Reading the Tables: Quantitative, Resolution-Aware Synthesis of How Microbes Reshape Roots

Plants that partner with soil fungi and bacteria often change the shape of their roots. Instead of growing long, carbon-expensive primary roots, they grow short, dense, highly branched zones that hand off nutrient foraging to fungal threads. These threads are far thinner than any root and can reach into soil pores the root cannot enter. Hundreds of pot and field studies measure this remodeling. Each paper picks its own host species, its own microbe, its own stress treatment (low phosphorus, drought, salt), and its own subset of root measurements. Many of these papers also profile the surrounding bacterial community with 16S rRNA gene sequencing, which identifies bacteria by reading a short barcode-like stretch of one gene. No one has pooled these studies, and the field cannot yet answer a basic question: across hosts and microbes, does colonization actually shift root investment from length toward surface area and branching, and how much?

We propose to answer that question by building an extraction system whose unit of output is a quantitative treatment-versus-control contrast. Each contrast carries a mean, a spread, a sample size, and the conditions of the experiment, read from full-text tables and figures, not from qualitative sentences in abstracts. The system also tags every reported microbiome claim with the taxonomic resolution its sequencing method can actually support. The bet is that, for this literature, a recent and influential finding that abstracts are enough for automated synthesis fails. If it fails, then any literature-synthesis system that reads abstracts will produce wrong pooled effect sizes.

Core Claim

The closest existing system, Affinage (Affinage; summary; Affinage: Genome-Scale Mechanistic Gene Annotation from the Published Literature), separates a skeptical fact-extraction pass from a synthesis pass that sees only the extracted facts. With that design it annotated every human protein-coding gene at 97.95% literature faithfulness. It also reported that full texts added only 16% more findings at fifteen times the cost and did not improve summaries. That result is reasonable for gene function, where the claim is qualitative ("protein X phosphorylates Y"). Root remodeling papers work differently. Their abstracts say "AMF improved root morphology," while the numbers needed for synthesis sit in the results tables: root surface area in cm², lateral root counts, sample sizes, and standard errors.

Our claim is that a meta-analysis built from abstracts would be biased in two ways, and that a contrast-level full-text extractor with deterministic arithmetic checks removes the bias at acceptable cost. First, the pooled estimates would be biased toward significant, positive traits. Second, they would be blind to traits that did not change, such as primary root length, which is the comparison the "outsourcing" hypothesis actually turns on.

The second ingredient is resolution awareness. Short-read sequencing of the V4 region of the 16S gene fails to classify 56% of sequences at species level. Full-length sequencing resolves species and strains (Evaluation of 16S rRNA gene sequencing for species and strain-level microbiome analysis). Papers nevertheless routinely state species-level rhizosphere shifts from short reads. The system should therefore record, for each microbiome claim, the region sequenced, whether reads were clustered into 97%-similarity groups (OTUs) or resolved as exact sequence variants (ASVs), and the rank of the claim. It should then refuse to pool claims finer than the method supports.

Research Program

The extraction target is a contrast record. It names the host species and genotype, the inoculant (a single fungus such as Piriformospora indica or Diversispora spurca, a fungus-plus-bacteria consortium, or native soil), and the stress condition. It gives the root or hormone trait in a controlled vocabulary with units, the inoculated and control means, their standard deviations or errors, the replicate counts, and a pointer to the exact table cell, figure panel, or sentence the values came from.

From each record we compute the log response ratio, the standard effect size in ecological meta-analysis:

$$\ln RR = \ln\frac{\bar X_I}{\bar X_C}, \qquad v = \frac{s_I^2}{n_I \bar X_I^2} + \frac{s_C^2}{n_C \bar X_C^2}$$

Here $\bar X_I$ and $\bar X_C$ are the inoculated and control means, $s_I$ and $s_C$ are their standard deviations, $n_I$ and $n_C$ are the replicate counts, and $v$ is the sampling variance used to weight the study. The extractor is judged by whether these numbers come out right, not by whether its prose sounds right.

