data acquisition and preprocessing module is activated, acquiring a batch of experimentally processed sample data from a publicly available omics database, totaling 927 multi-omics data samples. The sample data covers 927 multi-omics data samples including clinical data samples, mutation data samples, CNV data samples, and / or transcriptome data samples. Each sample is assigned a response status label, and the multi-omics data samples are classified and labeled. Samples that meet clinical characteristics are screened, categorical variables such as Gender / Race are statistically analyzed, and a labeled feature dataset is output.
[0058] Next, the multi-omics ensemble analysis module 20 performs parallel analysis on the preprocessed feature dataset to identify molecular features associated with the reaction state. This module performs the following key analyses:
[0059] When analyzing mutation data, the mutation load is analyzed using the maftools software package, the differences between groups are calculated using the tmb function, and a mutation heatmap is plotted.
[0060] In CNV differential analysis, the "loss" or "gain" status of marker genes was used to screen differentially expressed genes using the chi-square test, and the gene frequency distribution was calculated.
[0061] For transcriptome data analysis, the DESeq2 package was used to standardize the counts data, and differentially expressed genes (DEGs) with p < 0.05 and |log2FC| > 1 were screened. PCA dimensionality reduction and GO and Reactome enrichment analyses were then performed to construct a preliminary set of differentially expressed genes.
[0062] The monitoring panel screening and model building module 30 uses the aforementioned differentially expressed genes as a candidate feature pool, and combines the optimized XGBoost algorithm with SHAP interpretive analysis. The specific optimization strategies include:
[0063] Hyperparameter tuning: The optimal hyperparameters are determined by 5-fold cross-validation and grid search. The optimal hyperparameters are AK learning rate, max depth and subsample, etc. The pre-trained model is randomly seeded from 1 to 50 to exclude unsatisfactory results with HL p<0.05 or AUC<0.95.
[0064] Feature selection: Based on the descending order of the mean SHAP value, the top 5 genes were selected to construct a specific serum monitoring panel through independent assortment. Multiple candidate "response prediction biomarker panels" were constructed. Gene combinations with model performance index AUC values of 0.985-0.999 and stable results were selected as the final monitoring panel and the corresponding XGBoost model.
[0065] Furthermore, as an expandable function of this system, the hub gene screening module 40 further mines related core regulatory genes from a biological network perspective. This module imports the list of all DEGs obtained from the multi-omics integrated analysis module 20 into the STRING protein interaction network database to construct the interaction network of the proteins encoded by these genes. The connectivity of each node in the network is calculated, and genes with a Degree > 10 are screened out; these genes constitute the key subnetworks in the network.
[0066] The pivot gene was identified using the STRING protein interaction network and the Boruta algorithm. The specific experimental design included:
[0067] Stability is ensured by 50 random seed iterations, and the model's discrimination and calibration are evaluated using the AUC / HL test. The Boruta algorithm performs multiple rounds of training. In each round, it confirms or rejects a gene whose feature output state is "Confirmed" by comparing the importance of the original features with the randomly shuffled "shaded features." That is, a gene whose importance is consistently significantly higher than that of the shaded features in 50 rounds of random training is identified as a high-confidence pivot gene.
[0068] Gene knockdown and functional verification were performed using the SUP-B15 cell line. The knockdown procedure was as follows: shRNAs of genes such as CD74 / TGFB1 were designed, cloned into the pLKO.1 vector, lentiviruses were packaged in 293FT cells, and SUP-B15 cells were infected with puromycin to screen for stable strains.
[0069] Functional validation experiments: Gene expression levels were detected by qPCR, cell proliferation was continuously monitored using the CCK8 kit, BrdU incorporation rate (S phase ratio) and PI / APC apoptosis rate were analyzed by flow cytometry, and paired t-tests were used for statistical analysis.
[0070] Boruta algorithm robustness screening: Using the expression matrix of the aforementioned DEGs as features and sample state labels as target variables, the Boruta algorithm is run with the maxRuns parameter set to 50. The algorithm undergoes multiple training rounds. In each round, it confirms or rejects a feature by comparing the importance of the original features with randomly shuffled "shadow features." Ultimately, genes with an output state of "Confirmed"—those whose importance consistently and significantly exceeds that of the shadow features across 50 rounds of random training—are identified as high-confidence pivotal genes. This embodiment ultimately outputs a high-precision computational prediction model and a set of potential key biological genes, providing clear candidate targets for subsequent molecular biology mechanism research.
