Feature gene selection method and system based on deep learning attribution analysis and beam combination optimization

Abstract

The invention provides a feature gene selection method and system based on deep learning attribution analysis and beam combination optimization, and relates to the technical field of signal analysis, and the method comprises the following steps: obtaining single cell RNA sequencing data and batch RNA sequencing data; the method comprises the following steps: based on single-cell RNA sequencing data, obtaining an initial gene pool aiming at a plurality of disease subtypes through a consensus screening strategy fusing XGBoost multi-dimensional importance measurement and deep learning SHAP attribution analysis; performing first-stage optimization on the gene list of each subtype by taking the initial gene pool as a starting point and adopting beam combination search and based on a global discrimination objective function to obtain a first-stage optimized gene set; integrating the first-stage optimized gene set with the differential expression gene set of the batch RNA sequencing data, and carrying out second-stage optimization by adopting bundle combination search again to obtain a second-stage optimized gene set; and performing cross validation integration on the second-stage optimized gene set, and outputting a final feature gene set.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics analysis technology, and in particular to a method and system for selecting feature genes based on deep learning attribution analysis and bundle combinatorial optimization. Background Technology

[0002] Disease molecular subtyping is a core component of precision medicine, aiming to classify patient populations into subgroups with significantly different clinical phenotypes, disease progression risk, or treatment response based on differences in molecular characteristics. In recent years, the rapid development of high-throughput sequencing technologies (such as single-cell RNA sequencing and batch RNA sequencing) has provided high-resolution data resources for understanding disease heterogeneity.

[0003] However, the massive amounts of high-dimensional data generated by these technologies pose a serious challenge to feature selection methods. Genome-wide scale analysis (which typically covers more than 20,000 genes) is not only computationally complex, but also introduces a large amount of noise, leading to model overfitting and decreased generalization performance.

[0004] Traditional feature selection methods mainly include three strategies: filtering methods quickly screen genes based on statistical indicators (such as differential expression analysis), but ignore the combinatorial effects between genes; wrapping methods optimize classification performance through subset search, but have high computational complexity and are difficult to implement in high-dimensional spaces; and embedded methods (such as LASSO) select features during the modeling process, but their linear assumptions limit their applicability in complex biological mechanisms.

[0005] Furthermore, existing methods often focus on single-modal data (such as using only scRNA-seq or bulk RNA-seq) and lack a universal framework for effectively integrating multi-omics data, making it difficult to fully capture the complex synergistic regulatory networks and biological mechanisms behind disease subtypes. Summary of the Invention

[0006] To address at least one of the technical problems in the background art, this invention, based on a consensus screening strategy of multi-model fusion, a two-stage bundled combinatorial optimization framework, and a cross-modal data integration mechanism, solves the technical problems of feature selection bias, combinatorial explosion, and insufficient utilization of multi-omics data in high-dimensional genomic data processing.

[0007] The first aspect of this invention provides a feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization, comprising: Acquire single-cell RNA sequencing data and batch RNA sequencing data; Based on single-cell RNA sequencing data, an initial gene pool targeting multiple disease subtypes was obtained by integrating XGBoost multidimensional importance measurement and deep learning SHAP attribution analysis consensus screening strategy. Starting from the initial gene pool, a bundle combinatorial search is used and the gene list of each subtype is optimized in the first stage based on the global discriminant objective function to obtain the first-stage optimized gene set; The first-stage optimized gene set was integrated with the differentially expressed gene set from the batch RNA sequencing data, and then a bundle combinatorial search was used for the second-stage optimization to obtain the second-stage optimized gene set. The optimized gene set from the second stage is cross-validated and integrated to output the final feature gene set.

[0008] Furthermore, the consensus screening strategy includes: For each disease subtype, candidate differentially expressed genes were initially selected based on differential expression analysis using single-cell RNA sequencing data. An XGBoost model and a deep neural network model are trained. Gain, frequency, and coverage importance metrics are extracted from the XGBoost model and then standardized and weighted to obtain a first importance score. The SHAP value of each gene is calculated from the deep neural network model to obtain a second importance score. Based on the ranking of the first importance score and the second importance score, consensus genes are selected to form the initial gene pool.

