Skip to content

DrugCLIP

Summary

DrugCLIP embeds protein pockets and small molecules into a shared latent space so large compound libraries can be searched by dense retrieval (cosine similarity) rather than exhaustive docking. This runs up to 10 million times faster than docking. It consistently outperforms docking and deep-learning baselines on standard virtual-screening benchmarks.[1]

Framework

Both the pocket and molecule encoders are built on Uni-Mol, a transformer that takes 3D atomic features (rather than 1D sequence or 2D graph) as input. To pretrain the pocket encoder without relying solely on the limited number of experimentally resolved protein-ligand structures, the authors developed ProFSA (protein fragment-surrounding alignment): short peptide fragments are extracted from protein-only PDB structures and treated as "pseudo-ligands," with their surrounding structure treated as a "pseudo-pocket," exploiting the chemical similarity between intra-protein packing interactions (hydrogen bonds, ionic contacts, π-π stacking) and true protein-ligand interactions. This yielded 5.5 million pseudo-pocket-ligand pairs for contrastive-distillation pretraining, after which the model is fine-tuned on experimentally determined protein-ligand pairs (BioLiP2) using a bidirectional (pocket-to-molecule and molecule-to-pocket) contrastive loss.[1]

GenPack, a generative pocket-refinement component trained on PDBBind2020, improves pocket detection on predicted (rather than crystallographic) structures, supporting screening against targets modeled with tools such as AlphaFold 2 when no experimental structure exists.[1]

Performance and Validation

  • Outperformed selected docking (Glide-SP, Vina, Uni-Dock) and deep-learning baselines on the DUD-E and LIT-PCBA virtual-screening benchmarks (EF1% metric), and remained robust across sequence/molecule-identity cutoffs and pocket side-chain RMSD perturbations.[1]
  • Norepinephrine transporter (NET): of 100 DrugCLIP-selected compounds tested in a radio-ligand transport assay, 15% showed >60% inhibition, and cryo-EM structures of two inhibitors (0086-0043 and Y510-9709) in complex with NET were solved (PDB 9JF3 and 9JEL).[1]
  • Serotonin 2A receptor (5HT2AR): two identified agonists had EC50 <100 nM.[1]
  • Thyroid hormone receptor interactor 12 (TRIP12): a target with no reported holo structure or small-molecule ligand at the time; using only AlphaFold2-predicted structures with GenPack, DrugCLIP achieved a 17.5% hit rate by surface plasmon resonance, with two inhibitors further confirmed by enzymatic assay — the first reported small-molecule binders for this target.[1]
  • GenomeScreenDB: an open-access genome-wide screening resource covering ~10,000 human proteins (>20,000 pockets, using AlphaFold2-predicted structures and computational pocket detection) against >500 million public compounds, run in 24 hours on 8 GPUs and hosted at drugclip.com.[1]

See Also

  • AlphaFold 2 — supplies predicted structures for targets (like TRIP12) lacking experimental holo structures.
  • Boltz-2 and PPIFlow — pose/affinity co-prediction and binder-design approaches that complement DrugCLIP's retrieval-based screening.

Citations

[1] Jia et al. (2026), "Deep contrastive learning enables genome-wide virtual screening", Science. Supports: framework, ProFSA/GenPack, all benchmark and wet-lab validation figures above. Location: Abstract; Results (framework and ProFSA pretraining; benchmarking; NET and TRIP12 wet-lab validation; GenomeScreenDB).