PETA Boosts Virtual Screening Efficiency with Adaptation.

Virtual screening, a crucial step in drug discovery, involves accurately ranking potential drug molecules (ligands) for a specific protein target. While advanced models like DrugCLIP have accelerated this process, further performance gains typically require computationally expensive full model retraining for each new target protein. This research introduces PETA (Parameter-Efficient Test-Time Adaptation), a framework designed to specialize pretrained virtual screening models for individual protein pockets directly at test time. PETA significantly reduces computational overhead by adapting only a tiny fraction of the model's parameters, specifically the LayerNorm parameters, which constitute approximately 0.03% of the full model. PETA achieves this by constructing challenging pocket-specific negative examples through molecular diffusion and chemical validity filtering. It then uses embedding-space mixup to move these negatives closer to reference ligands, creating more effective ranking tasks. A specialized ranking objective prioritizes suppressing high-scoring invalid candidates, providing structured supervision for this lightweight adaptation. Experiments show PETA outperforms both pretrained and fully retrained baselines across diverse benchmarks.