ORAL 16S · EXPLORATORY ASSOCIATION ANALYSIS · 4 NOVEMBER 2026

Taxon associations without disease labels

What was analysed

The previous metadata audit supplied 434 qualifying human specimens. Controls, laboratory biofilms, taxa-empty samples and specimens with fewer than 1,000 classifier-used reads were already excluded. We selected one specimen per deposited participant ID where available. In PRJNA1306337, codes such as H077 and H077sl were conservatively grouped; this is a protection against possible repeat sampling, not a verified participant linkage. Unknown identities otherwise use individual runs as proxies.

Only studies with at least 20 distinct known/proxy units were included. The main model contains 280 units from 6 studies. The exact selected specimens are saved in tables/specimen_selection.tsv. Fifty alternative selections test sensitivity to which specimen represents a unit. Cross-study participant overlap and Alzheimer household relationships remain unresolved, so the model does not establish complete independence of all units.

We retained 23 genera detected at ≥0.1% abundance in at least 20% of selected units, and in at least 10% of units in three or more studies. This defines 253 possible genus pairs. Disease/clinical group labels were not used for feature selection, modelling, tuning, or robustness checks.

Results

Among the 1 links passing all checks, 1 are positive and 0 are negative. A positive regularized partial correlation means positive conditional association in the adjusted log-ratio model; a negative coefficient means the opposite. Neither coefficient measures absolute bacterial growth, direction of influence, cooperation or competition.

Bootstrap selection percentages describe how often a link reappears when resampling these data at the selected penalty. They are not p-values, posterior probabilities, or a false-discovery guarantee. “All checks passed” means internal robustness under this workflow, not independent replication.

How the model reduces misleading links

1. Work on genus compositions, without inventing integer read counts.
2. Add 10⁻⁵ to relative proportions and apply a centered log-ratio transformation across the retained genera.
3. Remove study-by-oral-site fixed effects and log10 classifier-used read depth, then standardize the residuals.
4. Estimate a sparse precision matrix using graphical lasso. Convert precision entries into regularized partial correlations, conditional on the other included genera.
5. Select the penalty using a StARS-style stability path: 16 decreasing penalties and 50 study-stratified 80% subsamples, stopping before the monotone average instability exceeds 0.05.

This custom workflow follows the broad compositional sparse-network rationale of Kurtz et al., using scikit-learn’s graphical lasso. It is not a run of the SPIEC-EASI package and does not claim equivalence to its full implementation. CLR graphical modelling is an approximation with assumptions, not an automatic cure for compositional bias.

Predefined evidence tiers

Bootstrap-stable: selected in the main model, |partial correlation| ≥0.10, present in at least 80% of 250 study-stratified bootstrap fits, and at least 95% sign agreement among fits where the edge is present. The penalty and feature set are fixed during these checks; this does not measure uncertainty from every modelling choice.

Passes all conservative checks: bootstrap-stable, plus:

• Same nonzero sign in at least 80% of 50 alternative specimen selections.
• Same nonzero sign after leaving out at least five of six entire studies.
• Same nonzero sign after increasing the zero constant tenfold.
• Same nonzero sign in the ≥10,000-used-read subset (n=221).
• Same-direction, study/site/depth-adjusted pairwise CLR correlation of magnitude ≥0.20 in at least three studies, with ≥70% sign agreement among eligible studies.

A study is eligible for the last check when both genera reach the within-study prevalence threshold and retain nonzero adjusted variation. These within-study correlations are marginal adjusted associations; they are supporting evidence, not study-level replications of conditional edges. No within-study significance or meta-analysis p-value is claimed.

What conclusions are justified?

The retained pairs are candidates whose association is not easily explained solely by pooled study/site differences, known/proxy repeat specimens, one single study, or the tested depth/zero-handling settings. They are more informative than simply connecting taxa detected in the same sample.

However, residual shared habitat, diet, medication, disease status, unmeasured technical effects, uncertain participant linkage, and classifier errors can still create associations. The model ignores disease labels by design; it does not remove disease-related confounding. Conditional dependence among transformed abundances is not proof of a direct interaction. Absence of an edge also does not prove independence: uncommon genera were filtered and sparse penalties can suppress real weak links.

Do not use this graph to identify pathogens, choose treatment, or claim cooperation/competition. External data and targeted experiments are required for mechanistic validation. Species-level networks were not presented as primary evidence because these classifier-derived species assignments have additional uncertainty.

Deliverables and reproduction

The standalone interactive map is taxon_association_map.html. It opens offline, distinguishes conservative/stable/exploratory tiers, allows genus and direction filtering, displays per-link evidence, and exports visible links and SVG. The earlier taxonomy/co-detection map remains a separate artifact.

All 253 evaluated pairs—not just retained links—are exported to tables/all_tested_pairs.tsv. Additional tables contain within-study evidence, specimen/feature selection, penalty tuning and raw-versus-adjusted coefficients. Raw pooled Spearman coefficients are included only to show why they should not substitute for the adjusted model.

The analysis manifest records thresholds, input hashes, software versions and random seed 20261105. Run python3 analysis/associations_507/analyze_associations.py, then validate_associations.py and build_outputs.py from the project workspace. Required Python packages and versions are in requirements.txt. The ZIP preserves the project layout and includes the required input snapshots.

The independent validation script checks the graphical-lasso coefficient against an analytical two-variable case, nuisance residuals, specimen selection, edge thresholds and solver convergence. This validates implementation, not biological truth.