FoldCraft enables fold-conditioning of binder structure, enabling design of binders with diverse folds like TIM-barrels, solenoid folds or Ig-like domains. Using VHH conditioned framework FoldCraft can succesfully design single domain nanobody binders against diverse tergets.
FoldCraft also significantly improves the de novo design of VHH nanobody binders, a challenging class of antibodies due to their complex fold and flexible loops. It achieved in silico success rates of up to 19.5% against four therapeutically relevant targets (PD-1, PD-L1, IFNAR2, EGFR) using AlphaFold3 evaluations, a substantial improvement over RFAntibody, where designs often failed to meet binding confidence thresholds.

5. The framework allows for flexible post-hallucination sequence optimization, with options like full-sequence redesign using ProteinMPNN or SolMPNN to enhance structural confidence, and interface-restricted optimization to preserve binding geometry.

6. While FoldCraft shows significant advancements, current limitations include a tendency for oracle models (AlphaFold3/RoseTTAFold) to predict side-docking interactions for VHH designs, possibly due to training data biases. Also, it doesn't generalize to scFv antibody topologies, indicating a potential need for more accurate structure prediction models within the hallucination workflow, such as open-source AlphaFold3 reproductions.