Intelligent analysis method and device for cancer-related genes based on model stacking, and medium

Abstract

The invention relates to the technical field of gene analysis, and discloses an intelligent cancer-related gene analysis method based on model stacking. In the existing cancer gene prediction method, the basic principle and the applicable scene of each computational biological tool are different, the performance of a machine learning model depends on the quality of training data, and the analysis reliability of cancer-related genes depending on a single tool is insufficient. According to the method, firstly, prediction scores of different tools are fused in a non-linear extension feature construction mode, and then the prediction scores, confidence coefficients and non-linear extension features of all to-be-measured variation points are used as comprehensive data to be subjected to prediction analysis by using multiple models. And synthesizing the analysis result of the primary base model by using the secondary model to obtain a final prediction result. According to the method, different computational biological tools and different artificial intelligence models are integrated, and analysis results of various analysis tools can be integrated in the face of complex cancer gene data, so that the reliability of the analysis results is improved on the whole.

Description

Technical Field

[0001] This invention relates to the field of gene analysis technology, and in particular to a method, device, and medium for intelligent analysis of cancer-related genes based on model stacking. Background Technology

[0002] Single nucleotide variants (SNPs / SNVs) in cancer-related genes (such as driver genes, tumor suppressor genes, and key genes in signaling pathways) have significant clinical implications for tumor development, prognostic assessment, and personalized treatment. With the interdisciplinary development of artificial intelligence and bioinformatics, several relatively mature computational biology tools have emerged, including SIFT (Sorting Intolerant From Tolerant), CADD (Combined Annotation Dependent Depletion), PredictSNP, and REVEL (Rare Exome Variant Ensemble Learner). These tools can predict the potential functional impact of SNPs in cancer-related genes, providing prediction scores and confidence levels. Furthermore, well-trained machine learning models can also predict the potential functional impact of SNPs in cancer-related genes. However, due to the different fundamental principles and applicable scenarios of each computational biology tool, and the fact that the performance of machine learning models depends on the quality of training data, the reliability of intelligent analysis of cancer-related genes relying on a single tool is insufficient. Summary of the Invention

[0003] This invention provides a method, device, and medium for intelligent analysis of cancer-related genes based on model stacking, which improves the reliability of intelligent analysis of cancer-related genes.

[0004] The first aspect of this invention discloses an intelligent analysis method for cancer-related genes based on model stacking, the method comprising: Acquire the gene data to be tested, and use a variety of computational biology tools to analyze the gene data to be tested, and obtain the prediction scores and confidence scores of all the variant points to be tested corresponding to the gene data to be tested, wherein the prediction scores are used to indicate the classification results of the corresponding variant points to be tested; Based on the predicted scores analyzed by various computational biology tools, a nonlinear extended feature of all test variants corresponding to the cancer-related gene data is constructed. The model stacking prediction stage includes: inputting the prediction scores, confidence levels, and nonlinear extension features of all the variants to be tested into multiple pre-trained first-level base models to obtain the output results of the multiple first-level base models; constructing a first-level result vector based on the output results of all the first-level base models; inputting the first-level result vector into a pre-trained second-level model to obtain the output results of the second-level model; and determining the functional prediction result corresponding to the gene data to be tested based on the output results of the second-level model.

[0005] A second aspect of this invention discloses a model-stacked intelligent analysis device for cancer-related genes, the device comprising: The data acquisition module is used to acquire the gene data to be tested, analyze the gene data to be tested using a variety of computational biology tools, and obtain the prediction scores and confidence levels of all the variant points to be tested corresponding to the gene data to be tested. The prediction scores are used to indicate the classification results of the corresponding variant points to be tested. The tool calculation module is used to construct nonlinear extended features of all test variants corresponding to the cancer-related gene data based on the predicted scores analyzed by various computational biology tools. The stacked prediction module is used to perform the operations of the model stacked prediction stage, specifically including: inputting the prediction scores, confidence levels, and nonlinear extension features of all the variants to be tested into multiple pre-trained first-level base models to obtain the output results of the multiple first-level base models; constructing a first-level result vector based on the output results of all the first-level base models; inputting the first-level result vector into a pre-trained second-level model to obtain the output results of the second-level model; and determining the functional prediction result corresponding to the gene data to be tested based on the output results of the second-level model.

[0006] As an optional implementation, in a second aspect of the invention, the first-level base model and the second-level model are obtained through the following steps: Obtain a first training dataset, which includes cancer-related gene data and variant classification labels corresponding to the cancer-related gene data; The first training dataset was analyzed using a variety of computational biology tools to obtain the prediction scores and confidence levels of all training variants corresponding to the cancer-related gene data. Based on the predicted scores analyzed by various computational biology tools, a nonlinear extended feature of all training variants corresponding to the cancer-related gene data is constructed. The stacked model training phase includes: using the prediction scores, confidence levels, and nonlinear extension features of all training mutation points as a second training dataset to train multiple machine learning models, resulting in multiple trained first-level base models; predicting the second training dataset based on the multiple first-level base models to obtain first-level pop prediction data composed of the output results of the multiple first-level base models; and using the first-level pop prediction data to train a preset initial machine learning model to obtain a trained second-level model.