[0009] Furthermore, the first stage of optimization includes: For each disease subtype, the starting point is the list of genes for that disease subtype in the initial gene pool; The number of genes in the gene combination is gradually increased through an iterative approach, at each iteration level: New gene combinations are generated based on the candidate combinations at the previous level; The gene list for other disease subtypes is fixed, and the discriminative ability of the newly generated gene combinations is evaluated using the global discriminative objective function. Based on the evaluation results, a preset number of optimal combinations are retained as candidate combinations for the current level; After all iterations are completed, the combination that maximizes the global discriminative objective function value is selected from the candidate combinations at all levels, and the gene list of the disease subtype is updated to obtain the first-stage optimized gene set.

[0010] Furthermore, the second phase of optimization includes: For each disease subtype, obtain the differentially expressed gene set from the batch RNA sequencing data; Based on the gene list of the disease subtype in the first-stage optimized gene set, the genes in the differentially expressed gene set are combined and expanded iteratively. In each iteration level: The newly added differentially expressed genes are merged with the gene combinations obtained in the previous round of optimization to generate new candidate combinations; The gene list for other disease subtypes is fixed, and the discriminative ability of the newly generated candidate combinations is evaluated using the global discriminative objective function; Based on the evaluation results, retain the optimal combination of the preset quantity; After all iterations are completed, the combination that maximizes the global discriminative objective function value is selected from all candidate combinations, and the gene list of the disease subtype is updated to obtain the second-stage optimized gene set.

[0011] Furthermore, the cross-validation integration includes: The entire process of consensus screening, first-stage optimization, and second-stage optimization is repeated multiple times within the K-fold cross-validation framework. Each execution yields a set of characteristic genes; Integrate all the feature gene sets obtained from the execution, select the gene with the highest frequency in each disease subtype, and form the final feature gene set.

[0012] Furthermore, the global discriminant objective function is used to quantify the discriminant ability of a gene set for multiple disease subtypes. The objective function calculates the average discriminant advantage of positive samples and the average discriminant advantage of negative samples based on the difference between the scores of each sample on the target subtype gene set and the scores on other subtype gene sets. The gene set is optimized by maximizing the average discriminant advantage of positive samples and minimizing the average discriminant advantage of negative samples.

[0013] A second aspect of the present invention provides a feature gene selection system based on deep learning attribution analysis and bundle combinatorial optimization, comprising: The data acquisition module is used to acquire single-cell RNA sequencing data and batch RNA sequencing data; The initial screening module is used to obtain an initial gene pool for multiple disease subtypes based on single-cell RNA sequencing data by using a consensus screening strategy that integrates XGBoost multidimensional importance measurement and deep learning SHAP attribution analysis. The optimization module is used to start from the initial gene pool, perform a first-stage optimization of the gene list of each subtype using bundle combinatorial search and based on a global discriminant objective function, to obtain a first-stage optimized gene set; integrate the first-stage optimized gene set with the differentially expressed gene set of batch RNA sequencing data, and perform a second-stage optimization again using bundle combinatorial search, to obtain a second-stage optimized gene set; The integration output module is used to perform cross-validation integration on the optimized gene set of the second stage and output the final feature gene set.

[0014] A third aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps in the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization as described in the first aspect of the present invention.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization as described in the first aspect of the present invention.

[0016] A fifth aspect of the present invention provides a computer program product comprising software code, wherein the program in the software code performs the steps of the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization as described in the first aspect of the present invention.

[0017] Compared with existing technologies, the feature gene selection method and system based on deep learning attribution analysis and bundle combinatorial optimization provided by this invention have the following beneficial effects: (1) This invention provides a multi-model fusion consensus screening strategy. By integrating XGBoost multi-dimensional importance measurement and deep learning SHAP attribution analysis, it realizes cross-validation of gene importance from different modeling perspectives, thereby improving the accuracy and reliability of initial gene screening.

