MAPLE: interpretable deep learning identifies selective antimicrobial peptides using joint...

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Antimicrobial peptides (AMPs) are crucial in innate immunity and are explored as alternatives to conventional antibiotics.
• A significant hurdle in AMP development is the perceived correlation between high antibacterial potency and mammalian toxicity (hemolysis).
• Existing predictive models often fail to identify sequence features responsible for selective toxicity.

Purpose of the Study:

• To develop an interpretable computational framework, MAPLE, for identifying and functionally profiling AMPs.
• To systematically analyze the relationship between AMP sequence, antibacterial activity, and hemolytic toxicity.
• To provide principles for engineering safer and more potent AMPs.

Main Methods:

• MAPLE utilizes a dual-stream framework combining protein language model embeddings and physicochemical descriptors for sequence-based prediction.
• The model predicts 14 activity categories, addressing severe label imbalance.
• Systematic k-mer enrichment analysis was performed to map motif-level selectivity.

Main Results:

• MAPLE achieved balanced performance across benchmark and independent validation datasets, including low-prevalence endpoints.
• The study demonstrated that the coupling between antibacterial potency and hemolysis is motif-regime-dependent.
• Antibacterial-selective motifs were identified, characterized by moderate cationicity, lower hydrophobicity, and higher amphipathicity, with reduced hemolytic overlap.

Conclusions:

• The perceived universal coupling between AMP potency and hemolysis is a misconception; selectivity is motif-dependent.
• MAPLE provides a framework for hypothesis generation and prioritization in AMP discovery.
• Findings support the engineering of potent and safer AMPs by focusing on specific sequence motifs and physicochemical properties.