Cross-species lung disease feature gene screening method, system, electronic system and storage device based on multivariate machine learning model

Abstract

The application provides a kind of screening method, system, electronic system and storage device of cross-species lung disease characteristic gene based on multi-element machine learning model, which comprises: obtaining human and mouse lung disease related single cell / transcriptome data from public database and preprocessing;Six machine learning algorithms are used to train the model and output gene importance score;According to the performance of the model, the weight is calculated and the score is normalized;The cross-species score is integrated by the weighted fusion formula, and the characteristic gene list and the visualization report are output.The system includes data acquisition and preprocessing module, multi-element machine learning model training module, weight calculation and normalization module, cross-species comprehensive scoring module and result output module.The screening accuracy, stability and generalization ability are improved by multi-algorithm integration and cross-species fusion, which can be widely used in the mechanism research, diagnostic marker development and drug target verification field of lung disease.

Description

Technical Field

[0001] This invention relates to the fields of bioinformatics and machine learning technology, specifically providing a method, system, electronic system, and storage device for screening characteristic genes of lung diseases across species based on a multivariate machine learning model. This method is used to accurately identify characteristic genes related to lung diseases in humans and mice, supporting disease mechanism research, diagnostic biomarker development, and drug target validation. Background Technology

[0002] Lung diseases, including silicosis, pulmonary fibrosis, and chronic obstructive pulmonary disease (COPD), pose a significant global health threat. Their pathological mechanisms involve multiple signaling pathways and cell type abnormalities. Precise screening of characteristic genes closely related to disease development is crucial for a deeper understanding of disease mechanisms, the development of early diagnostic biomarkers, and the discovery of therapeutic targets. However, current technologies suffer from two major shortcomings:

[0003] First, the stability of a single algorithm is insufficient. Traditional differentially expressed gene analysis or single machine learning models are easily affected by data noise, sample size limitations, and algorithm characteristics, resulting in poor robustness and weak generalization ability of screening results, inconsistent conclusions across different datasets, and seriously hindering clinical translation.

[0004] Second, cross-species differences are not effectively integrated. Existing methods do not fully consider the physiological and genetic differences between humans and model animals (such as mice), and lack a systematic cross-species data fusion framework, resulting in a lack of universality in gene screening and reduced transformation efficiency.

[0005] Specifically, traditional differential expression analysis methods rely on statistical differences, ignore complex intergene interaction networks, and are sensitive to batch effects, easily producing false positive or false negative results. Meanwhile, single machine learning algorithms exhibit significant biases, with different algorithms showing varying sensitivities to linear / nonlinear relationships, making them prone to overfitting with limited samples, and complex models suffer from poor interpretability. Furthermore, current technologies lack an effective framework for integrating cross-species information, making it difficult to leverage model animal data to improve the reliability of human disease research. Therefore, there is an urgent need to develop a high-precision gene screening scheme that integrates multiple algorithms and fuses cross-species data. Summary of the Invention

[0006] This application provides a method, system, electronic system, and storage device for screening characteristic genes of lung diseases across species based on a multivariate machine learning model. The aim is to provide a cross-species lung disease characteristic gene screening scheme. By integrating multiple machine learning algorithms and fusing human and mouse data, it solves the problems of poor stability of single models and neglect of species differences, significantly improving the accuracy, generalization ability, and interpretability of gene screening, and providing a reliable tool for disease mechanism research and precision medicine.

[0007] To solve the above-mentioned technical problems, the technical solution proposed in this application is as follows:

[0008] This invention provides a method for screening characteristic genes of lung diseases across species based on a multivariate machine learning model, comprising the following steps:

[0009] S1: Data Acquisition and Preprocessing — Batch acquisition of single-cell or transcriptome data of human and mouse lung diseases from public databases, quality control, normalization, clustering and cell type annotation, and generation of standardized gene expression matrices;

[0010] S2: Multivariate Machine Learning Model Training - The model is trained using six algorithms: LASSO, Random Forest, Decision Tree, Support Vector Machine, Gradient Boosting Machine, and XGBoost, and the gene importance score of each algorithm is output.

[0011] S3: Weight Calculation and Normalization - Calculate weights based on model performance using human and mouse datasets, and perform Min-Max normalization on the scores of each algorithm;

[0012] S4: Cross-species integrated scoring and output – Integrating human and mouse scores through a weighted fusion formula to obtain a comprehensive score for the target gene and output a list of characteristic genes.