[0018] (2) Based on a two-stage bundle combinatorial optimization framework, this invention generates candidate combinations iteratively and filters them based on a global discriminant objective function, thereby achieving efficient directional search in high-dimensional space and balancing search breadth and computational efficiency.

[0019] (3) Based on the cross-modal data integration mechanism, this invention optimizes single-cell data and batch RNA sequencing data in stages, realizes the complementary advantages of multi-omics data, and improves the biological interpretability of the feature gene set.

[0020] (4) Based on the cross-validation integration strategy, this invention ensures the robustness and generalization ability of the final feature gene set by repeatedly optimizing the process and integrating high-frequency genes. Attached Figure Description

[0021] The accompanying drawings, which form part of this disclosure, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0022] Figure 1 This is the overall flowchart of the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the Stage I bundle search optimization of the gene list for subtype i provided in Embodiment 1 of the present invention; Figure 3 This is a flowchart of Stage II integrated Bulk data optimization provided in Embodiment 1 of the present invention; Figure 4 This is a flowchart illustrating the specific implementation of the technology provided in Embodiment 1 of the present invention. Detailed Implementation

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. Furthermore, it should be understood that the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] All data acquisition in this embodiment is carried out in accordance with laws and regulations and with user consent, and the data is used legally.

[0027] Terminology Explanation: Bundle width: In optimization algorithms, the bundle width represents the number of candidate solutions retained in each iteration. and The beam widths of the first and second stages are controlled separately.

[0028] Example 1 This embodiment provides a feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization, including: Acquire single-cell RNA sequencing data and batch RNA sequencing data; Based on single-cell RNA sequencing data, an initial gene pool targeting multiple disease subtypes was obtained by integrating XGBoost multidimensional importance measurement and deep learning SHAP attribution analysis consensus screening strategy. Starting from the initial gene pool, a bundle combinatorial search is used and the gene list of each subtype is optimized in the first stage based on the global discriminant objective function to obtain the first-stage optimized gene set; The first-stage optimized gene set was integrated with the differentially expressed gene set from the batch RNA sequencing data, and then a bundle combinatorial search was used for the second-stage optimization to obtain the second-stage optimized gene set. The optimized gene set from the second stage is cross-validated and integrated to output the final feature gene set.

[0029] This invention constructs a multi-stage, progressive optimization pipeline. First, it acquires multimodal sequencing data (single-cell and batch RNA sequencing) to provide a rich information foundation for analysis. Then, it rapidly narrows down the candidate gene pool by strengthening the initial screening, laying the foundation for subsequent computationally intensive optimization. Furthermore, it utilizes the high resolution of single-cell data and the robustness of batch data for in-depth analysis through two-stage optimization. Finally, it ensures the universality of the final results through cross-validation integration.

[0030] Through this coherent process, the present invention decomposes the complex problem of high-dimensional feature selection into multiple manageable sub-tasks, thereby achieving the technical effect of efficiently and robustly selecting small-scale, highly discriminative feature gene sets from massive genes.

[0031] Specifically, the consensus screening strategy includes: For each disease subtype, candidate differentially expressed genes were initially selected based on differential expression analysis using single-cell RNA sequencing data. An XGBoost model and a deep neural network model are trained. Gain, frequency, and coverage importance metrics are extracted from the XGBoost model and then standardized and weighted to obtain a first importance score. The SHAP value of each gene is calculated from the deep neural network model to obtain a second importance score. Based on the ranking of the first importance score and the second importance score, consensus genes are selected to form the initial gene pool.