[0007] As an optional implementation, in the second aspect of the invention, the various computational biology tools include a variety of tools such as SIFT, CADD, PredictSNP, REVEL, and MCAP. The types of the multiple first-level base models are various of the following: random forest model, decision tree model, logistic regression model, SVM model, and feedforward neural network model; The types of secondary models are logistic regression models or gradient boosting tree models.

[0008] As an optional implementation, in a second aspect of the present invention, the first training dataset further includes pathological state labels corresponding to the cancer-related gene data; Furthermore, the stacked model training phase specifically includes: Based on the pathological state labels corresponding to the cancer-related genes, all training variants are divided into multiple state groups; the prediction scores, confidence levels, and nonlinear expansion features of all training variants in each state group are obtained. For each state group: the predicted scores, confidence levels, and nonlinear extension features of all training variants within the state group are used as the second training dataset to train multiple machine learning models, resulting in multiple trained first-level base models corresponding to the state group; predictions are made on the second training dataset based on the multiple first-level base models to obtain first-level pop prediction data composed of the outputs of the multiple first-level base models; the first-level pop prediction data is used to train a preset initial machine learning model to obtain a trained second-level model corresponding to the state group.

[0009] As an optional implementation, in a second aspect of the invention, the apparatus further includes: The pathology analysis module is used to input the predicted scores and confidence levels of all test variants analyzed by various computational biology tools into a pre-trained pathology state analysis model, obtain the output results of the pathology state analysis model, and determine the pathology state corresponding to the test gene based on the output results of the pathology state analysis model. The pathological state analysis model is obtained through the following steps: Determine the state label corresponding to each training variant point based on the pathological state label; The predicted scores, confidence levels, and state labels of all training variants are used as training data to train the preset initial pathological state analysis model, resulting in a trained pathological state analysis model.

[0010] As an optional implementation, in a second aspect of the present invention, the stacking prediction module performs the operations of the model stacking prediction stage, specifically including: Based on the pathological state corresponding to the gene to be tested, the state group corresponding to the gene to be tested is determined, and then multiple trained first-level base models corresponding to the state group are used as multiple target first-level base models, and the trained second-level models corresponding to the state group are used as target second-level base models. The predicted scores, confidence levels, and nonlinear extension features of all the variants to be tested are input into multiple target first-level base models to obtain the output results of the multiple target first-level base models; a target first-level result vector is constructed based on the output results of all the target first-level base models; the target first-level result vector is input into a target second-level model to obtain the output results of the target second-level model; and the functional prediction results corresponding to the gene data to be tested are determined based on the output results of the target second-level model.

[0011] As an optional implementation, in the second aspect of the invention, the predicted scores, confidence levels, and nonlinear extension features of all training variant points are used as a second training dataset to train multiple machine learning models, resulting in multiple trained first-level base models, including: The predicted scores, confidence levels, and nonlinear extension features of all training variants are used as the second training dataset to train multiple machine learning models, resulting in multiple well-trained initial first-level base models. Each initial first-level basis model is validated using a preset validation dataset to obtain a validation result for each initial first-level basis model. An evaluation score is then determined based on the validation result. From the plurality of initial primary basis models, the initial primary basis models whose evaluation scores exceed a preset score threshold are selected as primary basis models.

[0012] As an optional implementation, in a second aspect of the present invention, the step of training a preset initial machine learning model using the first-level pop-out prediction data to obtain a trained second-level model includes: Using the first-level pop prediction data, a preset initial machine learning model is trained using a K-fold cross-validation training device to obtain a trained second-level model, where K is a preset positive integer.

[0013] As an optional implementation, in a second aspect of the invention, the step of analyzing the first training dataset using various computational biology tools to obtain the prediction scores and confidence levels of all training variants corresponding to the cancer-related gene data includes: For each of the aforementioned computational biology tools: The first training dataset is analyzed using the computational biology tool to obtain the initial training variants analyzed by the computational biology tool and the prediction score corresponding to each initial training variant. For the cancer-related gene data, a filter window of a preset size is used to move along the cancer-related gene data once every preset amount of data. During each move: the initial training variant with the largest prediction score in the filter window is extracted as the training variant selected by the filter window. Based on all training variants selected by the filtering window during all movements, the prediction scores and confidence levels of all training variants corresponding to the cancer-related gene data are obtained.

[0014] A third aspect of this invention discloses an intelligent analysis system for cancer-related genes based on model stacking, the system comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute some or all of the steps in the intelligent analysis method for cancer-related genes based on model stacking as described in any of the first aspects of the present invention.

[0015] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute some or all of the steps in the intelligent analysis method for cancer-related genes based on model stacking as described in any of the first aspects of the present invention.

[0016] Compared with the prior art, the present invention has the following beneficial effects: In this invention, existing cancer gene prediction methods suffer from the following problems: due to the different fundamental principles and applicable scenarios of each computational biology tool, and the fact that the performance of machine learning models depends on the quality of training data, the reliability of intelligent analysis of cancer-related genes relying on a single tool is insufficient. Therefore, this invention first fuses the prediction scores of different tools by constructing nonlinear extended features. Then, it uses the prediction scores, confidence levels, and nonlinear extended features of all test variants as comprehensive data for prediction analysis using multiple models. Finally, a secondary model is used to synthesize the analysis results of the primary base model to obtain the final prediction result. This invention integrates different computational biology tools and different artificial intelligence models, enabling the synthesis of analysis results from various tools when facing complex cancer gene data, thereby improving the overall reliability of the analysis results. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an intelligent analysis method for cancer-related genes based on model stacking, as disclosed in an embodiment of the present invention. Figure 2 This is a flowchart illustrating a model training method disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a cancer-related gene intelligent analysis device based on model stacking disclosed in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of a cancer-related gene intelligent analysis system based on model stacking disclosed in an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0021] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0022] This invention discloses a method, device, and medium for intelligent analysis of cancer-related genes based on model stacking, which improves the reliability of intelligent analysis of cancer-related genes. These are described in detail below.

