Every taxon–disease pair extracted from 325 papers, aggregated across the papers that report it. Bar length is the number of papers, not effect size — the extractor returns direction only, and the underlying studies report incommensurable statistics (LEfSe LDA, fold-change, p-values) that cannot honestly be pooled into one magnitude.
Scroll to zoom · drag background to pan · drag a node to move and fix it · click to pin its neighbourhood · double-click to reset. Solid links are associations (blue enriched, red depleted, grey dashed = papers disagree); faint dotted links are taxonomic containment — a family and the genera inside it. Containment is not redundancy: Lachnospiraceae is depleted in Parkinson's in 8 of the 9 papers reporting it while Hungatella inside it is enriched in 6 of 7 — and one study reports both directions itself, so this is not an artefact of pooling separate cohorts. Across the graph, related taxa agree on direction 89% of the time within a single paper (vs 54% for unrelated taxa); the 11% that do not are the cases this containment layer exists to keep. 63 edges carry such a conflict — tick Rank conflicts only to see just those.
| Taxon | Rank | Direction | Enriched | Depleted | Papers | Consistency | Diseases | Specificity |
|---|
How to read a contested edge. 220 pairs have papers pointing
both ways, and they are kept, never averaged into a single direction. That is deliberate: the
microbiome replication literature reports roughly one taxon in three flipping sign between cohorts,
so disagreement is a finding about the evidence, not noise to be smoothed. Direction is encoded by
position as well as color, so the chart survives colorblindness, greyscale and print.
High evidence is not high information. Some taxa move the same way in every disease that
reports them — Streptococcus is enriched in all 12 here, Butyricicoccus depleted in
all 8 — so "enriched in disease X" for those restates a generic dysbiosis prior and says almost
nothing about X. 273 of 2,008 edges do exactly that; they are drawn hollow
and can be hidden outright. 170 edges are the opposite case: their taxon changes direction
across diseases, so the direction here is disease-specific. The remainder are either partly
consistent or reported in fewer than three diseases, which is too few to judge either way.
Measured across the corpus, roughly 70% of this graph's directional agreement with Disbiome and
Peryton is that shared prior rather than disease-specific content, which is why the loudest edges
are often the least informative — Prevotella and Bacteroides carry the most papers
(50 and 54) and are among the least directionally consistent taxa in the graph.
Source: extractions_screened.json ·
271 of 325 papers contributed at least one association.
Associations only — no causal claim.