Abstract
Abstract
Zero-shot prediction of mutation effects on viral proteins is hard because viral families evolve fast and are sparsely sampled, and protein language models are weaker on them. We evaluate AminoGrapes, an independent-site model from Afrium that scales a sequence-weighted log-odds conservation score from a homolog alignment by residue burial computed from a predicted structure. The formulation follows RSALOR [7]. The AminoGrapes score has no fitted parameters. On the 31 viral assays of ProteinGym v1.3, AminoGrapes reaches a mean Spearman correlation of 0.430, compared with 0.269 for Meta's ESM2-650M (+0.162, 95% CI 0.078 to 0.245), and is better on 25 of 31 assays. A blend with ESM2-650M, with the blend weight chosen only on non-viral assays, reaches 0.376 (+0.108, 95% CI 0.078 to 0.140) and is better on 27 of 31. AminoGrapes does not match the original RSALOR scores (0.480) or the strongest published methods on these assays, it adds nothing to an ensemble of the best public models, and its design was settled while viral results were visible (Section 4.2). This is an evaluation on a public benchmark, not an official leaderboard entry.