The pipeline reuses Affinage's separation of stages but changes what each stage does. A retrieval stage selects inoculation experiments that report root phenotypes. A skeptical reading stage parses full-text tables and figure captions into contrast records. It must leave fields empty rather than infer them. A deterministic audit layer then checks each record before anything is synthesized:

  • Percent changes: any "percent increase" stated in the text must match the extracted means within rounding.
  • Error bars: the paper's own legend determines whether a bar is a standard error or a standard deviation. Standard errors are converted using the replicate count.
  • Units: units must agree within a trait.
  • Replicates: replicate counts must agree with the methods section.

Records that fail the audit are flagged, not repaired. The microbiome readouts in the same papers are extracted into parallel records: diversity indices, the direction of change for named taxa, the marker gene (16S for bacteria, or ITS, the fungal barcode that 16S primers cannot detect), and the resolution tag.

Synthesis is a standard random-effects meta-regression over the audited records, with host lineage, microbe type, and stress type as moderators. It is deliberately not an LLM summary. Because the store is structured, new papers update the pooled estimates incrementally. This is the "living" behavior that Living Meta-Review Generation for Social Scientists and the LLM-evaluation database in Can LLMs Help Uncover Insights about LLMs? built for other fields, here applied to quantitative plant phenotypes.

The biological payoff is three pooled tests:

Evaluation Data

No benchmark exists for this domain. The research file confirms that extraction tools cover machine learning, clinical medicine, chemistry, and materials science, but not plant–microbe interactions. We will therefore build one.

The gold set is about 150 open-access full-text inoculation papers, including those cited above and the combined fungus-plus-bacteria study (combined AMF and PGPR inoculation). Two plant scientists will independently hand-extract every contrast and microbiome claim, and we will report their agreement. Following the leakage lesson of MedDistant19, we split by host–inoculant pair. That paper showed prior relation-extraction benchmarks reused 26% to 86% of test facts in training, and that a prior model's AUC fell from 82.6 to 11.8 once the overlap was removed. In our setup, no host–inoculant combination used during prompt development or audit-rule design appears in the test papers.

A hand-run meta-analysis on the gold set serves as the reference for the pooled estimates. A temporal split is held out to test living updates: papers published after a cutoff year are ingested last.

Baselines

Naive reference: abstract-only vote counting. For each trait, count the abstracts that claim an increase, decrease, or no change. This is what a reader skimming a search engine does. It produces no effect sizes, but it shows how far qualitative evidence alone departs from the quantitative answer, especially on primary root length.

Strongest existing method: Affinage in its documented configuration. - Setup: the two-stage skeptical pipeline reads abstracts only. We adapt its prompts to the contrast schema and use the same model family and token budget as our system. - Why it is the strongest competitor: it is the best documented faithful extraction architecture in the material. It reports 97.95% faithfulness and a 0.029% contradiction rate on human genes (Affinage). No reported result was found on Lacuna for Affinage, or any comparable system, on plant root or rhizosphere literature. - What counts as beating it: lower absolute error in pooled $\ln RR$ against the hand meta-analysis, at a cost we report rather than hide. Matching it on faithfulness is not enough, because an abstract-only extractor can be perfectly faithful and still miss most of the effect sizes.

Equal-information competitor: single-pass full-text extraction. The same LLM receives the full text with a careful schema prompt, but there is no separate skeptical stage and no audit. All systems share the same papers, the same model, and the same prompt-tuning budget, measured in development-set iterations.

Document retrieval-augmented generation (RAG). Passages are embedded and the model answers each meta-analytic question from the top retrieved chunks. This is the approach Scaling Scientific Knowledge Discovery with Neuro-Symbolic AI and LLMs found to refuse or fragment answers, compared with grounding in a structured knowledge graph. That paper reports no numeric hallucination rate.

Ablations. Two ablations isolate the key ingredients. The first removes the arithmetic audit. The second removes the resolution tags, letting species-level claims from short-read data pool freely. The first shows whether the audit, rather than full-text access, drives accuracy. The second is measured by how often pooled taxon claims exceed what their sequencing method can resolve.

Why not just…?

Why not just read abstracts, as Affinage showed is sufficient? The answer depends on the kind of claim. Affinage's sufficiency result concerned qualitative mechanisms, and its 16% gain from full text was measured as a count of findings, not as error in effect sizes. Our first decisive experiment tests this directly. It compares pooled estimates and moderator conclusions from the abstract-only and full-text arms. If they agree within the confidence intervals, the abstract result transfers. We would then report that, and the fifteen-fold cheaper configuration becomes the recommendation.

