What the oral microbiome results support

Reanalysis of 507 selected specimens from 13 public studies, using the existing Archaions v18.8 classifier outputs. This report audits metadata before testing hypotheses and uses 434 qualifying human specimens for primary summaries. These are specimens, not 434 verified independent people.

1. The metadata audit changes the interpretation

The original map displays all 475 specimens with reported taxa. That includes controls and laboratory-grown communities. For human-specimen analysis, we excluded 26 specimens from three in vitro/ex vivo biofilm studies, two records explicitly labelled Neg Control, and 13 specimens with fewer than 1,000 classifier-used reads. A further 32 selected specimens had no reported taxa (19 failed the length-window eligibility gate and 13 completed without taxa). The resulting primary set contains 434 specimens from nine studies, with 113 reported genera and 532 model-labelled species.

The minimum of 1,000 used reads is an analysis threshold, not a validated universal quality standard. Sensitivity analyses at 5,000 and 10,000 used reads are reported below. All 45 specimens in PRJNA1087153 fail either the taxa or minimum-used-read requirement; this analysis therefore cannot support a diabetes comparison for that study.

We retrieved public attributes for all 507 specimens: 424 from ENA and 83 using an NCBI BioSample fallback. The audit identifies repeated saliva/plaque specimens, longitudinal athlete samples, experimental biofilms, and different clinical recruitment settings. Study titles alone were insufficient: sample attributes specify tongue plaque for PRJNA1242173, even though its project title uses “supragingival.” PRJNA1243403 attributes specify supragingival swabs.

2. A common compositional backbone, not a universal healthy reference

Streptococcus averages 32.1% abundance and is detected at ≥0.1% in 98.8% of the primary specimens. Veillonella averages 12.8% and is detected in 95.2%. Together they account for about 44.9% of the sample-weighted mean composition. Equal weighting of the eight studies with at least five primary specimens gives similar means: 33.4% and 12.2%, respectively. Each meets 80% prevalence in seven of those eight studies.

This supports a broadly shared backbone in this particular collection. It does not define a normal range for healthy people: recruitment is heterogeneous, diseased participants and extracted-tooth specimens are included, and specimen counts differ sharply between studies.

The median specimen contains 59 model-labelled species at ≥0.1%, but its median Shannon effective species count is only 24.9. These are different quantities: detected richness counts taxa above a threshold, whereas effective diversity accounts for uneven abundances. Neither estimates absolute bacterial load or true total species richness.

Species labels should be interpreted more cautiously than genera. Closely related mitis-group streptococci can be difficult to distinguish using 16S; the classifier’s S. pneumoniae label is not evidence of infection. This is why genus composition is the primary testing level. See the primary study of mitis-group identification.

3. Source study is a major source of variation

Study membership accounts for 32.8% of variation in the primary genus centered-log-ratio (CLR) representation. The estimate is 34.8% with a larger zero-handling constant and 35.5% under a Hellinger representation. Selecting one specimen per known participant gives 34.5% (324 retained records; unknown participant identities cannot be deduplicated). These are descriptive R² values, not causal estimates.

Study labels combine oral site, clinical population, geography, primer/library protocols, sequencing characteristics and other technical effects. We cannot separate these components with the available metadata. A pooled disease-versus-control analysis across projects would confound them.

4. The clearest biological pattern is the paired site difference

We matched saliva and plaque specimens by the deposited participant identifier, excluded the two labelled controls, and restricted permutations within each participant. The models account for log10 classifier-used read depth. The “Parkinson cohort” comparison includes both PD and nonPD participants; the “Alzheimer cohort” comparison includes both patients and controls.

Partial R² refers to the variation remaining after the participant and depth terms, not all variation in the original profiles. The corresponding increments in total R² are 8.3% and 5.9%. Both site findings persist under the alternative CLR constant and Hellinger geometry.

