Machine learning-based model construction method and system for predicting osimertinib resistance of lung adenocarcinoma

Abstract

The present application relates to the technical field of medical data analysis, and more particularly to a model construction method and system for predicting lung adenocarcinoma osimertinib resistance based on machine learning. By integrating lung adenocarcinoma patient clinical data and public database mRNA data, after pretreatment such as Mann-Whitney U test and FPKM standardization, a basic model driven by clinical indicators is constructed using XGBoost, and a prognosis model driven by gene characteristics is constructed through R package Mime, and TYMS and UAP1L1 are selected as key genes. Through residual correction and stacked integration, the model performance is optimized, and finally the AUC is improved to 0.924. Serum RNA detection verification shows that TYMS and UAP1L1 are significantly highly expressed in the drug resistance group. This method replaces the traditional biopsy by non-invasive serum detection, reduces the detection cost, and the model generalization ability is verified by an independent queue, providing an efficient and accurate solution for early warning and personalized treatment of lung adenocarcinoma patients with osimertinib resistance.

Description

Technical Field

[0001] This invention relates to the field of medical data analysis technology, and in particular to a model construction method and system for predicting osimertinib resistance in lung adenocarcinoma based on machine learning. Background Technology

[0002] Lung cancer is a common malignant tumor in the Chinese population, with lung adenocarcinoma being its most common type. Activating mutations in the epidermal growth factor receptor (EGFR) gene are common mutations in lung adenocarcinoma. Osimertinib, as a third-generation EGFR tyrosine kinase inhibitor, is widely used as a first-line clinical treatment; however, drug resistance affects long-term efficacy for patients. Current drug resistance monitoring relies on imaging examinations and tissue biopsies, which have limitations such as being invasive and costly. There is an urgent need for an efficient, economical, and non-invasive predictive method. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a model construction method and system for predicting osimertinib resistance in lung adenocarcinoma based on machine learning, so as to solve the problems of existing technologies that rely on imaging examinations and tissue biopsies to monitor osimertinib resistance, which are invasive, costly and unable to predict early and accurately.

[0004] According to a first aspect of the present invention, a method for building a model for predicting osimertinib resistance in lung adenocarcinoma based on machine learning is provided, characterized in that the method includes:

[0005] Collect clinical data from lung adenocarcinoma patients and osimertinib resistance-related mRNA data from public databases;

[0006] The clinical data of the patients with lung adenocarcinoma were preprocessed to screen for indicators significantly related to drug resistance, and differentially expressed genes in osimertinib resistance-related mRNA data in the public database were screened using a preset algorithm.

[0007] Using the results of screening indicators significantly associated with drug resistance, a basic predictive model based on clinical indicators was constructed, and a prognostic model of gene expression characteristics was constructed by screening statistically significant gene data through pre-defined univariate Cox regression analysis.

[0008] Gene features selected from a pre-defined prognostic model based on the gene expression characteristics are combined with the basic predictive model based on clinical indicators, and the model performance is optimized through a pre-defined residual correction and stacking ensemble method; the gene features selected from the pre-defined prognostic model based on the gene expression characteristics include: TYMS and UAP1L1 gene expression levels.

[0009] The optimized model is used to predict the risk of osimertinib resistance in patients, and the predicted probability value is output.

[0010] Furthermore, the clinical data of the lung adenocarcinoma patients include gender, age, EGFR mutation site, complete blood count, tumor markers, tumor stage, and osimertinib treatment time information.

[0011] Furthermore, the osimertinib resistance-related mRNA data in the public database includes cell line mRNA data before and after osimertinib resistance and mRNA data from lung adenocarcinoma patients.

[0012] Furthermore, the process of preprocessing the collected clinical data of lung adenocarcinoma patients, screening for indicators significantly associated with drug resistance, and using a pre-defined algorithm to screen for differentially expressed genes in osimertinib resistance-related mRNA data in the public database includes:

[0013] The clinical data of the patients with lung adenocarcinoma were collected and the indicators significantly associated with drug resistance were screened using the Mann-Whitney U test.

[0014] The osimertinib resistance-related mRNA data in the public database were sequentially normalized using FPKM, transformed using log2, and de-batch processed using pyComBat. Differentially expressed genes were then screened using the pyDEG algorithm.

[0015] Furthermore, the process of constructing a basic predictive model based on clinical indicators using the results of screening indicators significantly associated with drug resistance, and constructing a prognostic model based on gene expression characteristics using pre-defined gene data screened through univariate Cox regression analysis, includes:

[0016] Using the results of screening indicators significantly associated with drug resistance, the data were processed and modeled using the Xgboost algorithm;

[0017] Gene data with statistical significance were screened using univariate Cox regression analysis based on preset parameters, and a prognostic model was established using the R software package Mime.

