Parameter-Efficient Adaptation of Protein Language Models for Dual-Target Antimicrobial Peptide Discovery Under Strict Homology Constraints

An end-to-end LoRA-adapted ESM-2 (150M) multi-task model under 60% maximal sequence identity achieves E. coli potency rank correlation ρ = 0.6132 and hemolytic activity ρ = 0.5374 on QMAP splits.

• Why it matters: AMR demands new antimicrobial peptide candidates, but computational designs often fail due to hemolytic off-target cytotoxicity and poor generalization under strict homology-controlled evaluations.
• What they did: The authors fine-tune only ~0.70% of ESM-2 weights with LoRA in self-attention and use Kendall homoscedastic uncertainty weighting for adaptive multi-task loss, evaluated on QMAP splits without sequence leakage.
• The result: The adapted model improves potency and hemolysis correlations versus frozen embeddings and other regressors, and screens 749 de-novo peptides to isolate selective candidates with Therapeutic Indices up to 144.03 while deprioritizing broadly lytic motifs (TI < 2.0).

AI summary of the abstract · may contain errors · not medical advice · read the paper.