Integrative genomic analysis of 21 orofacial diseases identifies shared genetic architecture with systemic diseases

Abstract

Orofacial diseases are closely linked to systemic health, yet their genetic architecture and causal relationships remain incompletely understood. Here we show that integrative genomic analyses of 21 orofacial diseases in the UK Biobank, combining genome-wide, transcriptome-wide, rare variant association studies and Mendelian randomization, identify 48 novel loci and prioritize 160 putative causal genes. Among these, a synonymous variant in ALDH1A2 associated with periodontitis, encoding a key enzyme in retinoic acid synthesis, reduces translational efficiency, as shown by CRISPR–Cas9 editing. Single-cell RNA sequencing of periodontitis tissues further reveals impaired retinoic acid signaling and a T helper 17–skewed immune phenotype, accompanied by reduced retinoic acid levels in gingival crevicular fluid. Mendelian randomization supports a causal effect of Sjögren’s syndrome on lung cancer and stroke, replicated in FinnGen. By contrast, dental caries and periodontitis share substantial genetic architecture with systemic diseases, including enrichment of nicotine-related pathways, without robust evidence of causality. Together, these findings map the genetic landscape of orofacial diseases and clarify their complex links to systemic pathophysiology.

The integrative genomic analysis of 21 orofacial diseases identifies shared genetic architecture with systemic diseases, supports causal relationships, and highlights retinoic acid deficiency in periodontitis.

Introduction

Orofacial diseases, such as dental caries, periodontal diseases, temporomandibular joint disorders (TMD), and Sjögren’s syndrome, are highly prevalent, contributing substantially to global morbidity. The Global Burden of Disease Study 2021 identified oral disorders as the most prevalent diseases, noting that they affect 3.69 billion people and ranking them 12th among the leading global causes of Years Lived with Disability¹. Despite their multifactorial etiology, recent studies have revealed a substantial genetic component, with heritability estimates of ~ 50% for dental caries and ~ 38% for periodontitis^2,3. Most genetic risk loci for orofacial diseases remain undiscovered, however, and the majority of identified variants lie in non-coding regions, complicating the identification of causal genes and regulatory mechanisms⁴.

Orofacial diseases significantly reduce quality of life, not only by impairing essential functions like speaking, chewing, and swallowing, but also through their interactions with major systemic non-communicable diseases (NCDs). For example, periodontitis, an inflammatory disorder of the gums, is associated with atherosclerotic cardiovascular disease (ASCVD), type 2 diabetes (T2D), rheumatoid arthritis, and chronic obstructive pulmonary disease (COPD)^5–8. But despite consistent data linking these diseases, the underlying causality remains uncertain⁵. Given that several orofacial conditions are partially preventable or treatable, clarifying their causal role in systemic disease has important clinical implications⁹. While randomized controlled trials are rarely feasible in this context, Mendelian randomization (MR) enables causal inference by using genetic variants associated with orofacial traits as instrumental variables (IVs) in population-based settings.

Shared risk factors may also underlie the observed associations between orofacial and systemic diseases. Many orofacial conditions and NCDs share environmental risk factors (e.g., tobacco use, alcohol consumption, and high-sugar diets)¹⁰. In addition to lifestyle factors, recent studies suggest a contribution from shared genetic architecture. Variants in CDKN2B-AS1, for example, are associated with ASCVD, T2D, cancers, and periodontitis^11–13. Furthermore, recent studies have reported significant genetic correlations between dental caries or periodontitis and systemic diseases, including coronary artery disease (CAD), lung cancer, and rheumatoid arthritis¹⁴. The specific pleiotropic genes and pathways underlying these correlations, however, remain poorly defined. To date, no study has systematically mapped the shared genetic architecture between orofacial and systemic diseases using an unsupervised approach.

To address these gaps, we conducted a comprehensive genomic investigation of 21 orofacial diseases using data from the UK Biobank and FinnGen. We began by performing a genome-wide association study (GWAS) to identify common variants associated with orofacial traits. Next, transcriptome-wide association studies (TWAS) combined with eQTL-based MR were used to prioritize putative causal genes underlying these conditions. In parallel, GWAS-derived genetic instruments were used to evaluate potential causal relationships between orofacial diseases and systemic outcomes. To capture the contribution of rare coding variants, we carried out exome-wide association studies (ExWAS). By integrating these results with single-cell transcriptomic data, pathway enrichment, and in vitro functional validation, we aimed to map the genetic architecture of orofacial diseases and uncover the molecular pathways linking them to systemic NCDs.

Results

Integrative genomic analyses identify causal genes for orofacial diseases

Figure 1 summarizes our study design and analytic workflow. We defined 21 orofacial disease phenotypes by integrating ICD-coded diagnoses, OPCS-coded procedures, and self-reported data from the UK Biobank (Supplementary Data 1). We conducted GWAS using 93,095,623 genotyped and imputed variants, employing SAIGE to account for case–control imbalance and relatedness^15,16. At a genome-wide significance threshold (P < 5 × 10⁻⁸), we identified 73 independent loci associated with 19 traits, including 44 novel associations, using clumping with a 10 Mb window and an linkage disequilibrium (LD) threshold of r² < 0.1 (Fig. 2a and Supplementary Data 2). Applying a Bonferroni correction for 21 phenotypes (P < 2.38 × 10⁻⁹), we identified 26 loci that remained significant across six traits.

Fig. 1. Schematic overview of the study design and analytical workflow.

WES, whole-exome sequencing; GWAS, genome-wide association study; ExWAS, exome-wide association study; TWAS, transcriptome-wide association study; CAD, coronary artery disease; COPD, chronic obstructive pulmonary disease; HTN, hypertension; RhA, rheumatoid arthritis; T2D, type 2 diabetes; GCF, gingival crevicular fluid; RA, retinoic acid.

Fig. 2. Integrative genomic analyses of 21 orofacial diseases.

a Manhattan plots of GWAS results for 21 orofacial diseases in the UK Biobank. The dashed line indicates the genome-wide significance threshold (two-sided P = 5 × 10⁻⁸). b Manhattan-style plots of 1355 TWAS-significant gene–disease–tissue associations that passed permutation testing (two-sided PERM.PV < 0.05) across 16 of the 21 orofacial diseases. The x-axis groups genes by associated trait, with multiple tissues for the same gene plotted at the same position. For dental caries, the top 10 genes are labeled. Points are colored by direction of association: red for positive Z-scores (risk-increasing), blue for negative Z-scores (protective). Genes within the MHC region (chr6:25–34 Mb) are not shown. Novel long non-coding RNAs are labeled using locus-based shorthand names (e.g., CCDC91-AS1, Adjacent to TNPO3). The corresponding Ensembl IDs and genomic coordinates are provided in Supplementary Data 6. c Manhattan-style plots for the Mendelian randomization (MR) analyses of tissue-specific gene expression and the risk of 8 of the 21 orofacial diseases. Each point represents a gene–tissue–disease association tested using two-sample MR (two-sided), with GTEx v8 cis-eQTLs as instruments. The points are colored by the direction of the causal effect: red for positive β (risk-increasing), blue for negative β (protective), and gray for associations identified as pleiotropic by MR-PRESSO. Associations within the MHC region are not shown in the plot. Abbreviations: Benign_bone, benign bone tumor; Benign_headneck, benign head and neck tumor; Benign_salivary, benign salivary gland tumor; CLP, cleft lip and/or palate; Malocclusion, dental malocclusion; Eruption, eruption disorder; Impacted3, impacted third molar; Malig_headneck, malignant head and neck tumor; Malig_salivary, malignant salivary gland tumor; Retrognathism, mandibular retrognathism; Ulcer, oral ulcer; Sjögren, Sjögren’s syndrome.

To assess robustness to subtle within-European population structure, we repeated the GWAS for each orofacial phenotype after (i) adjusting for additional PCs, (ii) removing PC outliers, and (iii) restricting analyses to White British and non–White British subsets. Across all traits, the primary findings remained highly consistent, with minimal deviations except in the non–White British–only subset, where the sample size was markedly smaller (Supplementary Data 3, 4 and Supplementary Fig. 1). To further evaluate the presence of residual population structure or systematic test-statistic inflation, we examined pairwise plots of the top 20 genetic principal components and computed genomic control λ_GC and LDSC intercepts. The PC plots showed no major substructure beyond expected patterns (e.g., separation of Irish/Welsh clusters; Supplementary Fig. 2). Across all 21 orofacial diseases, λ_GC values were close to 1, and LDSC intercepts showed no evidence of confounding beyond polygenicity (Supplementary Data 5).

Building on these GWAS findings, we performed TWAS for 21 diseases across 49 tissues using FUSION. This allowed us to identify 6263 gene–disease–tissue associations that exceeded tissue-specific Bonferroni thresholds (P < 0.05 / number of cis-heritable genes per tissue). After accounting for multiple testing across the 1,119,321 gene–disease–tissue pairs using a global Bonferroni threshold (P  <  4.47 × 10⁻⁸), we found 3829 associations remained significant. We then applied a conditional analysis to the original 6263 tissue-specific associations to prioritize independent signals, resulting in 2386 conditionally independent gene–disease–tissue pairs across 16 diseases (Supplementary Data 6). Of these, we found that 2009 remained significant under a joint significance threshold (P < 0.05 / 2386) and 1355 (67%) were robust to LD-driven artifacts based on the FUSION permutation test (PERM.PV < 0.05), representing 624 unique gene–disease pairs (Fig. 2b and Supplementary Data 6). When we performed a colocalization analysis of the GWAS and expression QTL (eQTL) signals using COLOC¹⁷, we identified shared association signals for 629 associations (PP4 > 0.5), 472 of which showed strong evidence (PP4 > 0.75).

