AmpGPT2 (IMI Münster)

We submit AmpGPT2, an autoregressive language model trained on antimicrobial peptide sequences, as a broad AMP generator for the AMP Challenge 2027. Because AmpGPT2 is not specific to antibacterial activity against the challenge pathogen panel, generated peptides are computationally enriched using APEX.

Valid, unique peptides of 8–50 amino acids are generated until 50,000 sequences with a mean APEX-predicted MIC <=256 µM are obtained. Exact matches to the challenge antibacterial reference set and the additional COMPASS reference set are excluded.

Top candidates are ranked by ascending mean APEX-predicted MIC. Candidates are then filtered using the challenge <=80% Levenshtein-similarity requirement against the antibacterial reference set, local-alignment novelty against the antibacterial and COMPASS reference sets, and sequence diversity within the selected set.

Generation and filtering are fully automated and reproducible from a fixed seed.