[0013] Further, S1 includes:

[0014] S11: Database retrieval strategy development – ​​Develop a search query covering silicosis, pulmonary fibrosis, and chronic obstructive pulmonary disease, with keywords including “lung disease”, “pulmonary fibrosis”, “COPD”, “silicosis”, “singlecell RNA sequencing”, and “transcriptome”, to ensure that the target disease data is a subset of the search results;

[0015] S12: Batch Acquisition of Cross-Species Data—Acquire single-cell transcriptome data of human (Homosapiens) and mouse (Mus musculus) lung diseases in batches from the GEO and EMBL-EBI databases, and filter and download qualified datasets according to data quality, sample size and experimental design;

[0016] S13: Data quality control and consolidation, including:

[0017] Quality assessment: Examine gene detection rate, cell viability, and mitochondrial gene expression ratio;

[0018] Remove low-quality samples and genes;

[0019] Multiple datasets are merged using inner join;

[0020] S14: Batch effect removal and standardization, including:

[0021] Dynamically select Harmony, SCTransform, or scVI to remove batch effects;

[0022] The processed data is standardized to generate a standardized gene expression matrix;

[0023] S15: Cell clustering and type annotation, including:

[0024] Unsupervised clustering analysis was performed using the Leiden algorithm;

[0025] By combining manual annotation with large language model-assisted annotation, the following two-level system is established:

[0026] Level 1: Classified into major cell types, including epithelial cells, immune cells, and stromal cells;

[0027] Level 2: Preserves cell subset information, including alveolar epithelial cell type I and II, and T cell subsets.

[0028] Furthermore, step S2 includes the following steps:

[0029] S21: Construction of LASSO regression model:

[0030] For binary classification problems: use a logistic regression model with L1 penalty, set solver='saga' and regularization strength C=1;

[0031] For multi-class classification problems: the OneVsRestClassifier framework is used in combination with L1 regularized logistic regression, with C=0.1 and a maximum number of iterations of 10000.

[0032] Gene importance scores are output based on the absolute value of the coefficients.

[0033] S22: Random Forest Model Training:

[0034] Construct a classifier containing 150 decision trees;

[0035] Set the minimum number of split samples to 10 and the minimum number of leaf node samples to 10.

[0036] Calculate the importance of features based on Gini impurity;

[0037] S23: Support Vector Machine Model Implementation:

[0038] A linear kernel function is used, with a regularization parameter C=0.1;

[0039] Stratified sampling is used when the sample size is >5000.

[0040] Feature importance is calculated using the absolute values ​​of linear SVM coefficients;

[0041] S24: XGBoost gradient boosting model configuration:

[0042] Parameter settings: Maximum depth = 5, Learning rate = 0.1, Number of estimators = 100;

[0043] The objective function for binary classification is 'binary:logistic', and for multi-class classification it is 'multi:softprob'.

[0044] Automatically switch to random forest when training fails;

[0045] S25: Gradient Boosting Machine Model Training:

[0046] We prioritize using LightGBM, with parameters of 100 estimators, a learning rate of 0.1, and a maximum depth of 3.

[0047] Stratified sampling is enabled when the sample size is greater than 10,000.

[0048] Based on the importance of gain calculation characteristics;

[0049] S26: Decision Tree Model Construction:

[0050] Set the maximum depth to 10, the minimum number of split samples to 5, and the minimum number of leaf node samples to 5.

[0051] The importance of features is calculated based on the reduction in Gini impurity.

[0052] Furthermore, in S3:

[0053] S31: Calculation of Model Performance Evaluation Indicators

[0054] Two performance metrics were calculated for six machine learning algorithms on both human and mouse datasets:

[0055] (a) Absolute inverse residual: obtained by calculating the absolute value of the difference between the model's predicted probability and the one-hot encoding of the actual label, reflecting the accuracy of the model's prediction;

[0056] (b) Area under the ROC curve (AUC): Calculated using the area under the receiver operating characteristic curve, where:

[0057] (a1) For binary classification problems, the area under the ROC curve (AUC) is calculated directly.

[0058] (b1) For multi-class classification problems, the macro-averaging strategy is used to calculate the area under the overall ROC curve (AUC).

[0059] S32: Weight Allocation Strategy Design

[0060] Design a weight allocation scheme based on S31 performance metrics:

[0061] (A) ROC curve area under the curve (AUC) weight allocation: The weight is allocated proportionally, and the algorithm with the higher the ROC curve area under the curve (AUC) value receives a larger weight.

[0062] (B) Residual weight allocation: The algorithm adopts reverse allocation, and the smaller the mean absolute residual, the greater the weight is obtained;

[0063] (C) Calculation of comprehensive weights: The final weights are determined by the arithmetic mean of the area under the ROC curve (AUC) weights and the residual weights;

[0064] S33: Min-Max Normalization Implementation

[0065] The gene importance scores output by each algorithm are standardized:

[0066] The normalization formula is:

[0067]

[0068] in,

[0069] The score of gene i in algorithm j;

[0070] The score of gene i in algorithm j after normalization;

[0071] The minimum score in algorithm j;

[0072] The maximum score in algorithm j;

[0073] The score range is standardized to the [0,1] interval to eliminate the difference in units between algorithms;

[0074] S34: Cross-species weight consistency verification

[0075] Calculate the weights of each algorithm on the human and mouse datasets respectively;

[0076] The consistency of cross-species weight allocation was verified through correlation analysis;

[0077] Establish a weight correction mechanism: When the performance difference of an algorithm between two species exceeds a threshold, a conservative strategy is adopted to adjust the weights to avoid the influence of single-species data bias on the results.