To refine putative causal genes among the TWAS-prioritized candidates, we performed MR analyses using tissue-specific cis-eQTLs from GTEx v8 for 814 TWAS-significant and permutation-supported gene–tissue–disease associations with adequate instrument strength (Supplementary Data 7). Using the inverse-variance weighted (IVW) or Wald ratio method, we found 605 associations remained significant after Bonferroni correction (P < 6.14 × 10⁻⁵). We noted that horizontal pleiotropy, as assessed with the MR-PRESSO global test, was absent in 314 associations (P > 0.05), present in 140 (P ≤ 0.05), and untestable in 151 due to insufficient numbers of IVs. Of the 140 pleiotropic associations, 130 (93%) were in the MHC, likely reflecting its high gene density, strong LD structure, and the multifunctional roles of immune-related genes. Among the 465 associations with no or untestable pleiotropy, collapsing gene–tissue–disease associations into gene–disease–level pairs yielded 121 associations for dental caries, 57 for Sjögren’s syndrome, and 12 for other orofacial conditions, totaling 190 unique gene–disease associations (Fig. 2c and Supplementary Data 7). These 190 associations corresponded to 160 unique causal genes prioritized through eQTL-based MR, of which 123 represent novel associations with orofacial diseases that, to our knowledge, have not been previously reported in the GWAS Catalog (Supplementary Data 8).

Of the 121 genes causally linked to dental caries, we found that 53 (44%) were located in the MHC, implicating immune-related pathways in caries susceptibility. Outside the MHC, we observed the strongest association with RBM6 (β = − 0.08; 95% confidence interval (CI) − 0.09 – − 0.07; P = 1.97 × 10⁻⁵²; Whole Blood), an RNA-binding protein involved in mRNA splicing. RBM6 is linked to inflammation and autoimmunity, suggesting an MHC-independent immunoregulatory role in dental caries¹⁸. For Sjögren’s syndrome, 56 of 57 causal genes were located in the MHC, which is consistent with known HLA associations¹⁹. We found the sole non-MHC signal was located in a transcript near TNPO3, which was previously implicated (β = 0.57; 95% CI 0.36 – 0.78; P = 1.81 × 10⁻⁷; testis)²⁰. Although they were previously implicated in disease risk, we provide additional support for SIGLEC5 and ZNF418 as causal genes for periodontitis and malignant head and neck tumor, respectively (β = 0.18; 95% CI 0.12 – 0.24; P = 1.67 × 10⁻⁸; stomach, β = 0.12; 95% CI 0.07 – 0.17; P = 8.24 × 10⁻⁷; spinal cord [C1] for SIGLEC5, β = − 0.23; 95% CI − 0.34 – − 0.12; P = 3.72 × 10⁻⁵; putamen [basal ganglia] for ZNF418). We also noted that increased expression of CCDC91 (β = 0.44; 95% CI 0.30 – 0.58; P = 1.86 × 10^⁻9; frontal cortex [BA9], β = 0.22; 95% CI 0.15 – 0.29; P = 2.49 × 10⁻⁹; cultured fibroblasts) and reduced levels of nearby long non-coding RNAs showed a causal association with periodontitis. Given that CCDC91 is implicated in elastin trafficking²¹, dysregulation of this pathway may impair extracellular matrix integrity, contributing to gingival instability and disease progression. Collectively, these findings delineate a genomic atlas of orofacial disease susceptibility, offering insights into underlying mechanisms and candidate therapeutic targets.

Rare coding variants identify genes associated with orofacial diseases

To further investigate the contribution of rare coding variants to orofacial disease risk, we first performed a gene-based ExWAS using whole-exome sequencing (WES) data from 470,000 UK Biobank participants and SAIGE-GENE+ to account for case-control imbalance (Table 1 and Supplementary Data 9)²². For genes that reached exome-wide significance (P < 2.5 × 10⁻⁶), we then examined single-variant association statistics using SAIGE to identify the variants driving each aggregated signal (Supplementary Data 10). This allowed us to identify four genes with significant associations across four diseases—ALDH1A2 (periodontitis), MGAT4A (malignant salivary gland tumor), TET2 (oral ulcer), and TMEM52B (TMD). None of these four genes appeared in our GWAS or TWAS analyses (Supplementary Data 2 and 6), and we did not identify prior reports linking these genes to these conditions.

Table 1.

ExWAS results for orofacial diseases

Phenotype	CHR	Gene	Group	max_MAF	P- value	Burden P- value	SKAT P- value
Malignant salivary gland tumor	2	MGAT4A	missense;lof;synonymous	1.00E-03	1.92E-07	1.94E-07	0.00046135
	2	MGAT4A	Cauchy	NA	1.53E-06	1.40E-06	0.00133353
Oral ulcer	4	TET2	missense;lof;synonymous	1.00E-04	3.75E-07	9.65E-07	1.28E-05
	4	TET2	Cauchy	NA	1.43E-06	2.49E-06	1.83E-05
Periodontitis	15	ALDH1A2	missense;lof;synonymous	1.00E-02	6.93E-08	6.70E-06	7.84E-08
	15	ALDH1A2	Cauchy	NA	6.24E-07	6.09E-05	7.05E-07
TMD	12	TMEM52B	missense;lof;synonymous	1.00E-03	1.08E-07	0.00432119	5.95E-08
	12	TMEM52B	Cauchy	NA	3.36E-07	0.00409576	1.28E-07

Results of a gene-based association test performed using SAIGE-GENE + , with two-sided P-values reported²². Each row represents the P-value for the most significant mask, representing a subset of variants based on the maximum MAF cutoff and functional annotation group, as well as the Cauchy combined P-value, which integrates the P-values from nine different masks. Genome-wide significance was defined as a Cauchy combined P-value < $2.5\times 10^{-6}$.

ALDH1A2 showed a strong gene-level association with periodontitis (P = 6.24 × 10⁻⁷) in SAIGE-GENE + . Single-variant analysis using SAIGE revealed that this signal was driven by the rare synonymous variant rs16939660 (AF = 0.0083; odds ratio (OR) = 1.36; 95% CI 1.21 – 1.53; P = 1.66 × 10⁻⁷) (Supplementary Data 10). Notably, the most significant common variant within the gene, rs138853906 (intronic; AF = 0.01; P = 4.04 × 10⁻⁶), identified in the GWAS summary statistics, did not reach genome-wide significance, emphasizing the power of rare variant aggregation. For malignant salivary gland tumors, MGAT4A showed a significant gene-level association (P = 1.53 × 10⁻⁶), although no individual variant—including the top missense variant rs141841148 (AF = 9.8 × 10⁻⁵; P = 3.88 × 10⁻⁴)—was suggestive on its own. For oral ulcers, the top gene TET2 reached gene-level significance (P = 1.43 × 10⁻⁶), despite no single variant showing a strong association (lead variant rs372862373; AF = 9.66 × 10⁻⁵; P = 6.52 × 10⁻³). Last, for TMD, TMEM52B showed the most significant gene-level association (P = 3.36 × 10⁻⁷), driven by a rare missense variant, rs772079638 (AF = 3.25 × 10⁻⁵; P = 5.27 × 10⁻⁷). Collectively, these findings underscore the utility of rare variant-based gene-level testing in identifying disease-associated genes that may not be detectable by conventional GWAS. Together with the 44 loci identified through GWAS, these findings define a total of 48 novel loci in this study.

Functional characterization of a periodontitis-associated synonymous variant in ALDH1A2

Of the four orofacial diseases for which we identified rare coding variant associations via ExWAS, we focused on periodontitis due to its high prevalence and clinical impact. Despite being a synonymous variant that does not alter its amino acid sequence, the lead ALDH1A2 SNP (rs16939660, c.453 A > G, p.A151A) showed significant association with periodontitis in our SAIGE analysis (OR = 1.36; 95% CI 1.21 – 1.53; P = $1.66\times {10}^{-7}$) (Supplementary Data 10). We decided to pursue it because synonymous mutations can affect gene function by altering mRNA stability or translational efficiency²³.

To assess the functional impact of the rs16939660 variant, we introduced it into HEK293T cells using CRISPR-Cas9 prime editing. We established three independent mutant clones, verifying each by Sanger sequencing (Fig. 3a, b). We selected HEK293T cells for their high transfection efficiency and suitability for precise genome editing, which permits the controlled evaluation of transcriptional and translational effects. While ALDH1A2 mRNA levels remained unchanged in the mutant cells (Fig. 3c), there was a significant reduction in ALDH1A2 protein levels (Fig. 3d), suggesting a translational defect. We confirmed reduced protein production from the variant transcripts via in vitro translation assays (Fig. 3e).

Fig. 3. Functional characterization of a periodontitis-associated synonymous variant in ALDH1A2.

a Schematic of the CRISPR prime editing strategy used to introduce rs16939660 (c.453 A > G, p.A151A) into HEK293T cells. b Sanger sequencing confirmation of the edited ALDH1A2 locus in mutant clones. c Quantitative PCR analysis of ALDH1A2 mRNA expression in wild-type (n = 11 biological replicates) and mutant (n = 3 biological replicates) cells. n.s., not significant. d Immunoblot analysis and quantification of ALDH1A2 protein levels in wild-type (n = 5 biological replicates) and mutant (n = 3 biological replicates) cells. α-tubulin was used as a loading control. ALDH1A2 and α-Tubulin were run on separate gels using identical protein samples and processed in parallel under the same experimental conditions. Equal amounts of protein were loaded for each sample. Quantification was performed within each gel and normalized to the corresponding loading control. e In vitro translation assay comparing ALDH1A2 protein output between wild-type (n = 6 biological replicates) and mutant (n = 6 biological replicates) constructs. f RNAfold-predicted centroid secondary structures for wild-type and rs16939660 mutant transcripts. Minimum free energy (MFE) values are indicated. For pairwise comparisons in panels (c–e), statistical significance was determined using an unpaired two-sided t test. Data are presented as mean ± SEM. *P < 0.05; ***P < 0.001.

