Predicting conformational changes induced by biomolecular interactions

Explicitly modeling the unbound-to-bound structural transition with K-Fold better resolves ligand/partner-induced conformational states than generating bound complexes alone, across multiple protein complex systems.

• Why it matters: Co-folding models often miss specific interaction-induced conformations, limiting accurate prediction of signaling, regulation, and drug-action states.
• What they did: K-Fold uses a generative learning framework that assembles independently predicted unbound priors into complexes via an ultrafast diffusion bridge, without multiple sequence alignments.
• The result: It accurately recovers non-coevolved interfaces in antibody-antigen and targeted-degrader complexes and ligand-directed regulatory states in GPCRs and kinases, enabling tractable sampling of multi-state landscapes without retraining.