[0078] Furthermore, the cross-species comprehensive scoring formula in S4 is as follows:

[0079]

[0080] in,

[0081] The performance weights of algorithm j on the mouse dataset;

[0082] The performance weights of algorithm j on the human dataset;

[0083] The number of algorithms that selected gene i in the mouse dataset;

[0084] The number of algorithms that selected gene i in the human dataset;

[0085] The score of gene i in algorithm j after normalization;

[0086] : Comprehensive score of cross-species gene i.

[0087] On the other hand, this application also claims protection for a cross-species lung disease characteristic gene screening system based on a multivariate machine learning model, comprising the following sequentially connected components:

[0088] The data retrieval module is used to retrieve single-cell or transcriptome data related to human and mouse lung diseases in batches from public databases;

[0089] The data preprocessing module is used to perform quality control, normalization, clustering, and cell type annotation on the raw data, and generate a standardized gene expression matrix.

[0090] The candidate gene scoring module uses six machine learning algorithms—LASSO, random forest, decision tree, support vector machine, gradient boosting machine, and XGBoost—to train the model and output gene importance scores.

[0091] The weight calculation module calculates weights based on the model performance of human and mouse datasets and performs Min-Max normalization on the scores of each algorithm.

[0092] The comprehensive scoring module integrates human and mouse data using a weighted fusion formula to generate a final comprehensive score for the target gene and outputs a visual evaluation chart.

[0093] The results output module outputs a list of highly reliable lung disease characteristic genes and a visualization report based on the comprehensive scoring threshold.

[0094] Furthermore, the data preprocessing module specifically includes:

[0095] The data quality control unit performs the following operations:

[0096] Gene detection rate >200 and <5000, mitochondrial gene expression rate <20%, and genes expressed in at least 3 cells are preserved;

[0097] The batch effect removal unit uses an adaptive batch effect correction strategy to dynamically select the Harmony, SCTransform, or scVI algorithm based on the characteristics of the dataset and the degree of batch effect, thereby eliminating systematic bias between technical repetitions and ensuring comparability between datasets.

[0098] Cell clustering unit: Unsupervised cell population identification is achieved based on Leiden graph clustering algorithm. By constructing k-nearest neighbor graph and optimizing modularity function, cell clustering results at optimal resolution are obtained.

[0099] Cell type annotation unit, establishing a two-level cell type annotation system:

[0100] Level 1 – Coarse-grained annotation: Integrating mainstream annotation database resources such as scType and scCATCH, and combining cell type marker genes from published literature, to achieve annotation of major cell types;

[0101] Level 2 – Fine-grained annotation: Based on the Level 1 annotation results, cell subpopulations and functional states are further annotated to meet research objectives and disease-specific needs, forming a high-resolution cell type atlas.

[0102] Furthermore, in the candidate gene scoring module:

[0103] The LASSO regression model uses L1 penalized logistic regression. For binary classification, solver='saga' and C=1 are set, while for multi-class classification, the OneVsRestClassifier framework is used with C=0.1.

[0104] The random forest model contains 150 decision trees, with a minimum number of split samples of 10 and a minimum number of leaf node samples of 10.

[0105] The support vector machine uses a linear kernel function, with a regularization parameter C=0.1 and a maximum sample size limit of 5000.

[0106] XGBoost settings: maximum depth = 5, learning rate = 0.1, number of estimators = 100;

[0107] The gradient boosting machine preferentially uses LightGBM with parameters of 100 estimators, a learning rate of 0.1, and a maximum depth of 3.

[0108] The decision tree is set to a maximum depth of 10 and a minimum number of split samples of 5.

[0109] Furthermore, this application also claims protection for an electronic system for enabling cross-species screening of genes characteristic of lung diseases, including:

[0110] processor;

[0111] Memory, which stores computer programs;

[0112] When the processor executes the computer program, it implements the steps of the method as described in any of the foregoing.

[0113] Furthermore, this application also claims protection for a storage device storing computer programs.

[0114] When the computer program is executed by a processor, it implements the steps of the method as described in any of the foregoing.

