The causal relationship between diabetes mellitus and hashimoto thyroiditis

Abstract

This study aims to investigate the genetic correlation and potential causal relationship between diabetes mellitus (DM) and Hashimoto thyroiditis (HT), and to assess the effects of type 1 diabetes mellitus (T1DM) and type 2 diabetes mellitus (T2DM) on HT risk, with the goal of informing thyroid-function monitoring in patients with DM and improving HT management. We used linkage disequilibrium score regression (LDSC) to estimate the genetic correlation between DM and HT. We then performed two-sample Mendelian randomization (TSMR) using data from genome-wide association studies (GWAS) to evaluate potential causal effects of DM subtypes on HT. We selected single-nucleotide polymorphisms (SNPs) as instrumental variables (IVs) and applied the inverse-variance weighted (IVW) method along with complementary sensitivity analyses to estimate causal effects. We assessed heterogeneity using Cochran’s Q test and identified outlier SNPs with MR-PRESSO (Mendelian Randomization Pleiotropy RESidual Sum and Outlier). We used the MR-Egger intercept to test for directional pleiotropy. Finally, we performed a leave-one-out analysis to evaluate whether any single SNP unduly influenced the MR estimates. LDSC showed a significant negative genetic correlation between DM and HT (rG = −0.3285, SE = 0.062, P = 1.20 × 10⁻⁷), with evidence of nonzero SNP heritability (h² P < .005). In two-sample MR analyses of DM overall, the IVW estimate suggested increased HT risk in UK Biobank (OR = 7.26, 95% CI = 1.95–27.11, P < .05) and was directionally consistent in FinnGen (OR = 1.08, 95% CI = 1.02–1.14, P < .05). Subtype-specific MR analyses supported a causal effect of T2DM on HT (OR = 1.10, 95% CI = 1.05–1.16, P < .05), whereas there was no evidence for a causal effect of T1DM (OR = 1.04, 95% CI = 0.97–1.12, P > .05). We provide genetic support for a causal relationship between T2DM liability and higher HT risk, with no evidence of a direct causal effect for T1DM. These findings motivate periodic thyroid-function assessment and integrated endocrine–metabolic care in T2DM, and warrant multi-ancestry and mechanistic studies to refine risk prediction and identify actionable biology.

1. Introduction

Hashimoto thyroiditis (HT), also referred to as chronic lymphocytic thyroiditis or autoimmune thyroiditis, is the most prevalent autoimmune thyroid disorder in iodine-sufficient regions worldwide.^[1,2] The incidence of HT has increased worldwide,^[3] with females affected 15–20 times more frequently than males. Current management of HT primarily involves levothyroxine (L-T4) replacement to control symptoms of hypothyroidism. Further research into the etiology of HT may enable the development of new therapeutic approaches.

DM and thyroid disorders are increasingly recognized to co-occur in both epidemiological studies and clinical practice.^[4,5] In T2DM, central obesity and insulin resistance – assessed using the triglyceride–glucose (TyG) index and related metrics – are strongly associated with increased HT risk, suggesting that metabolic disturbances may contribute to HT pathogenesis through inflammatory and immune mechanisms.^[6] Similarly, in T1DM, ~33.8% of patients have additional autoimmune diseases; HT is the most common, affecting up to 28.8%, particularly among women.^[4] Consistent with these findings, individuals with T2DM who are positive for glutamic acid decarboxylase 65 antibodies (GAD65Ab) show a higher likelihood of thyroid autoimmunity.^[7] Moreover, diabetic ketoacidosis may transiently disrupt thyroid function, with hormone abnormalities normalizing after metabolic control.^[8] Collectively, available evidence supports a strong epidemiological and clinical association between DM and HT; however, causality remains unestablished, underscoring the need for further investigation to inform prevention, diagnosis and treatment strategies for HT.

DM and HT may be genetically correlated. Genetic correlation refers to the association between phenotypes within a population that arises from shared genetic factors. LDSC is widely used to quantify genetic relationships among complex traits. It helps distinguish inflation in GWAS test statistics due to confounding from that due to polygenic effects.^[9] Mendelian Randomization (MR),^[10,11]analogous to a randomized controlled trial (RCT), uses germline genetic variants as instrumental variables to infer causal relationships between phenotypes. This approach has been widely applied to investigate associations between diverse risk factors and HT.^[12–15] Because genetic variants are fixed at conception and are typically independent of many environmental confounders, MR can strengthen causal inference and is less susceptible to confounding and reverse causation than conventional observational studies.^[16]