[0023] Example 1 like Figure 1 As shown, Embodiment 1 of the present invention discloses a model-stacked intelligent analysis method for cancer-related genes. This method can be integrated into a model-stacked intelligent analysis device for cancer-related genes, which can in turn be integrated into a local server or a cloud server. Specifically, the model-stacked intelligent analysis method for cancer-related genes disclosed in Embodiment 1 of the present invention may include: Step 101: Obtain the gene data to be tested, and use a variety of computational biology tools to analyze the gene data to obtain the prediction scores and confidence levels of all the variants to be tested corresponding to the gene data.

[0024] In this embodiment of the invention, for the gene data to be tested, all single nucleotide variant (SNP / SNV) records can be extracted from clinical data, retaining key fields such as chromosome, locus, reference / substitute nucleotide, and clinical variant classification. For example, a segment of data from the gene data to be tested is illustrated below:

[0025] In this embodiment of the invention, the predicted score is used to indicate the classification result of the corresponding variant to be tested; the classification result can be classified according to clinical significance (based on the ACMG / AMP guidelines) into the following five categories: pathogenic, possibly pathogenic, unclear significance, possibly benign, and benign. Alternatively, it can be classified according to the intensity of functional impact into high impact, moderate impact, low impact, and modifying factors, etc.

[0026] In this embodiment of the invention, optional computational biology tools may include multiple tools selected from SIFT, CADD, PredictSNP, REVEL, and MCAP. These tools are authoritative or commonly used in bioinformatics for predicting the functional impact of gene variations; they can score and classify variations based on different principles.

[0027] The principles and output results of the above computational biology tools are shown in the table below:

[0028] Step 102: Based on the predicted scores analyzed by various computational biology tools, construct nonlinear extended features for all test variants corresponding to cancer-related gene data.

[0029] Because different computational biology tools have different analytical principles and output results, the output results of different tools cannot be directly fused and used as input data for the next stage. In this embodiment of the invention, the outputs of different tools are used to construct nonlinear extended features. For example, nonlinear extended features such as square and cube are constructed on the predicted scores of five tools to enhance the fitting of complex decision boundaries and to better fuse the outputs of different tools.

[0030] For example, in an alternative embodiment, the nonlinear extension feature can be constructed as follows: Let S be the score predicted by the i-th computational biology tool. i The score predicted by the j-th computational biology tool is S. j Then, exemplary nonlinear extension features include: Single-tool nonlinear term: S i 2 S i 3 ; Inter-tool interaction project: S i ·S j ; Global statistics: max(S1...S5), std(S1...S5).

[0031] Step 103, Model Stacking Prediction Stage, includes: inputting the prediction scores, confidence levels, and nonlinear extension features of all test variants into multiple pre-trained first-level base models to obtain the output results of multiple first-level base models; constructing a first-level result vector based on the output results of all first-level base models; inputting the first-level result vector into a pre-trained second-level model to obtain the output results of the second-level model; and determining the functional prediction results corresponding to the test gene data based on the output results of the second-level model.

[0032] In this embodiment of the invention, the types of the multiple first-level base models can be various of the following: random forest model, decision tree model, logistic regression model, SVM model, and feedforward neural network model; the types of the second-level models can be logistic regression model or gradient boosting tree model.

[0033] In this embodiment of the invention, existing cancer gene prediction methods suffer from the following problems: due to the different fundamental principles and applicable scenarios of each computational biology tool, and the fact that the performance of machine learning models depends on the quality of training data, the reliability of intelligent analysis of cancer-related genes relying on a single tool is insufficient. Therefore, this embodiment of the invention first fuses the prediction scores of different tools by constructing nonlinear extended features. Then, the prediction scores, confidence levels, and nonlinear extended features of all test variants are used as comprehensive data for prediction analysis using multiple models. Finally, a secondary model is used to synthesize the analysis results of the primary base model to obtain the final prediction result. This embodiment of the invention integrates different computational biology tools and different artificial intelligence models, enabling the synthesis of analysis results from various tools when facing complex cancer gene data, thereby improving the overall reliability of the analysis results.

[0034] To verify the reliability of the intelligent analysis method for cancer-related genes in this embodiment of the invention, the following experimental verification was conducted: Evaluation principles: Primarily based on accuracy, recall, and AUC, supplemented by precision, F1 score, PR-AUC, and confusion matrix; emphasis is placed on minority class recall. Single-model experimental results: accuracy 0.985, recall 0.985, AUC 0.995.

[0035] Experimental results of the intelligent analysis method for cancer-related genes in this embodiment of the invention: accuracy 0.986, recall 0.986, AUC 0.996.

