Beyond the Red Complex: De Novo Marker Discovery Uncovers Novel Periodontitis-Associated Taxa and Enables Non-Invasive Machine Learning Diagnosis

Background: Periodontitis affects over 1 billion people worldwide, yet diagnosis relies on clinical measures that capture tissue destruction rather than underlying microbial dysbiosis. Most microbial-biomarker studies use 16S rRNA sequencing or reference-database mapping, systematically under-detecting uncultivated or divergent taxa. Methods: We assembled 341 supra- and subgingival shotgun metagenomes (218 periodontitis, 123 health) across nine countries/regions. Using MetaMarker, a de novo, reference-free pipeline, we identified conserved genomic markers directly from reads in a 305-sample discovery pool without database mapping. Markers were taxonomically annotated against the Human Oral Microbiome Database, functionally annotated with Prodigal/eggNOG-mapper, and used for eight machine-learning classifiers, externally validated on three independent held-out cohorts (36 samples). Results: We recovered 2,142 significant markers (1,999 periodontitis-enriched, 143 health-enriched; q<0.05), recapitulating the canonical red/orange-complex dysbiotic shift. Beyond established pathogens, 128 periodontitis markers (6.4%) were novel, including an uncultivated Paludibacteraceae genus and divergent Fretibacterium fastidiosum strains. Case markers encoded a coherent virulence programme spanning proteolysis, haem/iron acquisition, and Type IX secretion. On external validation, boosted-tree classifiers generalized best (XGBoost and gradient boosting, AUC=0.96), and a compact SHAP-ranked 20-marker panel spanning five taxa reproduced full-set directionality. Conclusions: Reference-free metagenomic marker discovery recovers known periodontal pathobiology while revealing unrecognized candidate biomarkers and supports an accurate, externally validated, non-invasive classifier with translational potential as a compact diagnostic panel.