To explore the underlying mechanism, we used RNAfold to model secondary structures of the mutant and wild-type transcripts²⁴. Although the minimum free energy (MFE) of the optimal structure for each group was similar (Supplementary Fig. 3), the centroid structure of the mutant transcript exhibited a higher MFE (–175.60 vs. – 199.17 kcal/mol), indicating reduced structural stability and increased ensemble diversity (Fig. 3f). These changes may underlie the impaired translational efficiency we observed. Collectively, these findings indicate that the periodontitis-associated synonymous ALDH1A2 variant rs16939660 represents a translation-impairing synonymous variant, as its altered mRNA structure reduces translation efficiency.

Impaired RA biosynthesis and signaling underlies immune dysregulation in periodontitis

Aldehyde dehydrogenase (ALDH) is the rate-limiting enzyme in the biosynthesis of all-trans retinoic acid (RA), an anti-inflammatory metabolite^25,26. After identifying the translation-impairing synonymous variant in ALDH1A2 associated with periodontitis, we next investigated whether RA signaling plays a role in periodontal pathogenesis by performing single-cell RNA sequencing (scRNA-seq) on gingival tissues from chronic periodontitis (PD) patients and healthy controls (HC). From nine human samples, we analyzed 65,315 cells, representing ten major cell types, including fibroblasts, epithelial cells, and various immune and vascular populations (Fig. 4a, b). In the PD samples, we observed a relative reduction in the population of structural cells, such as fibroblasts and epithelial cells, and an expansion of immune cells, especially NK/T and plasma cells (Fig. 4c), reflecting chronic inflammation²⁷.

Fig. 4. Reduced ALDH1A2 expression and retinoic acid deficiency in periodontitis patients.

a Uniform manifold approximation and projection (UMAP) plot of 65,315 single cells from the gingival tissues of healthy controls (HC, n = 3) and periodontitis patients (PD, n = 6), identified using scRNA-seq. b DotPlot showing marker gene expression used to define each cell type. c Proportions of the major cell types, showing the reduction of structural cells and expansion of immune cells in PD. HC: Healthy control; PD: Chronic periodontitis. d Feature Plot for ALDH1A2 expression, showing the highest levels of expression in endothelial and lymphatic endothelial cells (orange dashed box). e DotPlot showing endothelial and lymphatic endothelial cell ALDH1A2 expression (blue) and total ALDH1A2 expression (red) in HC vs PD. f UMAP plot depicting the endothelial and lymphatic endothelial cell subclusters. g DotPlot showing ALDH1A2 expression across endothelial subclusters. h Immunofluorescence staining of ALDH1A2 (green) and CD31 (magenta) in paraffin-embedded human gingival tissues from the HC and PD groups. Scale bars, 20 µm. i Quantification of ALDH1A2 fluorescence intensity in CD31⁺ endothelial cells from panel h. A total of nine region-of-interest (ROI)-level measurements per group were obtained from two independent donors (two images per donor, multiple ROIs per image). Statistical comparisons were performed at the ROI level using an unpaired two-sided t-test (P = 0.0056). Data are presented as mean ± SEM. j Schematic depicting gingival crevicular fluid (GCF) collection and retinoic acid (RA) extraction. Refer to the Materials and Methods section. Created in BioRender. Eom, B. S. (2026) [https://BioRender.com/prwbluj]. k RA concentrations in eluted GCF samples (HC, n = 6; PD, n = 6), analyzed with an unpaired two-sided t-test (P = 6.35 × 10⁻⁵). Data are presented as mean ± SEM.

Among all cell types, ALDH1A2 was most highly expressed in endothelial and lymphatic endothelial cells, accounting for 18.9% and 2.08% of expression, respectively. This was compared to < 0.53% for other lineages (Fig. 4d). In the PD group, ALDH1A2 expression was markedly reduced ( − 78.4% vs. HC), along with that of its other family members (i.e., ALDH1A1 and ALDH1A3) (Fig. 4e and Supplementary Fig. 4a). To further characterize endothelial heterogeneity, we re-clustered endothelial and lymphatic endothelial cells in an unsupervised analysis, identifying seven transcriptionally distinct subclusters (Fig. 4f and Supplementary Fig. 4b)²⁸. Among the subclusters, we found ALDH1A2 expression predominantly in clusters 0 and 1, which were also enriched for the leukocyte adhesion and transendothelial migration pathways (Fig. 4g and Supplementary Fig. 4c). We also confirmed via immunofluorescence staining that CD31⁺ endothelial cells of PD tissues showed significantly reduced ALDH1A2 protein levels (P = 0.0056) (Fig. 4h, i).

To functionally assess RA synthesis, we measured RA concentrations in gingival crevicular fluid (GCF) via ELISA, finding an 88% reduction in PD tissues (HC: 14.14 ng/ml; PD: 1.69 ng/ml, P < 0.0001) (Fig. 4j, k). In parallel, we found that the key RA-responsive genes ITGB7²⁹, ITGAE³⁰, SLC1A2³¹, IL2RA, IL2RB³², and TGM2³³ were all markedly downregulated in PD immune cells (Supplementary Fig. 4d, e).

Given RA’s role in suppressing T-helper 17 (Th17) cell responses and promoting regulatory T (Treg) cell activation²⁵, we wanted to assess Th17/Treg imbalance in periodontitis. We did so by quantifying IL17A⁺ and FOXP3⁺ cells, which were localized to cluster 4 in a UMAP-based subclustering of NK/T cells (Supplementary Fig. 5a–c). Compared to HC samples, PD samples exhibited increased expression of Th17 markers (IL17A, IL17F) and decreased expression of Treg markers (FOXP3, IL2RA, TNFRSF18) (Supplementary Fig. 5d, e). This was accompanied by a decline in FOXP3⁺ cells from 23.5% to 13.9% and an increase in IL17A⁺ cells from 4.6% to 16.8% (Supplementary Fig. 5f, g). Using Search Tool for the Retrieval of Interacting Genes (STRING) network analysis, we observed a broad downregulation of RA-regulated genes in neutrophils, NK/T cells, and plasma cells (Supplementary Fig. 6a–c). Together, these findings demonstrate that impaired RA biosynthesis and signaling contribute to a Th17/Treg imbalance and persistent inflammation in periodontitis—strong evidence supporting a causal role for ALDH1A2 in periodontal pathogenesis.

Causal relationships between orofacial and systemic diseases

Epidemiological studies have implicated orofacial diseases in increasing the risk of systemic conditions. To systematically evaluate these associations, we analyzed the relationships between 21 orofacial and 9 systemic diseases using individual-level UK Biobank data. Fourteen orofacial diseases were associated with elevated risk of at least one systemic condition, highlighting the possibility of biological links between oral and systemic health (Supplementary Data 11).

To determine whether these associations reflect underlying causality, we conducted MR analyses using genetic instruments derived from GWAS of the four orofacial diseases with sufficient significant variants—dental caries, periodontitis, Sjögren’s syndrome, and malignant head and neck tumor. We used the UK Biobank as the discovery cohort (Supplementary Data 12), and FinnGen for replication (Supplementary Data 13). We conducted the MR analyses across nine systemic disease outcomes, using a Bonferroni-corrected significance threshold (P  <  1.43 × 10⁻³). For Sjögren’s syndrome, both exposure and outcome effect sizes were obtained from FinnGen, enabling full cross-cohort replication. In contrast, for dental caries and periodontitis, comparable exposure GWAS results were not available in FinnGen; therefore, genetic instruments and exposure effect sizes were derived from the UK Biobank, while FinnGen data were used solely for outcome associations. For malignant head and neck tumor, the genome-wide significant instruments identified in the UK Biobank were not present in the FinnGen summary statistics; therefore, both the genetic instruments and corresponding effect size estimates were derived directly from FinnGen GWAS summary statistics.

Sjögren’s syndrome showed consistent causal effects on multiple systemic diseases in both cohorts, including lung cancer (OR, 1.103; 1.092) and stroke (OR, 1.030; 1.035), with no evidence of horizontal pleiotropy (Fig. 5). Dental caries also showed evidence of causal associations with CAD, COPD, and lung cancer in the UK Biobank analyses without detectable pleiotropy; however, these associations were not replicated in FinnGen, potentially reflecting reduced instrument strength, population-specific genetic architecture, or environmental differences between the two populations. In contrast, we found little evidence supporting a causal role for periodontitis or malignant head and neck tumor on any systemic diseases. Periodontitis was not significantly associated with any systemic outcome in either cohort, and the initial signals for malignant head and neck tumor were attenuated after excluding outcome-associated variants, suggesting a pleiotropic bias.

Fig. 5. Mendelian randomization analysis of the causal effects of orofacial diseases on systemic diseases.

Two-sample Mendelian randomization (MR) was conducted using two-sided tests to evaluate the causal effects of four orofacial diseases (i.e., dental caries, periodontitis, malignant head and neck tumor, and Sjögren’s syndrome) on nine systemic conditions. Each point represents an exposure–outcome pair. The plotted MR results are based on analyses using filtered IVs. To mitigate horizontal pleiotropy, outliers identified by MR-PRESSO and variants significantly associated with the outcome (Bonferroni-corrected P < 0.05) were excluded. Circles represent results derived from the UK Biobank, and triangles represent those replicated in FinnGen. Red symbols indicate positive associations (β > 0), and blue symbols indicate negative associations (β < 0). Gray dots denote non-significant associations or those flagged for pleiotropy by MR-PRESSO. The horizontal dashed line indicates the Bonferroni significance threshold. For Sjögren’s syndrome, beta coefficients for the exposure were obtained from FinnGen summary statistics, representing a fully independent replication. In contrast, for periodontitis and dental caries, beta coefficients were obtained from the UK Biobank due to the lack of corresponding FinnGen exposure summary statistics, and FinnGen data were used only for outcome associations, representing cross-cohort validation. For malignant head and neck tumors, both the genetic instruments and beta coefficients were derived from FinnGen data. Abbreviations: CAD, coronary artery disease; Cancer, all-type cancer; COPD, chronic obstructive pulmonary disease; RhA, rheumatoid arthritis; T2D, type 2 diabetes.

