Microbial bioremediation has a reporting problem, not just a biology problem. A reader who wants to know whether a bacterium can detoxify a pollutant must dig through unstructured prose to assemble, by hand, the full experimental picture: which strain, which contaminant, at what concentration, what temperature and pH, what inoculum size, for how long, with what degradation percentage, into which metabolites, and with what safety caveats. The raw ingredients exist in papers like Isolation, Identification and Characteristics of an Endophytic Quinclorac Degrading Bacterium Bacillus megaterium Q3, which reports 93% quinclorac degradation at 20 mg/L within seven days under an exact condition set (30°C, pH 8, 6% inoculum) — but these facts are scattered across Methods, Results, and Discussion, and the field's honest limitations (no field validation, unsequenced enzymes, missing metabolite standards) are relegated to closing sentences. Meanwhile, machine learning publishing solved an analogous problem with standardized documentation cards and pipelines that generate them automatically. We propose to import that machinery — and to fix where it breaks — for bioremediation.
The core deliverable is the Strain Card: a structured, machine-readable record for each bioremediation isolate, auto-generated from its source paper by a retrieval-grounded extraction pipeline, and shipped with a statistical guarantee on how often its fields are unfaithful to the source. The hard technical problem is not card formatting; it is that bioremediation experiments are multi-variable and cross-sectional. Existing hierarchical extractors assume a value sits near its parameter and its artifact, and that assumption fails here.
Core Claim
Current publication-logistics tools cannot represent a biodegradation experiment, because they extract flat or shallow-hierarchical parameters rather than condition-scoped observations. We define the unit of extraction as a tuple
$$o = (\text{strain},\ \text{substrate},\ \text{matrix},\ C,\ y,\ M)$$
where $C$ is the full set of experimental conditions (temperature, pH, initial concentration, inoculum fraction, duration), $y$ is the measured outcome (e.g., percent degradation), and $M$ is the identified metabolite set. The claim is threefold. First, a schema built around $o$ makes bioremediation literature machine-aggregable for the first time, enabling queries like "all strains degrading ochratoxin A above 90% within 24 hours at 37°C" that today require manual review. Second, extracting $o$ requires cross-section linking — anchoring a strain introduced in Materials and Methods to outcomes reported pages later — which breaks the spatial-distance heuristics that existing extractors rely on. Third, generated cards need a conformal guarantee, because unguarded extraction of wet-lab parameters hallucinates at unacceptably high rates.
Research Program
Three components, each adapting a cited technique to a regime it was never designed for.
Schema and benchmark. We design the Strain Card schema covering the variables the bioremediation literature itself identifies as inconsistently reported: initial pollutant concentration, degradation kinetics, abiotic optima, enzymatic localization (intracellular versus extracellular, as contrasted by Isolation and Characterization of Brevibacillus parabrevis S09T2, a Novel Ochratoxin A-Degrading Strain with Application Potential versus Wang et al.), intermediate metabolites, repository accession numbers, biosafety screens (hemolysis, enterotoxin genes), and explicit limitation fields. We then annotate BioCardBench: a corpus of bioremediation papers with gold $o$-tuples, seeded by the assay-rich studies in this literature (Isolation, Identification and Characteristics of an Endophytic Quinclorac Degrading Bacterium Bacillus megaterium Q3; Isolation and Characterization of Brevibacillus parabrevis S09T2, a Novel Ochratoxin A-Degrading Strain with Application Potential; Yang; Jiang et al.) and extended to several hundred papers by the same annotation protocol. This mirrors how HyperPIE: Hyperparameter Information Extraction from Scientific Publications built its 444-paragraph gold set for ML hyperparameters; no such resource exists for environmental microbiology, and no reported result found on Lacuna for extraction F1 in this domain.
Cross-section extraction pipeline. CARDGEN (Automatic Generation of Model and Data Cards: A Step Towards Responsible AI) showed that a two-step retrieval-augmented pipeline — splitting card questions into sub-queries, generating pseudo-answers to bridge jargon, then retrieving with those pseudo-answers as embeddings, with role-based personas — beats human authors decisively on completeness and objectivity. We reuse that skeleton but replace its flat question-answering with a two-stage $o$-tuple extractor: a first stage anchors entities (strain, substrate) wherever they appear, and a second stage propagates a condition scope from each anchor, binding each reported value ("93%") to the nearest governing condition set even when that set was defined in a different section. This is the specific fix for the failure mode HyperPIE documents: its spatial signals (artifacts precede parameters 78% of the time, parameters precede values 93% of the time) are exactly what bioremediation prose violates. We adopt its YAML prompting grammar, which improved entity F1 by 5.5% over JSON.