[0071] This invention divides cell line samples into a 'sensitive group' and a 'resistant group'. Using the system of this invention, transcriptomic and genomic data of the two groups of cells are integrated and analyzed to screen for differentially expressed genes. Predictive models are constructed using these genes, ultimately obtaining a 'compound Y resistance prediction panel' consisting of several genes and a list of potential pivot genes. These results are used to guide subsequent molecular biology experiments to verify the function of these genes and further explore drug resistance mechanisms.
[0072] This example demonstrates the application of a biomarker prediction system based on multi-omics data integration in a cell line sensitivity study scenario. Using this system, researchers can efficiently converge a high-precision prediction panel consisting of a small number of genes from multi-omics data, while simultaneously obtaining a list of potential pivot genes. This significantly narrows down the range of candidate genes that would otherwise require extensive trial and error and validation, providing clear and prioritized targets for subsequent design of targeted gene knockout / overexpression experiments, thereby accelerating the molecular biology research process of drug action mechanisms.
[0073] This application also provides a biomarker prediction method based on multi-omics data integration. Please refer to the appendix. Figure 5 As shown, it includes the following steps:
[0074] S1: Acquire multi-omics data samples, classify and label these samples, and output feature data. Here, "multi-omics data samples" broadly refers to a multi-dimensional data set encompassing different molecular levels, such as genomics, transcriptomics, epigenetics, and proteomics, generated from the same biological sample or cohort. By cleaning, standardizing, and assigning "state labels" related to the biological state to these raw data, this step provides a well-structured and quality-controlled feature dataset for subsequent analysis.
[0075] S2: Perform multi-omics integrated analysis on the feature data to identify differentially expressed genes associated with biomarker states. This step involves performing differential analysis on features from different omics levels in parallel or sequentially to identify molecular features, particularly differentially expressed genes, that show significant statistical differences between different state groups. This multi-dimensional integration strategy overcomes the limitations of a single omics data perspective, more comprehensively capturing key molecular events associated with the target biological state, thereby forming an initial set of candidate biomarkers with broader coverage and stronger biological relevance.
[0076] S3: Based on the differentially expressed genes, the XGBoost algorithm is used for model training. Grid search and cross-validation are combined to optimize model parameters. The SHAP algorithm is used to interpret feature importance. Through multiple randomized repeated training and validation, the optimal feature combination is selected as the prediction panel. In this step, the differentially expressed gene set is used as the feature pool, and the XGBoost algorithm is used as the basic classifier to build the prediction model. Grid search and K-fold cross-validation are used to systematically optimize the model's hyperparameters, aiming to improve the model's generalization performance on unknown data and effectively prevent overfitting. After model training, the SHAP algorithm is introduced to quantify the contribution of each gene feature in the model, thus making the model's prediction decision process transparent. Based on the ranking of feature SHAP values, several genes with the highest importance are selected. Through a robustness evaluation process of multiple randomized repeated training and validation, the optimal feature combination that performs well and stably on the independent validation set is finally determined, which is the final "prediction panel". This method achieves the technical effect of automatically, accurately, and interpretably discovering small but precise combinations of key biomarkers from multi-omics big data, providing an efficient and reliable computational tool for precision medicine and biomarker research.
[0077] This application also provides an electronic device, such as... Figure 6 The diagram shown is a structural schematic of an electronic device according to an embodiment of this application, including: a processor 1401, a memory 1402, and a bus. The memory 1402 stores machine-readable instructions executable by the processor 1401. For example, Figure 1 The device contains data acquisition and preprocessing modules, monitoring panel screening and model building modules 30, and hub gene screening modules 40, and the corresponding execution instructions. When the computer device is running, the processor 1401 and the memory 1402 communicate via a bus. When the machine-readable instructions are executed by the processor 1401, they perform the processing of a biomarker prediction system based on multi-omics data integration.
