These binders were generated starting from a batch of candidate sequences obtained with BoltzGen and ranked sequences using ipSAE (boltz2). ipSAE proved to be extremely variable from seed to seed therefore we selected one best candidate that maintained a high mean ipSAE_min and a similar interface across 10 Boltz2 seeds.

Interestingly the best candidate had different binding sites depending on the seed (on both sides of the Nipah protein), but re-used the same interface in all cases. This candidate was then mutated with in-house interface statistical potentials based on the MuSiC suite (Dehouck, Y., Kwasigroch, J. M., Gilis, D., & Rooman, M. (2011) BMC Bioinformatics). This is a novel physic-based approach currently under development in our lab which shows promising performance in predicting the effect of mutations on variant effect datasets for binding such SKEMPI2 and DMS datasets.

Our predictions highlighted many positions where mutations had a highly positive impact on binding affinity. This behavior is not usually observed in protein and might suggest that de-novo sequences generated with Boltz are not optimal and rankings based on ipSAE might miss some sequences with high binding affinity.

This set of protein is only double and triple mutants. For this we selected positions where mutations would not be in close proximity and combined single mutants that had positive effects on binding affinity. This represents a trade-off as multiple mutations in the interface are susceptible to have a negative effect on ipSAE.