Publication 14-03-2025
High-throughput drug sensitivity screening of cancer cell lines (CCLs) offers a promising avenue for discovering new anti-tumor therapies. In this study, we leverage large-scale datasets and cell line transcriptomics to predict drug responses, emphasizing model interpretability and real-world applicability. By using large language models (LLMs) to associate drugs with mechanism-of-action (MOA) pathways, we identify key predictive genes enriched in MOA-related processes. Our approach improves accuracy by focusing on LLM-curated MOA genes and enhances translatability by aligning RNAseq data from CCLs with patient samples. Validating our method on TCGA data, we show that predicted top-scoring drugs align with actual treatments. This research was conducted in collaboration with researchers from Fondazione Pisana per la Scienza, whose expertise contributed to the experimental validation of effective therapies. Finally, we identify and experimentally validate potential treatments for pancreatic cancer and glioblastoma, demonstrating the clinical relevance of our approach.