[0018] Furthermore, the step of combining gene features selected from a pre-defined prognostic model based on the gene expression characteristics with the basic predictive model based on clinical indicators, and optimizing its model performance through a pre-defined residual correction and stacking ensemble method, includes:

[0019] Gene features selected from a pre-defined prognostic model based on the gene expression characteristics; the gene features include: TYMS and UAP1 L1;

[0020] RNA expression levels of TYMS and UAP1 L1 were examined using pre-defined serum samples.

[0021] Based on the detection results, the preset expression levels of TYMS and UAP1 L1 genes are used as new features, merged with the residual data, and a secondary correction model is trained to obtain the first processing result.

[0022] Using the first processing result, the outputs of multiple base models are taken as input and integrated through a meta-learner to generate the final prediction result, thereby optimizing the model performance.

[0023] Furthermore, the pre-defined method for measuring the expression levels of the TYMS and UAP1 L1 genes includes:

[0024] Total RNA was extracted using a serum or plasma RNA extraction kit;

[0025] cDNA was synthesized by reverse transcription, and gene expression levels were detected using SYBR Green quantitative PCR.

[0026] Using GAPDH as an internal reference gene, the relative expression level was calculated using the ΔΔCt method.

[0027] Furthermore, the construction of the basic predictive model based on clinical indicators also includes:

[0028] Five-fold cross-validation was used to optimize the parameters of the Xgboost model, including the learning rate, the number of trees, and the maximum depth.

[0029] Key clinical indicators, including NLR, lymphocyte percentage, and CEA, were identified through feature importance analysis.

[0030] Furthermore, the prognostic model of gene expression characteristics also includes:

[0031] Utilize various algorithms in the MIME package to construct prognostic models;

[0032] Evaluate model performance using C-index and AUC, and select the optimal algorithm combination;

[0033] The Mime package contains various algorithms, including: Random Forest, Elastic Network, Stepwise Cox Regression, CoxBoost, Partial Least Squares Regression (Cox), Generalized Boosting Regression Model, Survival Support Vector Machine, Ridge Regression, and Lasso Regression.

[0034] According to a second aspect of the present invention, a model building system for predicting osimertinib resistance in lung adenocarcinoma based on machine learning is provided, applied to the model building method for predicting osimertinib resistance in lung adenocarcinoma based on machine learning described in any one of the above embodiments, characterized in that the method includes:

[0035] The data acquisition module is used to collect clinical data of lung adenocarcinoma patients and osimertinib resistance-related mRNA data from public databases;

[0036] The data preprocessing module is used to preprocess the collected clinical data of lung adenocarcinoma patients, screen for indicators significantly related to drug resistance, and use a preset algorithm to screen differentially expressed genes in osimertinib resistance-related mRNA data in the public database.

[0037] The model building module is used to construct a basic predictive model based on clinical indicators by using the results of screening indicators that are significantly associated with drug resistance, and to construct a prognostic model of gene expression characteristics by using pre-defined gene data that are screened through univariate Cox regression analysis to select statistically significant gene data.

[0038] The model fusion module is used to combine gene features selected from a pre-defined prognostic model based on the gene expression characteristics with the basic prediction model based on clinical indicators, and optimize its model performance through a pre-defined residual correction and stacking integration method; the gene features selected from the pre-defined prognostic model based on the gene expression characteristics include: TYMS and UAP1L1 gene expression levels.

[0039] The predictive analytics module is used to predict the risk of osimertinib resistance in patients using an optimized model and outputs the predicted probability value.

[0040] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0041] By integrating clinical data from lung adenocarcinoma patients with mRNA data from public databases, and after preprocessing such as Mann-Whitney U test and FPKM standardization, a clinical indicator-driven basic model was constructed using XGBoost, and a gene feature-driven prognostic model was constructed using the R package Mime. TYMS (Thymidylate Synthetase) and UAP1L1 were screened out.

[0042] (UDP-N-Acetylglucosamine Pyrophosphorylase 1-like 1) was used as a key gene. Model performance was optimized through residual correction and stacked ensemble, ultimately achieving an AUC of 0.924 for accurate prediction. Serum RNA detection validation showed that TYMS and UAP1L1 were significantly overexpressed in the drug-resistant group. This method, using non-invasive serum detection to replace traditional biopsy, and with model generalization ability validated in independent cohorts, provides an efficient and accurate solution for early warning and personalized treatment of osimertinib resistance in lung adenocarcinoma patients.

[0043] Specifically, it is non-invasive: based on routine blood tests and serum RNA detection, it avoids the trauma of tissue biopsy and reduces patient suffering.

[0044] High efficiency: It can quickly process large amounts of data through machine learning algorithms to accurately predict drug resistance and provide a basis for timely adjustment of treatment plans in clinical practice.

