ORALBIOME: Reproducible 16S rRNA Analysis of the Oral Microbiome in Periodontitis

A reproducible Python pipeline that asks whether the salivary bacterial community differs between periodontally healthy people and people with periodontitis, with pre-registered predictions, negative and positive controls, and a replication cohort.

Associations only — no clinical claims

Associations in a small, cross-sectional dataset, not causal or clinical findings. Nothing here should be used to diagnose, predict or treat periodontitis.

PERMANOVA R², healthy vs periodontitis: 0.082 (p = 0.015)
Classifier AUC, without its top taxon: 0.92 → 0.64
Classifier AUC, replication cohort: 0.52 [0.32, 0.72]

Box plots of centred log-ratio abundance for Porphyromonas gingivalis, Tannerella forsythia and Treponema denticola in healthy (n = 16) and periodontitis (n = 18) saliva; all three are higher in periodontitis (q = 0.005, 0.012 and 0.004). — FIG. 02.2 — Pre-declared red-complex species test · Figure from the project repository

01Problem

Does the salivary bacterial community differ between periodontally healthy people and people with periodontitis? The pipeline runs on a real public 16S rRNA dataset, tests predictions written in HYPOTHESIS.md before any analysis, includes negative and positive controls, and checks whether the findings replicate in a second group of participants.

02Question

Can community composition alone predict periodontitis status — and does the signal replicate?

03Data

Unstimulated saliva, 16S rRNA V3–V4, Illumina NovaSeq; periodontitis diagnosed with the 2018 AAP/EFP classification. Source study: Guo Z., Yu X., Liu Y., Hu Q., Zhang Z., Zhang C., Li J. (2026), “A comparative analysis of oral microbial communities in hypertensive patients with and without chronic periodontitis,” BMC Oral Health 26(1):836, doi:10.1186/s12903-026-08144-6. Processed feature table: Guo, Ziyin (2025), figshare, doi:10.6084/m9.figshare.29897750.v1 (CC BY 4.0). 25,540 features, 67 samples and 4,503,386 reads; primary comparison 16 healthy vs 18 periodontitis, replication 17 hypertension vs 16 hypertension + periodontitis.

04Methods

Data layer: downloads and caches the figshare deposit, parses the BIOM table, writes a QC report, and flags samples where more than 50% of reads come from features seen in no other sample.
Preprocessing: aggregation to genus, a prevalence filter (≥ 10% of samples and mean relative abundance ≥ 0.01%), rarefaction to 44,786 reads for alpha diversity, and CLR with pseudocount 0.5 for everything else.
Alpha diversity: Shannon, Gini-Simpson, observed richness, Pielou's evenness, Chao1 and ACE; Mann-Whitney U with rank-biserial r and bootstrap CIs, BH correction across metrics, and two sensitivity analyses.
Beta diversity: Bray-Curtis and Jaccard, PCoA and NMDS with 95% ellipses; PERMANOVA and PERMDISP implemented from scratch with 999 permutations.
Differential abundance: per-genus Mann-Whitney U on CLR with BH-FDR and a label-shuffled null, plus a pre-declared species-level red-complex test.
Co-occurrence network: Spearman correlation on CLR abundances (|rho| ≥ 0.6, q < 0.05), hubs by degree.
Classifier: L1 logistic regression (nested CV) and random forest with 10× repeated stratified 5-fold CV, preprocessing inside folds, bootstrap CIs, a 100-shuffle permuted-label null and an ablation.
Replication: the same pipeline on hypertension vs hypertension + periodontitis; a synthetic Dirichlet-multinomial dataset with planted differences serves as the positive control.

05Tools

Python
NumPy
pandas
SciPy
scikit-learn
matplotlib
seaborn
NetworkX
Streamlit
Altair
pytest

06Visualizations

Two PCoA plots (Bray-Curtis and Jaccard) of healthy and periodontitis saliva samples with 95% ellipses; flagged samples are shown as open circles. — FIG. 02.2.1Figure from the project repository ·PCoA of Bray-Curtis and Jaccard distances with 95% ellipses; Bray-Curtis composition differs by group (PERMANOVA R² = 0.082, p = 0.015) with borderline unequal spread (PERMDISP p = 0.065).

ROC curves for the L1 logistic regression and random forest classifiers with bootstrap confidence bands. — FIG. 02.2.2Figure from the project repository ·Out-of-fold ROC curves from 10x repeated 5-fold CV with bootstrap 95% bands; L1 logistic regression reaches AUC 0.92 [0.81, 1.00] and random forest 0.81 [0.65, 0.94].

Charts of how often each genus is selected by the classifiers across folds, led by Unclassified Bacilli. — FIG. 02.2.3Figure from the project repository ·Genera used by the classifiers across the 50 fold-level models; one unidentified taxon (“Unclassified Bacilli”) is selected in every L1 fold, and dropping it lowers the AUC to 0.64 (L1) and 0.72 (random forest).

Screenshot of the ORALBIOME Streamlit dashboard overview: summary metrics (16 / 18 samples, PERMANOVA R² 0.082, 2 of 354 genera passing FDR, L1-LR AUC 0.92) above sequencing-depth and rarefaction plots. — FIG. 02.2.4Figure from the project repository ·The Streamlit dashboard’s overview tab: headline results and data-quality plots, with filters for group and flagged samples.

07Findings

All three pre-declared red-complex species are more abundant in periodontitis saliva: P. gingivalis q = 0.005, T. forsythia q = 0.012, T. denticola q = 0.004.
An L1 logistic regression separates 16 healthy from 18 periodontitis samples above a permuted-label null (AUC 0.92 [0.81, 1.00], permutation p = 0.0099), but it leans heavily on one unidentified taxon, “Unclassified Bacilli”: removing it drops the AUC to 0.64.
Did not replicate: in the hypertensive participants (17 vs 16), PERMANOVA p = 0.402, the classifier AUC is 0.52 [0.32, 0.72] (permutation p = 0.495) and the red-complex species give q = 0.53–0.76.
Overall composition differs modestly (Bray-Curtis PERMANOVA R² = 0.082, p = 0.015; PERMDISP p = 0.065), while alpha diversity does not (Shannon p = 0.796; no metric passes BH across six).
The top healthy-leaning genera (Subdoligranulum, Enterobacter, Nordella, Psychrobacter) are detected in all 6 unusually high-richness healthy samples but rarely elsewhere, so they may reflect contamination or batch rather than biology.

08Limitations

16S resolution: a ~470 bp V3–V4 amplicon usually identifies bacteria reliably only to genus, and says nothing about genes, function or live versus dead cells.
The deposit does not state which reference database assigned the taxonomy; names follow SILVA 138-era conventions, not eHOMD, so species labels carry more uncertainty than genus-level results.
“Unclassified Bacilli” could not be identified: the deposit contains no representative sequences, so a technical component cannot be excluded.
Sequencing gives proportions, not absolute amounts; CLR transforms reduce but do not remove this problem.
Batch effects and contamination cannot be separated from biology: no negative controls or processing-batch information are available.
Small (16 vs 18), cross-sectional and unadjusted for covariates; saliva from one study and one sequencing run, so the replication is internal only.
No clinical or diagnostic value: nothing here should be used to diagnose, predict or treat periodontitis.