AMP-SEMiner

Antimicrobial resistance is outpacing the antibiotic pipeline, and antimicrobial peptides (AMPs) are one of the few classes of alternative with broad-spectrum activity and diverse mechanisms of action. The microbiome is an obvious place to look for them, but most computational searches only find AMPs that exist as standalone short genes. A large fraction hide elsewhere: in small open reading frames that gene callers skip, and as encrypted peptides embedded inside larger proteins that were never annotated as antimicrobial at all.

AMP-SEMiner — the Antimicrobial Peptide Structural Evolution Miner — was built by Zhejiang Lab's Research Center for Life Sciences Computing with collaborators at Tongji University and the University of Nebraska–Lincoln. Its core move is to reframe AMP discovery as residue-level labelling rather than whole-sequence classification: a protein language model tags each amino acid as AMP or non-AMP, so an antimicrobial fragment is recoverable even when the protein containing it is not itself an AMP. Around that classifier the framework wraps structural clustering and evolutionary analysis, which is what lets it ask not just what the peptides are but how they have been shaped over time.

Applied across metagenome-assembled genomes from a wide range of habitats, the pipeline surfaced more than 1.6 million AMP candidates. Unusually for a mining study of this scale, a subset went to the bench.

#Key Features

• Residue-level classification: The model emits a per-token label, which is what makes encrypted peptides inside larger proteins discoverable rather than invisible to a sequence-level classifier.
• smORF and encrypted-peptide coverage: Both classes of hidden AMP are handled by the same pass, substantially widening the searchable space compared to pipelines restricted to annotated short genes.
• Parameter-efficient adaptation: Low-rank adapters are inserted into every transformer layer of the backbone, so the released artifact is a 19 MB adapter that reproduces the tuned model on top of public ESM-2 weights.
• Structure-aware grouping: Candidates are folded and each predicted structure reduced to descriptors such as radius of gyration and solvent-accessible surface area, then grouped by community detection, so families are defined by shape rather than sequence identity.
• Wet-lab validated output: Candidates were synthesized and assayed rather than left as in-silico predictions, and sequence variants derived from the active hits were tested alongside them.

#Technical Details

Three checkpoints are released. The headline configuration, Tok_CLS_LoRA, adapts facebook/esm2_t33_650M_UR50D for token classification with rank-8 rank-stabilized LoRA applied to the feed-forward output projection of all 33 encoder layers, with a dropout of 0.05; a fully fine-tuned Tok_CLS variant and a two-step design that chains a sequence classifier into a token classifier are also provided. Inference truncates at 300 residues. Mining ran over metagenome-assembled genomes drawn from MGnify's human gut, human oral, cow rumen, pig gut, fish gut, zebrafish faecal and marine catalogues, the Chinese Gut Microbial Reference, Hadza and ancient human gut cohorts, a human oral cohort, and glacier microbiomes — the habitat spread that gives the resulting candidate set its ecological range. Of 20 candidates tested experimentally, 9 showed antimicrobial activity and 5 exceeded the effectiveness of reference antibiotics. The earlier preprint reported a larger validation panel; the peer-reviewed Cell Reports version is the figure to cite. Analysis of the human gut subset shows AMPs under both conservative and adaptive evolutionary regimes, consistent with peptides that must stay functional in a competitive, shifting environment.

#Applications

For antimicrobial discovery groups the output is a ranked, structurally organized candidate list from environments that have not been systematically mined, which is a shortcut past the usual bottleneck of deciding what to synthesize. For microbiome researchers the framework can be pointed at their own assemblies: the checkpoints are fixed, so scoring new genomes needs no retraining, and a Colab notebook runs the prediction step without local setup. The evolutionary analysis component is separately useful to anyone studying how microbial communities maintain chemical warfare capacity.

#Impact

Published in Cell Reports after a bioRxiv preprint, AMP-SEMiner is one of the few large-scale peptide-mining efforts to close the loop with synthesis and antimicrobial assays, which is what distinguishes it from candidate lists that remain unvalidated. Code is on GitHub under the MIT license and the weights are archived on Zenodo under CC-BY-4.0, with example datasets for both prediction and retraining. The honest caveats are that the validated fraction is small relative to 1.6 million candidates, that activity was assessed in vitro with no toxicity, stability or in vivo data reported, and that the repository provides no model card beyond the README.