Verification and guard. Generated cards are checked with a SciFact-style verdict pipeline. Fact or Fiction: Verifying Scientific Claims exposed the directionality trap — models confusing "impaired" with "increased" — and bioremediation has an acute version of it: a strain may degrade a parent herbicide while accumulating a toxic intermediate, exactly as Isolation, Identification and Characteristics of an Endophytic Quinclorac Degrading Bacterium Bacillus megaterium Q3 shows quinclorac passing through three chlorinated intermediates. We build a directionality test set of claim pairs that swap degrade/accumulate and swap condition values (pH 8 versus pH 6). Finally, because Evaluating and Guarding Citation Faithfulness in Agentic Scientific Synthesis found that automated verifiers disagree wildly (negative-specific agreement of only 0.27–0.30) and unsupported-claim estimates range from 3% to 18% on identical outputs, we attach its split-conformal guard: calibrating our imperfect verifier on a gold-anchored subset of BioCardBench so that the rate of ungrounded card fields is bounded below a chosen tolerance with a finite-sample, distribution-free guarantee.
Baselines
The naive reference point is the status quo: human-authored narrative reporting with no card at all. Its cost is documented, not hypothetical — on CardBench's 1–10 audits, human-written cards averaged 1.92 for completeness and 2.03 for objectivity (Automatic Generation of Model and Data Cards: A Step Towards Responsible AI), and the bioremediation papers themselves confirm the pattern, with field-trial failures and unsequenced enzymes buried in final paragraphs. We will audit a sample of bioremediation papers with the same 1–10 rubric to establish the domain's own "do nothing" number.
The strongest existing competitor is CARDGEN itself: its two-step retrieval-augmented pipeline with persona prompts, which achieved 7.28 completeness and 7.16 objectivity (1–10 scale) on ML artifacts (Automatic Generation of Model and Data Cards: A Step Towards Responsible AI). We will run it unchanged on BioCardBench papers — same corpus, same gold tuples, same compute and tuning budget as our pipeline — and score it on both its native rubric and our tuple-level metrics. Beating it means higher $o$-tuple field F1 (stratified by whether the value and its condition set live in different sections) at equal or better completeness, with the cross-section stratum as the decisive split. We expect CARDGEN to look strong on single-fact fields and collapse on condition-scoped tuples, and the experiment is designed to show exactly where.
Two further comparisons isolate the ingredients. A flat hierarchical extractor in the HyperPIE style — fine-tuned PL-Marker BERT over Artifact→Parameter→Value triples, the architecture that reached 79% entity F1 but only 38.8% relation F1 on its own ML benchmark (HyperPIE: Hyperparameter Information Extraction from Scientific Publications) — retrained on BioCardBench isolates the value of condition scoping and cross-section linking. And a dictionary-plus-regex extractor over standard units (mg/L, °C, pH) establishes the floor; the BioIE literature predicts it, with rule-based interaction extraction recall around 30% and severe semantic drift (A Biomedical Information Extraction Primer for NLP Researchers).
Why not just…?
Why not just make authors fill out a mandatory checklist at submission? This is the cheapest fix and the one journals would actually adopt, so we take it seriously. The evidence against it is direct: when thousands of human authors filled structured cards, completeness averaged 1.92/10 and objectivity 2.03/10, against 7.28 and 7.16 for automated generation (Automatic Generation of Model and Data Cards: A Step Towards Responsible AI), and Reusable Templates and Guides For Documenting Datasets and Models for Natural Language Processing and Generation documents the months of labor and abandonment that manual templates cause. Bioremediation adds a structural twist: most of the relevant literature is already published in prose, and no checklist retrofits it. Our pipeline is the only option that works on the existing corpus. For future papers, we would happily endorse a hybrid — authors get a pre-filled card to correct — and our audit experiment quantifies how much the human pass adds or subtracts.
Why not just regex and dictionary NER over standard units? It is deterministic and cheap, and it will be run as a baseline rather than argued away. But the BioIE primer shows why it fails: rule-based extraction of biological relations captured only about 30% of true interactions, and biological nomenclature suffers synonymy, ambiguous abbreviations, and semantic drift (A Biomedical Information Extraction Primer for NLP Researchers). The deeper problem is relational: a regex finds "93%" but cannot say whether it is quinclorac degradation by strain Q3 under pH 8 or a control-group value, which is precisely the relation task where even fine-tuned BERT managed only 38.8% F1 (HyperPIE: Hyperparameter Information Extraction from Scientific Publications). The regex baseline in our experiments will make this concrete rather than rhetorical.