[0045] Accuracy: The optimized model has significantly improved prediction performance, with the residual correction model achieving an AUC of 0.873 and the stacked ensemble model achieving an AUC of 0.924, which are superior to traditional monitoring methods.

[0046] Economic efficiency: By utilizing routine laboratory test indicators, testing costs are reduced, and medical resources are saved.

[0047] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0048] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0049] Figure 1 This is a schematic diagram illustrating a model construction method for predicting osimertinib resistance in lung adenocarcinoma based on machine learning, according to an exemplary embodiment.

[0050] Figure 2 This is a schematic diagram of the ROC curves illustrating the model in a test set that distinguishes between osimertinib-sensitive and osimertinib-resistant patients according to an exemplary embodiment;

[0051] Figure 3 This is a schematic diagram illustrating the identification of common DEGs for osimertinib resistance from GEO data according to an exemplary embodiment;

[0052] Figure 4 This is a schematic diagram illustrating the sorting of the C-index of each model in different cohorts of a prognostic model based on osimertinib resistance-related genes according to the average C-index in the cohort, as shown in an exemplary embodiment.

[0053] Figure 5 This is a schematic diagram illustrating the relationship between risk scores calculated using the optimal combination model StepCox[forward]+RSF and outcomes in different patient cohorts, according to an exemplary embodiment.

[0054] Figure 6 This is a schematic diagram illustrating the top ten results with statistically significant differences (P<0.05) after (A)GO analysis, according to an exemplary embodiment.

[0055] Figure 7This is a schematic diagram of the cluster analysis heatmap representation of the target gene in the GEO dataset for target gene expression verification (A) according to an exemplary embodiment;

[0056] Figure 8 This is a schematic diagram of ROC curves for a base model, a residual correction model, and a stacked model, according to an exemplary embodiment.

[0057] Figure 9 This is a schematic diagram illustrating the system composition of a machine learning-based model for predicting osimertinib resistance in lung adenocarcinoma, according to an exemplary embodiment. Detailed Implementation

[0058] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0059] Example 1

[0060] Please see Figure 1 , Figure 1 This is a schematic diagram illustrating a model construction method for predicting osimertinib resistance in lung adenocarcinoma based on machine learning, according to an exemplary embodiment. The method includes:

[0061] S1. Collect clinical data of lung adenocarcinoma patients and osimertinib resistance-related mRNA data from public databases;

[0062] S2. Perform data preprocessing on the collected clinical data of lung adenocarcinoma patients, screen for indicators significantly related to drug resistance, and use a preset algorithm to screen differentially expressed genes in osimertinib resistance-related mRNA data in the public database;

[0063] S3. Using the results of screening indicators that are significantly associated with drug resistance, construct a basic predictive model based on clinical indicators, and use pre-defined gene data that are statistically significant through univariate Cox regression analysis to construct a prognostic model of gene expression characteristics.

[0064] S4. Using gene features selected from the pre-defined prognostic model based on the gene expression characteristics, and combining them with the basic prediction model based on clinical indicators, the model performance is optimized through a pre-defined residual correction and stacking integration method; the gene features selected from the pre-defined prognostic model based on the gene expression characteristics include: TYMS and UAP1L1 gene expression levels.

[0065] S5. Use the optimized model to predict the risk of osimertinib resistance in patients and output the predicted probability value.

[0066] In one embodiment, as described in step S1, a comprehensive search was conducted at a certain hospital for patients diagnosed with lung adenocarcinoma whose medical records explicitly mentioned "osimertinib." After rigorous screening and careful identification, information on 245 patients was retrieved. To further investigate the resistance to osimertinib in the treatment of lung adenocarcinoma, 90 patients meeting the resistance criteria were selected as the resistance group based on established rigorous inclusion and exclusion criteria. Given that the study design needed to ensure the balance and scientific rigor of the comparison between the two groups, 90 patients were correspondingly selected from the remaining eligible non-resistant patients as the sensitive group to ensure consistency in sample size between the two groups, thus laying a solid foundation for subsequent in-depth and scientifically valuable comparative analysis. Therefore, a total of 180 patients from a certain hospital who underwent surgical resection or lung biopsy and were pathologically confirmed to have lung adenocarcinoma, and whose lung cancer driver gene test results were EGFR positive, were included. After being numbered, the following information was collected through the hospital's inpatient management system (HIS): gender, age, EGFR mutation site, complete blood count, serum levels of tumor markers such as carcinoembryonic antigen (CEA), cancer antigen 125, cytokeratin 19 fragment (CYFRA21-1), and neuron-specific enolase (NSE), tumor stage, and osimertinib treatment time. Simultaneously, complete blood counts and their derived indicators were calculated and recorded, such as the neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), lymphocyte-to-monocyte ratio (LMR), Systemic Inflammation Response Index (SIRI): absolute number of neutrophils × absolute number of monocytes / absolute number of lymphocytes, and Systemic Immune Response Index (SII): absolute number of neutrophils × absolute number of platelets / absolute number of lymphocytes.Data entry points: (1) Baseline data recording: Before patients start taking osimertinib, collect baseline clinical information, including imaging, laboratory tests and related molecular markers; (2) Efficacy confirmation: After patients take osimertinib, record relevant data when the follow-up assessment achieves objective response (PR) or stable disease (SD) (according to RECIST 1.1 criteria) (follow-up time points are based on clinical protocols, such as 6 weeks or 12 weeks); (3) Drug resistance confirmation: When patients are found to have drug resistance based on imaging assessment (PD) or gene testing (such as EGFR T790M disappearance or C797S mutation appearance), record relevant data.

