Our sequences were generated through an iterative, in-silico maturation pipeline inspired by natural B-cell maturation in the germinal center. This approach integrates natural-like mutation processes with structure-guided optimization, yielding antibody candidates that maintain evolutionary plausibility while targeting improved binding.

Starting from an initial antibody lead, we applied two complementary computational models to design better binders. First, we trained and applied an in-house LLM-based somatic hypermutation model to introduce biologically plausible, natural-like antibody mutations aimed at improving intrinsic properties such as stability and developability. Next, we folded the antibody candidate using AlphaFold 3 and applied LigandMPNN to impose antigen-driven selection pressure by refining the antibody CDRs to enhance interaction with the Nipah virus target. These two steps were repeated until the designs converged, meaning that no further mutations were predicted or previously favored sequences reappeared, indicating stabilization of the design landscape.