StructAMP: A Residual Multi-Modal Framework for Leakage-Controlled Antimicrobial Peptide Prediction

Reliable evaluation of antimicrobial peptide (AMP) predictors is complicated by train-test sequence-similarity leakage and mismatched comparisons with strong protein language model baselines. We present a leakage-controlled framework that quantifies auxiliary-modality value through zero-initialized residual corrections to a frozen, fold-matched ESM2-650 M predictor. StructAMP combines GVP-encoded ESMFold geometry, 457 physicochemical descriptors, and Label-Wise Attention. Five shared homology-grouped folds were constructed from 21,094 internal peptides; independent fold-specific filtering at 40% identity and 80% bidirectional coverage retained 14,456 test peptides. An analogously filtered external cohort retained 3,135 peptides. Internally, StructAMP achieved an AUC of $0.9227\pm 0.0125$, an MCC of $0.7335\pm 0.0325$, and an AMP-only Macro-F1 of $0.6220\pm 0.0114$. It did not significantly improve binary discrimination over sequence-only and produced a statistically significant but negligible Macro-F1 reduction of 0.0026. Across uncertainty, perturbation, external, and ablation analyses, pretrained sequence information supplied most predictive signal. StructAMP therefore provides a matched framework for evaluating incremental multimodal value rather than evidence that additional modalities necessarily improve AMP prediction. Conclusions are limited to ESM2-650 M, the tested static ESMFold/GVP representations and descriptors, and frozen-sequence residual training; they do not imply that three-dimensional peptide structure is generally uninformative.