[0067] Inclusion criteria: (1) All patients were confirmed by pathological examination to have lung adenocarcinoma with only EGFR mutation and who received first-line or second-line osimertinib treatment. (2) The patients had signed the "Informed Consent Form for Sample Collection" and agreed that the remaining samples after testing would be used for medical research.

[0068] Exclusion criteria: (1) patients who are participating in other clinical studies; (2) patients with incomplete data.

[0069] Grouping criteria: Patients are grouped based on resistance to osimertinib, into a resistant group and a sensitive group. The criteria for determining resistance are as follows:

[0070] Symptoms worsen: such as increased cough, chest tightness, and shortness of breath, and even the appearance of new symptoms such as bone pain and headaches that have spread to distant sites.

[0071] Advances in imaging: CT, MRI or PET-CT may reveal tumor enlargement, new lesions or progression of existing lesions.

[0072] Gene mutations, gene amplifications, and fusions can be detected through tumor tissue biopsy or liquid biopsy (ctDNA): such as EGFR-related mutations (C797S, L718Q, L792F mutations, etc., EGFR amplification), bypass activation (MET gene amplification, HER2 amplification, PI3KCA mutation, BRAF mutation, KRAS mutation), histological transformation (small cell lung cancer transformation, squamous cell carcinoma transformation), other drug resistance mechanisms (EMT (epithelial-mesenchymal transition), new gene fusions (such as RET, ALK, NTRK); among them, the first criterion for suspected diagnosis can only be confirmed after the second and third pieces of evidence are met.

[0073] Furthermore, regarding public data: mRNA data of three HCC827 cell lines before and after osimertinib resistance were downloaded from the GEO database: GSE223006 (platform GPL24676), GSE243565 (platform GPL16791), and GSE249721 (platform GPL21697); clinical and mRNA data of 442 patients diagnosed with lung adenocarcinoma (GSE72094, platform GPL15048); mRNA data of biopsy tissues from four matched patients diagnosed with lung adenocarcinoma (GSE253742, platform GPL24676) before and after osimertinib resistance treatment; and pre-processed data, including mRNA data, phenotypic characteristics, and clinical data, from the LUAD cohort in the TCGA database, downloaded from UCSC Xena (https: / / xena.ucsc.edu / ), a comprehensive cancer genomics data analysis platform developed by the University of California, Santa Cruz, UCSC.

[0074] In the specific implementation, as described in steps S2-S3, using the clinical data of patients in the drug-resistant and drug-sensitive groups as features, exploratory data analysis (EDA) is first performed to understand the data distribution, preliminarily explore data patterns and outliers, and prepare for subsequent data cleaning. The Mann-Whitney U test is performed, and those with significant differences between the two groups (P<0.05) are included as model features. The model is built using Python software libraries Pandas, Numpy, matplotlib, Scikit-Learn, and Xgboost. Scikit-Learn provides a simple and efficient tool for data mining and data analysis, offering rich algorithms and functions, supporting various machine learning tasks such as classification, regression, clustering, and dimensionality reduction, and is suitable for both supervised and unsupervised learning scenarios. Xgboost is a highly efficient gradient boosting algorithm. Its core idea is to combine many "weak learners" (usually simple decision trees) into a powerful "strong learner," thereby improving the model's predictive ability. It has advantages such as high efficiency, high accuracy, and flexibility. This study utilizes the XGBoost algorithm to process and model data for predicting drug resistance in patients, providing better precision medicine support. After establishing and initially evaluating the model, this study further introduces new features by establishing a Residual Correction Model (RCM) and model stacking to achieve this goal. The RCM aims to improve overall prediction accuracy by learning and correcting the error (residual) of the initial model's prediction. Specifically, a base model is first trained to make preliminary predictions about the target; then, a secondary model is trained based on the residuals of this model to capture information that the base model failed to learn. The final prediction result is obtained by adding the prediction value of the base model to the prediction value of the secondary model for the residuals. Model stacking is an ensemble learning method that combines the prediction results of multiple different base learners to construct a secondary learner (meta-model) to improve the model's generalization ability and prediction performance. Unlike other ensemble methods (such as Bagging and Boosting), the stacking method allows for the heterogeneity of base learners, meaning that different types of models can be used. In this study, to comprehensively evaluate the model's performance and ensure its generalization ability on unseen data, the dataset was divided into a training set and a test set (in a 7:3 ratio). K-fold cross-validation (K=5) was used, dividing the dataset into K subsets. Each time, K-1 subsets were used for training, and the remaining subset was used for validation. This process was repeated K times, and the final performance metric was the average.The evaluation metrics after model building include: accuracy, precision, recall, F1 score, ROC curve (Receiver Operating Characteristic curve), and AUC value. Finally, the best model is selected by comparing the multiple models built.

