Highlights

Integrates residue-level ESM-2 embeddings with interpretable biochemical descriptors.

Cross-attention + Transformer aligns property cues to salient sequence positions.

Returns calibrated probabilities with ACC, Sn, Sp, MCC, and AUC.

Benchmark-validated on ACP135 and ACP99 with consistent gains over strong baselines.

Supports saliency visualizations for mechanism-aware interpretation.

Efficient, GPU-ready inference for large-scale peptide screening.