[0036] In an optional embodiment, the first-level base model and the second-level model are obtained through the following steps: Obtain the first training dataset, which includes cancer-related gene data and the variant classification labels corresponding to the cancer-related gene data; The first training dataset was analyzed using a variety of computational biology tools to obtain the prediction scores and confidence levels of all training variants corresponding to cancer-related gene data. Based on the predicted scores analyzed by various computational biology tools, a nonlinear extended feature of all training variants corresponding to cancer-related gene data is constructed. The stacked model training phase includes: using the prediction scores, confidence levels, and nonlinear extension features of all training mutation points as the second training dataset, training multiple machine learning models to obtain multiple trained first-level base models; predicting the second training dataset based on the multiple first-level base models to obtain first-level pop prediction data composed of the output results of multiple first-level base models; and using the first-level pop prediction data to train the preset initial machine learning model to obtain a trained second-level model.

[0037] Regarding the above model training process, as follows: Figure 2 As shown in the example below: Data source: All SNPs / SNVs were extracted from NCBI clinical data of 1196 cancer-related genes, retaining chromosomal, genomic loci, reference / substitute bases, and variant classification (pathogenic / benign / unclear significance, etc.).

[0038] Tool scores: Aggregate prediction scores and confidence levels from SIFT, CADD, PredictSNP, REVEL, and MCAP, and align coordinates with IDs.

[0039] Cleaning and standardization: Remove missing samples for key fields, and use median or multiple imputation for imputable fields; unify the different subdivisions and directions to a consistent interval and semantics.

[0040] Effective sample size: A total of 300,228 valid data points were obtained for training and validation.

[0041] Basic features: raw scores and confidence levels of the five tools; encoding of basic variation information (chromosome, base type, etc.).

[0042] Nonlinear extension: Generate square and cube terms for each tool's score, and add interaction terms if necessary; perform standardization / robust scaling and anomaly handling.

[0043] Optional tumor annotation: Information such as whether it is located in a known driver region, protein conservation, or domain can be introduced to improve the ability to identify cancer-related sites.

[0044] Optional features include feature selection: using LightGBM or permutation importance to filter the top 30 features, providing a basis for subsequent model simplification.

[0045] Single-model training and parameter tuning steps: Model set: Random Forest, Decision Tree, Logistic Regression, SVM (RBF / Linear Kernel), Feedforward Neural Network.

[0046] Parameter tuning strategy: K-fold cross-validation combined with grid / Bayesian search to optimize tree depth, regularization coefficient, kernel parameters, number of layers / width, Dropout, etc.

[0047] Evaluation metrics: primarily accuracy, recall, and AUC, supplemented by precision, F1 score, PR-AUC, and confusion matrix; emphasis is placed on minority class recall.

[0048] Optimal results: The best single-model accuracy and recall is 0.985, and the AUC is 0.995.

[0049] The three best-performing single models on the validation set are selected as first-level learners; first-level pop predictions are generated using the K-fold method to form second-level inputs; second-level learners are trained using logistic regression or LightGBM to obtain the final model.

[0050] In another optional embodiment, the functional prediction of variants in cancer-related genes should not be limited to the gene itself, but should also consider the pathological state corresponding to the cancer gene. Unlike the classification of variants, the classification of variants in this embodiment refers to single nucleotide variants of cancer-related genes, while the pathological state refers to the entire cancer-related gene. All single nucleotide variants on the cancer gene collectively determine the pathological state of the entire cancer-related gene. For example, the pathological state can be divided into: GX: grade cannot be assessed; G1: highly differentiated - cancer cells are close to normal cells, grow slowly, and have low invasiveness; G2: moderately differentiated - between G1 and G3; G3: poorly differentiated - cancer cells are very different from normal cells, grow rapidly, and have high invasiveness; G4: undifferentiated - cancer cells are extremely primitive, their origin cannot be identified, and the malignancy is the highest. Therefore, different state groups can be set for different pathological states, and then a corresponding model can be trained for each state group to make the trained model more targeted, thereby improving the reliability of the prediction.

[0051] Therefore, in yet another optional embodiment, the first training dataset also includes pathological state labels corresponding to cancer-related gene data; Furthermore, the stacked model training phase specifically includes: Based on the pathological state labels corresponding to cancer-related genes, all training variants are divided into multiple state groups; the prediction scores, confidence scores and nonlinear expansion features of all training variants in each state group are obtained. For each state group: the predicted scores, confidence levels, and nonlinear extension features of all training variants within the state group are used as the second training dataset to train multiple machine learning models, resulting in multiple trained first-level base models corresponding to the state group; predictions are made on the second training dataset based on the multiple first-level base models to obtain first-level pop prediction data composed of the outputs of the multiple first-level base models; the first-level pop prediction data is used to train a preset initial machine learning model to obtain a trained second-level model corresponding to the state group.

[0052] As can be seen, this optional embodiment sets different state groups for different pathological states, and then trains a corresponding model for each state group to make the trained model more targeted, thereby improving the reliability of the prediction.

[0053] In this optional embodiment, the method may further include: The predicted scores and confidence levels of all the variants to be tested, analyzed by various computational biology tools, are input into a pre-trained pathological state analysis model to obtain the output results of the pathological state analysis model. Based on the output results of the pathological state analysis model, the pathological state corresponding to the gene to be tested is determined.

[0054] The prediction of pathological status is not limited to the functional impact of a single nucleotide variation, but rather takes the entire cancer-related gene as the object of analysis and prediction. This enables the intelligent analysis method of cancer-related genes based on model stacking in the embodiments of the present invention to also predict and analyze pathological status information.

