Co-folding affinity heads underrespond to single-substitution binding-site mutations

Several co-folding models now support binding affinity prediction and have reported performance approaching established free-energy perturbation benchmarks in selected evaluation settings. How much of this performance depends on the protein structure? We evaluated the affinity heads of three such models on 163 curated mutation pairs from HiQBind, and found that their predicted $\Delta\Delta G$ values yield little practical improvement over a predictor that always outputs zero, despite performing well on absolute affinity prediction. We further investigate the propagation of mutation-associated differences through Boltz-2 and find that these differences are detectable in the internal representations. However, their relationship to measured $\Delta\Delta G$ does not strengthen through the affinity module. Code and data are available at https://anonymous.4open.science/r/cofolding-affinity-mutation-ddg-F45E.