MAPLE deep learning model identifies selective antimicrobial peptides, revealing motif-dependent potency-hemolysis coupling

The global escalation of Antimicrobial Resistance (AMR) intensifies the urgent need for novel antibiotics. Antimicrobial peptides (AMPs) offer a promising alternative due to their unique mechanisms of action. However, a major hurdle in developing new AMPs is the perceived trade-off between antibacterial potency and mammalian toxicity, particularly hemolysis. This assumption complicates prioritizing effective and safe candidates, as existing predictive models often fail to pinpoint the specific sequence features or motifs responsible for AMP selectivity. Understanding where selectivity resides in sequence space is critical for engineering safer, more effective AMPs.

Researchers developed MAPLE (Multifunctional AMP Learning Engine), an interpretable dual-stream deep learning framework. This model takes peptide sequences as input and predicts across 14 activity categories, encompassing both antibacterial potency and mammalian toxicity. MAPLE integrates protein language model embeddings with explicit physicochemical descriptors to achieve robust, task-specific predictions, even under severe label imbalance. They performed systematic k-mer enrichment analysis to map motif-level selectivity and conducted structural modeling to visualize conformations, particularly for antibacterial-selective motifs. The model was benchmarked on a dataset and an independent, sequence-non-overlapping validation set.

MAPLE achieved consistently well-balanced predictive performance across 14 activity categories, including low-prevalence but clinically relevant endpoints, on both benchmark and independent validation datasets. The model's interpretability allowed for detailed analysis of sequence features.

Systematic k-mer enrichment analysis revealed that the perceived coupling between antibacterial potency and hemolysis is motif-regime-dependent rather than universal. Motifs most strongly enriched for antibacterial activity exhibited significantly reduced hemolytic overlap compared to broadly active motifs. These selective motifs occupied a distinct physicochemical regime characterized by moderate cationicity, lower hydrophobicity, and higher amphipathicity. Furthermore, proof-of-concept prioritization workflows leveraging these antibacterial-selective motifs, combined with structural modeling, yielded conformations consistent with amphipathic α-helices, suggesting specific structural features underpin the observed selectivity.