[0032] This invention introduces a multi-model, multi-perspective feature importance fusion evaluation mechanism. Specifically, the XGBoost model can provide statistics (gain, frequency, coverage) reflecting the importance of features in split nodes based on a tree model structure, while deep neural networks (DNNs) combined with SHAP attribution analysis can extract the marginal contribution of each feature to the prediction results from complex nonlinear models. Ranking and integrating these two evaluation methods, which originate from different theoretical foundations, is equivalent to conducting an "expert review".

[0033] The aforementioned fusion evaluation mechanism addresses the issues of traditional filtering methods neglecting gene combination effects and potential biases in single-model evaluations. This consensus strategy identifies candidate genes that are important across multiple machine learning paradigms, demonstrating greater universality and robustness. This provides a high-quality "seed" gene pool for subsequent optimization, effectively reducing the search space and improving the quality of the starting point.

[0034] Specifically, the first stage of optimization includes: For each disease subtype, the starting point is the list of genes for that disease subtype in the initial gene pool; The number of genes in the gene combination is gradually increased through an iterative approach, at each iteration level: New gene combinations are generated based on the candidate combinations at the previous level; The gene list for other disease subtypes is fixed, and the discriminative ability of the newly generated gene combinations is evaluated using the global discriminative objective function. Based on the evaluation results, a preset number of optimal combinations are retained as candidate combinations for the current level; After all iterations are completed, the combination that maximizes the global discriminative objective function value is selected from the candidate combinations at all levels, and the gene list of the disease subtype is updated to obtain the first-stage optimized gene set.

[0035] This invention employs a controlled, iterative beam search algorithm. Starting with a single gene, it constructs more complex gene combinations by iteratively increasing the number of genes in the combination layer by layer. At each layer, it does not exhaust all possibilities, but rather, based on the evaluation result of the global discriminative objective function, it retains only a predetermined number of currently optimal candidate combinations for the next round. This is an effective strategy for balancing the breadth and depth of the search.

[0036] Based on the aforementioned bundle search algorithm, the combinatorial explosion problem faced by wrapper-style methods in high-dimensional spaces is solved. Bundle search enables targeted exploration of a vast combinatorial space, avoiding the impossibility of searching the entire space, and is less prone to getting trapped in local optima compared to greedy algorithms. Ultimately, this step can identify the most effective locally optimal gene combinations for distinguishing specific subtypes from the initial gene pool, thus significantly improving the discriminative power of the feature gene set.

[0037] Specifically, the second phase of optimization includes: For each disease subtype, obtain the differentially expressed gene set from the batch RNA sequencing data; Based on the gene list of the disease subtype in the first-stage optimized gene set, the genes in the differentially expressed gene set are combined and expanded iteratively. In each iteration level: The newly added differentially expressed genes are merged with the gene combinations obtained in the previous round of optimization to generate new candidate combinations; The gene list for other disease subtypes is fixed, and the discriminative ability of the newly generated candidate combinations is evaluated using the global discriminative objective function; Based on the evaluation results, retain the optimal combination of the preset quantity; After all iterations are completed, the combination that maximizes the global discriminative objective function value is selected from all candidate combinations, and the gene list of the disease subtype is updated to obtain the second-stage optimized gene set.

[0038] Building upon the optimized gene set obtained in the first phase, differentially expressed genes from bulk RNA sequencing data are introduced for cross-data modality integration and expansion. This process optimizes the first phase, but its core lies in incremental optimization—exploring the possibility of incorporating genes from new data sources without discarding the results of the first phase. Candidate genes from the bulk data are used as new elements and merged with the optimal combinations obtained in the first phase. The efficacy of these new combinations is then evaluated again using bundle search.

[0039] This technology addresses the lack of a universal framework for effectively integrating multi-omics data in existing technologies. The second-stage optimization does not simply merge data from different sources; instead, it employs a structured combinatorial optimization process that organically combines the high-resolution features of single-cell data with the population-level robust features of batch data. This leverages the complementary advantages of multimodal data and holds promise for discovering characteristic genes with greater biological consistency and clinical application potential.