[0055] The pathological state analysis model is trained through the following steps: Determine the state label corresponding to each training variant point based on the pathological state label. The predicted scores, confidence levels, and state labels of all training variants are used as training data to train the preset initial pathological state analysis model, resulting in a trained pathological state analysis model.

[0056] In this optional embodiment, the pathological state analysis model may optionally rely on an XGBoost or Transformer classifier; the input features of the pathological state analysis model may include variable point statistics (such as pathogenicity ratio).

[0057] For the above embodiments that include pathological transformation analysis, the operational examples are as follows: TP53 gene variation analysis workflow: (1) Input data: Variant site: chr17:7573987 (G>T) Pathogenicity conclusion: ClinVar ★★★ (2) Tool output:

[0058] (3) Pathological grouping: The model identifies it as group G3 → call the G3 dedicated stacking model. (4) Prediction result: pathogenic variant (probability 98.6%) In yet another optional embodiment, the above-described model stacking prediction stage specifically includes: Based on the pathological state corresponding to the gene to be tested, the state group corresponding to the gene to be tested is determined. Then, multiple trained first-level basis models corresponding to the state group are used as multiple target first-level basis models, and the trained second-level models corresponding to the state group are used as target second-level basis models. The predicted scores, confidence levels, and nonlinear extension features of all the variants to be tested are input into multiple target first-level basis models to obtain the output results of multiple target first-level basis models; based on the output results of all target first-level basis models, a target first-level result vector is constructed; the target first-level result vector is input into a target second-level model to obtain the output results of the target second-level model; based on the output results of the target second-level model, the functional prediction results corresponding to the gene data to be tested are determined.

[0059] In this optional embodiment, since the trained primary and secondary models are trained in multiple sets according to different pathological states, and each set of models is trained specifically, when performing single nucleotide variant analysis, the current state group corresponding to the gene to be tested is first determined based on the pathological analysis results. Then, the primary and secondary models corresponding to this state group are used as the models for this analysis and prediction. Through the above operations, this optional embodiment can use more targeted primary and secondary models to perform subsequent analysis and prediction operations, thereby further improving the reliability of the prediction.

[0060] In another optional embodiment, the predicted scores, confidence levels, and nonlinear extension features of all training variants are used as a second training dataset to train multiple machine learning models, resulting in multiple trained first-level base models, which may include: The predicted scores, confidence levels, and nonlinear extension features of all training variants are used as the second training dataset to train multiple machine learning models, resulting in multiple well-trained initial first-level base models. Each initial first-level basis model is validated using a pre-defined validation dataset to obtain the validation results for each initial first-level basis model. Based on the validation results, the evaluation score for each initial first-level basis model is determined. From multiple initial primary basis models, the initial primary basis models whose evaluation scores exceed a preset score threshold are selected as primary basis models.

[0061] In this optional embodiment, accuracy, recall, and AUC may be the primary metrics, supplemented by precision, F1 score, PR-AUC, and confusion matrix; with emphasis placed on minority class recall as the basis for obtaining the evaluation score.

[0062] In this optional embodiment, the number of initial primary base models can be greater. After one round of training, the performance of the initial primary base models is evaluated by a validation dataset. Then, several primary base models with better performance are selected as the final primary base models used, thereby obtaining primary base models with better prediction performance.

[0063] In another optional embodiment, training a pre-defined initial machine learning model using first-level pop prediction data to obtain a trained second-level model may include: Using the first-level pop prediction data, a pre-set initial machine learning model is trained using the K-fold cross-validation training method to obtain a trained second-level model, where K is a pre-set positive integer.

[0064] In this optional embodiment, a pre-defined initial machine learning model is trained using K-fold cross-validation. K-fold cross-validation is a statistical method for evaluating the performance of machine learning models. It divides the dataset into K mutually exclusive subsets of similar size, each subset is used sequentially as the test set, and the remaining K-1 subsets are used as the training set. This allows for a more comprehensive use of the data, ensuring that each sample has the opportunity to serve as both a training and test set. Furthermore, by employing K-fold cross-validation, a more stable performance estimate can be obtained through multiple training and testing iterations, thereby reducing evaluation variance and avoiding overfitting.

[0065] In another alternative embodiment, the overall data quality of the first training dataset may be poor, and there may be a lot of invalid or inefficient data. When the first training dataset has a large amount of data, the low-quality data may lead to a decrease in training results, or even cause the trained model to give incorrect analysis results when there is a lot of inefficient data.

[0066] To address the aforementioned issues, in this optional embodiment, various computational biology tools are used to analyze the first training dataset to obtain prediction scores and confidence levels for all training variants corresponding to cancer-related gene data. This may include: For each computational biology tool: The computational biology tool was used to analyze the first training dataset to obtain the initial training variants and the prediction scores corresponding to each initial training variant. For cancer-related gene data, a filter window of a preset size is used to move along the cancer-related gene data once every preset amount of data. During each move: the initial training variant with the largest prediction score in the filter window is extracted as the training variant selected by the filter window. Based on all training variants selected by the filtering window during all moves, the prediction scores and confidence levels of all training variants corresponding to cancer-related gene data are obtained.