[0075] Furthermore, the Reads Count data obtained from the database for each group were first standardized. The FPKM (Fragments Per Kilobase of exon per Million reads mapped) value was calculated consistently based on gene length and sequencing depth, and then transformed using log2. Genes with low or no expression in both drug-resistant and drug-sensitive groups were filtered out. To unify the gene IDs across groups, the R package bioomaRt was used to uniformly identify the gene tags as EnsemblID. The pyComBat method built into the Bluk module of the Omicverse Python library was used to de-batch process the obtained data. Differentially Expressed Genes (DEGs) analysis was performed using the built-in pyDEG method, with all parameters maintained at the default values ​​given in the official documentation (https: / / omicverse.readthedocs.io / en / latest / ). The differential analysis compared the groups before and after osimertinib resistance, and the screening threshold was set to an absolute logarithm of fold change > 1.5 with P < 0.05. After obtaining the DEGs, we used RRA (RobustRank Aggregation) analysis, which was performed using the R package RobustRankAggreg. Genes with scores less than 0.005 were selected, and the results were presented as a heatmap.

[0076] In practice, lung cancer patient data from the TCGA and GEO databases were used as training data. The R software package Mime was used to build a prognostic model. Mime integrates many practical algorithms for building common clinical machine learning prognostic models, including data standardization, feature selection, and model training. These algorithms include Random Forest, Elastic Net, Stepwise Cox Regression, CoxBoost, Partial Least Squares Regression for Cox, Generalized Boosted Regression Models, Survival Support Vector Machine, Ridge Regression, and Lasso Regression. Furthermore, K-fold cross-validation was performed through a combination of multiple algorithms to select the optimal model (including 10 algorithms, which can be combined to build a total of 101 models). First, univariate Cox regression analysis is performed on the input genes to screen for statistically significant genes. Then, the screened genes are input into various models for training. At the same time, the performance validation methods provided by this integration package are used to validate the model performance, such as C-index and AUC. The optimal model is selected, and the features used by the optimal model are used as the target genes for the next step of the research.

[0077] As described in steps S4-S5, in one embodiment:

[0078] This study included information on 180 patients, with 90 patients in each of the drug-resistant and drug-sensitive groups. The clinical characteristics of the two groups are shown in Table 1 (comparison of clinical characteristics between drug-sensitive and drug-resistant patients). There were no statistically significant differences in clinical characteristics between the two groups (P>0.05).

[0079] Table 1

[0080]

[0081] Patient test results showed that the percentage of neutrophils, percentage of lymphocytes, absolute number of lymphocytes, CEA, CA125, CA153, NLR, PLR, LMR, SIRI, and SII had an FDR of less than 0.01 according to the Mann-Whitney U test, indicating that they were associated with the patient's drug resistance status. Other comparative results are also presented in Table 2 (comparison of test results between sensitive and resistant patients).

[0082] Table 2

[0083]

[0084]

[0085] In practice, indicators that showed statistical significance in the U-test were considered effective features for identifying drug resistance and were included as indicators in the model construction. The established model achieved a precision of 0.65 and a recall of 0.64 in distinguishing between osimertinib-sensitive and resistant patients. The area under the cross-validation receiver operating characteristic (AUC) was 0.716 ± 0.051, and the validation set AUC was 0.64 (see [link to model construction]). Figure 2 (Receiver Operating Characteristic curve).

[0086] Furthermore, differential analysis was performed on data from three groups of drug-resistant cell lines from different laboratories. The DEGs identified by dataset GSE223006 contained 253 upregulated genes and 199 downregulated genes. Figure 3 (A~C); DEGs confirmed by dataset GSE243565 contain 1205 upregulated genes and 1278 downregulated genes; DEGs confirmed by dataset GSE249721 contain 2162 upregulated genes and 2089 downregulated genes. Specifically, the figure shows a volcano plot of differentially expressed genes from each GEO database. RRA analysis yielded a total of 1001 co-expressed DEGs in the three cell lines before and after drug resistance, of which 473 were upregulated and 528 were downregulated as target genes. The figure shows the top 20 genes ranked in the RRA analysis (…). Figure 3 D).