[0040] Specifically, the cross-validation integration includes: The entire process of consensus screening, first-stage optimization, and second-stage optimization is repeated multiple times within the K-fold cross-validation framework. Each execution yields a set of characteristic genes; Integrate all the feature gene sets obtained from the execution, select the gene with the highest frequency in each disease subtype, and form the final feature gene set.

[0041] The entire feature selection process described above (from initial screening to two-stage optimization) is repeatedly executed within a stable evaluation framework—K-fold cross-validation. Each execution is performed on a different subset of data. Finally, the most stable and frequently selected genes are chosen by statistically analyzing the frequency of each gene being selected in the final set across all runs. This solves the common overfitting problem in machine learning, as well as the instability of feature selection results. Through multiple repetitions and frequency statistics, unstable genes that might be selected due to random fluctuations in the training data are effectively filtered out, ensuring that the final output feature gene set possesses high robustness and reproducibility, thereby significantly improving its generalization performance on independent datasets.

[0042] Specifically, the global discriminant objective function is used to quantify the discriminant ability of a gene set for multiple disease subtypes. The objective function calculates the average discriminant advantage of positive samples and the average discriminant advantage of negative samples based on the difference between the scores of each sample on the target subtype gene set and the scores on other subtype gene sets. The gene set is optimized by maximizing the average discriminant advantage of positive samples and minimizing the average discriminant advantage of negative samples.

[0043] This invention employs a design philosophy that maximizes inter-class discriminative power. It focuses not only on whether a sample scores highly on its true subtype, but more importantly, on whether that score is significantly higher than its highest score across all other subtypes. By calculating the average discriminant power of positive samples (samples belonging to the subtype) and the average discriminant power of negative samples (samples not belonging to the subtype), and setting the objective to maximize the former and minimize the latter, the performance of the gene set in distinguishing different subtypes is optimized. This forces the optimization process to find gene combinations that clearly separate different subclass samples in the feature space, thereby ensuring that the final feature gene set has extremely strong class discriminative power, closely aligned with the ultimate application goal of molecular subtyping.

[0044] In one specific embodiment, the present invention systematically integrates complementary information from single-cell RNA sequencing (scRNA-seq) and bulk RNA sequencing (bulk RNA-seq) data by constructing a coherent multi-stage optimization framework, and innovatively introduces deep learning interpretability technology (SHAP) to guide high-quality initial gene screening. Finally, an efficient beam search is guided by a carefully designed objective function aimed at maximizing cross-subtype discriminancy, thereby selecting a small-scale, highly discriminative, and highly robust set of optimal feature genes.

[0045] The core of this invention lies in a complete technical framework consisting of enhanced initial screening, two-stage optimization, and cross-validation integration, the overall process of which is as follows: Figure 1 As shown.

[0046] Phase 1: Strengthening Initial Screening First, in the enhanced initial screening stage, this invention abandons the traditional approach of solely relying on differential expression analysis and proposes a multi-model fusion consensus screening strategy. For each disease subtype, a large number (e.g., 500) of candidate differentially expressed genes are initially selected based on scRNA-seq data using non-parametric tests (such as the Wilcoxon test). Then, a binary classification problem (target subtype vs. other subtypes) is constructed, and XGBoost and deep neural network (DNN) models are trained separately.

[0047] From the XGBoost model, three importance metrics—Gain, Weight, and Cover—are extracted, and then standardized and weighted to obtain a comprehensive importance score. The formula is as follows: set up , , They represent genes respectively. The three categories of original importance scores, after standardization, are as follows:

[0048] Then the overall score It can be represented as:

[0049] in Let be the weight coefficient, and satisfy... . Meanwhile, for the DNN model, the KernelExplainer from SHAP (SHapley Additive exPlanations) is used to calculate the contribution of each feature gene to the model's predicted output. Another independent importance metric is obtained by calculating the mean absolute value of the SHAP values ​​for each gene. .

[0050] Ultimately, through integration Ranking and The rankings were determined, and consensus genes that ranked highly in both evaluation systems were selected to form a high-quality initial gene pool for each subtype. ,in Representing the An initial gene list for each subtype.

