Reproduce “PETA:Parameter-Efficient Test-Time Adaptation for Virtual Screening” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jia-Qi Lin, Yinghua Yao, Chang-Dong Wang, Yew-Soon Ong, Yuangang Pan
PETA adapts a pretrained retrieval model to rank molecules for protein pockets by constructing pocket-specific ranking examples and updating only ligand-encoder LayerNorm parameters. It uses a reference ligand and generated invalid candidates for supervision. On DUD-E, it reports EF₀.₅% of 42.34, versus 37.90 for DrugCLIP and 39.82 for BindCLIP; this benchmark result does not establish faster screening or validated drugs.
Key insights
- Chemically invalid molecules that are highly scored by the pretrained model (hard negatives) serve as effective, annotation-free adaptation signals when generated specifically for a target pocket.
- Adapting only a minimal subset of parameters, specifically LayerNorms in the ligand encoder, allows for efficient and targeted model specialization to individual protein pockets.
- A top-heavy ranking optimization objective, combined with latent space mixup for negatives, effectively prioritizes the correction of high-risk false positives and improves early enrichment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.