Because lifestyle and socioeconomic factors such as smoking, alcohol intake, and adiposity strongly influence orofacial diseases, we evaluated the robustness of our MR findings using GWAS summary statistics additionally adjusted for smoking status, body mass index, alcohol consumption, annual household income, and educational attainment. Although the covariate-adjusted models reduced effective sample sizes due to missing covariate information (Supplementary Data 14, 15), resulting in fewer genome-wide significant variants, the effect size estimates and P- values remained highly correlated with those from the primary GWAS, with consistent effect directions (Supplementary Data 16 and Supplementary Fig. 7). Re-running the MR analyses with the same genetic instruments but using exposure effect sizes from the covariate-adjusted GWAS yielded causal estimates that were essentially unchanged from the primary models (Supplementary Data 17, 18). Notably, in the periodontitis analysis, one variant (rs6495163) achieved genome-wide significance only in the covariate-adjusted GWAS (Supplementary Data 14); however, MR analysis using this variant as the sole instrument yielded non-significant p-values for all outcomes (Supplementary Data 19, 20). Together, these results indicate that the primary MR conclusions are robust to covariate adjustment in the exposure GWAS and to alternative instrument specifications, suggesting that major lifestyle or socioeconomic factors are unlikely to explain the observed findings.

To assess reverse causality, we performed MR using 9 systemic diseases as exposures and 21 orofacial diseases as outcomes. Of 189 possible exposure–outcome pairs (9 systemic diseases × 21 orofacial diseases), those involving stroke were excluded due to the lack of genome-wide significant instruments. Four additional pairs were removed during IV filtering based on significant outcome associations, resulting in 164 pairs tested. Two of these—the effect of COPD and oral cancer on dental caries—initially surpassed the Bonferroni-corrected threshold (P  <  2.98 × 10⁻⁴), suggesting genetic liability. Both effects, however, were attenuated after filtering, suggesting pleiotropic bias (Supplementary Data 21).

Collectively, these results support a causal role for Sjögren’s syndrome in systemic disease susceptibility, whereas for dental caries, periodontitis and malignant head and neck tumor, the MR analyses did not provide evidence for a causal effect.

Pleiotropic loci implicate nicotine signaling in shared susceptibility to orofacial and systemic diseases

Shared genetic factors are another plausible explanation for orofacial and systemic disease associations. To investigate, we performed genetic correlation analyses using LD Score Regression (LDSC) across 21 orofacial and 9 systemic diseases. We identified significant positive genetic correlations between dental caries or periodontitis and multiple systemic conditions—including CAD, COPD, lung cancer, rheumatoid arthritis, stroke, and T2D—all surviving Bonferroni correction (P < 2.65 × 10⁻⁴; Supplementary Data 22). LDSC-based heritability partitioning revealed shared genetic factors accounted for 9–24% of phenotypic correlations for dental caries and 15–25% for periodontitis (Supplementary Data 23).

To identify shared genetic loci in an unsupervised manner, we conducted cross-phenotype meta-analyses using CPASSOC for genetically correlated disease pairs. This allowed us to identify 275 genome-wide significant SNPs for dental caries–systemic disease pairs (mapping to 251 genes) and 34 SNPs for periodontitis-related pairs (32 genes) (Fig. 6a, b and Supplementary Data 24, 25). All SNPs reached P < 5 × 10⁻⁸ in CPASSOC and showed at least suggestive association (P < 1 × 10⁻³) in the relevant individual GWAS, supporting the contribution of pleiotropic effects to comorbidity through shared biology.

Fig. 6. Shared genetic architecture and pathway enrichment between orofacial and systemic diseases.

a, b Cross-phenotype association (CPASSOC) results for dental caries (a) and periodontitis (b) across nine systemic diseases. Manhattan-style plots display loci with significant heterogeneity-values (two-sided P_Het < 5 × 10⁻⁸), concatenated across all outcomes. Loci are labeled with their nearest coding genes. For dental caries, only the top three genes per outcome are annotated. c Pathway enrichment analysis of CPASSOC-significant SNPs mapped to eQTL genes using FUMA and analyzed with g:Profiler. Gene Ontology Biological Process (GO:BP) and Reactome terms are shown for dental caries and periodontitis. KEGG pathways are not shown as no terms reached significance. For dental caries, significant GO:BP terms were grouped and simplified into representative categories using the R package rrvgo, with the top 10 representative terms shown. For periodontitis, all significant GO:BP and Reactome terms (adjusted P < 0.05) are displayed without simplification. Abbreviations: CAD, coronary artery disease; COPD, chronic obstructive pulmonary disease; RhA, rheumatoid arthritis; T2D, type 2 diabetes; NS, not significant.

For further insight, we mapped the CPASSOC-significant SNPs to eQTL target genes via FUMA and analyzed pathway enrichment using g:Profiler (Supplementary Data 26). For dental caries, the enriched GO:BP terms included biological regulation, anatomical structure morphogenesis, regulation of cell-matrix adhesion, and response to nicotine. In a reactome analysis, we found interactions with nicotinic acetylcholine receptor signaling and the butyrophilin family of proteins, implicating tissue remodeling and neuroimmune pathways in the shared disease risk. For periodontitis, fewer pathways showed enrichment, which was consistent with the smaller SNP set. Still, both GO and reactome analyses consistently indicated a role for nicotine-related signaling (behavioral response to nicotine and nicotinic receptor activity). We note that all CPASSOC, gene-mapping, and pathway enrichment analyses were performed separately for each orofacial disease, and the nicotine-related pathways identified for dental caries and periodontitis therefore reflect disease-specific pleiotropic architecture, rather than any combined analysis.

This enrichment of nicotine-related pathways and the link between smoking and oral diseases led us to assess the association of these pleiotropic variants with smoking. We identified 10 SNPs within nicotine signaling pathways—8 from analyses of dental caries and 2 from analyses of periodontitis (Supplementary Data 27). Logistic regression using individual-level UK Biobank data showed that 7 of the 10 SNPs tested were significantly associated with smoking (P < 5 × 10⁻³, Bonferroni-corrected). Notably, 4 of these associations went in the opposite direction of the effect for smoking versus orofacial disease risk. These findings suggest a nicotine signaling-dependent but smoking-independent mechanism contributing to the shared genetic basis of orofacial and systemic conditions.

Discussion

This study, enabled by the scale and phenotypic resolution of the UK Biobank, provides a comprehensive genomic investigation spanning a broad range of orofacial diseases. Unlike prior efforts focused on single conditions, our analysis covered 21 traits and identified susceptibility loci, including loci for traits that have been difficult to study due to limited power or heterogeneous definitions. By integrating GWAS, TWAS, and eQTL-based MR, we addressed challenges in causal gene prioritization driven by LD and non-coding variation, identifying 160 putative causal genes across eight traits. Additionally, we identified four genes through gene-based rare variant association testing with whole-exome sequencing, leveraging the interpretability of coding variants. To ensure robustness under phenotype imbalance, we applied SAIGE and SAIGE-GENE + , which calibrate type I error in large-scale biobank analyses.

Although often considered functionally neutral, synonymous variants can regulate gene expression and translation. We identified a synonymous variant in ALDH1A2 (rs16939660) associated with periodontitis that disrupts translational efficiency via altered mRNA secondary structure. This finding highlights the pathogenic potential of synonymous mutations and supports their systematic assessment in sequencing studies. Using an integrative framework combining population-scale association testing, genome editing, single-cell transcriptomics, protein-level assays, and metabolite profiling, we established a mechanistic link between this variant and impaired RA signaling.

Our results further implicate dysregulated RA metabolism in periodontitis pathogenesis. RA modulates Th17/Treg balance and protects against experimental periodontitis in mice³⁴, but its role in human disease remained unclear. We found ALDH1A2, the rate-limiting enzyme for RA biosynthesis, was genetically associated with disease risk and transcriptionally downregulated in gingival endothelial cells from affected individuals. We found reduced levels of RA in periodontitis patient GCF with an accompanying suppression of canonical RA target genes and an immune shift characterized by decreased Tregs and increased Th17 cells. Notably, RA depletion in GCF suggests local RA levels may serve as a non-invasive biomarker of periodontal inflammation. These findings suggest that disruption of the ALDH1A–RA axis is a key molecular feature of periodontitis and implicate RA metabolism as an immunoregulatory pathway in the oral mucosa.

Although the causal relationship between orofacial and systemic diseases has long been debated, we show here that Sjögren’s syndrome confers a modest but significant increase in the risk of lung cancer and stroke. While prior studies have reported associations between Sjögren’s syndrome and a range of malignancies³⁵, evidence for causality has been largely restricted to specific cancer subtypes, particularly non-Hodgkin lymphoma³⁶. By leveraging genetic instruments derived from individual-level UK Biobank data and replicating our findings with FinnGen data, we identified consistent causal effects of Sjögren’s syndrome on lung cancer. In addition, our use of individual-level data rather than summary-level resources allowed us to confirm a causal effect on stroke, previously suggested by epidemiologic and MR studies^37,38. These findings expand the causal scope of Sjögren’s syndrome beyond hematologic malignancies and underscore the systemic relevance of orofacial diseases, especially given that xerostomia—a hallmark of Sjögren’s—often triggers its diagnosis in dental settings.