[0067] In this optional embodiment, it was found that the occurrence of variant points tends to be concentrated in segments of gene data that are close to each other. That is, multiple locations identified as variant points may be clustered together in a segment of data. However, these variant points may be caused by a single nucleotide variation. Therefore, within a continuous segment of gene data, selecting the data with the highest training quality can improve the overall data quality. To address this, this optional embodiment employs a filtering window approach, moving along the data and performing a filtering operation each time it moves. This leverages the principle of concentrated variant point occurrences to filter out low-quality data.

[0068] Optionally, the data filtering operations described above can be applied both during the model training phase and during the analysis of the gene to be tested. In this case, various computational biology tools can be used to analyze the gene data to obtain the prediction scores and confidence levels of all the variants corresponding to the gene data. This can include: The computational biology tool was used to analyze the gene data to be tested, and the initial variants to be tested and the prediction scores corresponding to each initial variant were obtained. For the gene data to be tested, a filter window of a preset size is used to move along the gene data to be tested once at a preset data interval. During each move: the initial variant point to be tested with the largest prediction score in the filter window is extracted as the variant point to be tested selected by the filter window. Based on all the test variants selected by the filtering window during all the moves, the prediction scores and confidence levels of all test variants corresponding to the test gene data are obtained.

[0069] Example 2 Based on the same inventive concept, Embodiment 2 of this invention discloses a cancer-related gene intelligent analysis device based on model stacking, such as... Figure 3 As shown, the device may include: The data acquisition module 201 is used to acquire the gene data to be tested, and to analyze the gene data to be tested using a variety of computational biology tools to obtain the prediction scores and confidence levels of all the variant points to be tested corresponding to the gene data to be tested. The prediction scores are used to indicate the classification results of the corresponding variant points to be tested. The tool calculation module 202 is used to construct nonlinear extended features of all test variants corresponding to cancer-related gene data based on the predicted scores analyzed by various computational biology tools. The stacked prediction module 203 is used to perform the operations of the model stacked prediction stage, which may include: inputting the prediction scores, confidence levels and nonlinear extension features of all the variants to be tested into multiple pre-trained first-level base models to obtain the output results of multiple first-level base models; constructing a first-level result vector based on the output results of all first-level base models; inputting the first-level result vector into a pre-trained second-level model to obtain the output results of the second-level model; and determining the functional prediction results corresponding to the gene data to be tested based on the output results of the second-level model.

[0070] In an optional embodiment, the first-level base model and the second-level model are obtained through the following steps: Obtain the first training dataset, which may include cancer-related gene data and the variant classification labels corresponding to the cancer-related gene data; The first training dataset was analyzed using a variety of computational biology tools to obtain the prediction scores and confidence levels of all training variants corresponding to cancer-related gene data. Based on the predicted scores analyzed by various computational biology tools, a nonlinear extended feature of all training variants corresponding to cancer-related gene data is constructed. The stacked model training phase may include: using the prediction scores, confidence levels, and nonlinear extension features of all training variants as a second training dataset to train multiple machine learning models, resulting in multiple trained first-level base models; using the multiple first-level base models to predict the second training dataset, resulting in first-level pop prediction data composed of the outputs of the multiple first-level base models; and using the first-level pop prediction data to train a preset initial machine learning model, resulting in a trained second-level model.

[0071] In another alternative embodiment, the various computational biology tools may include multiple tools such as SIFT, CADD, PredictSNP, REVEL, and MCAP. The types of first-level base models include various types of random forest models, decision tree models, logistic regression models, SVM models, and feedforward neural network models; The types of secondary models are logistic regression models or gradient boosting tree models.

[0072] In yet another optional embodiment, the first training dataset may further include pathological state labels corresponding to cancer-related gene data; Furthermore, the stacked model training phase may specifically include: Based on the pathological state labels corresponding to cancer-related genes, all training variants are divided into multiple state groups; the prediction scores, confidence scores and nonlinear expansion features of all training variants in each state group are obtained. For each state group: the predicted scores, confidence levels, and nonlinear extension features of all training variants within the state group are used as the second training dataset to train multiple machine learning models, resulting in multiple trained first-level base models corresponding to the state group; predictions are made on the second training dataset based on the multiple first-level base models to obtain first-level pop prediction data composed of the outputs of the multiple first-level base models; the first-level pop prediction data is used to train a preset initial machine learning model to obtain a trained second-level model corresponding to the state group.

[0073] In yet another alternative embodiment, the device may further include: The pathology analysis module is used to input the predicted scores and confidence levels of all the variants to be tested analyzed by various computational biology tools into a pre-trained pathology state analysis model, obtain the output results of the pathology state analysis model, and determine the pathology state corresponding to the gene to be tested based on the output results of the pathology state analysis model. The pathological state analysis model is trained through the following steps: Determine the state label corresponding to each training variant point based on the pathological state label. The predicted scores, confidence levels, and state labels of all training variants are used as training data to train the preset initial pathological state analysis model, resulting in a trained pathological state analysis model.

[0074] In yet another optional embodiment, the stacked prediction module 203 performs operations in the model stacked prediction phase, which may specifically include: Based on the pathological state corresponding to the gene to be tested, the state group corresponding to the gene to be tested is determined. Then, multiple trained first-level basis models corresponding to the state group are used as multiple target first-level basis models, and the trained second-level models corresponding to the state group are used as target second-level basis models. The predicted scores, confidence levels, and nonlinear extension features of all the variants to be tested are input into multiple target first-level basis models to obtain the output results of multiple target first-level basis models; based on the output results of all target first-level basis models, a target first-level result vector is constructed; the target first-level result vector is input into a target second-level model to obtain the output results of the target second-level model; based on the output results of the target second-level model, the functional prediction results corresponding to the gene data to be tested are determined.

