To design VHH (nanobodies) against the Nipah virus glycoprotein G, a highly truncated version of the 2VSM structure was used as the antigen input for the public Germinal model. Germinal generates VHH designs using AF2-multimer hallucination jointly optimized with the antibody language model IgLM. Several nanobody frameworks were tested, including the default humanized VHH provided in the Germinal repository, the PDB:7XL0 framework, and the TPL1158_01_C09 framework from PDB:9GCN (a nanobody targeting cobra toxin). The resulting designs were evaluated using an initial AF3 filtering step followed by Boltz-2 ipSAE filtering to identify top candidates.

Most designed VHHs passed Germinal’s internal filters, meaning they are predicted to be binding, however these filters rely more on iPTM and PTM scores than on ipSAE. In this dataset, the ipSAE values are distributed over a wide range, with of 9GCN_FW_s977282 (“golden-tiger-granite”), 9GCN_FW_s315086 (deep-kiwi-rose) and 9GCN_FW_s258580 (brisk-jaguar-willow) having the highest ipSAE values of 0.82, 0.77 and 0.80 respectively.

These results might suggest that either (1) ipSAE may not be an optimal standalone metric for assessing de novo VHH designs, (2) the ipSAE threshold may need to be relaxed when evaluating VHH (perhaps depending on the VHH framework used).