In contrast, we found little evidence supporting a causal role for dental caries or periodontitis in systemic disease risk. Instead, we found significant overlap in genetic architecture that suggests the prior epidemiologic associations reflect shared genetic or behavioral risk factors rather than direct causality. Notably, we found a strong enrichment for nicotine signaling among the loci common to orofacial and systemic diseases. This raises two plausible mechanisms. First, several of the shared loci were associated with smoking propensity, suggesting a genetic predisposition to nicotine use may increase risk for both disease categories. The directionality of some SNP effects, however, argues against this being the sole explanation. Second, nicotinic acetylcholine signaling—particularly via the α7 receptor on immune cells—modulates inflammation through the cholinergic anti-inflammatory pathway³⁹. Dysregulation of this pathway may influence both local and systemic inflammatory responses. While our results do not support direct causality for dental caries or periodontitis, the strong genetic and immunologic overlap highlights their potential as early indicators of systemic disease susceptibility.

Despite these insights, several limitations should be considered. First, orofacial disease phenotypes in the UK Biobank may be substantially underdiagnosed, particularly for mild cases unlikely to prompt clinical attention or be reliably captured in electronic health records. This may lead to non-differential outcome misclassification and reduced statistical power, potentially obscuring additional associations. Nevertheless, previous cross-cohort benchmarking has demonstrated that UK Biobank proxy definitions for periodontitis and dental caries capture essentially the same heritable component as clinically ascertained disease, with high genetic correlations reported between GLIDE traits and their UK Biobank counterparts¹⁴. These findings support the validity of proxy-based phenotypes in large-scale genetic studies, although more clinically phenotyped cohorts will be crucial for refining trait definition.

In addition, interpretation of the MR analyses requires consideration of instrument-related characteristics. Although the mean and median F-statistics for all MR analyses substantially exceeded the conventional threshold of 10, indicating statistically strong instruments (Supplementary Data 28), periodontitis and malignant head and neck tumor were associated with only a limited number of genome-wide significant loci, resulting in a relatively small proportion of phenotypic variance explained. This inherently restricts the achievable statistical power of MR, even when F-statistics are high. Consistent with this, our power calculations showed that only relatively large causal effects would be detectable for these phenotypes under current sample sizes and instrument availability (Supplementary Data 29, 30). Therefore, the absence of significant MR associations should not be interpreted as evidence of biological non-causality but rather as a reflection of insufficient statistical power.

Furthermore, phenotype definitions were not identical between UK Biobank and FinnGen. In several instances, UK Biobank employed broader definitions incorporating self-reported diagnoses, procedure codes, or biomarker-based criteria, whereas FinnGen relied primarily on strictly registry-based ICD coding. Although these differences were carefully documented in Supplementary Data 31 and 32 and considered during interpretation, residual heterogeneity in case definition across cohorts may have contributed to attenuated or inconsistent associations in cross-cohort and MR analyses. In addition, case and control counts differed between the two cohorts (Supplementary Data 32), which may have affected statistical power and cross-cohort replication.

Moreover, because the primary discovery analyses were conducted in UK Biobank participants of European ancestry, the generalizability of our findings to non-European populations is limited. Although FinnGen also consists of individuals of broadly European ancestry, the Finnish population represents a genetically distinct founder group characterized by extended LD and unique allele frequency distributions. These features can reduce the tagging accuracy of UK Biobank–identified lead variants and attenuate their observed effects in replication analyses. Environmental exposures—particularly smoking-related behaviors known to influence risk for several orofacial diseases—also differ between populations and may further modulate genetic effect sizes. Consequently, incomplete replication in FinnGen may reflect population-specific LD structure and environmental interactions rather than the absence of genuine underlying biological effects.

In conclusion, our findings provide a genetic framework for improving our understanding of orofacial diseases and their links to systemic health. By integrating GWAS, TWAS, rare variant analysis, and causal inference, we identified susceptibility loci, prioritized candidate effector genes, and revealed mechanistic insights—including impaired RA metabolism in periodontitis and the systemic impact of Sjögren’s syndrome. The shared genetic architecture between orofacial and systemic diseases may explain long-standing epidemiologic associations. It also supports the potential of dental phenotypes as early indicators of systemic risk. These results lay the foundation for future translational efforts in disease prediction, prevention, and integrated care across the dental and medical domains.

Methods

Ethics

All analyses involving human genetic and phenotypic data from the UK Biobank (https://www.ukbiobank.ac.uk) were conducted under an approved UK Biobank application and in accordance with the policies and regulations of the UK Biobank resource. UK Biobank received ethical approval from the North West Multi-center Research Ethics Committee (MREC), and all participants provided written informed consent at the time of enrollment. Human gingival tissue and GCF samples used for functional experiments were collected with approval from the Institutional Review Board of Yonsei University Dental Hospital (approval number 2-2024-0003). All participants provided written informed consent prior to sample collection. All procedures involving human participants were conducted in accordance with the principles of the Declaration of Helsinki and relevant institutional and national guidelines.

GWAS and ExWAS of 21 orofacial diseases

For the single-variant association study, we used called and imputed genotype data from 457,548 European participants in the UK Biobank. The UK Biobank is a UK-based prospective cohort of ∼500,000 individuals aged 40 to 69 at enrollment. We conducted GWAS for 21 orofacial diseases and 9 systemic diseases, each defined according to criteria detailed in Supplementary Data 1, 31 and 32. Phenotypes were based on ICD codes, OPCS-coded procedures, and self-reported questionnaire items. For periodontitis and dental caries, we adopted proxy definitions previously shown to have high genetic validity in the UK Biobank: cross-cohort benchmarking by Shungin et al. ¹⁴ demonstrated strong concordance between clinically ascertained GLIDE traits and analogous UK Biobank proxy traits, including a near-unity genetic correlation between periodontitis and “loose teeth” (r_g ≈ 1.0) and a high correlation between DMFS and dentures (r_g= 0.82). These findings support the use of registry- and self-report–based phenotypes as genetically meaningful proxies in large-scale association studies. To minimize collider bias, “dentures” was removed from the exclusion criteria when defining periodontitis.

We performed a single-variant GWAS using a linear mixed model implemented in SAIGE¹⁶ (version 1.3.0) to maximize statistical power while accounting for sample relatedness and correcting for type I errors arising from case-control imbalances. In step 1, we constructed a genetic relatedness matrix (GRM) using 215,514 variants selected by LD pruning with PLINK2 (version 2.00a4LM), applying a window size of 50 kb, a step size of 5, and an r² threshold of 0.05. All analyses were adjusted for age, sex, and the top 10 genetic principal components (PCs). For the GWAS of dental caries, we also included periodontitis status as a covariate to reduce potential confounding arising from their shared genetic architecture. After GWAS, we conducted a clumping analysis using PLINK2 for the variants with P- values less than $5\times {10}^{-8}$, a window size of 10 Mb, and a LD threshold of 0.1 to identify independent genome-wide significant loci.

To evaluate the robustness of our GWAS results to subtle within-European population structure, we conducted a series of sensitivity analyses using alternative ancestry adjustment and sample filtering strategies. First, we repeated all GWAS analyses with an expanded set of genetic principal components, adjusting for the top 15 and 20 PCs instead of the standard 10 PCs. Second, to remove individuals exhibiting extreme ancestry deviation along individual PCs, we excluded samples lying beyond ±6 standard deviations from the mean for any single PC⁴⁰ and additionally computed Mahalanobis distances using the top 10 PCs and removed individuals with P < 0.001⁴¹, corresponding to the tail of the χ² distribution with 10 degrees of freedom. Third, we restricted analyses to White British and non-White British subsets. In addition, we quantified the extent of potential population stratification in each GWAS by computing the genomic control lambda (λGC) and the LD score regression intercept. To visualize population structure and assess potential substructure within the European sample, we generated pairwise scatter plots of the top 20 genetic principal components, with sample colorings based on assessment center and country of birth (Supplementary Fig. 2).

For the gene-level association analysis, we employed SAIGE-GENE + (version 1.3.0)²², a method designed for set-based rare variant association testing, to maintain statistical power while controlling for type 1 errors arising from case-control imbalance and the presence of ultra-rare variants. We used the same set of variants for step 1, with the addition of an indicator for exome sequencing batch⁴². In this analysis, we used WES data from 440,125 European participants in the UK Biobank but excluded variants with a missingness rate greater than 0.1 across individuals, variants with a Hardy-Weinberg equilibrium (HWE) P-value of less than ${10}^{-15}$, and variants that were monomorphic. Using the Loss-Of-Function Transcript Effect Estimator (LOFTEE), we created group files to define the list of variants within genes along with their functional annotations⁴³. Then, we classified only variants labeled with high-confidence (HC) as loss-of-function (LoF) variants, considering those labeled low-confidence (LC) as missense variants. For genes that reached exome-wide significance (Gene-level P-value < $2.5\times {10}^{-6}$) in the SAIGE-GENE + analysis, we additionally performed single-variant association testing using SAIGE to obtain variant-level P-value within each significant gene.

We then searched the GWAS catalog⁴⁴ (https://www.ebi.ac.uk/gwas/) for existing associations within $\pm$ 1 Mb from the lead variant to determine whether an associated locus should be considered novel. We used Experimental Factor Ontology (EFO) traits to harmonize trait names reported under different terms. In addition, we conducted an exhaustive search for previous reports of associations with the same or similar phenotypes.

