henhao
Sign in
MicrobiologybioRxiv
Preprint

ResLit: A Large-Scale Automated Literature Mining Database for Antimicrobial Resistance

Anargyros SkoulakisiD, Hanyu Xiao, Kimon Antonios Provatas, Αλέξανδρος ΓαλάραςiD, Georgios A. PavlopoulosiD, Ilias Georgakopoulos‐SoaresiD

Abstract

Antimicrobial resistance generates a vast, rapidly growing literature, yet no resource offers a comprehensive, evidence-linked repository of AMR findings at scale. We present ResLit, an automated pipeline and public database that mines the AMR literature for resistance genes, mutations, organisms, and mechanisms. From 2 million candidate PubMed records, BioMistral-7B screened abstracts to 356,000 relevant papers; multi-tier retrieval yielded 117,000 full texts, from which Qwen3-30B performed two-step extraction. ResLit contains 3,120 genes and 13,593 mutations, cross-linked to CARD, ResFinder, and NCBI Reference Gene Catalog across four evidence tiers. It further supports community-driven curation of automated outputs and reference databases. Freely available at www.reslit.info.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
0
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (21 Aug 2026) · CC BY 4.0

ResLit couples BioMistral-7B screening with two-pass Qwen3-30B extraction to build an antimicrobial-resistance database of 3,120 genes and 13,593 mutations, evidence-tiered against CARD, ResFinder, and NCBI RGC.

Curated antimicrobial resistance (AMR) references such as CARD, ResFinder, and the NCBI Reference Gene Catalog rely on expert curation that lags a fast-growing literature and covers few species, and no resource had applied language models to the full AMR literature with systematic links to them. ResLit supplies an updatable, literature-linked evidence layer that exposes uncurated determinants and flags curated entries whose cited papers may not support them.

Key results

  1. ShownResLit's gene catalogue, built from the literature, is larger than any single curated reference and recovers most of the genes in each one. It also adds many genes that no reference records. The curated references themselves overlap poorly.

    ResLit 2,546 genes, the largest single catalogue; recovered 885/1,272 CARD (70%), 840/1,138 NCBI RGC (74%), 537/638 ResFinder (84%); of 3,120 pooled genes, 480 (15%) in all four and 1,941 (62%) in one; 1,540 in no reference; 1,179 of 1,580 curated genes recovered (74.6%). · Gene coverage across the databases; Evidence tiers

  2. ShownFor mutations, the literature catalogue is far larger than the curated references, but ResLit recovers a much smaller share of curated mutations than of curated genes. The authors attribute this to inconsistent mutation notation and numbering.

    ResLit 10,939 unique mutations vs 2,283 CARD, 1,424 NCBI RGC, 1,236 ResFinder; recovered 838 (37%), 651 (46%), 490 (40%) respectively; 1,738 of 4,093 curated mutations overall (42.5%); 69 shared by all four catalogues; 9,500 found in no reference. · Mutation coverage across the databases; Evidence tiers

  3. ShownMetadata filtering and recall-tuned BioMistral-7B screening kept most of the benchmark AMR papers. The need for downloadable full text was the main cause of lost recall.

    AmrProfiler benchmark (2,628 gene, 365 mutation PMIDs): pre-filter 91.1% (2,394/2,628) and 97.8% (357/365); BioMistral-7B screening 86.9% (2,285/2,628) and 97.3% (355/365); after full-text retrieval 61.5% (1,617/2,628) and 65.5% (239/365). · Database content

  4. ShownSuccessive metadata, language, LLM and access filters shrank the PubMed candidate pool to a full-text corpus in which only a minority of papers yielded audited gene or mutation evidence.

    2,057,493 PMIDs; 1,543,369 after the four-signal filter; 1,420,587 after removing 79,380 non-English and 43,402 non-bacterial records; 356,262 classed relevant; 117,112 full texts; 19,566 papers with verified evidence. The released database holds 54,175 gene-paper and 26,094 mutation-paper records. · Database content

  5. SuggestedChecking the cited sources of curated entries that ResLit could not recover exposed possible errors in the reference databases. Most fall into three patterns: allele misattribution, ambiguous numbering schemes, and resistance claimed where the paper did not show it.

    Representative cases in Table 2, e.g. a blaOXA-143 entry whose cited paper characterized blaOXA-499 variants; cases arose during routine cross-referencing; authors present them as candidates for expert review, not confirmed errors, and give no overall error count. · Surfacing candidate discrepancies in reference…; Tables

Methods, data and reproducibility

Pipeline of 87 Entrez queries, a four-signal PubTator3/MeSH metadata filter, recall-tuned five-shot BioMistral-7B abstract screening, multi-tier full-text retrieval, and Qwen3-30B-A3B extraction plus an auditing pass and deterministic post-processing; entries are matched to CD-hit-merged CARD/ResFinder/NCBI RGC records and assigned four evidence tiers.