[0087] Specifically, Figure 3 This indicates the common DEGs for osimertinib resistance identified from GEO data. (A) to (C) are differentially expressed gene volcano plots from multiple sets of data in the GEO database, respectively. (D) is the top 20 co-expressed genes with the highest scores obtained from the three sets of differentially expressed gene data through RRA analysis.

[0088] Furthermore, to identify predictive indicators of drug resistance, commonly expressed DEGs were used as features, and a prognostic model was constructed with patient prognosis as the outcome. Feature selection for a univariate Cox model was performed using increasingly popular and scalable algorithmic architectures on two large lung adenocarcinoma datasets (TCG A lung adenocarcinoma cohort of 542 cases and GEO lung adenocarcinoma cohort of 372 cases). Then, 101 models were constructed using the selected genes, and the best model was StepCox[forward]+GBM, which had a C-index of 0.72 on the TCGA training dataset, a C-index of 0.64 on the GEO validation dataset, an average C-index of 0.68 across cohorts, and an average C-index of 0.64 on the validation set. Other model parameters are detailed below. Figure 4 The model ultimately selected features (78 genes in total) as the target genes.

[0089] Based on the risk score calculated by the optimal model, patients were divided into high-risk and low-risk groups. Survival curves were plotted with survival time as the outcome, and log-rank tests were performed. In the TCGA dataset, the hazard ratio (HR) was 13.52, with a 95% confidence interval of 10.02–18.23; in the GEO dataset, the hazard ratio was 2.21, with a 95% confidence interval of 1.49–3.28. Survival analyses in both datasets were statistically significant (P<0.05). Results are shown below. Figure 5 .in, Figure 5 This represents the relationship between the risk score calculated by the optimal combination model StepCox[forward]+RSF and the outcomes of patients in different cohorts.

[0090] Please see Figure 6 Differential gene analysis was performed on the high-risk group within the risk grouping, and the DEGs obtained from this analysis were used for enrichment analysis. The results showed that the high-risk group in the drug resistance prognostic model was significantly enriched in biological processes such as cell cycle, mitotic cell cycle, and chromosome segregation; in cellular component aggregation regions, chromatin and cell membrane were observed; and in molecular function, nucleotide binding and drug binding were enriched. Furthermore, KEGG enrichment analysis indicated that drug resistance mainly occurs in cell cycle, cell senescence, p53 signaling pathway, FoxO signaling pathway, pyrimidine metabolism, and drug metabolism.

[0091] against Figure 6 (A) The top ten results with statistically significant differences (P<0.05) after GO analysis; (B) The top five results with statistically significant differences (P<0.05) after KEGG pathway enrichment analysis.

[0092] Furthermore, using features from previous prognostic models as target genes, we validated whether these genes have the function of predicting drug resistance and prognosis in patient tissues. We validated related genes using tissue datasets of osimertinib before and after resistance downloaded from the GEO database. TYMS, UAP1 L1, OLFM2A, and CL EC7A showed differences in the database; the comparison results of the remaining genes are presented in a heatmap format. Figure 7 A. Serum RNA was detected in 73 serum samples from our hospital to identify genes showing significant differences in the database. The results showed that TYMS and UAP1 L1 were increased in the drug-resistant group (P<0.05), while no significant differences were observed in the expression of other genes in serum and body fluids. Figure 7 BC).

[0093] Figure 7 The expression of the target gene is verified; (A) cluster analysis heatmap of the target gene in the GEO dataset; (B) and (C) serum RNA detection to verify the expression of TYMS and UAP1 L1 genes in the sensitive and drug-resistant groups.

[0094] After identifying the target gene, the remaining samples were measured. New features were added to the original drug resistance prediction model to enhance its performance. The residual correction model achieved an AUC of 0.924, while the stacked ensemble model achieved an AUC of 0.873. The ROC curves for each model are shown in the figure. Figure 8 Represents the ROC curves of the basic model, the residual correction model, and the stacked model.

[0095] The embodiments in this application retrospectively collected clinical data from 180 patients with EGFR mutation-positive lung adenocarcinoma at a certain hospital from March 2020 to December 2024, and grouped them according to drug resistance status. XGBoost and other machine learning algorithms were used for feature selection and model construction. The initial model used routine blood tests, biochemical examinations, and tumor marker results as input features. Subsequently, differential gene analysis was performed using public databases (GEO, TCGA) to screen for drug resistance-related genes. A prognostic model was established based on clinical data, and features with high clinical value were selected as target genes. Further verification of their association with drug resistance was performed using public databases, and serum RNA detection was conducted on drug-resistant samples to clarify the expression level of target genes. Residual correction and stacking methods were used to optimize and update the basic model using target genes. Finally, the predictive ability of the model was evaluated through cross-validation, AUC, accuracy, and other indicators.

