AResKGLM

A resistome survey returns candidate resistance genes; deciding which gene–drug pairings are mechanistically plausible is the harder half of the job. Knowledge-graph embedding methods compress the relevant evidence into a scalar score with nothing to inspect, while a general-purpose language model writes a fluent explanation that need not correspond to any curated fact. AResKGLM (Antimicrobial Resistance Knowledge Graph Language Model) makes the graph evidence part of the prompt instead: a constrained breadth-first search extracts 2–6-hop paths connecting the queried gene and drug, and a topology-preserving serialization renders them as text the model must read before it answers.

The serialization is where the design work sits. Each prompt carries three blocks — Context Information (entity names with descriptions from RefSeq, UniProtKB, NCBI Gene and Wikipedia), Reasoning Path (each retrieved path emitted verbatim, one per line, with relation labels in positional order and intermediate node identities preserved), and Question. Writing edges in order rather than as a bag of relations, and keeping intermediate entities addressable across paths, is what distinguishes the input from a flattened triple list. Llama-3-8B-Instruct and DeepSeek-R1-7B are adapted with QLoRA to emit a short reasoning trace ending in a binary decision.

AResKGLM was developed by the Tencent AI for Life Science Lab in Shenzhen and published in Briefings in Bioinformatics in September 2026. It is unrelated to BacteReason, a separately developed AMR reasoning model from a different group that predicts isolate-level antimicrobial susceptibility rather than gene–drug links.

#Key Features

• Path serialization that keeps graph topology: Relation direction, edge order and repeated intermediate nodes survive the conversion to text, so the model can attend to the same entity across several supporting paths.
• A rationale with every prediction: Each decision arrives with a short trace citing the entities and relations used, checkable against the supplied path rather than taken on trust.
• One checkpoint, unseen entities: The adapter is trained once per fold under a group-based entity holdout, then applied without retraining to held-out entities, gene-family holdouts, matched-negative regimes and post-2022 literature pairs.
• Controlled probes over the evidence: A six-condition battery over 32,910 instances removes context, symbolizes names, rewrites templates and corrupts paths, quantifying how far predictions track the supplied evidence rather than pretrained priors.

#Technical Details

The primary knowledge graph is KIDs, anchored to the Escherichia coli K-12 MG1655 reference genome (RefSeq GCF_000005845.2) and supplemented from UniProt, RegulonDB and DrugBank: 8,047 entities (81.3% of them carrying textual descriptions), 12 relation types and 25,668 positive triples. Splits use GroupKFold over prefix-based entity clusters, holding out about 20% of triple-active entities per fold; evaluation negatives corrupt the tail to an entity unseen in training. Fine-tuning is QLoRA (4-bit NF4) over 2,048-token sequences in bf16, at LoRA rank 16 / α 32 for Llama-3 and 32 / α 64 for DeepSeek. On KIDs the Llama-3-8B configuration reaches F1 = 0.8482 against KG-BERT at 0.7213, NBFNet at 0.5260 and ULTRA at 0.2541, and degrades gently with depth, from 0.9197 at two hops to 0.8148 at six. Replacing entity and relation names with anonymous symbols while preserving topology drops F1 from 0.8470 to 0.7312; the unadapted backbone reaches 0.3596. Cross-species graphs for three WHO priority pathogens give F1 = 0.8811 (M. tuberculosis), 0.8141 (S. aureus) and 0.8097 (P. aeruginosa). On 81 gene–drug associations drawn from 2023–2026 PubMed abstracts and absent from training, ranking by the mean log-probability margin between the yes and no completions gives Precision@20 = 100% and AUC-PR = 0.855.

#Applications

The intended use is prioritization rather than diagnosis: resistome profiling and metagenomic surveillance surface many candidate genes, and AResKGLM ranks which gene–drug associations are worth experimental follow-up while showing the multi-hop path behind each call. It composes with curated resources such as CARD, and the traces give curators something concrete to argue with over whether an association belongs in a knowledge base. The framework operates over curated isolate- and pathogen-level graphs; metagenomic assemblies or longitudinal cohorts would require further entity normalization.

#Impact

Path-level graph evidence, not just retrieved text, changes what a fine-tuned language model does with a biological query: a BM25 retrieval control matched aggregate F1 but was markedly less precise. The authors are candid about where the interpretability claim stops. With generation capped at 64 tokens, 44.8% of cached traces are a bare "The answer is yes/no." naming no entity, and among the rest the entity-hallucination rate is 0.34; sanitizing the descriptions shows positive predictions lean partly on surface AMR keywords alongside the relational path. The K. pneumoniae cross-species F1 of 0.7167 is a positivity artifact — specificity 0.0028, MCC 0.039 — not a fourth working pathogen. Code, probes and baselines are released under Apache-2.0; the repository ships training and inference scripts that expect a locally supplied backbone and a locally trained adapter rather than distributing fine-tuned weights.