Bioactivity foundation model for hit identification and hit-to-lead optimization from sequence data

PSICHIC+ predicts binding-affinity probability distributions from protein sequence and ligand SMILES, trained on ~82 million records, enabling millisecond virtual screening and hit-to-lead optimization without target-specific data.

• Why it matters: Large bioactivity archives are hard to learn from because assay readouts differ and data reliability varies, limiting how well sequence-only models can exploit heterogeneous measurements.
• What they did: The authors train PSICHIC+ on ~82 million protein–ligand records with a distributional objective that preserves quantitative resolution while accounting for reliability variation, then benchmark zero-shot screening and lab-in-the-loop acquisition.
• The result: In zero-shot virtual-screening, it improves enrichment versus regression/classification methods and matches leading structure-based approaches; distribution-guided acquisition boosts recovery under fixed budgets, and combining with Boltz-2 improves hit-to-lead affinity prediction.

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