[0115] Compared with existing technologies, the present invention provides a method, system, electronic system, and storage device for screening characteristic genes of lung diseases across species based on a multivariate machine learning model, achieving the following beneficial technical effects:

[0116] This invention significantly overcomes the bias of single models by integrating six complementary machine learning algorithms, greatly improving the stability and anti-interference ability of feature gene screening; it innovatively integrates human and mouse cross-species data, breaking through the limitation of traditional methods that ignore biological conservation, ensuring that the screening results have strong generalization and translational value; at the same time, it constructs a visual evaluation system, providing an efficient and reliable computing platform for the analysis of lung disease mechanisms, the development of diagnostic biomarkers, and the discovery of drug targets. Attached Figure Description

[0117] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0118] Figure 1 This is a flowchart illustrating a cross-species lung disease feature gene screening method based on a multivariate machine learning model, as provided in an embodiment of the present invention.

[0119] Figure 2 This is a schematic diagram of a cross-species lung disease feature gene screening system based on a multivariate machine learning model, provided as an embodiment of the present invention.

[0120] Figure 3 This is a structural block diagram of an electronic system provided in an embodiment of the present invention. Detailed Implementation

[0121] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0122] like Figure 1 As shown, this application discloses a method for screening characteristic genes of lung diseases across species based on a multivariate machine learning model, comprising the following steps:

[0123] S1: Data Acquisition and Preprocessing — Batch acquisition of single-cell or transcriptome data of human and mouse lung diseases from public databases, quality control, normalization, clustering and cell type annotation, and generation of standardized gene expression matrices;

[0124] In the embodiments of this application, single-cell / transcriptome data related to human and mouse lung diseases are obtained in batches from public databases such as GEO and EMBL-EBI. The raw data are then subjected to quality control, normalization, clustering, and cell type annotation to form a standardized gene expression matrix.

[0125] S2: Multivariate Machine Learning Model Training - The model is trained using six algorithms: LASSO, Random Forest, Decision Tree, Support Vector Machine, Gradient Boosting Machine, and XGBoost, and the gene importance score of each algorithm is output.

[0126] S3: Weight Calculation and Normalization - Calculate weights based on model performance using human and mouse datasets, and perform Min-Max normalization on the scores of each algorithm;

[0127] S4: Cross-species integrated scoring and output - Integrating human and mouse scores through a weighted fusion formula to obtain a comprehensive score for the target gene and output a list of characteristic genes;

[0128] In the embodiments of this application, a weighted fusion formula is used to integrate the scores of human and mouse six-model cross-species analysis to obtain the final comprehensive score of the target gene. Based on the comprehensive score threshold, a list of high-confidence lung disease characteristic genes and a visualization report are output.

[0129] Furthermore, in the embodiments of this application, the specific steps of data acquisition and preprocessing S1 include:

[0130] S11: Database search strategy development. Based on the research objectives for lung diseases, develop search queries covering disease types such as silicosis, pulmonary fibrosis, and chronic obstructive pulmonary disease. Search keywords include, but are not limited to, "lung disease," "pulmonary fibrosis," "COPD," "silicosis," "single cell RNA sequencing," and "transcriptome," ensuring that the search queries are sufficiently broad so that the target disease data constitutes a subset of the search results.

[0131] S12: Batch Acquisition of Cross-Species Data. Single-cell transcriptome data related to lung diseases in humans (Homo sapiens) and mice (Mus musculus) were acquired in batches from public databases such as GEO (Gene Expression Omnibus) and EMBL-EBI (European Molecular Biology Laboratory-European Bioinformatics Institute). The search results were filtered according to criteria such as data quality, sample size, and experimental design, and single-cell transcriptome datasets that met the research theme and were of acceptable quality were downloaded in batches.

[0132] S13: Data Quality Control and Merging. The downloaded raw data undergoes quality assessment, including checks on indicators such as gene detection rate, cell viability, and mitochondrial gene expression ratio, removing low-quality samples and genes. Multiple datasets are merged using inner join to ensure consistency and integrity of the merged data.

[0133] S14: Batch effect removal and standardization. Based on data characteristics and the severity of batch effects, batch effect removal methods such as Harmony, SCTransform, and scVI (single-cell Variational Inference) are dynamically selected and applied to eliminate technical differences caused by different experimental batches, sequencing platforms, and experimental conditions; the batch effect-removed data are then standardized to generate a standardized gene expression matrix.

[0134] S15: Cell Clustering and Type Annotation. The Leiden algorithm was used to perform unsupervised clustering analysis on the preprocessed single-cell data to identify different cell populations. Cell type identification was performed by combining manual annotation and large language model-assisted annotation, establishing a two-level annotation system: Level 1 annotation grouped cell subpopulations into major cell types, such as epithelial cells, immune cells, and mesenchymal cells; Level 2 annotation preserved cell subpopulation information, such as alveolar epithelial cell type I and II, and T cell subpopulations. The appropriate annotation level was selected for analysis based on the required accuracy of subsequent analyses.

