ESMARG

A metagenome from an environmental sample yields millions of predicted protein-coding sequences, of which only a tiny fraction confer antibiotic resistance. Alignment tools such as BLAST and DIAMOND recover an antibiotic resistance gene (ARG) only when a close homolog is already catalogued in a reference database like CARD or SARG, and earlier deep-learning detectors were trained against negative sets far smaller than the real environmental background, so they flag ordinary proteins as resistance genes at high rates. ESMARG attacks both problems at once: it reads sequences through a pretrained protein language model rather than by alignment, and it floods training with authentic non-ARG sequences from real metagenomes.

Posted as a bioRxiv preprint in September 2025 by researchers at Beijing University of Posts and Telecommunications, Capital Medical University, and Tsinghua University, ESMARG is a four-module pipeline built on top of ESM-1v. ESMARG-1 decides whether a sequence is an ARG, ESMARG-2 assigns it to a drug class, ESMARG-3 predicts its resistance mechanism, and ESMARG-4 attaches a gene-level annotation by DIAMOND alignment against the DeepARG database.

The framework occupies a distinct slot among resistance models. resLens fine-tunes genomic language models on nucleotide sequence and AMRscope scores individual point mutations; ESMARG classifies whole predicted proteins, which is what a metagenomic assembly actually delivers. The authors then run the fixed pipeline, unchanged, over 26 cyanobacterial-aggregate metagenomes from Lake Taihu.

#Key Features

• Frozen-backbone classifier heads: ESM-1v embeddings are extracted once per sequence and consumed by small multilayer perceptrons, so the released checkpoints are a few megabytes each and inference is one pass through a shallow network on cached features.
• Metagenome-realistic negatives: Non-ARG training sequences come from UniProt and the Global Microbial Gene Catalog, spanning food, livestock, compost, soil, and water samples, with known ARGs stripped out by DIAMOND at an e-value threshold of 1e-5.
• Three prediction levels in one pipeline: binary identification, 41-way drug-class assignment, and 7-way resistance-mechanism assignment, followed by alignment-based gene naming.
• External validation: On 420 manually curated ARGs from the Oxford database, the identification module recovered 302 sequences, against 296 for ARGNet and 268 for BLAST.
• Throughput: Roughly a 10-fold speedup over BLAST and DIAMOND on the same sequence set, since the expensive step is a single embedding pass rather than a database search.

#Technical Details

ESM-1v produces a 1,280-dimensional mean-pooled representation per protein, and each ESMARG module is a compact head over that vector: 1280→128→1 with dropout and a sigmoid for identification, 1280→128→64→41 for drug class, and 1280→512→64→7 for mechanism. The positive training set is 15,624 ARG sequences pooled from CARD, SARG, MEGARes, and ResFinder. Negative set size was swept from 70,000 to 200,000 UniProt sequences to measure its effect on false positives; the configuration used for the environmental survey combined 30,000 UniProt with 40,000 Global Microbial Gene Catalog sequences.

At a decision threshold of 0.6, ESMARG-1 reached 0.998 precision, 0.939 recall, and 0.968 F1 on the held-out test set, against 0.778 / 0.916 / 0.842 for ARGNet evaluated on the same data. ESMARG-2 reached 0.986 accuracy and a weighted F1 of 0.981, ahead of ARG-SHINE, BLAST, and DIAMOND, with aminoglycoside and nucleoside its weakest classes. ESMARG-3 exceeded 0.95 F1 for reduced permeability, target alteration, and target protection or replacement, while antibiotic efflux — the most heterogeneous mechanism — was hardest to call.

#Applications

ESMARG is aimed at environmental and clinical resistome surveillance, where the input is an open-ended set of predicted open reading frames rather than a curated gene list. In the accompanying survey of Lake Taihu across a full annual cyanobacterial bloom cycle, it detected 110 distinct ARGs spanning 24 drug classes; efflux genes accounted for 38.2% of total abundance, bacA, macA, and vanY were the three most abundant genes, and mean ARG load ranged from 3.21 copies per cell in summer to 0.43 in winter. The same workflow transfers to wastewater, soil, livestock, and clinical metagenomes, and the alignment step keeps results interpretable as named genes for groups acting on specific resistance determinants.

#Impact

ESMARG's substantive argument is that in metagenomic ARG detection the binding constraint is class imbalance rather than sequence representation: holding the protein language model fixed and expanding the realistic negative pool is what drives precision from ARGNet's 0.778 to 0.998. That result is unusually easy to check, because the code, all three checkpoints, and the training FASTA files are in the project repository. Readers should weight the numbers accordingly. The work is a preprint that has not been peer reviewed, most benchmarking is on the authors' own test split with a single 420-sequence external set, and the frozen backbone caps performance at whatever ESM-1v already encodes about resistance function. The repository README declares MIT licensing but ships no LICENSE file, and the source databases carry their own commercial restrictions.