Summary
A bioRxiv preprint reports that targeted finetuning helped the co-folding model Boltz-1 predict a ligand-bound, inactive conformation of Werner helicase. The authors used ten previously unseen X-ray structures and report generalisation to other chemical series and related helicases.
A bioRxiv preprint reports that targeted finetuning enabled the protein–ligand co-folding model Boltz-1 to represent a ligand-bound, inactive state of Werner (WRN) helicase. The authors say they trained the model using ten previously unseen X-ray structures from a drug-discovery programme, and that it retained accuracy on the ATP-bound state.
Co-folding models aim to predict a protein’s structure together with a small molecule bound to it, using the protein sequence and the molecule’s structure as inputs. The new work addresses a challenge for these systems: a protein can adopt different shapes when different molecules bind, and a model may not have encountered the relevant binding site or shape in its training data.
What the finetuning changed
The authors report that Boltz-1 learned an allosteric binding site on WRN and a substantial conformational change associated with the enzyme’s inactive state. An allosteric site is a place where a molecule binds away from a protein’s main functional site and can affect its activity. In this case, the model was also reported to preserve accuracy for WRN in its ATP-bound state—a distinct state of the helicase.
The preprint further reports that the finetuned model generalized to different chemical series and transferred the conformational pattern to other members of the RecQ helicase family. That transfer depended on the sequence of the binding site, according to the authors.
The results support the authors’ view that limitations in training data, rather than a fixed architectural barrier, can prevent co-folding models from capturing ligand-induced protein states. Targeted finetuning offers a way to adapt a model when new structural and mechanistic data become available, potentially making it more useful for studying protein regulation and drug discovery.
Evidence and scope
The work is a bioRxiv preprint, posted on September 27, 2026. Its abstract describes a modelling study based on ten X-ray structures, but does not provide numerical accuracy measures or details of the evaluation design. The reported generalisation should therefore be read as the authors’ finding within the systems and tests described in the preprint, rather than as evidence that the approach will work for proteins broadly.