[0075] In another optional embodiment, the predicted scores, confidence levels, and nonlinear extension features of all training variants are used as a second training dataset to train multiple machine learning models, resulting in multiple trained first-level base models, which may include: The predicted scores, confidence levels, and nonlinear extension features of all training variants are used as the second training dataset to train multiple machine learning models, resulting in multiple well-trained initial first-level base models. Each initial first-level basis model is validated using a pre-defined validation dataset to obtain the validation results for each initial first-level basis model. Based on the validation results, the evaluation score for each initial first-level basis model is determined. From multiple initial primary basis models, the initial primary basis models whose evaluation scores exceed a preset score threshold are selected as primary basis models.

[0076] In another optional embodiment, training a pre-defined initial machine learning model using first-level pop prediction data to obtain a trained second-level model may include: Using the first-level pop prediction data, a pre-set initial machine learning model is trained with a K-fold cross-validation training device to obtain a trained second-level model, where K is a pre-set positive integer.

[0077] In another optional embodiment, analyzing the first training dataset using various computational biology tools to obtain prediction scores and confidence levels for all training variants corresponding to cancer-related gene data may include: For each computational biology tool: The computational biology tool was used to analyze the first training dataset to obtain the initial training variants and the prediction scores corresponding to each initial training variant. For cancer-related gene data, a filter window of a preset size is used to move along the cancer-related gene data once every preset amount of data. During each move: the initial training variant with the largest prediction score in the filter window is extracted as the training variant selected by the filter window. Based on all training variants selected by the filtering window during all moves, the prediction scores and confidence levels of all training variants corresponding to cancer-related gene data are obtained.

[0078] Example 3 Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of a model-stacked intelligent analysis system for cancer-related genes disclosed in an embodiment of the present invention. The model-stacked intelligent analysis system for cancer-related genes may include: Memory 301 storing executable program code; Processor 302 coupled to memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute some or all of the steps in any of the intelligent analysis methods for cancer-related genes based on model stacking in Embodiment 1 of the present invention.

[0079] Example 4 This invention discloses a computer storage medium storing computer instructions. When executed by a processor, these computer instructions implement some or all of the steps in any of the model-stacking-based intelligent analysis methods for cancer-related genes in Embodiment 1 of this invention.

[0080] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented using software plus necessary general-purpose hardware platforms, or of course, using hardware. Based on this understanding, the above technical solutions, in essence or in terms of their contribution to the prior art, can be embodied in the form of a software product, which can be stored in a computer-readable storage medium.

Claims

1. A method for intelligent analysis of cancer-related genes based on model stacking, characterized in that, The method includes: Acquire the gene data to be tested, and use a variety of computational biology tools to analyze the gene data to be tested, and obtain the prediction scores and confidence scores of all the variant points to be tested corresponding to the gene data to be tested, wherein the prediction scores are used to indicate the classification results of the corresponding variant points to be tested; Based on the predicted scores analyzed by various computational biology tools, a nonlinear extended feature of all test variants corresponding to the cancer-related gene data is constructed. The model stacking prediction stage includes: inputting the prediction scores, confidence levels, and nonlinear extension features of all the variants to be tested into multiple pre-trained first-level base models to obtain the output results of the multiple first-level base models; constructing a first-level result vector based on the output results of all the first-level base models; inputting the first-level result vector into a pre-trained second-level model to obtain the output results of the second-level model; and determining the functional prediction result corresponding to the gene data to be tested based on the output results of the second-level model.

2. The intelligent analysis method for cancer-related genes based on model stacking according to claim 1, characterized in that, The first-level base model and the second-level model are obtained through the following steps: Obtain a first training dataset, which includes cancer-related gene data and variant classification labels corresponding to the cancer-related gene data; The first training dataset was analyzed using a variety of computational biology tools to obtain the prediction scores and confidence levels of all training variants corresponding to the cancer-related gene data. Based on the predicted scores analyzed by various computational biology tools, a nonlinear extended feature of all training variants corresponding to the cancer-related gene data is constructed. The stacked model training phase includes: using the prediction scores, confidence levels, and nonlinear extension features of all training mutation points as a second training dataset to train multiple machine learning models, resulting in multiple trained first-level base models; predicting the second training dataset based on the multiple first-level base models to obtain first-level pop prediction data composed of the output results of the multiple first-level base models; and using the first-level pop prediction data to train a preset initial machine learning model to obtain a trained second-level model.

3. The intelligent analysis method for cancer-related genes based on model stacking according to claim 1 or 2, characterized in that, The various computational biology tools include multiple tools such as SIFT, CADD, PredictSNP, REVEL, and MCAP. The types of the multiple first-level base models are various of the following: random forest model, decision tree model, logistic regression model, SVM model, and feedforward neural network model; The types of secondary models are logistic regression models or gradient boosting tree models.