[0096] In practice, the initially constructed predictive model performed poorly, with an area under the receiver operating characteristic (AUC) of 0.64 and a recall of 0.64, indicating that the model's predictive performance needs improvement. Through bioinformatics analysis, we identified 1001 differentially expressed genes, mainly enriched in cell cycle regulation, drug metabolism, and pyrimidine metabolism pathways. Using machine learning methods, we screened 73 prognostic-related genes from these differentially expressed genes and constructed a prognostic predictive model. GEO dataset and serum RNA detection results showed that thymidylate synthase (TYMS) and UDP-N-acetylglucosamine pyrophosphorylase 1-like 1 (UAP1 L1) were significantly overexpressed in osimertinib-resistant patients, suggesting their potential role in the resistance mechanism. Using TYMS and UAP1 L1 to optimize the model, the AUC of the residual correction model was improved to 0.873, and the AUC of the stacked ensemble model reached 0.924.

[0097] Furthermore, this application constructs a machine learning-based osimertinib resistance prediction model and screens out potential resistance biomarkers (TYMS, UAP1 L1), providing a new approach for non-invasive resistance monitoring. Future model optimization can be combined with multi-center data to enhance its clinical application value.

[0098] Please see Figure 9 , Figure 9 This is a schematic diagram illustrating the system composition of a machine learning-based model for predicting osimertinib resistance in lung adenocarcinoma, according to an exemplary embodiment. The system includes:

[0099] Data acquisition module 10 is used to collect clinical data of patients with lung adenocarcinoma and osimertinib resistance-related mRNA data in public databases;

[0100] The data preprocessing module 20 is used to preprocess the collected clinical data of lung adenocarcinoma patients, screen for indicators significantly related to drug resistance, and use a preset algorithm to screen differentially expressed genes in osimertinib resistance-related mRNA data in the public database.

[0101] The model building module 30 is used to build a basic predictive model based on clinical indicators by using the results of screening indicators that are significantly related to drug resistance, and to build a prognostic model of gene expression characteristics by using pre-defined gene data that are screened through univariate Cox regression analysis to select statistically significant gene data.

[0102] The model fusion module 40 is used to combine gene features selected from the pre-defined prognostic model based on the gene expression characteristics with the basic prediction model based on clinical indicators, and optimize its model performance through a pre-defined residual correction and stacking integration method; the gene features selected from the pre-defined prognostic model based on the gene expression characteristics include: TYMS and UAP1 L1 gene expression levels.

[0103] The predictive analysis module 50 is used to predict the risk of osimertinib resistance in patients using an optimized model and outputs the predicted probability value.

[0104] More specifically, data acquisition module 10 integrates clinical data of lung adenocarcinoma patients (such as blood routine and tumor markers) with gene data from public databases (such as mRNA before and after osimertinib resistance), covering phenotypic and molecular features, providing multi-dimensional data sources for the model, and ensuring the comprehensiveness and accuracy of predictions.

[0105] Data preprocessing module 20: Through statistical tests (such as the Mann-Whitney U test) and algorithm screening (such as pyDEG), redundant information is removed, key features (such as NLR and TYMS genes) are retained, data quality is improved, model training complexity is reduced, and the reliability of subsequent analysis is ensured.

[0106] Model building module 30:

[0107] Basic prediction model: Based on routine clinical indicators, a preliminary screening model can be quickly built to achieve a non-invasive preliminary assessment of drug resistance risk, which is suitable for primary healthcare scenarios.

[0108] Prognostic models: By mining deep-seated mechanisms of drug resistance (such as cell cycle regulation pathways) through gene expression data, and screening key biomarkers (such as TYMS and UAP1L1), we can provide direction for mechanism research and targeted therapy.

[0109] Model fusion module 40: Combines gene features (TYMS, UAP1 L1) with clinical models, improves prediction accuracy through residual correction and stacking integration (AUC increases from 0.64 to 0.924), achieves early and accurate warning of drug resistance risk, and reduces patient trauma and cost through non-invasive serum testing.

[0110] Predictive Analysis Module 50: Outputs quantitative prediction results (such as drug resistance probability values) to assist doctors in dynamically adjusting treatment strategies (such as combination therapy), promoting the implementation of personalized medicine, and improving patients' quality of life and treatment efficiency.

[0111] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0112] It should be noted that in the description of this invention, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this invention, unless otherwise stated, "a plurality of" means at least two.