[0051] Phase Two: Two-Phase Optimization Process Next, the core two-stage optimization process begins. The optimization process of this invention is based on a carefully designed global discriminant objective function. Driven by [the function name]. This function is used to quantize a given set of genes. The overall discriminative ability of [the system] is defined as follows:

[0052] in Represents a gene set. Represents the total sample set. Indicates subtype The sample set, This indicates the total sample excluding subtypes. The sample set, Quantization subtype The discriminative advantage of gene lists, Indicates the target subtype The gene list is as follows: 1. Determine the dominant quantity :

[0053] Indicates sample In target subtype Gene set score The difference between the highest score (calculated using a scanpy-like scoring function) and the scores of all other subtypes. ,but More likely to belong to a subtype (Correct classification); If ,but It is more likely to belong to other subtypes, indicating a classification error. Quantified the sample Correctly identified as a subtype Rather than the degree of dominance of other subtypes.

[0054] 2. Advantages of positive sample discrimination:

[0055] Indicates subtype All real samples The average discriminant advantage. Maximizing this value can improve the target subtype. The positive sample scores were significantly higher than those of other subtypes.

[0056] 3. Negative samples are used to identify disadvantages:

[0057] Indicates non-subtype All samples The average discriminant advantage. Minimizing this value can reduce the target subtype. Interference scores on other subtype samples.

[0058] The indicator requires that, for any subtype s: 1) s-type samples score significantly higher on the s list than on other subtype lists; 2) non-s-type samples score significantly lower on the s list than on their corresponding true subtype lists. This indicator emphasizes the relative discriminative power between gene lists rather than absolute score levels.

[0059] In summary, The core idea of ​​the function is to maximize the average discriminative advantage of positive samples while minimizing the average discriminative advantage of negative samples (i.e., reducing interference), thereby forcing the optimized gene set to be optimized. It has a strong ability to distinguish between different subtypes.

[0060] Phase Optimization The first stage of optimization (see its search strategy) Figure 2 The initial gene set obtained by single-cell differential expression analysis As the starting point (where Indicates subtype The initial gene list, (This is the length of the list), and the goal is to optimize the gene list for each subtype based on scRNA-seq data.

[0061] Specifically, the algorithm employs a greedy strategy to optimize the gene list subtype by subtype. Initialize the single-gene combination set The iterative process starts from to Each round generates a new combination. , Fixed list of other subtypes ,Evaluate And filter the top with the highest discrimination. Combinations constitute After the iteration terminates, select candidates from all levels of the candidate pool. Choose the global optimal solution Update the subtype list to create an optimized gene set. .

[0062] Phase Optimization The second phase of optimization (see the data integration process for details) Figure 3 The aim is to integrate differentially expressed gene sets from bulk RNA-seq data. Integrate into the optimized gene set (of which Indicates subtype The initial gene list, (Length of the list). For subtypes Initialize the single-element set of newly added genes. and inherit the results of the first phase. The iterative expansion process starts from... to Each round generates combinations ,fixed Evaluate And retain the previous The optimal combination Ultimately from Select Update the list and output. .

[0063] This framework selects the globally optimal solution from candidates across all levels to update the gene list. Furthermore, it can be extended to multi-omics data such as epigenomics and proteomics, revealing synergistic regulatory mechanisms through cross-modal combinatorial optimization.

[0064] Phase 3: Cross-validation and final integration Finally, to ensure the robustness of the results, this invention places the entire process within a hierarchical K-fold cross-validation framework for repeated execution. The final output gene set is the set of genes that appear most frequently in all cross-validation folds, thereby ensuring the stability and reliability of its feature selection.

[0065] In this embodiment, the molecular subtyping process for breast cancer based on single-cell data is as follows: 1. Data Sources and Preprocessing The data used in this embodiment include scRNA-seq data and Bulk RNA-seq data, each data point carrying a subtype tag (ER+, HER2+ER+, HER2+, TN) confirmed by immunohistochemistry (IHC).