4. The intelligent analysis method for cancer-related genes based on model stacking according to claim 2, characterized in that, The first training dataset also includes pathological state labels corresponding to the cancer-related gene data; Furthermore, the stacked model training phase specifically includes: Based on the pathological state labels corresponding to the cancer-related genes, all training variants are divided into multiple state groups; the prediction scores, confidence levels, and nonlinear expansion features of all training variants in each state group are obtained. For each state group: the predicted scores, confidence levels, and nonlinear extension features of all training variants within the state group are used as the second training dataset to train multiple machine learning models, resulting in multiple trained first-level base models corresponding to the state group; predictions are made on the second training dataset based on the multiple first-level base models to obtain first-level pop prediction data composed of the outputs of the multiple first-level base models; the first-level pop prediction data is used to train a preset initial machine learning model to obtain a trained second-level model corresponding to the state group.

5. The intelligent analysis method for cancer-related genes based on model stacking according to claim 4, characterized in that, The method further includes: The predicted scores and confidence levels of all the variants to be tested, analyzed by various computational biology tools, are input into a pre-trained pathological state analysis model to obtain the output results of the pathological state analysis model. Based on the output results of the pathological state analysis model, the pathological state corresponding to the gene to be tested is determined. The pathological state analysis model is obtained through the following steps: Determine the state label corresponding to each training variant point based on the pathological state label; The predicted scores, confidence levels, and state labels of all training variants are used as training data to train the preset initial pathological state analysis model, resulting in a trained pathological state analysis model.

6. The intelligent analysis method for cancer-related genes based on model stacking according to claim 5, characterized in that, The model stacking prediction stage specifically includes: Based on the pathological state corresponding to the gene to be tested, the state group corresponding to the gene to be tested is determined, and then multiple trained first-level base models corresponding to the state group are used as multiple target first-level base models, and the trained second-level models corresponding to the state group are used as target second-level base models. The predicted scores, confidence levels, and nonlinear extension features of all the variants to be tested are input into multiple target first-level base models to obtain the output results of the multiple target first-level base models; a target first-level result vector is constructed based on the output results of all the target first-level base models; the target first-level result vector is input into a target second-level model to obtain the output results of the target second-level model; and the functional prediction results corresponding to the gene data to be tested are determined based on the output results of the target second-level model.

7. The intelligent analysis method for cancer-related genes based on model stacking according to claim 2, characterized in that, The predicted scores, confidence levels, and nonlinear extension features of all training variants are used as the second training dataset to train multiple machine learning models, resulting in multiple trained first-level base models, including: The predicted scores, confidence levels, and nonlinear extension features of all training variants are used as the second training dataset to train multiple machine learning models, resulting in multiple well-trained initial first-level base models. Each initial first-level basis model is validated using a preset validation dataset to obtain a validation result for each initial first-level basis model. An evaluation score is then determined based on the validation result. From the multiple initial primary basis models, the initial primary basis models whose evaluation scores exceed a preset score threshold are selected as primary basis models.

8. The intelligent analysis method for cancer-related genes based on model stacking according to claim 2, characterized in that, The step of training a preset initial machine learning model using the first-level pop-out prediction data to obtain a trained second-level model includes: Using the first-level pop prediction data, a preset initial machine learning model is trained using a K-fold cross-validation training method to obtain a trained second-level model, where K is a preset positive integer.

9. The intelligent analysis method for cancer-related genes based on model stacking according to claim 2, characterized in that, The analysis of the first training dataset using multiple computational biology tools yields prediction scores and confidence levels for all training variants corresponding to the cancer-related gene data, including: For each of the aforementioned computational biology tools: The first training dataset is analyzed using the computational biology tool to obtain the initial training variants analyzed by the computational biology tool and the prediction score corresponding to each initial training variant. For the cancer-related gene data, a filter window of a preset size is used to move along the cancer-related gene data once every preset amount of data. During each move: the initial training variant with the largest prediction score in the filter window is extracted as the training variant selected by the filter window. Based on all training variants selected by the filtering window during all movements, the prediction scores and confidence levels of all training variants corresponding to the cancer-related gene data are obtained.

10. A smart analysis device for cancer-related genes based on model stacking, characterized in that, The device includes: The data acquisition module is used to acquire the gene data to be tested, analyze the gene data to be tested using a variety of computational biology tools, and obtain the prediction scores and confidence levels of all the variant points to be tested corresponding to the gene data to be tested. The prediction scores are used to indicate the classification results of the corresponding variant points to be tested. The tool calculation module is used to construct nonlinear extended features of all test variants corresponding to the cancer-related gene data based on the predicted scores analyzed by various computational biology tools. The stacked prediction module is used to perform the operations of the model stacked prediction stage, specifically including: inputting the prediction scores, confidence levels, and nonlinear extension features of all the variants to be tested into multiple pre-trained first-level base models to obtain the output results of the multiple first-level base models; constructing a first-level result vector based on the output results of all the first-level base models; inputting the first-level result vector into a pre-trained second-level model to obtain the output results of the second-level model; and determining the functional prediction result corresponding to the gene data to be tested based on the output results of the second-level model.

11. A cancer-related gene intelligent analysis system based on model stacking, characterized in that, The system includes: a memory storing executable program code; a processor coupled to the memory; the processor calling the executable program code stored in the memory to execute the intelligent analysis method for cancer-related genes based on model stacking as described in any one of claims 1-9.

12. A computer storage medium, characterized in that, The computer storage medium stores computer instructions, which, when executed by a processor, implement the intelligent analysis method for cancer-related genes based on model stacking as described in any one of claims 1-9.

Images