Comparative Evaluation of Drug Response Metrics for Predicting Cancer Sensitivity Using Transcriptomic Profiles

Background: Pharmacogenomic modeling aims to predict cancer drug sensitivity from molecular features such as gene expression. However, the choice of drug response metric can critically affect model performance. This study systematically evaluates three commonly used metrics—area under the dose-response curve (AUC), Z-score, and half-maximal inhibitory concentration (IC50)—to determine their relative suitability for machine learning-based prediction using transcriptomic data. Methods: We assembled an integrated dataset comprising 636 cancer cell lines and drug response profiles for 169 compounds, along with RNA-seq-based gene expression features. XGBoost regression models were trained separately using AUC, Z-score, and IC50 as response variables. Model performance was assessed using R2, Pearson correlation, and mean absolute error. Additionally, gene-level correlation analyses were conducted to evaluate linear associations between gene expression and drug sensitivity. Results: AUC-based models consistently outperformed those based on Z-score and IC50 in terms of predictive accuracy and robustness. The highest-performing drugs under the AUC framework included Nelarabine (R2 = 0.83), Sorafenib, and Venetoclax—all of which have established clinical relevance. In contrast, gene-wise Pearson correlation analysis revealed that most genes exhibited weak linear relationships with drug sensitivity across all metrics (|PCC| < 0.1), suggesting that response prediction depends on complex, multigenic interactions. Conclusion: AUC is a more reliable and informative drug response metric for transcriptome-based prediction of cancer sensitivity. The findings support the application of multivariate machine learning models and emphasize the importance of metric selection in pharmacogenomic modeling pipelines.