scDrugMap: Benchmarking Large Foundation Models for Drug Response Prediction

Area of Science:

• Computational biology
• Genomics
• Pharmacology

Background:

• Drug resistance hinders cancer therapy effectiveness.
• Single-cell profiling reveals cellular heterogeneity driving resistance.
• Foundation models show promise for single-cell analysis but need evaluation for drug response prediction.

Purpose of the Study:

• To develop and evaluate large-scale foundation models for single-cell drug response prediction.
• To introduce scDrugMap, an integrated framework with a Python tool and web server.
• To benchmark eight single-cell foundation models and two large language models (LLMs).

Main Methods:

• Developed scDrugMap, a framework for evaluating foundation models on single-cell drug response.
• Utilized curated datasets (326,751 primary cells, 18,856 validation cells).
• Conducted pooled-data and cross-data evaluations with layer freezing and Low-Rank Adaptation (LoRA) fine-tuning.

Main Results:

• scFoundation achieved top performance in pooled-data evaluation (mean F1: 0.971).
• UCE excelled in cross-data evaluation on tumor tissue (mean F1: 0.774).
• scGPT showed strong zero-shot performance (mean F1: 0.858).

Conclusions:

• Presents the first comprehensive benchmark of foundation models for single-cell drug response prediction.
• scDrugMap provides a flexible platform for drug discovery and translational research.
• Highlights the potential of foundation models in personalized cancer therapy.