Metabolomic machine learning-based model predicts efficacy of chemoimmunotherapy for advanced lung squamous cell carcinoma.

Background: Unlike lung adenocarcinoma, patients with advanced squamous carcinoma exhibit a low proportion of driver gene positivity, with fewer effective treatment strategies available. Chemoimmunotherapy has now become the standard first-line treatment for individuals diagnosed with advanced lung squamous carcinoma. Serum metabolomics holds significant potential for application in predicting responses to chemoimmunotherapy and is capable of identifying and validating potential biomarkers. The aim of our study was to establish a model that can predict the prognosis of chemoimmunotherapy in patients with advanced lung squamous cell carcinoma, integrating metabolomics with machine learning techniques. Methods: We collected 79 serum samples from patients with advanced lung squamous cell carcinoma before receiving combined immunotherapy and performed untargeted metabolomics analysis. Patients were divided into non-response (NR) and response (R) groups according to overall survival (OS), and prognostic models were constructed and validated using different machine learning methods. The patients were further categorized into high-risk and low-risk groups based on the median risk score, to assess the model's predictive performance. Results: There were significant differences in metabolites and metabolic pathways between NR and R groups, and 117 differential metabolites were preliminarily screened (p < 0.05, VIP > 1). Further, least absolute shrinkage and selection operator (LASSO) and random forest (RF) were used to identify metabolites, and then their common metabolites were used as the best biomarkers to build a prediction model containing 8 differential metabolites. Based on these biomarkers, RF, support vector machine (SVM) and logistic regression were used to randomly divide patients into training and validation sets in a 7:3 ratio, respectively. We found that the RF method resulted in area under curves (AUCs) of 0.973 and 0.944 for the training and validation sets, respectively, with the best predictive performance. Subsequently, both OS and progression-free survival (PFS) were notably reduced in the high-risk group when contrasted with the low-risk group. Conclusions: We developed a model containing 8 metabolites based on metabolomics and machine learning that may predict survival outcomes in patients with advanced lung squamous cell carcinoma undergoing chemoimmunotherapy, helping to more accurately assess efficacy and prognosis in clinical practice.