TWAS of 21 orofacial diseases

For the tissue-specific transcriptome-level association analysis, we used the Functional Summary-based Imputation (FUSION; commit e1ba5f7f3907e6f586f7fb5bb115b35cc0d3c0c2) pipeline to perform TWAS on GTEx v8 multi-tissue expression data (https://gtexportal.org/home/) from European ancestry samples and our GWAS summary statistics of 21 orofacial diseases⁴⁵. This included a total of 49 TWAS analyses per disease, one tissue–disease pair at a time, using precomputed gene expression weights provided by the Mancuso laboratory⁴⁶. Transcriptome-wide significance was determined by applying a Bonferroni correction within each reference panel, defining significant associations as those with nominal P-values below 0.05 divided by the number of tested genes. This analysis allowed us to identify 6263 gene–tissue–disease associations that surpassed the Bonferroni-corrected threshold. To identify conditionally independent transcriptome-wide associations, we conducted a joint/conditional analysis on these 6263 significant TWAS signals using the post-processing module in FUSION with precomputed LD reference panels (1000 Genomes Project Phase 3 European ancestry), downloaded from the Broad Institute server (https://data.broadinstitute.org/alkesgroup/FUSION/LDREF.tar.bz2) in March 2025. We combined the expression weights for each gene by consolidating overlapping loci within 100 kb and then evaluating for joint significance. In this joint/conditional analysis, we identified 2386 conditionally independent genes, 2009 of which remained significant after applying the Bonferroni correction. To empirically evaluate the possibility that the TWAS signals we observed arose from a correlation with local GWAS signals rather than a true expression–trait association, we performed FUSION permutation tests with 1000 permutations per gene by randomly shuffling expression weights to generate a null distribution of TWAS Z-scores conditional on the observed GWAS statistics. Finally, for each gene–tissue–disease pair with a significant TWAS association, we performed a colocalization analysis to assess whether the GWAS and predicted expression signals were driven by shared causal variants. After computing posterior probabilities for the five COLOC hypotheses (PP0–PP4), we prioritized genes with high posterior probabilities of colocalization (PP4) as putative targets being supported by both expression and trait association signals.

Mendelian randomization

To investigate causal relationships between gene expression and disease, as well as between orofacial and systemic diseases, we performed Mendelian randomization (MR) analyses using the TwoSampleMR (version 0.6.6) and MR-PRESSO (version 1.0) packages in R^47,48. These included two types of MR analyses: (1) gene expression–disease MR assessing the causal effects of genetically predicted gene expression (based on TWAS-significant genes) on disease risk, and (2) disease–disease MR evaluating bidirectional causal relationships between orofacial and systemic diseases by alternately assigning each trait as the exposure or outcome. To assess instrument strength, we calculated mean/median F-statistics for all MR analyses, including sensitivity models with additional covariates. Across traits, F-statistics substantially exceeded the conventional threshold of 10, indicating strong instruments and reducing concerns about weak-instrument bias (Supplementary Data 28)^49,50. The number of overlapping participants for each orofacial–systemic disease pair is provided in Supplementary Data 33.

For the gene expression–disease MR, we analyzed 1,355 gene–tissue–disease pairs identified by TWAS and supported by the FUSION permutation test (PERM.PV < 0.05). Of these, 816 pairs were included in the MR analyses because they had at least one valid instrumental variable with a minor allele frequency (MAF) > 0.01. Then, we extracted independent cis-eQTLs from GTEx v8 tissue-specific datasets using LD clumping with an r² threshold of 0.1. When only one instrument was available, we applied the Wald ratio method. Otherwise, we used five complementary methods: inverse-variance weighted (IVW), MR-Egger, weighted median, simple mode, and weighted mode^51–54. All MR results from the five methods appear in Supplementary Data 7. To visualize the results in Manhattan-style plots, we used P-values from the IVW or Wald methods. We then assessed horizontal pleiotropy using the MR-PRESSO global test. When we did detect pleiotropy (P < 0.05), we removed outlier variants and performed the MR again to obtain corrected estimates.

In the disease–disease MR analyses, we initially considered all 21 orofacial diseases as exposures. After filtering for variants with minor allele frequencies (MAF) ≥ 0.01, only four traits (i.e., malignant head and neck tumor, Sjögren’s syndrome, dental caries, and periodontitis) retained sufficient IVs. Among these IVs, we applied LD clumping using an R² threshold of 0.1 within a 250 kb window to ensure their independence. We performed Steiger filtering to ensure the correct direction of effect between the exposure and outcome; however, no variants were removed at this step. Only Sjögren’s syndrome and dental caries had enough IVs to evaluate horizontal pleiotropy using the MR-PRESSO global test. Pleiotropy was detected in some pairs involving both traits, but we identified and removed outlier variants only for dental caries. Despite outlier removal, pleiotropy persisted in certain pairs. In accordance with the assumptions of MR, we further excluded variants with Bonferroni-significant associations with the outcome (P < 0.05 / (# initial IVs)) for both traits. Supplementary Data 13 presents the unfiltered results and the results after excluding both outlier variants and outcome-associated instruments. The Manhattan plots in the figure panels display results based on outcome-associated variant filtering in cases where pleiotropy was detected or could not be reliably evaluated, and unfiltered results when no evidence of pleiotropy was observed.

Power for the IVW MR analyses was estimated using analytical calculations based on the expected noncentrality parameter (NCP) of the IVW test statistic⁵⁵. For each exposure–outcome pair, we incorporated (i) the total sample size of the outcome GWAS ($N_{out}$), (ii) the variance explained by the set of instrumental variables (R²), derived from the exposure GWAS, and (iii) a prespecified causal effect size under the alternative hypothesis (${\beta }_1$). The NCP was computed as shown in Eq. (1).

$$\mathrm{NCP}=\frac{N_{out}\times R^2\times {\beta }_1^2}{1-R^2}$$ 1

Statistical power was then obtained as defined in Eq. (2).

$$\mathrm{Power}=P\left({\chi }_1^2>{\chi }_{1,0.05}^2\mid \mathrm{NCP}\right)$$ 2

MR replication in FinnGen

To assess the replicability of our disease–disease MR findings, we performed an independent replication analysis using summary statistics from the FinnGen study (release 12; https://www.finngen.fi/en), which includes genetic association data from up to 500,000 individuals of Finnish ancestry⁵⁶. The definitions of the orofacial diseases we used as exposures and the systemic diseases we used as outcomes followed those of the discovery stage and are detailed in Supplementary Data 1 and 31, respectively.

For MR replication in FinnGen, we applied a standardized variant-level harmonization procedure to ensure consistent allele definitions across UK Biobank and FinnGen and to minimize errors introduced by population-specific allele frequency and LD differences. Exposure and outcome summary statistics were first aligned by genomic position and allele coding. Effect alleles were matched across datasets, and strand alignment was performed using the harmonise_data function in the TwoSampleMR package. Palindromic SNPs (A/T or C/G) were handled according to the default TwoSampleMR criteria: palindromic variants with sufficiently low minor allele frequency (MAF < 0.42) were retained because allele orientation can be inferred unambiguously, whereas palindromic SNPs with MAF ≥ 0.42 were automatically excluded due to strand ambiguity. To further account for allele-frequency and LD differences between populations, we compared MAF between UK Biobank and FinnGen for every candidate instrument and removed variants with large allele frequency discordance (absolute MAF difference > 0.05). After these steps, harmonized instruments were finalized for each exposure–outcome pair.

Where possible, we used the same set of IVs identified in the UK Biobank discovery analysis. For Sjögren’s syndrome, both exposure and outcome summary statistics were available in FinnGen, enabling a fully independent replication. For other traits, we applied a complementary approach tailored to the available data. In the case of malignant head and neck tumors, none of the IVs identified in the UK Biobank were present in the FinnGen dataset. We therefore selected a genome-wide significant variant (P < 5 × 10⁻⁸) from FinnGen as the IV. For periodontitis and dental caries, directly comparable exposure phenotypes were unavailable in FinnGen. Accordingly, we used the UK Biobank exposure summary statistics in combination with FinnGen outcome data to enable cross-cohort validation. In our framework, we define a full independent replication as an MR analysis in which both exposure and outcome GWAS originate from an external cohort (FinnGen), whereas a cross-cohort validation uses exposure statistics from the discovery cohort (UK Biobank) and outcome statistics from FinnGen, thereby assessing primarily the replicability of outcome associations rather than instrument replication.

To strengthen robustness under this cross-cohort design, we additionally performed Steiger directionality tests and leave-one-out MR analyses for all exposure-outcome pairs evaluated in FinnGen (Supplementary Data 30 and Supplementary Fig. 8). When evidence of horizontal pleiotropy was detected, we excluded outcome-associated variants after Steiger filtering; when no pleiotropy was observed, the Steiger-filtered instruments were retained. Accordingly, the Manhattan plots in the figure panels reflect post-outcome-filtering results when pleiotropy was detected or could not be reliably assessed, and pre-outcome-filtering (Steiger-filtered) results when no pleiotropic signal was detected.

Logistic regression analysis between orofacial diseases and systemic diseases

We examined associations between 21 orofacial diseases and 9 major systemic diseases using logistic regression via the logistf R package (version 1.26.0) with Firth bias correction⁵⁷ to account for any bias arising from case-control imbalances. We estimated adjusted odds ratios that incorporated a unified set of covariates considered key factors for each disease. The specific covariates used in each model are detailed in Supplementary Data 34.

