Interrogating contrastive learning embeddings for structure-based virtual screening: a case study on DrugCLIP

DrugCLIP pocket embeddings enable pocket similarity search, achieving state-of-the-art performance while running over 10,000 times faster than existing structural descriptors.

• Why it matters: It is unclear what contrastive embeddings encode and how they map to conventional structural and chemical similarity, limiting reliable use in structure-based virtual screening.
• What they did: The study systematically dissects DrugCLIP’s latent space, evaluating pocket and ligand embeddings with a de-leakage benchmark across 223 pockets and a library of 48,999 compounds.
• The result: For pockets with SI ≤ 30% and TS ≤ 0.7, the crystallographic ligand is retrieved within the top 1% for 56.5% of pockets, but performance drops under realistic conditions due to sidechain and pocket-residue mismatches.

AI summary of the abstract · may contain errors · not medical advice · read the paper.