Abstract
Drug resistance remains a major challenge in cancer treatment. While single-cell profiling offers unprecedented resolution for uncovercovering resistance mechanisms, the potential of emerging foundation models for drug response prediction at the single-cell level is still largely unknown. Here, we introduce scDrugMap, a unified framework featuring both Python toolkits and an interactive web server for benchmarking and predicting drug responses with state-of-the-art foundation models. scDrugMap evaluates eight single-cell foundation models and two large language models across 495,000 cells from 60 datasets, spanning diverse tissues, drugs, cancer types, and treatment conditions. In pooled-data evaluation, scFoundation delivered the strongest performance, particularly in tumor tissue. In cross-data analysis, UCE performed best after fine-tuning, while in zero-shot settings, scGPT achieved the highest accuracy. Together, scDrugMap provides the first systematic benchmark of foundation models for single-cell drug response prediction and offers a powerful, user-friendly platform to accelerate drug discovery and translational precision oncology.