[0066] Data preprocessing included: using the Scanpy library to perform cell / gene filtering (removing low-quality cells and low-expression genes), total count normalization (normalize_total), and logarithmic transformation (log1p) on both datasets. Subsequently, the gene naming format of the two datasets was standardized, and the intersection genes were identified to ensure that subsequent analyses were based on the common gene set.

[0067] 2. Parameter Settings The main parameters of this invention are set as follows: cross-validation fold number is 5; number of target genes per subtype is 20; first-stage bundle width is 10; second-stage bundle width is 5.

[0068] 3. Optimization Process and Results The algorithm automatically performs 5-fold cross-validation. In each fold, it sequentially completes the initial screening for reinforcement, the first stage of optimization, the second stage of optimization, and the evaluation on the test set. Figure 4 The process shown was executed five times. Finally, all gene lists generated in the 5-fold optimization were integrated, and the 20 most frequent genes in each subtype were selected to form the final feature gene set.

[0069] The genes screened in this invention exhibit clear subtype-specific patterns, demonstrating good biological interpretability and providing important clues for subsequent mechanistic studies.

[0070] Example 2 This embodiment provides a feature gene selection system based on deep learning attribution analysis and bundle combinatorial optimization, including: The data acquisition module is used to acquire single-cell RNA sequencing data and batch RNA sequencing data; The initial screening module is used to obtain an initial gene pool for multiple disease subtypes based on single-cell RNA sequencing data by using a consensus screening strategy that integrates XGBoost multidimensional importance measurement and deep learning SHAP attribution analysis. The optimization module is used to start from the initial gene pool, perform a first-stage optimization of the gene list of each subtype using bundle combinatorial search and based on a global discriminant objective function, to obtain a first-stage optimized gene set; integrate the first-stage optimized gene set with the differentially expressed gene set of batch RNA sequencing data, and perform a second-stage optimization again using bundle combinatorial search, to obtain a second-stage optimized gene set; The integration output module is used to perform cross-validation integration on the optimized gene set of the second stage and output the final feature gene set.

[0071] Example 3 Embodiment 3 of the present invention provides an electronic device.

[0072] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps in the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization as described in Embodiment 1 of the present invention.

[0073] The detailed steps are the same as the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization provided in Example 1, and will not be repeated here.

[0074] Example 4 Embodiment 4 of the present invention provides a computer-readable storage medium.

[0075] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization as described in Embodiment 1 of the present invention.

[0076] The detailed steps are the same as the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization provided in Example 1, and will not be repeated here.

[0077] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0078] A computer program product includes software code, wherein the program in the software code performs the steps of the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization as described in Embodiment 1 of the present invention.

[0079] The detailed steps are the same as the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization provided in Example 1, and will not be repeated here.

[0080] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0081] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0082] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0083] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0084] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0085] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0086] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization, characterized in that, include: Acquire single-cell RNA sequencing data and batch RNA sequencing data; Based on single-cell RNA sequencing data, an initial gene pool targeting multiple disease subtypes was obtained by integrating XGBoost multidimensional importance measurement and deep learning SHAP attribution analysis consensus screening strategy. Starting from the initial gene pool, a bundle combinatorial search is used and the gene list of each subtype is optimized in the first stage based on the global discriminant objective function to obtain the first-stage optimized gene set; The first-stage optimized gene set was integrated with the differentially expressed gene set from the batch RNA sequencing data, and then a bundle combinatorial search was used for the second-stage optimization to obtain the second-stage optimized gene set. The optimized gene set from the second stage is cross-validated and integrated to output the final feature gene set.