[0078] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the biomarker prediction system based on multi-omics data integration described above.
[0079] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.
[0080] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. If the functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.
[0081] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
[0082] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.
Claims
1. A biomarker prediction system based on multi-omics data integration, characterized in that, include: Data acquisition and preprocessing module: used to acquire multi-omics data samples, classify and label the multi-omics data samples, and output feature data; Multi-omics integrated analysis module: used to perform gene mutation, CNV difference, and transcriptome expression profile analysis on the feature data, identify differentially expressed genes related to biomarker status, and obtain a set of differentially expressed genes; Monitoring panel selection and model building module: Based on the selected differentially expressed gene set, the XGBoost algorithm is used for model training, and the parameters are optimized by combining grid search and K-fold cross-validation, where K is a positive integer. The SHAP algorithm is used to interpret the feature importance, and the optimal feature combination is selected as the monitoring panel through multiple random repeated training and validation.
2. The biomarker prediction system based on multi-omics data integration according to claim 1, characterized in that: The data acquisition and preprocessing module acquires multi-omics data samples including clinical data samples, mutation data samples, CNV data samples, and / or transcriptome data samples; it performs state annotation on the clinical data samples, mutation data samples, CNV data samples, and / or transcriptome data samples, and outputs a feature dataset containing annotation information.
3. The biomarker prediction system based on multi-omics data integration according to claim 2, characterized in that, The preprocessing operations performed by the data acquisition and preprocessing module include: A correlation analysis was performed on the clinical information data of the clinical data sample. Mutation load calculation is performed on the mutation data samples; Differential analysis was performed on the CNV data samples to screen out CNV genes that showed significant differences between different state groups; The transcriptome data samples were standardized, and differentially expressed genes were screened.
4. The biomarker prediction system based on multi-omics data integration according to claim 1, characterized in that, The XGBoost algorithm is used for model training, and its optimization process includes: The hyperparameters of the XGBoost algorithm are optimized by combining grid search with K-fold cross-validation. During model training, multiple iterations with different random seeds are used to evaluate model stability.
5. The biomarker prediction system based on multi-omics data integration according to claim 4, characterized in that, The steps of feature selection using the SHAP algorithm in the monitoring panel screening and model building module include: Calculate the SHAP value of each gene feature for the model prediction result; Genes were sorted according to the mean of their SHAP values, and the top N genes in terms of importance were selected. The selected Top N genes are combined to construct a feature gene combination that can distinguish different sample states.
6. The biomarker prediction system based on multi-omics data integration according to claim 5, characterized in that, The monitoring panel screening and model building module is further used for: Multiple prediction models constructed based on different gene combinations were validated, and the model whose performance index reached a predetermined threshold was selected as the optimized model. The model performance metrics include the area under the receiver operating characteristic (AUC).
7. The biomarker prediction system based on multi-omics data integration according to claim 1, characterized in that, It also includes a hub gene screening module, which is used for: Based on the differentially expressed genes, a gene interaction network was constructed using a protein interaction network database; Key subnetworks are extracted from the gene interaction network, and hub genes are screened based on a preset algorithm.
8. The biomarker prediction system based on multi-omics data integration according to claim 7, characterized in that, The specific methods by which the hub gene screening module screens hub genes include: Based on the connectivity of genes in the network, genes with connectivity greater than a preset threshold are selected to form key subnetworks; The Boruta feature selection algorithm was used, and after multiple random training sessions, genes that were confirmed a number of times reached a preset threshold were selected as hub genes.
9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the functions of each module of the biomarker prediction system based on multi-omics data integration as described in any one of claims 1 to 8.
10. A biomarker prediction method based on multi-omics data integration, characterized in that, Includes the following steps: Acquire multi-omics data samples, classify and label the multi-omics data samples, and output feature data; Multi-omics integrated analysis was performed on the feature data to identify differentially expressed genes associated with biomarker status; Based on the differentially expressed genes, the XGBoost algorithm was used for model training. The model parameters were optimized by combining grid search and cross-validation. The SHAP algorithm was used to interpret the importance of features. Through multiple random repetitions of training and validation, the optimal feature combination was selected as the prediction panel.