XAMP: debiased dual-engine AI framework enables global discovery of antimicrobial peptides from deep-sea microbiomes

Abstract

The global crisis of multidrug-resistant pathogens necessitates innovative antimicrobial peptide (AMP) discovery. However, current AMP predictors are severely compromised by three overlooked systematic data biases, including sequence length imbalance, N-terminal methionine artifacts, and lack of microbial optimization. To overcome these fundamental bottlenecks, we developed XAMP, a dual-engine predictor integrating XAMP-E (ESM-2-based for high accuracy) and XAMP-T (a one-layer Transformer for high speed), trained on rigorously debiased datasets. Consequently, XAMP achieved a median AUC of 0.972, representing an improvement of up to 21.9% in AUC while operating 5 to 40 times faster. Beyond accurate classification, we found that AMPs inherently possess a significantly higher net charge (+3.9 vs. +0.5) despite similar hydrophobicity, along with distinct cooperative patterns such as W-K/R pairs. Applying this pipeline to 238 deep-sea metagenomes (>1000m depth), we identified 2,355 promising AMP candidates from microbial dark matter, and metaproteomic analysis confirmed in-situ expression for a subset of these candidates. Notably, these deep-sea AMPs exhibited a significant 1.53-to 2.93-fold enrichment in K/R/W/Y residues compared to known AMPs. Experimental validation demonstrated that six synthesized peptides exhibit potent, broad-spectrum activity against ESKAPE pathogens, with particular efficacy against Gram-negative bacteria that dominate deep-sea ecosystems and pose major clinical challenges. This study establishes that correcting algorithmic data biases is a prerequisite for reliable mining, providing a robust framework to unlock the therapeutic potential hidden within microbial dark matter, offering new avenues to combat the antibiotic resistance crisis.