Machine Learning-Driven Integration of Cancer Cell Phenotypes Predicts Cisplatin Sensitivity

Precision medicine has personalized anticancer therapies and has been considered standard practice. Although current cancer genomic profiling tests are powerful tools to predict the efficacy of molecular targeted drugs or immune checkpoint inhibitors, they are not readily applicable for classical anticancer agents. In this study, we report a novel concept of phenotype-based classification using machine learning analysis of gene expression patterns to predict the effectiveness of anticancer agents.

Hierarchical clustering of IC₅₀ values distinguished cisplatin-sensitive and resistant cell lines. Differentially expressed gene (DEG) analysis and SHAP value-based machine learning identified 26 key genes, and the cisplatin sensitivity predictor using 26 genes (CSP26G) model was developed.

Cisplatin-resistant A549CR cells experimentally confirmed the external validity of the CSP26G model. The model also classified patients with non-small cell lung cancer in The Cancer Genome Atlas (TCGA) clinical database into cisplatin-sensitive and cisplatin-resistant groups. The predicted sensitive group showed significantly longer survival than the predicted resistant group. Furthermore, CSP26G predicts not only cisplatin efficacy but also responsiveness to other DNA-damaging agents.

These findings indicate that the sensitivity prediction model constructed through the integration of DEG and machine learning analyses can forecast drug sensitivity, thereby contributing to the advancement of effective and personalized precision medicine in classical chemotherapies.