Designing high-affinity binders against viral glycoproteins requires methods that account for the diverse architectures and evolutionary trajectories of antiviral antibody responses. Traditional structure-prediction metrics, such as pLDDT, ipTM, and derived scores like ipSAE are not yet fully reliable for modelling these immune complexes. This is especially true for long, highly variable CDRH3 loops, a hallmark of potent neutralizing antibodies targeting pathogens like HIV-1, Ebola, and influenza. These loops often produce sparse MSAs and therefore are scored with low confidence, even when their predicted structures are accurate, simply because they lack evolutionary depth.

These limitations motivate strategies that incorporate immunological priors rather than relying solely on standard structural confidence metrics. In our workflow, structural refinement is guided not by raw scores but by a convergence-based measure (ssRMSD) that captures agreement across AlphaFold prediction trajectories, offering a more dependable proxy for antigen compatibility under MSA scarcity. Sequence diversification is performed using an inverse-folding model conditioned on the evolving structural ensemble, effectively mimicking cycles of somatic hypermutation. To further optimize the binding interfaces, we also employed structEVO and ligand MPNN.

Through iterative rounds of structure-to-sequence generation and ssRMSD-driven selection, this approach yields antibody candidates predicted to engage the Nipah virus attachment glycoprotein with high affinity. By integrating immunological bias with a convergence-informed structural filter, the workflow provides a principled framework for discovering antiviral binders in scenarios where classical structure-prediction metrics are consistently uninformative.