Pathogen context reshapes antimicrobial peptide generation

Antimicrobial peptides are a promising source of new anti-infective agents, but their discovery is limited by a practical bottleneck: selecting a small set of candidates to synthesize and test for a specific pathogen. Existing peptide generators can expand antimicrobial-like sequence space, but most leave pathogen specificity to predictors or filters after generation. Here, we use AMPHORA, a pathogen-conditioned sequence-structure generator, to build antimicrobial peptide libraries shaped by target context. AMPHORA couples a short-peptide representation model to latent-flow generation conditioned on target class, genome-derived features and strain-description text. In matched, partial and shuffled controls, pathogen context redirected generated peptide pools beyond broad activity labels, with the strongest effects when genome and text conditions were combined. Counterfactual generation from the same initial noise showed that strain descriptions primarily changed amino-acid choices, whereas genome-derived features contributed more strongly to predicted structural properties. Across species, AMPHORA produced target-dependent enrichment, and matched bacterial context shifted APEX-predicted potency-score distributions compared with class-only generation. The generated peptides remained diverse, novel and peptide-like by sequence and predicted-structure analyses. Collectively, our results establish pathogen context as an active design signal for antimicrobial peptide generation and provide a target-aware route to pre-synthesis library design.