KMLEE: A multimodal transformer for antimicrobial-peptide discovery

Multi-drug resistance is accelerating, sharpening the search for fresh antimicrobial scaffolds. We present KMLEE —Knitted Multi-modal Layer Efficient Encoders—an interpretable transformer that knits together primary sequence, dihedral angles, intrinsic disorder and solvent accessibility to build holistic peptide representations. Contact-map positional encoding, PIPA-Norm and attention-guided pooling enable an integrated multimodal transformer framework to outperform state-of-the-art antimicrobial peptide (AMP) predictors (AUC = 0.883) within a single network. Attention maps expose a proline-centred “bend–balance–pack” motif unseen by frequency analysis, suggesting a unifying structural blueprint for AMP activity. Screening seven metagenomes uncovers habitat-specific AMP signatures, later corroborated by motif enrichment tests. KMLEE therefore couples high predictive power with mechanistic insight, offering a tractable platform for rational peptide design and ecological surveillance. • KMLEE integrates sequence, structure and dynamics for holistic AMP prediction • PIPA-Norm and Attention-Guided Pooling independently boost performance • Outperforms existing methods, including an ensemble model while using a single network • Attention maps reveal a proline-centred ‘bend-balance-pack’ blueprint • Identifies habitat-specific AMP signatures across seven metagenomes