Saliva has higher Shannon diversity in both cohorts. Mean plaque-minus-saliva genus Shannon differences are −0.360 (95% paired-bootstrap interval −0.443 to −0.279) and −0.242 (−0.364 to −0.120). Species-level differences are −0.516 and −0.347. Diversity is not a health score: higher saliva diversity does not mean healthier saliva or diseased plaque.

5. Clinical comparisons: signals and limits

Parkinson’s status: exploratory association

Within the Parkinson study, genus composition differs between PD and nonPD participants after adjusting for recorded sex and read depth. Saliva: n=39, partial R²=5.4%, p=0.014, q=0.0288. Tongue plaque: n=40, partial R²=9.5%, p=0.002, q=0.00583. Both findings persist with alternative composition transforms.

The tongue-plaque groups also differ in multivariate dispersion (unadjusted diagnostic p=0.023). The group comparison may therefore reflect differences in variability as well as average composition. Age, medication, oral hygiene and other potential confounders were not available in a usable sample-level form. Neither result is a validated diagnostic signature.

Exploratory genus tests highlight Fusobacterium and Ruminococcus in tongue plaque, and Ruminococcus and Dialister in saliva. Fusobacterium mean abundance is 0.19% in PD tongue plaque versus 4.35% in nonPD; this is an association in this cohort, not a protective or harmful effect. The Ruminococcus finding depends on low-abundance detection (roughly 0.10–0.23% means in PD and zero reported in nonPD), making it especially sensitive to classification and zero handling. Genus-level multiplicity correction does not validate the underlying taxonomic assignments.

Smoking: no clear overall community separation here

Each smoking study contains four groups of ten specimens: non-smokers, cigarettes, medwakh and shisha. The overall community tests are not significant after correction: supragingival p=0.094, q=0.143; subgingival p=0.0865, q=0.143. Adding recorded periodontitis severity to the subgingival model gives p=0.066, q=0.116. Alternative transforms do not change this conclusion.

A few individual genus tests (Granulicatella, and Porphyromonas in the periodontitis-adjusted model) pass the exploratory feature-level correction. They are candidates for validation, not evidence of a general smoking signature. Small group sizes and unmeasured covariates limit power; failure to reject the global null does not establish that smoking has no effect.

Alzheimer’s disease: descriptive only

The source study used cohabiting patient–partner households. We recovered participant identities and can validly match each participant’s oral sites, but did not reconstruct the household matching needed for a patient-versus-partner test. We therefore report effect sizes without disease p-values: depth-adjusted partial R² is 4.3% in saliva and 5.9% in supragingival plaque. These values do not establish significance. The study’s published design makes independent-group inference inappropriate without further matching information.

Other clinical hypotheses

No obesity/IHD labels were inferred from sample codes; no diabetes contrast was performed after its samples failed the primary threshold; and no athlete concussion effect was claimed from repeated time points without a dedicated longitudinal model. This avoids treating missing labels or repeated measurements as independent evidence.

6. Sequencing depth still matters

After study membership is included, log read depth adds only about 0.5 percentage points to genus CLR explained variation (0.3 points for Hellinger). However, detected species richness remains depth-associated: pooled Spearman ρ=0.465 and study-adjusted rank correlation=0.549. Broad composition and rare-taxon detection are therefore not interchangeable outcomes.

Raising the minimum classifier-used count from 1,000 to 5,000 leaves 430 specimens; raising it to 10,000 leaves 351. Study R² remains about 32.7–33.1%. This supports the broad study-structure conclusion but does not show that every species-level or clinical result is depth-independent. The depth sensitivity here repeats global structure and diversity summaries, not every clinical test.

7. What we can and cannot conclude

• Supported in this collection: a Streptococcus/Veillonella-dominated average composition; strong study-associated variation; consistent within-person saliva–plaque differences in two cohorts; depth-sensitive taxon detection.
• Exploratory: Parkinson-status associations within one study, and a small number of clinical genus candidates. They need independent validation and better covariates.
• Not established: a universal healthy oral profile; a disease classifier; a causal microbial role in neurological disease; smoking having no biological effect; absolute microbial loads; pathogen infection; or true species-level interactions.