[0135] Furthermore, in the embodiments of this application, the specific steps of training the multivariate machine learning model S2 include:

[0136] S21: LASSO Regression Model Construction. The appropriate model architecture is selected based on the number of categories in the classification problem: For binary classification problems, a logistic regression model with L1 penalty is used, the solver is set to 'saga', and the regularization strength C=1; for multi-class classification problems, the OneVsRestClassifier framework combined with L1 regularized logistic regression is used, with a regularization strength C=0.1 and a maximum number of iterations set to 10,000 to ensure model convergence; feature selection is automatically performed based on the sparsity of LASSO regression, and the absolute value of the coefficients is extracted as the gene importance score.

[0137] S22: Random Forest Model Training. A random forest classifier with 150 decision trees was constructed, with a minimum split sample count of 10 and a minimum leaf node sample count of 10. Parallel computation was used to accelerate the training process. Model performance was evaluated using the Out-of-Bag (OOB) error rate, and a curve showing the OOB error rate as a function of the number of trees was plotted. Feature importance scores for each gene were calculated based on Gini impurity.

[0138] S23: Support Vector Machine (SVM) model implementation. To improve computational efficiency, stratified sampling is performed on the large-scale dataset, with a maximum sample size limit of 5000. A linear kernel SVM model is adopted, with a regularization parameter C=0.1, and class weight balancing and probability estimation functions are enabled. The OneVsRestClassifier framework is used to handle multi-class classification problems, and feature importance is calculated through the absolute value of the coefficients of the linear SVM.

[0139] S24: XGBoost Gradient Boosting Model Configuration. Model parameters are set as follows: maximum depth 5, learning rate 0.1, number of estimators 100, subsample ratio 0.8, feature subsampling ratio 0.8; the objective function is automatically selected based on the number of classification categories: 'binary:logistic' for binary classification and 'multi:softprob' for multi-class classification; when XGBoost training fails, it automatically switches to Random Forest as an alternative model to ensure the stability of the analysis process.

[0140] S25: Gradient Boosting Machine Model Training. LightGBM is preferred as an initial implementation, with parameters including 100 estimators, a learning rate of 0.1, and a maximum depth of 3. If LightGBM is unavailable, GradientBoostingClassifier from sklearn is used. For large datasets (over 10,000 samples), a stratified sampling strategy is employed to reduce computational complexity. Gain-based feature importance scores are extracted.

[0141] S26: Decision Tree Model Construction. The maximum depth of the decision tree is set to 10, the minimum number of split samples to 5, and the minimum number of leaf node samples to 5 to prevent overfitting. Five-fold cross-validation is used to evaluate the model's generalization performance. Feature importance scores are calculated based on the reduction in Gini impurity, generating a highly interpretable gene importance ranking.

[0142] Furthermore, in the embodiments of this application, the weight calculation and normalization S3 includes the following steps:

[0143] S31: Calculation of Model Performance Evaluation Metrics. The absolute inverse residual and area under the ROC curve (AUC) are calculated for six machine learning algorithms on both human and mouse datasets as performance evaluation metrics. The absolute inverse residual is obtained by calculating the absolute value of the difference between the model's predicted probability and the one-hot encoding of the true label, reflecting the accuracy of the model's prediction. The ROC-AUC value is calculated using the area under the receiver operating characteristic curve. For binary classification problems, the AUC is calculated directly; for multi-class classification problems, a macro-averaging strategy is used to calculate the overall AUC.

[0144] S32: Weight Allocation Strategy Design. A weight allocation scheme is designed based on model performance metrics: AUC weights are allocated proportionally, with algorithms having higher AUC values ​​receiving greater weights; residual weights are allocated in reverse order, with algorithms having smaller mean absolute residuals receiving greater weights; the overall weight is calculated using the arithmetic mean of the AUC weights and residual weights to ensure a balance between the two evaluation dimensions.

[0145] S33: Min-Max Normalization Implementation. The gene importance scores output by each algorithm are normalized using Min-Max normalization, standardizing the score range to the [0,1] interval; the normalization formula is:

[0146]

[0147] The score of gene i in algorithm j;

[0148] The score of gene i in algorithm j after normalization;

[0149] The minimum score in algorithm j;

[0150] The maximum score in algorithm j;

[0151] Normalization eliminates differences in scoring dimensions and scales between different algorithms, ensuring the effectiveness of subsequent weight fusion;

[0152] S34: Cross-species weight consistency verification. The weights of each algorithm on both the human and mouse datasets are calculated, and correlation analysis is used to verify the consistency of cross-species weight allocation. A weight correction mechanism is established; when an algorithm shows significant differences in performance between the two species, a conservative strategy is adopted to adjust the weights, avoiding the influence of single-species data bias on the final results.