Genetic correlation analysis between orofacial diseases and systemic diseases

We performed LD score regression using LDSC version 1.0.1 to estimate SNP-based heritability and genetic correlation between 21 orofacial diseases and 9 systemic diseases^58,59. To quantify the proportion of phenotypic correlation attributable to shared genetic factors, we applied the following formulas as defined in Eqs. (3)–(5).

$$\mathrm{Phenotypic}\mathrm{correlation}{r}_p=\frac{\mathrm{Cov}\left(P_1,P_2\right)}{\sqrt{\mathrm{Var}\left(P_1\right)\mathrm{Var}\left(P_2\right)}}$$ 3

$$\mathrm{Genetic}\mathrm{contribution}\mathrm{to}{r}_p=r_g\cdot \sqrt{h_{g,1}^2{h}_{g,2}^2}$$ 4

$$\mathrm{Genetic}\mathrm{contribution}\mathrm{proportion}=\frac{r_g\cdot \sqrt{h_{g,1}^2{h}_{g,2}^2}}{r_p}$$ 5

In Eq. (3), $P_1$ and $P_2$ denote the two phenotypes. In Eq. (4), $r_g$ represents the genetic correlation, and $h_{g,1}^2$ and $h_{g,2}^2$ refer to the SNP-based heritabilities of the two traits, all calculated on the observed scale for binary outcomes⁶⁰.

Prediction of mRNA secondary structure

We used RNAfold (in the ViennaRNA package version 2.6.3)⁶¹ to predict changes in mRNA secondary structure caused by the target SNP (rs16939660). For this analysis, we assumed all other bases in the relevant 1,557-base pair exon matched the reference sequence, substituting only the target SNP from T to C. We then compared the predicted secondary structures before and after the substitution to assess the structural impact of the target SNP.

Generation of an ALDH1A2 mutant cell line with rs16939660

We began by designing a pegRNA sequence using DeepPrime⁶². The top-scoring DeepPrime result suggested a guide sequence of 5’-GAAACCTTTCGATATTACGC-3’ and a PBS-RTT sequence of 5’-CCTTTCGATATTACGCGGGCTGGGCT-3’ (Fig. 3a). Therefore, we cloned the gRNA, gRNA scaffold, and PBS-RTT sequence into the pU6-pegRNA-GG-acceptor vector (#132777, Addgene). To enhance editing efficiency, we introduced an additional nick approximately 90 bp downstream of the pegRNA-induced nick using the pRG2 plasmid (#104174, Addgene) with the guide sequence 5’- GGGCCAAAGCGCATTTCTGG-3’.

HEK293T cells (ATCC CRL-3216; human embryonic kidney, female origin) were cultured in DMEM (#11995-065, Gibco) supplemented with 10% fetal bovine serum (#26140079, Gibco) and 1% penicillin/streptomycin (#15140-122, Gibco) and then seeded into a 6-well plate one day prior to transfection. We transfected a total of 4 μg of pCMV-PEmax-P2A-BSD (#174821, Addgene), 1 μg of the pegRNA vector, and 1 μg of the pRG2 vector using the X-tremeGENE™ HP DNA Transfection Reagent (XTGHP-RO, Roche) according to the manufacturer’s protocol. Four days post-transfection, we trypsinized the cells and seeded them into a 96-well plate via serial dilution to achieve one cell per well. After monitoring the wells for the presence of single-cell clones, we expanded only verified single clones. After transferring the clones to a 24-well plate, we extracted genomic DNA using the LaboPass™ DNA Purification Kit (CME0112, Cosmogenetech). We performed PCR using forward (5’-GTGGTTACTGGAAGCACAGGA-3’) and reverse primers (5’-TCATACCTACCCCAGCACCT-3’) and purified the amplified DNA using the QIAquick PCR Purification Kit (#28106, Qiagen). Finally, we performed Sanger sequencing to confirm the introduction of the target mutation. The engineered ALDH1A2 mutant cell lines generated in this study are available from the corresponding author upon request for non-commercial academic research purposes, subject to institutional approval and applicable material transfer agreements.

RNA isolation and quantitative real-time RT-PCR

After extracting total RNA from cells using the RNeasy Mini Kit (#74104, Qiagen) according to the manufacturer’s protocol, we performed reverse transcription (RT) using 2 μg of RNA and oligo(dT)₁₈ primers with the RevertAid RT Kit (EP0441, Thermo Fisher Scientific). We then performed quantitative real-time RT-PCR (qPCR) using the SensiFAST SYBR Hi-ROX Kit (BIO-92020, Bioline) on the QuantStudio 3 Real-Time PCR System (Applied Biosystems) with specific primers for ALDH1A2 (forward: 5’-CTTTGACCCCACCACTGAGC-3’, reverse: 5’-CGTTGGAAAACACTGTGGGC-3’) and the housekeeping gene GAPDH (forward: 5’-GAGTCAACGGATTTGGTCGT-3’, reverse: 5’-TTGATTTTGGAGGGATCTCG-3’), normalizing the results to GAPDH expression.

Western blot

After lysing cells in RIPA buffer (50 mM Tris-HCl, pH 7.6, 150 mM NaCl, 1% Triton X-100, 1% sodium deoxycholate, 0.1% SDS, 2 mM EDTA) supplemented with a 1x protease inhibitor cocktail (cOmplete Mini, EDTA-free, 11836170001, Sigma-Aldrich), we measured protein concentrations in the lysates using a BCA assay kit. The protein samples were prepared by mixing the lysates with 4x Laemmli Sample Buffer (#1610747, Bio-Rad) supplemented with 10% β-mercaptoethanol (M6250, Sigma-Aldrich), followed by incubation at 95 °C for 5 minutes. After loading and separating equal amounts of protein (20 μg) on 8% SDS-polyacrylamide gels and transferring to Immobilon-P PVDF membranes (IPVH00010, Merck Millipore), we blocked the membranes for 1 hour in 5% skim milk in 1x Phosphate-Buffered Saline (PBS) with 0.1% Tween 20 (PBST). We then incubated them overnight at 4 °C with primary antibodies against ALDH1A2 (rabbit polyclonal, 1:1000, #83805, Lot# 1, Cell Signaling Technology) or α-tubulin (1:5000, #12G10, Developmental Studies Hybridoma Bank) diluted in 1x PBS. After three 30 min washes with 1 × PBST, membranes were incubated with HRP-conjugated secondary antibodies (goat anti-mouse IgG (whole molecule)–peroxidase, 1:10000, A4416, Sigma-Aldrich; goat anti-rabbit IgG (whole molecule)–peroxidase, 1:5000, A6154, Sigma-Aldrich) for 1 h at room temperature. Then, after three additional 30 min washes with 1 x PBST, we visualized the protein bands using the ChemiDoc MP Imaging System (Bio-Rad).

In vitro translation assay

We constructed the ALDH1A2 expression vector by PCR amplification of the ALDH1A2 coding sequence from wild-type and mutant HEK293T cells, followed by cloning into the pcDNA3.1(+) vector using its EcoRI and NotI restriction sites. We then used the MEGAscript™ T7 Transcription Kit (AM1334, Thermo Fisher Scientific) according to the manufacturer’s protocol to generate ALDH1A2 transcripts from the linearized expression vector after digesting it with NotI. Then, we purified the transcripts using the Invitrogen™ MEGAclear™ Transcription Clean-Up Kit (AM1908, Thermo Fisher Scientific) according to the manufacturer’s protocol.

Using the Retic Lysate IVT™ Kit (AM1200, Thermo Fisher Scientific) according to the manufacturer’s protocol, we then performed in vitro translation using equal amounts (2.5 μg) of wild-type and mutant ALDH1A2 transcripts. In brief, this meant incubating a mixture of RNA template and Retic Lysate supplemented with L-methionine (final concentration: 50 μM, M5308, Sigma-Aldrich) at 30 °C for 75 min and then on ice for 5 minutes. We then prepared protein samples by mixing the lysates with 4x Laemmli Sample Buffer (#1610747, Bio-Rad) supplemented with 10% β-mercaptoethanol (M6250, Sigma-Aldrich), followed by incubation at 95 °C for 5 minutes. We then loaded the entire protein samples onto gels for western blotting using primary antibodies against ALDH1A2 (rabbit polyclonal, 1:1000, #83805, Lot# 1, Cell Signaling Technology).

Patient cohorts and sample collection

A total of nine individuals (six patients with periodontitis and three healthy controls) were included in the single-cell RNA sequencing (scRNA-seq) analysis. Participants comprised five females and four males, aged 24–74 years. Gingival tissues from one patient with periodontitis and three healthy controls were directly processed for scRNA-seq, and the resulting data were integrated with publicly available scRNA-seq datasets from patients with periodontitis (GSE164241; https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE164241). For the immunohistological analysis, gingival tissue samples were obtained from two patients with periodontitis and two healthy controls (two females and two males, aged 26–74 years). Tissues were collected during resective periodontal flap surgery for patients with periodontitis and during crown-lengthening procedures for healthy controls. Samples were transferred to the laboratory within one hour of surgery. After rinsing in sterile phosphate-buffered saline (PBS), tissues were either stored at − 80 °C for subsequent scRNA-seq analysis or fixed in 4% paraformaldehyde for immunohistological analysis.

For the GCF analysis, samples were obtained from 12 individuals (six patients with periodontitis and six healthy controls; six females and six males; age range, 24–68 years). GCF was collected by inserting sterile paper points into the gingival sulcus for 20 seconds at the site with the deepest probing pocket depth.

Sex was recorded based on self-report in the clinical records at the time of enrollment. Sex was not used as a stratification variable in downstream analyses owing to the limited sample size in functional experiments. No financial compensation was provided to participants. All samples were used exclusively for the analyses described in this study, and no remaining material is available for further distribution.

