Genotypic Trigger: Backdoor Attack on Antimicrobial Peptide Models

A recent study published on arXiv (2608.06779) reveals a backdoor attack known as 'Genotypic Trigger', aimed at generative models used for designing antimicrobial peptides (AMPs). This attack alters the generative distribution, leading to the creation of peptides that exhibit a significantly higher predicted immunogenicity risk for individuals carrying a specific HLA allele, which is a gene variant critical for immune presentation. In widely used peptide generation models, the attack resulted in an average increase of 743% in the predicted immunogenicity risk score for these target-allele carriers. This research exposes a flaw in current validation processes for peptide generation models, which fail to consider that some drugs may primarily endanger individuals with particular genetic backgrounds. The authors emphasize the urgent need for pharmacogenomic considerations in validating generative models in drug development.