[0153] Furthermore, in the embodiments of this application, the cross-species integrated scoring and output S4 includes:

[0154] S41: Homologous gene mapping establishment. A human-mouse gene homology mapping table was established based on the NCBI HomoloGene database and the Ensembl Compara database. A one-to-one homologous gene matching strategy was adopted to ensure the biological significance of cross-species comparisons. In the case of multiple homologous genes, gene pairs with the highest sequence similarity and consistent functional annotations were selected first.

[0155] S42: Cross-species comprehensive score calculation. A weighted fusion strategy is used to integrate gene importance scores from human and murine data. The comprehensive score calculation formula is as follows:

[0156]

[0157] The performance weights of algorithm j on the mouse dataset;

[0158] The performance weights of algorithm j on the human dataset;

[0159] The number of algorithms that selected gene i in the mouse dataset;

[0160] The number of algorithms that selected gene i in the human dataset;

[0161] The score of gene i in algorithm j after normalization;

[0162] : Comprehensive score of cross-species gene i.

[0163] S43: Validation of scoring validity. Verify the biological relevance of high-scoring genes through gene function enrichment analysis; calculate the correlation coefficient between human and mouse scores to assess cross-species consistency; establish a dynamic adjustment mechanism for scoring thresholds, adaptively determining the feature gene screening threshold based on disease type and data quality.

[0164] This solution significantly improves the accuracy and stability of feature gene screening through multi-algorithm integration; enhances the model's generalization ability through cross-species information fusion; provides a complete visualization evaluation system, enhancing the model's interpretability; and offers an efficient and reliable computational tool for precision diagnosis and treatment of lung diseases and new drug development.

[0165] According to another aspect of the embodiments of this application, a cross-species lung disease feature gene screening system based on a multivariate machine learning model is also provided for implementing the above-described method for screening cross-species lung disease feature genes based on a multivariate machine learning model. Figure 2 This is a schematic diagram of an optional cross-species lung disease feature gene screening system based on a multivariate machine learning model, according to an embodiment of this application. Figure 2 As shown, the system may include:

[0166] The data retrieval module is used to retrieve single-cell or transcriptome data related to human and mouse lung diseases in batches from public databases;

[0167] The data preprocessing module is used to perform quality control, normalization, clustering, and cell type annotation on the raw data, and generate a standardized gene expression matrix.

[0168] The candidate gene scoring module uses six machine learning algorithms—LASSO, random forest, decision tree, support vector machine, gradient boosting machine, and XGBoost—to train the model and output gene importance scores.

[0169] The weight calculation module calculates weights based on the model performance of human and mouse datasets and performs Min-Max normalization on the scores of each algorithm.

[0170] The comprehensive scoring module integrates human and mouse data using a weighted fusion formula to generate a final comprehensive score for the target gene and outputs a visual evaluation chart.

[0171] The results output module outputs a list of highly reliable lung disease characteristic genes and a visualization report based on the comprehensive scoring threshold.

[0172] According to another aspect of the embodiments of this application, a server or terminal for a cross-species lung disease feature gene screening method based on a multivariate machine learning model is also provided.

[0173] Figure 3 This is a structural block diagram of a terminal according to an embodiment of this application, such as... Figure 3 As shown, the terminal may include: one or more (only one is shown in the figure) processors 401, memory 403, and transmission devices 405, such as... Figure 3 As shown, the terminal may also include input / output devices 407.

[0174] The memory 403 can be used to store software programs and modules, such as the program instructions / modules corresponding to the cross-species lung disease feature gene screening method and system based on a multivariate machine learning model in this embodiment. The processor 401 executes various functional applications and data processing by running the software programs and modules stored in the memory 403, thereby realizing the aforementioned cross-species lung disease feature gene screening method based on a multivariate machine learning model. The memory 403 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 403 may further include memory remotely located relative to the processor 401, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0175] The aforementioned transmission device 405 is used to receive or send data via a network, and can also be used for data transfer between the processor and memory. Specific examples of the network described above may include wired networks and wireless networks. In one example, the transmission device 405 includes a Network Interface Controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 405 is a radio frequency (RF) module used for wireless communication with the Internet.

[0176] Specifically, memory 403 is used to store application programs.

[0177] The processor 401 can invoke the application program stored in the memory 403 via the transmission device 405 to perform the following steps:

[0178] (a) Obtain single-cell data of human / mouse lung diseases from public databases and generate a standardized gene expression matrix;

[0179] (b) Gene importance scores are output using six machine learning algorithms;

[0180] (c) Calculate the cross-species comprehensive score based on model performance weights;

[0181] (d) Output a list of characteristic genes and a visualization report.

[0182] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments, and will not be repeated here.

[0183] Those skilled in the art will understand that Figure 3 The structure shown is for illustrative purposes only. The terminal can be a smartphone (such as an Android phone, an iOS phone, etc.), a tablet computer, a PDA, a mobile internet device (MID), a PAD, or other terminal devices. Figure 3 This does not limit the structure of the aforementioned electronic device. For example, the terminal may also include components that are more... Figure 3 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 3 The different configurations shown.