  • PubMed records from 87 MeSH, substance-term and free-text Entrez eSearch queries, published 1960 to 2025 · n = 2,057,493 unique PMIDs
  • Records passing the four-signal PubTator3/MeSH filter and language and subject filters, screened by five-shot BioMistral-7B · n = 1,420,587 retained; 356,262 classed AMR-relevant
  • Full texts from PMC OA, Europe PMC, PMC eFetch, CrossRef/Unpaywall locations and publisher TDM services (Elsevier, Wiley, Springer) · n = 117,112 full texts; 19,566 with verified evidence
  • AmrProfiler ground-truth PMID sets for stagewise recall benchmarking · n = 2,628 gene-associated; 365 mutation-associated PMIDs
  • CARD reference genes and mutations, downloaded 01 May 2026 · n = 1,272 genes; 2,283 mutations
  • ResFinder reference genes and mutations, downloaded 01 May 2026 · n = 638 genes; 1,236 mutations
  • NCBI Reference Gene Catalog genes and mutations, downloaded 01 May 2026 · n = 1,138 genes; 1,424 mutations
  • Released ResLit catalogue pooling literature-extracted entries (filtered to recognized bacterial organisms and antibiotics) with reference-derived entries · bacteria (multiple species) · n = 3,120 genes; 13,593 mutations
Methods & tools
NCBI Entrez eSearch and eFetch APIs, PubTator3 entity and relation annotations, BioMistral-7B (five-shot relevance classifier), Qwen3-30B-A3B (two-pass extraction and audit), NCBI PMC ID Converter API, CrossRef and Unpaywall open-access lookup, CD-hit (100% identity protein clustering), CARD, ResFinder, NCBI Reference Gene Catalog, AmrProfiler (ground-truth PMID sets), Next.js 16, SQLite, Docker on Railway (web portal)
Code
https://github.com/Georgakopoulos-Soares-lab/ResLithttps://huggingface.co/Qwen/Qwen3-30B-A3B
Data & accessions
https://www.reslit.info/https://europepmc.org/downloads
Software & environment
Models: BioMistral-7B and Qwen3-30B-A3B; CD-hit at 100% identity; reference databases downloaded 01 May 2026; portal on Next.js 16 with embedded SQLite in a single Docker container on Railway. No pipeline package versions stated.
Reproducibility gaps
Prompts, query and term lists, output schema and post-processing are in Supplementary Sections S1 to S6, not in the main text; some full texts came via publisher TDM under an institutional subscription; AmrProfiler benchmark sets lack a stated link.
Limitations (6)
  • Recovery of curated mutations reached only 43%, because mutation notation varies and positions are ambiguous without a stated reference sequence, especially for rrs and rrl rRNA genes numbered under organism-native versus E. coli schemes. · stated by the authors
  • Indels are under-represented relative to substitutions, since they are often described in prose (frameshift, truncation, named deletion) without coordinates and are extracted less reliably. · stated by the authors
  • Extraction needs full text, so paywalled papers were missed, and because paywalls cluster by publisher the gap is systematic; benchmark recall dropped to 61.5% (genes) and 65.5% (mutations) at this stage. · stated by the authors
  • Coverage is restricted to English-language PubMed-indexed papers, and reference matching relies on names, so some candidate-tier entries may be synonyms of already characterized determinants rather than new ones. · stated by the authors
  • Only recall against AmrProfiler PMIDs is reported; neither the precision of BioMistral-7B screening nor the accuracy of Qwen3 extraction and auditing is measured against manual annotation, so the false-positive rate among the 9,500 literature-only mutations is unknown. · noted in this summary
  • Reported counts do not fully reconcile: ResLit catalogue totals (2,546 genes, 10,939 mutations) differ from recovered-plus-literature-only tallies (1,179 and 1,540 genes; 1,738 and 9,500 mutations), suggesting different matching rules across analyses. · noted in this summary

Who should read:Read this if you curate AMR reference databases such as CARD or ResFinder, need literature evidence for resistance genes or mutations, or design LLM pipelines for literature-scale information extraction.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISFeldgarden '19Edalatmand '26Arango-Argo '20Hunt '17Berends '21Ghosh '26
66% semantic matchMicrobiology
M. Feldgarden, V. Brover, D. H. Haft+14 more·bioRxiv·Feb 15, 2019