The taxonomy map’s shared-detection edges are not statistical interactions. We did not reinterpret pooled co-occurrence as ecological cooperation or competition. A credible interaction analysis would require appropriate compositional modelling, study/site stratification, substantially better covariates and independent validation; even then, observational association would not demonstrate interaction.

8. Methods and reproducibility

Input abundances are the existing model-derived percentages, reconciled to per-sample JSON and TSV outputs. They are not raw integer ASV counts. We did not round fractional estimates into invented counts, apply DESeq2-style count models, calculate count-based Chao richness, or claim rarefaction saturation.

Primary diversity: Shannon entropy using natural logarithms, effective diversity exp(Shannon), maximum relative abundance, and detected richness at ≥0.1%. Primary composition: genus CLR after adding 10⁻⁵ to proportions (0.001 percentage points). Global features require ≥0.1% abundance in ≥5% of primary specimens; comparison-specific features require the same abundance in ≥10% of specimens. CLR filtering defines the reference subcomposition. Sensitivities use a 10⁻⁴ constant and Hellinger-transformed all-genus proportions. The rationale for compositional treatment is described by Gloor et al.

Multivariate testing uses nested multivariate linear models (Euclidean PERMANOVA equivalent for the specified transformed coordinates), with 1,999 Freedman–Lane residual permutations for primary tests and 999 for transform sensitivities. Paired-site permutations are restricted within participant. No unrestricted study-label significance test is claimed because independent-person sampling is not assured. The nuisance-variable procedure follows Winkler et al. Exchangeability and model specification remain assumptions.

Group dispersion is checked using distances to Euclidean group centroids, corrected by √(n/(n−1)), with 999 label permutations. These diagnostic p-values are shown unadjusted and are not proof of equal or unequal dispersion. Paired diversity also has Wilcoxon tests and 5,000 participant-pair bootstrap resamples for mean-difference intervals. These intervals are unadjusted 95% intervals. They do not propagate classifier uncertainty or household dependence between different participants.

Benjamini–Hochberg correction is applied over all 35 multivariate/alpha/sensitivity tests; separately over all 213 primary genus-feature tests, including the additional periodontitis-adjusted comparison; and separately over the eight paired-diversity Wilcoxon tests. The seed is 20261104. The smallest primary Monte Carlo p is 0.0005. Species assignment uncertainty, first-read sampling, absence of measured FASTQ qualities in the transferred FASTA inputs, and technical study differences remain limitations.

Important next steps: verify the two control records; reconstruct Alzheimer household pairs and obtain age, medications, oral health and other covariates; validate candidate taxa against an independent oral reference workflow; repeat key findings with uniform random read subsampling from original FASTQ and appropriate read-level quality control; and test disease associations in an independent cohort. Increasing read depth alone will not resolve metadata or study confounding.

9. Files and audit trail

The analysis bundle contains the downloaded public metadata, sample-by-sample exclusions, diversity and taxon tables, all test results, six figures in PNG/PDF, numerical validation, and Python scripts. The complete raw classifier exports remain in the separate combined-results archive.

• analyze.py: full statistical analysis; fetch_attributes.py: public metadata retrieval.
• tables/: all numerical results, including nonsignificant tests.
• analysis_manifest.json: settings, software versions and input SHA-256 hashes.
• validation.json: numerical and sample-integrity checks.
• build_report.py: this report; validate_statistics.py: reproducible numerical checks.

Run from the project workspace: OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1 python3 analysis/statistics_507/analyze.py, then the validation and report scripts. Required packages: NumPy, pandas, SciPy and Matplotlib. The project’s combined_507 and taxonomy_map inputs are required, with hashes in the manifest.

Public archive sources: Parkinson study; Alzheimer study; supragingival smoking study; subgingival smoking study. Per-sample XML records are preserved in metadata/.