Abstract
<title>Abstract</title> <p>Predicting which SARS-CoV-2 spike mutations enable antibody escape requires training labels whose biological relevance can be independently verified. We built a two-stage pipeline linking 905 Indian SARS-CoV-2 genome sequences to solved antibody-spike protein structures: an exhaustive similarity comparison against 311 candidate structures (281,455 pairwise comparisons), followed by three-dimensional contact-sphere analysis, producing 898 valid sequence-structure pairs. Validation revealed a mutation-calling artifact in 351 pairs (implausible substitution counts contradicting independently computed similarity scores), traced to an unhandled indel; we corrected this by indel-aware realignment rather than exclusion, recovering reliable labels for 844 of 898 pairs. On the corrected dataset, verified escape-relevant positions occurred in antibody-contact spheres at 3.29 times the rate expected by chance. Cross-referencing corrected hits against an independent deep mutational scanning dataset, hit positions showed significantly higher measured escape potential than background-drift mutations at the unique-position level (n = 54 versus 74 positions; rank-based p = 0.0054; permutation-on-median p = 0.0144; bootstrap 95% confidence interval on the median difference [0.006, 0.508]), corroborating the labels using data external to model construction. A graph attention network trained on the pre-correction dataset reached F1 0.98 and AUROC 0.998 on held-out strains of already-seen structures, but showed only a modest, size-dependent signal (Matthews correlation coefficient 0.21, AUROC 0.57 on adequately sampled folds) under a stricter leave-one-structure-out test simulating novel antibody targets; structural similarity between complexes did not predict this gap, and transfer instead depended on whether verified contact positions overlapped between structures. These model results predate the labeling correction and are reported with this caveat. These findings support structural contact verification as a meaningful, though not yet fully sufficient, proxy for functional antibody escape, and identify both structural representativeness of the reference set and label-pipeline correctness as primary determinants of reliable escape-mutation prediction.</p>