ETD

• Cancer
• Machine Learning
• Pharmacogenomics
• Precision Medicine

In precision oncology, accurate drug sensitivity prediction is critical for treatment selection. Many current drug response models restrict input features to known Mechanism of Action (MOA) pathway genes, overlooking poorly annotated genes and diverse pathways.
Here, we developed a two-step, MOA-agnostic pipeline combining XGBoost feature selection with TabPFN regression to predict sensitivity for 286 drugs across 946 cancer cell lines.
Our approach outperformed state-of-the-art MOA-driven method for over 90% of drugs, achieving a median Pearson correlation >0.6. MOA-associated genes comprised only ~25% of selected features, while pathway coverage more than doubled. Gene Set Enrichment Analysis (GSEA) revealed that fully data-driven TabPFN captures a broader, more coherent biological signaling landscape, identifying more pathways and critical survival-related axes missed by traditional models. We identified 60 recurrently relevant genes lacking Reactome annotation (T-dark), including several highly predictive candidates. Pan-cancer survival analysis using TCGA data demonstrated that expression of these T-dark genes significantly correlates with patient overall survival across multiple cancer types, highlighting their prognostic potential as novel biomarkers in oncology.