Why not just give a strong zero-shot LLM the PDF? A real threat, since modern models read papers well. Three cited results bound it: open-domain verification collapses from 60.6 F1 with oracle-provided evidence to 39.5 F1 when the model must find the evidence itself (Fact or Fiction: Verifying Scientific Claims); unguarded LLM synthesis emits unsupported claims at 3–18% rates with verifiers that cannot agree on which ones (Evaluating and Guarding Citation Faithfulness in Agentic Scientific Synthesis); and even prompting format shifts entity F1 by 5.5 points (HyperPIE). We will run this baseline honestly — same model family, same corpus, generous context windows — and if it matches our pipeline on condition-scoped tuple F1 and passes the conformal calibration, the proposal's extraction contribution collapses and we would report that. Our working hypothesis, grounded in the retrieval-collapse evidence, is that it will hallucinate condition bindings at rates the guard will expose.
Why not just reuse CARDGEN off the shelf and write the schema? If the transfer were trivial, the contribution would shrink to a domain schema, which is real but thin. The experiments are structured to settle this: CARDGEN is our strongest baseline on identical data, and the cross-section stratum of BioCardBench is the test it was never built for. Its authors explicitly note it struggles with parameters buried in tables and complex interdependencies requiring multi-step reasoning — exactly the bioremediation regime.
Proposed Experiments
First, extraction: all systems (ours, CARDGEN, flat HyperPIE-style, regex, zero-shot LLM) evaluated on BioCardBench with the same tuning budget, reporting $o$-tuple field F1 overall and stratified by same-section versus cross-section values, plus entity F1 for strains, substrates, and metabolites. HyperPIE's own benchmark numbers (79% entity, 38.8% relation F1) serve as the reference point for what flat hierarchical extraction achieves in an easier domain; our target is relation-level F1 meaningfully above the flat baseline, with the cross-section gap as the headline. Second, verification: directionality accuracy on the degrade/accumulate and condition-swap claim pairs, reported against VERISCI's 75.7% label accuracy given correct rationales (Fact or Fiction: Verifying Scientific Claims) as the calibration-free reference. Third, the guard: measured slip-through of ungrounded card fields versus the conformal bound at several tolerance levels, following the calibration protocol of Evaluating and Guarding Citation Faithfulness in Agentic Scientific Synthesis; the check passes if the empirical violation rate stays at or below the stated tolerance across fresh papers. Fourth, downstream utility, the test that matters most: given only the generated Strain Cards, an expert must reconstruct the exact reported condition sets for the seed papers — for instance, recovering Q3's 20 mg/L, 30°C, pH 8, 6% inoculum, seven-day, 93% record (Isolation, Identification and Characteristics of an Endophytic Quinclorac Degrading Bacterium Bacillus megaterium Q3) and S09T2's over-93% OTA degradation of 5–8 µg/mL within 24 hours at 37°C (Isolation and Characterization of Brevibacillus parabrevis S09T2, a Novel Ochratoxin A-Degrading Strain with Application Potential) — and answer cross-paper aggregation queries that the prose alone makes impractical. The long-term outlet is semantic submission: the RASH framework (The RASH Framework: Enabling HTML+RDF Submissions in Scholarly Venues) already shows how HTML plus embedded RDF triples can carry machine-readable metadata into publisher workflows, and a Strain-Card ontology is the domain vocabulary it lacks — but that extension only pays off once extraction and verification are trustworthy, which is what this proposal establishes first.
References
- Isolation, Identification and Characteristics of an Endophytic Quinclorac Degrading Bacterium Bacillus megaterium Q3
- Isolation and Characterization of Brevibacillus parabrevis S09T2, a Novel Ochratoxin A-Degrading Strain with Application Potential
- Characterization and genomic analysis of a bensulfuron methyl-degrading endophytic bacterium Proteus sp. CD3 isolated from barnyard grass (Echinochloa crus-galli)
- Fermentation of ginkgo biloba kernel juice using Lactobacillus plantarum Y2 from the ginkgo peel: Fermentation characteristics and evolution of phenolic profiles, antioxidant activities in vitro, and volatile flavor compounds
- Screening and Characterization of Nitrite Degrading <i>Lactobacillus plantarum</i> in Chinese Traditional Pickles
- HyperPIE: Hyperparameter Information Extraction from Scientific Publications
- Automatic Generation of Model and Data Cards: A Step Towards Responsible AI
- Fact or Fiction: Verifying Scientific Claims
- Evaluating and Guarding Citation Faithfulness in Agentic Scientific Synthesis
- A Biomedical Information Extraction Primer for NLP Researchers
- Reusable Templates and Guides For Documenting Datasets and Models for Natural Language Processing and Generation: A Case Study of the HuggingFace and GEM Data and Model Cards
- The RASH Framework: Enabling HTML+RDF Submissions in Scholarly Venues