[0184] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0185] Embodiments of this application also provide a storage medium. Optionally, in this embodiment, the storage medium can be used to execute program code for a cross-species lung disease feature gene screening method based on a multivariate machine learning model.

[0186] Optionally, in this embodiment, the storage medium may be located on at least one of the multiple network devices in the network shown in the above embodiment.

[0187] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0188] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for screening characteristic genes of lung diseases across species based on a multivariate machine learning model, characterized in that, Includes the following steps: S1: Data Acquisition and Preprocessing — Batch acquisition of single-cell or transcriptome data of human and mouse lung diseases from public databases, quality control, normalization, clustering and cell type annotation, and generation of standardized gene expression matrices; S2: Multivariate Machine Learning Model Training - The model is trained using six algorithms: LASSO, Random Forest, Decision Tree, Support Vector Machine, Gradient Boosting Machine, and XGBoost, and the gene importance score of each algorithm is output. S3: Weight Calculation and Normalization - Calculate weights based on model performance using human and mouse datasets, and perform Min-Max normalization on the scores of each algorithm; In S3: S31: Calculation of Model Performance Evaluation Indicators Two performance metrics were calculated for six machine learning algorithms on both human and mouse datasets: (a) Absolute inverse residual: obtained by calculating the absolute value of the difference between the model's predicted probability and the actual one-hot encoding of the label, reflecting the accuracy of the model's prediction; (b) Area under the ROC curve (AUC): Calculated using the area under the receiver operating characteristic curve, where: (a1) For binary classification problems, the area under the ROC curve (AUC) is calculated directly. (b1) For multi-class classification problems, the macro-averaging strategy is used to calculate the area under the overall ROC curve (AUC). S32: Weight Allocation Strategy Design Design a weight allocation scheme based on S31 performance metrics: (A) ROC curve area under the curve (AUC) weight allocation: The weight is allocated proportionally, and the algorithm with the higher the ROC curve area under the curve (AUC) value receives a larger weight. (B) Residual weight allocation: The algorithm adopts reverse allocation, and the smaller the mean absolute residual, the greater the weight is obtained; (C) Calculation of comprehensive weights: The final weights are determined by the arithmetic mean of the area under the ROC curve (AUC) weights and the residual weights; S33: Min-Max Normalization Implementation The gene importance scores output by each algorithm are standardized: The normalization formula is: in, The score of gene i in algorithm j; The score of gene i in algorithm j after normalization; The minimum score in algorithm j; The maximum score in algorithm j; The score range is standardized to the [0,1] interval to eliminate the dimensional differences between algorithms; S34: Cross-species weight consistency verification Calculate the weights of each algorithm on the human and mouse datasets respectively; The consistency of cross-species weight allocation was verified through correlation analysis; Establish a weight correction mechanism: When the performance difference of an algorithm between two species exceeds a threshold, a conservative strategy is adopted to adjust the weights to avoid the influence of single-species data bias on the results; S4: Cross-species integrated scoring and output - Integrating human and mouse scores through a weighted fusion formula to obtain a comprehensive score for the target gene and output a list of characteristic genes; The cross-species comprehensive scoring formula in S4 is as follows: in, The performance weights of algorithm j on the mouse dataset; The performance weights of algorithm j on the human dataset; The number of algorithms that selected gene i in the mouse dataset; The number of algorithms that selected gene i in the human dataset; The score of gene i in algorithm j after normalization; : Comprehensive score of cross-species gene i.

2. The method according to claim 1, characterized in that, S1 includes: S11: Database retrieval strategy development – ​​Develop a search query covering silicosis, pulmonary fibrosis, and chronic obstructive pulmonary disease, with keywords including “lung disease”, “pulmonary fibrosis”, “COPD”, “silicosis”, “single cell RNA sequencing”, and “transcriptome”, to ensure that the target disease data is a subset of the search results; S12: Batch Acquisition of Cross-Species Data—Acquire single-cell transcriptome data of human (Homo sapiens) and mouse (Mus musculus) lung diseases in batches from the GEO and EMBL-EBI databases, and filter and download qualified datasets according to data quality, sample size and experimental design; S13: Data quality control and consolidation, including: Quality assessment: Examine gene detection rate, cell viability, and mitochondrial gene expression ratio; Remove low-quality samples and genes; Multiple datasets are merged using an inner join method; S14: Batch effect removal and standardization, including: Dynamically select Harmony, SCTransform, or scVI to remove batch effects; The processed data is standardized to generate a standardized gene expression matrix; S15: Cell clustering and type annotation, including: Unsupervised clustering analysis was performed using the Leiden algorithm. By combining manual annotation with large language model-assisted annotation, the following two-level system is established: Level 1: Classified into major cell types, including epithelial cells, immune cells, and stromal cells; Level 2: Preserves information on cell subsets, including alveolar epithelial cells type I and II, and T cell subsets.

