PlasRAG

Metagenomic sequencing has pushed the number of catalogued plasmids past a million, and almost none carry usable annotation. That matters because a plasmid's properties determine whether it is a clinical concern: whether it carries antimicrobial resistance or virulence factor genes, how mobile it is, which hosts it can occupy, which ecosystems it circulates in. Existing tools each address one of these questions, were built on the comparatively small and biased set of NCBI plasmids, and — being alignment-based or narrowly supervised — degrade precisely on the novel sequences metagenomics keeps producing.

PlasRAG, from the City University of Hong Kong with collaborators at MIT and the Chinese University of Hong Kong, treats plasmid annotation as a cross-modal retrieval problem instead. Property annotations from three reference databases are rewritten as complete English sentences, and a contrastive model learns to place a plasmid's sequence embedding near the embeddings of the sentences that describe it. Because both directions of that alignment are usable, the same model answers two questions: given a plasmid, which properties describe it, and given a description, which plasmids match. A retrieval-augmented generation layer then feeds the predicted properties and the closest annotated literature into an instruction-tuned language model so users can ask about a plasmid in plain English.

Where PlasmidLM generates plasmid sequences, PlasRAG interprets them, and the trained retrieval model at its core is what distinguishes it from pipelines that merely wrap a general language model around a fixed index.

#Key Features

• Bidirectional sequence-text alignment: A single contrastively trained model supports both sequence-to-text characterization and text-to-sequence retrieval, rather than requiring separate classifiers per property.
• Ten property facets: Antimicrobial resistance, host range, ecological host, ecosystem, fundamental properties, heavy metal resistance, incompatibility group, mobility, risk index, and virulence factors each get a dedicated sub-model with a fixed sentence vocabulary.
• Boolean retrieval queries: Property sentences can be composed with AND, OR, and NOT to filter either the bundled reference database or a user-supplied plasmid set.
• Grounded question answering: Predicted properties are combined with abstracts from the nearest annotated reference plasmid and passed to Llama-3 at low temperature, so answers are conditioned on retrieved evidence rather than generated from the model's priors.
• Cheap to train and to run: Protein and text embeddings are precomputed and frozen, leaving only a compact resampler to optimize.

#Technical Details

A plasmid is encoded by translating its genes with Prodigal, embedding each protein with a frozen ESM-2 650M model, and passing the resulting variable-length set of 1,280-dimensional vectors through a Perceiver Resampler that emits one 768-dimensional embedding per property sentence. Property text is encoded by the frozen MedCPT query encoder (110M parameters). Only the resampler is trained: the ten facet sub-models together total 89.3 million parameters, trained for 20 epochs at batch size 2,048 on a single A100.

Training data was drawn from PIPdb, IMG/PR, and PLSDB, yielding 1,521,635 plasmids after excluding megaplasmids above 250 kbp; 100,000 were held out for validation and testing, with splits enforced at the plasmid segment cluster level so sequences sharing above 76% average nucleotide identity never straddle a split. Prodigal produced 37,050,904 encoded proteins, reduced to 4,925,205 representatives by MMseqs2 clustering at the UniRef90 threshold.

On the held-out test set, the full configuration reaches weighted text-centric F1 scores of 97.4 for mobility, 96.5 for heavy metal resistance, 94.0 for antimicrobial resistance, 90.7 for virulence factors, 83.1 for host range, and 77.7 for ecosystem, with P@1000 following the same pattern. Substituting the Perceiver Resampler with a Transformer, Bi-LSTM, RNN, or CNN encoder lowers every facet, as does replacing MedCPT with BioMedBERT, Llama-3, or Word2Vec. A BLASTN baseline trails badly on the facets that depend on gene content — 55.6 F1 on antimicrobial resistance against PlasRAG's 94.0 — which is the concrete form of the novel-plasmid problem the paper set out to solve.

#Applications

PlasRAG is aimed at microbial genomics and antimicrobial resistance surveillance: annotating plasmid contigs recovered from metagenomes, flagging high-risk plasmids in clinical isolates, and assembling cohorts for comparative analysis by querying for combinations such as broad host range without adherence-related virulence factors. The paper's own application to 68,350 plasmids recovered from human gut metagenomes by PlasX found that broad host range and encoded resistance genes both associate with wider dissemination.

#Impact

The design choice worth noting beyond plasmid biology is the use of natural-language property sentences as the label space: because facets with different problem types share one contrastive objective, adding a property means adding sentences rather than building a new classifier, and inconsistent vocabularies across source databases are reconciled in text rather than by schema merging. The evaluation is retrieval-metric based and inherits whatever biases the crowd-sourced database annotations carry, and the Llama-3 answer-generation layer is assessed qualitatively rather than benchmarked. Code is released under GPL-3.0 with pretrained weights and the reference database on Zenodo; the work was published in Genome Biology after preprint.