Escape Cost as a Target-Selection Criterion for De Novo Epitope-Scaffold Immunogen Design

Epitope-scaffold immunogens re-present a broadly neutralizing antibody (bNAb) epitope on a stable, minimal protein scaffold, but which epitope to scaffold is almost always chosen from structural or biochemical intuition rather than a quantitative evolutionary criterion. We propose \emph{escape cost}, the amount of fitness a virus must give up to mutate away from an epitope, as an explicit, data-derived target-selection signal, and ask whether it also predicts how \emph{designable} that epitope is as a de novo scaffold. We estimate escape cost for three HIV-1 Env bNAb epitopes (PGT121/N332, PGDM1400/V1V2-apex, VRC07-523LS/CD4bs) by combining a resistance classifier trained on real neutralization data (CATNAP) with a protein language model's zero-shot fitness prior, and we validate the resulting cheap-vs-costly route predictions against real escape routes observed in an independent HIV rebound cohort. We then run a controlled four-arm RFdiffusion/ProteinMPNN scaffolding experiment across epitopes spanning this escape-cost range. Among three structurally rigid epitopes, scaffold designability rises monotonically. This points to a real tension in immunogen design: the epitopes most valuable as vaccine targets, precisely because escape is costly there, are also the structurally hardest to present on a de novo scaffold.