[0113] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0114] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0115] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0116] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0117] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0118] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0119] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for constructing a model for predicting osimertinib resistance of lung adenocarcinoma based on machine learning, characterized in that, The method comprises: Collecting clinical data of lung adenocarcinoma patients and mRNA data of osimertinib-resistant cell lines from a GEO database; Data preprocessing is performed on the collected clinical data of lung adenocarcinoma patients, and indexes significantly related to drug resistance are screened; A basic prediction model based on clinical indexes is constructed using the results of screening indexes significantly related to drug resistance; The mRNA data of the osimertinib-resistant cell lines are preprocessed and analyzed for differentially expressed genes, and a plurality of common differentially expressed genes are obtained by screening; Using the plurality of common differentially expressed genes as features, combining the mRNA data of lung adenocarcinoma patients and their prognosis information from the TCGA and GEO public databases, performing feature screening by single-factor COX regression analysis, and training and verifying by a prognosis model construction process containing multiple machine learning algorithms, target genes related to prognosis are screened; Using the target genes containing TYMS and UAP1L1 gene expression, expression verification is performed in independent clinical serum samples; The TYMS and UAP1L1 gene expression verified by the above method is used as new features, combined with the basic prediction model based on clinical indexes, and the model performance is optimized by residual correction and stacking integration method; The optimized model is used to predict the osimertinib drug resistance risk of patients, and the prediction probability value is output.

2. The method of claim 1, wherein, The clinical data of lung adenocarcinoma patients include gender, age, EGFR mutation site, blood routine, tumor marker, tumor stage, and osimertinib medication time information.

3. The method of claim 1, wherein, The mRNA data of the osimertinib-resistant cell lines are preprocessed and analyzed for differentially expressed genes, including FPKM standardization, log2 conversion, and batch processing of the mRNA data, and differentially expressed genes are screened by pyDEG algorithm.

4. The method of claim 1, wherein, The data preprocessing of the clinical data includes Mann-Whitney U test for screening.

5. The method of claim 1, wherein, The basic prediction model based on clinical indexes is constructed by using the results of screening indexes significantly related to drug resistance and processing data by XGBoost algorithm.

6. The method of claim 1, wherein, The training and verification by the prognosis model construction process containing multiple machine learning algorithms include: using multiple algorithms in the R software package Mime to construct a prognosis model; the model performance is evaluated by C-index and AUC, and the optimal algorithm combination is selected; the optimal algorithm combination includes the combination of stepwise Cox regression and generalized boosting regression model.

7. The method of claim 1, wherein, The expression verification in independent clinical serum samples includes testing the RNA expression levels of TYMS and UAP1L1 by serum samples; Based on the detection results, the TYMS and UAP1L1 gene expression is used as a new feature, combined with residual data, a secondary correction model is trained, and a first processing result is obtained; Using the first processing result, the outputs of multiple base models are used as inputs, and a final prediction result is generated by a meta-learner to optimize the model performance.

8. The method of claim 7, wherein, The method for measuring the expression amount of the TYMS and UAP1L1 genes comprises the following steps: extracting total RNA by using a serum or plasma RNA extraction kit; synthesizing cDNA through reverse transcription; detecting the gene expression level by using SYBRGreen fluorescent quantitative PCR; taking GAPDH as a reference gene, and calculating the relative expression amount by using the ΔΔCt method.

9. The method of claim 5, wherein, The construction of the XGBoost model further comprises the following steps: optimizing model parameters by using 5-fold cross-validation, including a learning rate, a number of trees, and a maximum depth; determining key clinical indexes by performing feature importance analysis, including a neutrophil / lymphocyte ratio, a lymphocyte percentage, and a carcinoembryonic antigen.

10. The model construction system for predicting the drug resistance of lung adenocarcinoma to osimertinib based on machine learning, applied to the model construction method for predicting the drug resistance of lung adenocarcinoma to osimertinib based on machine learning according to any one of claims 1-9, characterized in that, The system comprises: a data acquisition module configured to collect clinical data of lung adenocarcinoma patients and mRNA data of an osimertinib-resistant cell line in a public database; a data preprocessing module configured to perform data preprocessing on the collected clinical data of lung adenocarcinoma patients, screen indexes significantly related to drug resistance, and screen differentially expressed genes in the mRNA data of the osimertinib-resistant cell line in the public database by using a preset algorithm; a model construction module configured to construct a basic prediction model based on clinical indexes by using the results of screening indexes significantly related to drug resistance, and construct a prognosis model based on gene expression characteristics by using gene data with statistical significance screened by single-factor COX regression analysis; a model fusion module configured to combine gene features screened in a preset prognosis model based on the gene expression characteristics with the basic prediction model based on clinical indexes, and optimize the model performance by using a preset residual correction and stacking integration method; the gene features screened in the preset prognosis model based on the gene expression characteristics comprise the expression amounts of TYMS and UAP1L1 genes; a prediction analysis module configured to predict the risk of osimertinib resistance of a patient by using the optimized model, and output a prediction probability value.

Images