Why not just give a strong model the full text in one prompt? This is the real threat, since modern models read tables well. Affinage's evidence against single-pass prompting concerns overclaiming mechanisms in prose, which is a different failure from mis-transcribing a standard error. The single-pass baseline and the audit ablation settle it. If single-pass extraction plus the deterministic audit matches the staged pipeline on contrast accuracy, we drop the second LLM stage. The contribution then narrows to the contrast schema, the audit, and the benchmark, which we would consider an acceptable outcome.

Why not just use retrieval-augmented question answering over the papers? Meta-analysis requires every eligible contrast, weighted and pooled. Retrieval returns the top few passages and silently drops the rest. Schmidt et al.'s comparison already found document retrieval fragmented and poorly cited relative to structured grounding. Our RAG baseline will measure the cost as coverage (the fraction of gold contrasts reflected in each answer) and pooled-estimate error.

Why not just pool the raw 16S data with topic models instead of extracting reported claims? Structural topic modeling over amplicon count matrices (themetagenomics) groups co-occurring bacteria into subcommunities and links them to host traits. But it needs deposited raw reads with compatible primers, and it was evaluated only on human clinical cohorts. We will measure what fraction of gold papers deposit usable reads. Where enough do, linking extracted root contrasts to topic-model subcommunities is a natural follow-up, not a substitute.

Proposed Experiments and Decisive Checks

Extraction accuracy. The first experiment measures contrast-level precision and recall on the inductive test split. A contrast counts as correct only if the trait, direction, and both means fall within 2% of gold, and the variance and sample size are correct. We also report the fraction of values traceable to a source cell or sentence, and dollar cost per paper for every arm.

Pooled-estimate error. The second experiment, the decisive one, compares each system's pooled $\ln RR$ for root surface area, lateral density, primary root length, root volume, and root IAA with the hand meta-analysis. It also checks whether each system reaches the same verdict on the outsourcing contrast, surface area minus length. The hypothesis fails if abstract-only Affinage reaches the same verdicts with overlapping intervals.

Resolution violations. The third experiment counts how many extracted species- or strain-level microbiome claims rest on short-read data. It reports how much pooled taxon conclusions change once such claims are coarsened to genus.

Living updates. The fourth experiment ingests the post-cutoff papers. It checks that the incremental updates match a fresh hand re-analysis, and that the resulting changes in moderator estimates are reported with the new evidence attached.

References

  1. Affinage: Genome-Scale Mechanistic Gene Annotation from the Published Literature
  2. Evaluation of 16S rRNA gene sequencing for species and strain-level microbiome analysis
  3. Living Meta-Review Generation for Social Scientists: An Interface and A Case Study on Human Cooperation
  4. Can LLMs Help Uncover Insights about LLMs? A Large-Scale, Evolving Literature Analysis of Frontier LLMs
  5. The effect of arbuscular mycorrhizal fungi on the growth of wheat seedlings with contrasting phosphorus use efficiencies under low phosphorus stress
  6. Mycorrhizal roles in plant growth, gas exchange, root morphology, and nutrient uptake of walnuts
  7. The Influence of Piriformospora indica Colonization on the Root Development and Growth of Cerasus humilis Cuttings
  8. Promoting Effects of Piriformospora indica on Plant Growth and Development of Tissue-Cultured Cerasus humilis Seedlings
  9. Growth Performance and Osmolyte Regulation of Drought-Stressed Walnut Plants Are Improved by Mycorrhiza
  10. Metabolomics Revealed the Tolerance and Growth Dynamics of Arbuscular Mycorrhizal Fungi (AMF) to Soil Salinity in Licorice
  11. Effects of combined inoculation of arbuscular mycorrhizal fungi and plant growth-promoting rhizosphere bacteria on seedling growth and rhizosphere microecology
  12. MedDistant19: Towards an Accurate Benchmark for Broad-Coverage Biomedical Relation Extraction
  13. Scaling Scientific Knowledge Discovery with Neuro-Symbolic AI and Large Language Models
  14. Exploring thematic structure and predicted functionality of 16S rRNA amplicon data