I used my custom version of BindCraft modified such that you can give a partial structure as input, as well as a target, and added an ipSAE loss that was up-weighted compared to the other losses. As input I used a known nanobody scaffold (https://www.nature.com/articles/s41594-018-0028-6), and allowed only the CDR loops to be redesigned. The starting structure was based on the placement of an existing Hemagglutinin-neuraminidase antibody binder (PDB: 8TQI) where only the heavy chain is in contact with the viral protein, making it more nanobody-like. The Nipah target protein was heavily truncated for more efficient iterations.

Unlike BoltzGen which uses Boltz, BindCraft uses ColabFold with AlphaFold2 weights, and therefore does not hyperoptimize for this competition's Boltz-2 ipSAE score. As was shown by other users, the Boltz-2 and AlphaFold2/3 ipSAE values do not always correlate, therefore it would be interesting to see if high AlphaFold2 ipSAE designs (which these submissions had) also show binding affinity, even if their Boltz-2 ipSAE scores are not as high.