Single-cell RNA sequencing

Library preparation

We processed gingival tissues designated for scRNA-seq using the Multi Tissue Dissociation Kit 2 (#130-110-203, Miltenyi Biotec) to generate a single-cell suspension according to the manufacturer’s instructions. After removing red blood cells with Red Blood Cell Lysis Solution (#130-094-183, Miltenyi Biotec), we filtered the resulting cell suspension through a 70 µm filter (#130-110-916, Miltenyi Biotec). Then, we loaded this single-cell suspension onto a 10X Chromium Controller (10X Genomics) for partitioning and barcoding of individual RNA transcripts using the Chromium Next GEM Single Cell 3’ Library Kit v3.1 (#1000121, 10X Genomics). We diluted approximately 10,000 cells in nuclease-free water, mixed them with the master mix, and co-loaded them with Single Cell 3’ v3.1 gel beads and partitioning oil onto a Chromium Next GEM chip G. This allowed the droplets containing individual cells to undergo reverse transcription, generating uniquely barcoded cDNAs. We then constructed the cDNA libraries according to the 10X Genomics protocol. This protocol includes steps for end-repair, single ‘A’ base addition, adapter ligation, and PCR amplification. After assessing the libraries for quality using a 4200 TapeStation (Agilent Technologies) and quantity via qPCR, we subjected them to sequencing on a HiSeq X system (Illumina) with the read configuration recommended by 10X Genomics.

Read alignment and quality control

We processed the raw sequencing data with CellRanger v6.1.2 (10X Genomics), aligning the reads to the ENSEMBL GRCh38 human transcriptome and generating raw gene expression matrices. We then carried out subsequent analyses in Seurat (v4.4.0)⁶³. We filtered low-quality or dying cells with fewer than 200 or more than 6000 genes detected, as well as those with over 25% mitochondrial gene content. We identified and removed putative doublet cells with DoubletFinder (v2.0)⁶⁴ and then performed normalization and variance stabilization with SCTransform (v0.4.1). Then, we addressed batch effects and multiple sample integration using the Harmony package (v1.0)⁶⁵.

Clustering, visualization, and annotation

Following data integration, dimensionality reduction was performed using Uniform Manifold Approximation and Projection (UMAP) via the RunUMAP function in Seurat (v4.4.0) using default settings⁶³. We used Seurat’s graph-based clustering methodology to identify clusters and FeaturePlot to visualize marker gene expression. A subsequent analysis of known marker gene expression with reference to the relevant literature allowed us to allocate biological annotations.

Immunohistological analysis

After fixing the tissue samples in 4% paraformaldehyde at 4 °C for 12 h and then washing them with PBS, we processed them using an open-type autotechnicon for paraffin embedding and sequential dehydration. We then sectioned the paraffin-embedded tissues at 5 μm thickness using a microtome, trimming the blocks to 30 μm as needed. We immersed the resulting sections in 10% EDTA under cold conditions, floated them on xylene, and transferred them to a 45 °C distilled water bath before mounting them on glass slides and drying the slides on a 45 °C slide warmer. For immunohistochemistry, we preheated the slides in an oven at 55 °C for 30 minutes and then equilibrated them at room temperature for 10 min. We deparaffinized the slides with three changes of xylene for 5 min each, rehydrated them through a graded ethanol series, and then immersed them in distilled water for 15–30 minutes. We performed antigen retrieval using a 10-minute incubation in preheated 95–100 °C citrate buffer (#C9999, Sigma-Aldrich). After cooling the slides for 10 minutes at room temperature, we washed them with PBST three times for 2 min each. Then, we used an ImmEdge pen (#H-4000, Vector Laboratories) to delineate a hydrophobic barrier around the sections on each slide. We permeabilized the sections with 0.1% Triton X-100 (T8787, Sigma Aldrich) in PBS for 10 min. After washing the slides with PBST, we blocked the sections with 5% skim milk (#232100, DIFCO) in PBST for 1 hour at room temperature. We then incubated the slides overnight at 4 °C in primary antibodies against ALDH1A2 (rabbit polyclonal, #83805, Lot# 1, Cell Signaling Technology) and CD31/PECAM-1 (mouse monoclonal, #SC-376764, H-3, Lot# F0721, Santa Cruz Biotechnology) at dilutions of 1:500 and 1:200, respectively. The following day, after washing the slides with PBST, we incubated them at room temperature for 2 hours in the dark in anti-mouse (donkey polyclonal, #A21203, Lot# 1918277, Invitrogen) and anti-rabbit (goat polyclonal, #A11008, Lot# 2179202, Invitrogen) secondary antibodies at a dilution of 1:1000. Then, we incubated the slides for 30 min in DAPI (#10236276001, Lot# 77788321, Roche), diluted 1:200. After applying xylene around the tissue sections, we mounted coverslips using EcoMount (#EM897L, Biocare Medical) and allowed the slides to dry in a sterile environment until imaging.

A total of nine ROI-level fluorescence intensity measurements were obtained for each group (HC, n = 9; PD, n = 9). These ROI values represent technical replicates derived from two biological donors per group (two images per donor, multiple ROIs per image). Statistical comparisons were performed at the ROI level using an unpaired two-sided t test (t = 3.194, df = 16, P = 0.0056), with mean difference − 873.0 ± 273.3 and 95% CI − 1452 to −293.6. Data are presented as mean ± SEM.

Quantifying retinoic acid in human samples

Each sample comprised three paper points used to absorb GCF, which we eluted individually with 200 µL of 0.1% DPBST (DPBS with 0.1% Tween 20) and then agitated on a tabletop vortexer at room temperature for 30 min before meticulously extracting the paper point from the tubes. We then measured GCF retinoic acid concentrations using the human retinoic acid ELISA kit (#CSB-E16712h, Cusabio) according to the manufacturer’s instructions.

CPASSOC result analysis and pathway enrichment analysis

We conducted a pairwise cross-phenotype meta-analysis using CPASSOC (v1.01) software and the S_Het statistic, which applies a sample size-weighted fixed-effects meta-analysis across traits. Given the expected heterogeneity across the phenotypes included in our analysis, we chose S_Het over S_Hom, as it offers higher power in the presence of trait-specific effect size variability.

Next, we mapped the SNPs we identified in our CPASSOC analysis (P_Het < 5 × 10⁻⁸) and that were suggestive in the trait-specific GWAS (P < 1 × 10⁻³) to specific genes. We did this via eQTL mapping using FUMA (v1.8.0) and incorporating GTEx v8 expression profiles across all available tissues. Then, we subjected the resulting gene list to pathway enrichment analysis using the gprofiler2 R package v0.2.3, querying the GO Biological Processes, KEGG, and Reactome databases. In this analysis, we assessed statistical significance using the g:SCS multiple testing correction method, considering pathways with an FDR-adjusted P-value < 0.05 as exhibiting significant enrichment.

Statistics & reproducibility

All statistical analyses were performed using R (v4.3.3) and the software packages described in the Methods. All tests were two-sided unless otherwise specified. No statistical method was used to predetermine sample size for in vitro experiments. Sample sizes for functional assays were based on standard practice in the field. No data were excluded from the analyses. Sample sizes for the GWAS and MR analyses were determined by the number of eligible participants available in the respective datasets. Statistical power was calculated for both GWAS and MR analyses as described in the Methods. The experiments were not randomized, and the investigators were not blinded to allocation during experiments and outcome assessment. Biological replicates and sample sizes for each experiment are indicated in the corresponding figure legends. Genetic variants were excluded only according to predefined quality control criteria and instrument selection procedures described in the Methods.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

K.N. performed the genome-wide and exome-wide association studies using UK Biobank and FinnGen data, as well as the Mendelian randomization analyses between orofacial and systemic diseases. B.S.E. conducted the single-cell RNA sequencing analyses and the quantification of retinoic acid levels. J.Y.K. carried out the transcriptome-wide association studies, the eQTL-based Mendelian randomization, and the pathway enrichment analyses. M.G.L. generated the CRISPR-Cas9-edited mutant cell lines and performed various in vitro functional experiments. K.L. and J.-M.L. contributed to data processing and statistical analysis. J.-K.C. performed the single-cell RNA sequencing of patient gingival tissues. K.N. and B.S.E. wrote the initial draft of manuscript. S.-J.K. wrote the final manuscript and supervised all aspects of the study. J.M.K., S.L., and S.-J.K. conceptualized the study, oversaw the analyses, and interpreted the results. All authors reviewed and approved the final manuscript.

Peer review

Peer review information

Nature Communications thanks Qing Li, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Data availability

The GWAS summary statistics generated in this study have been deposited in Zenodo under the identifier 10.5281/zenodo.18707014 [10.5281/zenodo.18707014] and are also available in the GWAS Catalog under accession numbers GCST90837174–GCST90837203 [https://www.ebi.ac.uk/gwas/]. The full list of GWAS Catalog accession numbers, together with their corresponding traits and links, is provided in Supplementary Data 35. The single-cell RNA sequencing data generated in this study (raw and processed) are available in the Gene Expression Omnibus (GEO) under accession number GSE262668. Publicly available single-cell RNA sequencing data used in this study are available under accession number GSE164241. UK Biobank data were accessed under application number 45227 and are available under controlled access due to participant privacy and data protection regulations. Access requires submission of an application through the UK Biobank Access Management System [https://www.ukbiobank.ac.uk] and is subject to approval and the terms of the UK Biobank Data Access Agreement. Publicly available summary statistics from the FinnGen study (release 12) used for replication analyses are accessible via the FinnGen consortium website [https://www.finngen.fi] in accordance with their data access policies. Access to individual-level data is subject to separate application and approval by FinnGen. All source data underlying the figures are provided with this paper.

Code availability

The code used to perform the analyses and generate the results in this study is publicly available in Zenodo under the Creative Commons Attribution 4.0 International (CC-BY 4.0) license: 10.5281/zenodo.17773337 [10.5281/zenodo.17773337]⁶⁶. The specific version associated with this publication is v1.0.0.

Competing interests

S.-J.K., J.-M.K., and S.G.L. are listed as inventors on a Korean patent application (application number 10-2025-0123762) filed by Seoul National University R&DB Foundation, relating to the use of retinoic acid as a biomarker for inflammatory oral diseases. The patent application is pending. The application was filed after the initial submission of this manuscript. The other authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-73925-0.