scDrugMap: benchmarking large foundation models for drug response prediction

Area of Science:

• Computational biology
• Genomics
• Pharmacology

Background:

• Drug resistance is a significant hurdle in cancer therapy.
• Single-cell profiling reveals resistance mechanisms, but foundation models for drug response prediction are under-explored.

Purpose of the Study:

• To introduce scDrugMap, a framework for benchmarking and predicting drug responses using single-cell foundation models.
• To systematically evaluate the performance of various foundation models in single-cell drug response prediction.

Main Methods:

• Evaluated eight single-cell foundation models and two large language models.
• Analyzed 495,000 cells from 60 diverse datasets covering various tissues, drugs, and cancer types.
• Assessed model performance in pooled-data, cross-data (fine-tuned), and zero-shot settings.

Main Results:

• scFoundation demonstrated strong performance, especially in tumor tissues.
• UCE achieved the best results after fine-tuning in cross-data analysis.
• scGPT exhibited the highest accuracy in zero-shot predictions.

Conclusions:

• scDrugMap offers the first comprehensive benchmark of foundation models for single-cell drug response prediction.
• The platform accelerates drug discovery and supports translational precision oncology by providing user-friendly tools.