3. The method according to claim 1, characterized in that, S2 includes the following steps: S21: Construction of LASSO regression model: For binary classification problems: adopt a logistic regression model with L1 penalty, set solver='saga', regularization strength C=1; For multi-class classification problems: the OneVsRestClassifier framework is used in combination with L1 regularized logistic regression, with C=0.1 and a maximum number of iterations of 10000. Gene importance scores are output based on the absolute value of the coefficients. S22: Random Forest Model Training: Construct a classifier containing 150 decision trees; Set the minimum number of split samples to 10 and the minimum number of leaf node samples to 10. Calculate the importance of features based on Gini impurity; S23: Support Vector Machine Model Implementation: A linear kernel function is used, with a regularization parameter C=0.1; Stratified sampling is used when the sample size is >5000. Feature importance is calculated using the absolute values ​​of linear SVM coefficients; S24: XGBoost gradient boosting model configuration: Parameter settings: Maximum depth = 5, Learning rate = 0.1, Number of estimators = 100; The objective function for binary classification is 'binary:logistic', and for multi-class classification it is 'multi:softprob'. Automatically switch to random forest when training fails; S25: Gradient Boosting Machine Model Training: We prioritize using LightGBM, with parameters of 100 estimators, a learning rate of 0.1, and a maximum depth of 3. Stratified sampling is enabled when the sample size is greater than 10,000. Based on the importance of gain calculation characteristics; S26: Decision Tree Model Construction: Set the maximum depth to 10, the minimum number of split samples to 5, and the minimum number of leaf node samples to 5. The importance of features is calculated based on the reduction in Gini impurity.

4. A system for implementing the cross-species lung disease characteristic gene screening method based on a multivariate machine learning model as described in claim 1, characterized in that, Including those connected sequentially: The data retrieval module is used to retrieve single-cell or transcriptome data related to human and mouse lung diseases in batches from public databases; The data preprocessing module is used to perform quality control, normalization, clustering, and cell type annotation on the raw data, and generate a standardized gene expression matrix. The candidate gene scoring module uses six machine learning algorithms—LASSO, random forest, decision tree, support vector machine, gradient boosting machine, and XGBoost—to train the model and output gene importance scores. The weight calculation module calculates weights based on the model performance of human and mouse datasets and performs Min-Max normalization on the scores of each algorithm. The comprehensive scoring module integrates human and mouse data using a weighted fusion formula to generate a final comprehensive score for the target gene and outputs a visual evaluation chart. The results output module outputs a list of highly reliable lung disease characteristic genes and a visualization report based on the comprehensive scoring threshold.

5. The system according to claim 4, characterized in that, The data preprocessing module specifically includes: The data quality control unit performs the following operations: Gene detection rate >200 and <5000, mitochondrial gene expression rate <20%, and genes expressed in at least 3 cells are preserved; The batch effect removal unit uses an adaptive batch effect correction strategy to dynamically select the Harmony, SCTransform, or scVI algorithm based on the characteristics of the dataset and the degree of batch effect, thereby eliminating systematic bias between technical repetitions and ensuring comparability between datasets. Cell clustering unit: Unsupervised cell population identification is achieved based on Leiden graph clustering algorithm. By constructing k-nearest neighbor graph and optimizing modularity function, cell clustering results at optimal resolution are obtained. Cell type annotation unit, establishing a two-level cell type annotation system: Level 1 – Coarse-grained annotation: Integrating mainstream annotation database resources such as scType and scCATCH, and combining cell type marker genes from published literature, to achieve annotation of major cell types; Level 2 – Fine-grained annotation: Based on the Level 1 annotation results, cell subpopulations and functional states are further annotated to meet research objectives and disease-specific needs, forming a high-resolution cell type atlas.

6. The system according to claim 4, characterized in that, In the candidate gene scoring module: The LASSO regression model uses L1 penalized logistic regression. For binary classification, solver='saga' and C=1 are set, while for multi-class classification, the OneVsRestClassifier framework is used with C=0.

1. The random forest model contains 150 decision trees, with a minimum number of split samples of 10 and a minimum number of leaf node samples of 10. The support vector machine uses a linear kernel function, with a regularization parameter C=0.1 and a maximum sample size limit of 5000. XGBoost settings: maximum depth = 5, learning rate = 0.1, number of estimators = 100; The gradient booster prioritizes LightGBM with the following parameters: number of estimators 100, learning rate 0.1, and maximum depth 3. The decision tree is set to a maximum depth of 10 and a minimum number of split samples of 5.

7. An electronic system for screening genes characteristic of lung diseases across species, characterized in that, include: processor; Memory, which stores computer programs; When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-3.

8. A storage device storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-3.

Images