2. The method as described in claim 1, characterized in that, The consensus filtering strategy includes: For each disease subtype, candidate differentially expressed genes were initially selected based on differential expression analysis using single-cell RNA sequencing data. An XGBoost model and a deep neural network model are trained. Gain, frequency, and coverage importance metrics are extracted from the XGBoost model and then standardized and weighted to obtain a first importance score. The SHAP value of each gene is calculated from the deep neural network model to obtain a second importance score. Based on the ranking of the first importance score and the second importance score, consensus genes are selected to form the initial gene pool.

3. The method as described in claim 1, characterized in that, The first phase of optimization includes: For each disease subtype, the starting point is the list of genes for that disease subtype in the initial gene pool; The number of genes in the gene combination is gradually increased through an iterative approach, at each iteration level: New gene combinations are generated based on the candidate combinations at the previous level; The gene list for other disease subtypes is fixed, and the discriminative ability of the newly generated gene combinations is evaluated using the global discriminative objective function. Based on the evaluation results, a preset number of optimal combinations are retained as candidate combinations for the current level; After all iterations are completed, the combination that maximizes the global discriminative objective function value is selected from the candidate combinations at all levels, and the gene list of the disease subtype is updated to obtain the first-stage optimized gene set.

4. The method as described in claim 1, characterized in that, The second phase of optimization includes: For each disease subtype, obtain the differentially expressed gene set from the batch RNA sequencing data; Based on the gene list of the disease subtype in the first-stage optimized gene set, the genes in the differentially expressed gene set are combined and expanded iteratively. In each iteration level: The newly added differentially expressed genes are merged with the gene combinations obtained in the previous round of optimization to generate new candidate combinations; The gene list for other disease subtypes is fixed, and the discriminative ability of the newly generated candidate combinations is evaluated using the global discriminative objective function; Based on the evaluation results, retain the optimal combination of the preset quantity; After all iterations are completed, the combination that maximizes the global discriminative objective function value is selected from all candidate combinations, and the gene list of the disease subtype is updated to obtain the second-stage optimized gene set.

5. The method as described in claim 1, characterized in that, The cross-validation integration includes: The entire process of consensus screening, first-stage optimization, and second-stage optimization is repeated multiple times within the K-fold cross-validation framework. Each execution yields a set of characteristic genes; Integrate all the feature gene sets obtained from the execution, select the gene with the highest frequency in each disease subtype, and form the final feature gene set.

6. The method as described in claim 1, characterized in that, The global discriminant objective function is used to quantify the discriminant ability of a gene set for multiple disease subtypes. The objective function calculates the average discriminant advantage of positive samples and the average discriminant advantage of negative samples based on the difference between the scores of each sample on the target subtype gene set and the scores on other subtype gene sets. The gene set is optimized by maximizing the average discriminant advantage of positive samples and minimizing the average discriminant advantage of negative samples.

7. A feature gene selection system based on deep learning attribution analysis and bundle combinatorial optimization, characterized in that, include: The data acquisition module is used to acquire single-cell RNA sequencing data and batch RNA sequencing data; The initial screening module is used to obtain an initial gene pool for multiple disease subtypes based on single-cell RNA sequencing data by using a consensus screening strategy that integrates XGBoost multidimensional importance measurement and deep learning SHAP attribution analysis. The optimization module is used to start from the initial gene pool, perform a first-stage optimization of the gene list of each subtype using bundle combinatorial search and based on a global discriminant objective function, to obtain a first-stage optimized gene set; integrate the first-stage optimized gene set with the differentially expressed gene set of batch RNA sequencing data, and perform a second-stage optimization again using bundle combinatorial search, to obtain a second-stage optimized gene set; The integration output module is used to perform cross-validation integration on the optimized gene set of the second stage and output the final feature gene set.

8. A feature gene selection device based on deep learning attribution analysis and bundle combinatorial optimization, characterized in that, The device includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement, when executing the computer program, the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization as described in any one of claims 1 to 6.

10. A computer program product, comprising software code, characterized in that, The program in the software code executes the steps of the feature gene selection method based on deep learning attribution analysis and bundle combinatorial optimization as described in any one of claims 1 to 6.

Images