Recent MR studies have examined vascular and blood-pressure outcomes in autoimmune diseases, including rheumatoid arthritis and inflammatory bowel disease, and have explored broader autoimmune profiles in relation to hypertensive disorders.^[17–19] However, causal evidence for endocrine comorbidity between diabetes and autoimmune thyroiditis remains limited, and few studies have explicitly contrasted T1DM and T2DM within a unified MR framework. Here, we evaluate overall and subtype-specific effects of DM (T1DM vs T2DM) on HT, integrate LDSC with two-sample MR to triangulate shared genetic architecture and causality, and validate findings across independent large-scale GWAS resources, including UK Biobank and FinnGen, using comprehensive sensitivity analyses.

We first applied LDSC to assess the genetic correlation between DM and HT. We then conducted two-sample MR using large-scale GWAS summary statistics to evaluate the potential causal effects of DM subtypes on HT. To ensure robustness, we performed multiple sensitivity analyses, including MR-Egger regression, Cochran’s Q test, MR-PRESSO outlier detection and leave-one-out analyses. Integrating these genetic approaches, we aimed to clarify the causal landscape of DM–HT comorbidity and to inform future strategies for early detection and prevention of autoimmune thyroid disease.

2. Materials and methods

2.1. Research principles

The basic principle of MR is to use genetic variants associated with the exposure as IVs to infer whether the exposure has a causal effect on the outcome. In brief, MR typically involves obtaining GWAS summary statistics, selecting valid SNP instruments, estimating causal effects using appropriate models, and performing sensitivity and quality-control analyses to assess robustness. Because alleles are randomly assorted at meiosis and fixed before disease onset, MR can reduce bias from confounding and reverse causation that often affect observational associations.

The validity of MR depends on three core assumptions. First, the IVs must be robustly associated with the exposure (relevance). Second, the IVs must be independent of confounders of the exposure–outcome relationship (independence). Third, the IVs must affect the outcome only through the exposure and not via alternative pathways (exclusion restriction; Fig. 1). Together, these assumptions reduce bias and enable causal interpretation of the MR estimates.

Figure 1.

Research process.

2.2. Data sources

Diabetes GWAS summary statistics were obtained from UK Biobank and FinnGen (Table 1). The UK Biobank GWAS (ukb-b-10753) comprised 22,340 cases and 439,238 controls (n = 461,578) with 9851,867 SNPs. The FinnGen GWAS (finn-b-E4_DIABETES) included 35,607 cases and 183,185 controls (n = 218,792) with 16,380,466 SNPs. All GWAS were restricted to participants of European ancestry (UK Biobank European-ancestry subset; FinnGen participants were Finnish). GWAS summary statistics for DM subtypes were also obtained from FinnGen. The T1DM GWAS (finngen_R10_T1D) comprised 4320 cases and 335,112 controls (n = 339,432) with 19,344,249 SNPs. The T2DM GWAS (finngen_R10_T2D) comprised 65,085 cases and 335,112 controls (n = 400,197) with 19,345,488 SNPs. HT GWAS summary statistics were obtained from IEU OpenGWAS (GCST90018855). This GWAS included 15,654 cases and 379,986 controls (n = 395,640) with 24,146,037 SNPs. This HT GWAS was based on participants of European ancestry.

Table 1.

Data Sources.

Database	Phenotype	Sample Size (n)	Participants		SNPs (n)	ID
			Case (n)	Control (n)
UK Biobank	DM	461,578	22,340	439,238	9,851,867	ukb-b-10753
FinnGen	DM	218,792	35,607	183,185	16,380,466	finn-b-E4_DIABETES
FinnGen	T1DM	339,432	4320	335,112	19,344,249	finngen_R10_T1D
FinnGen	T2DM	400,197	65,085	335,112	19,345,488	finngen_R10_T2D
IEU Open GWAS	HT	395,640	15,654	379,986	24,146,037	GCST90018855

2.3. LDSC

We used LDSC (v1.0.1) to estimate the genetic correlation between DM and HT. LDSC estimates genetic correlation from GWAS summary statistics by leveraging per-variant association signals and their linkage disequilibrium (LD) scores.^[20] We first harmonized GWAS summary statistics using munge_sumstats.py, aligning effect and non-effect alleles across datasets. We then ran ldsc.py with LD scores precomputed from the European panel of the 1000 Genomes Project to estimate the genetic correlation between DM and HT. The entire analysis was conducted using the tool available at https://github.com/bulik/ldsc.

2.4. SNP selection and evaluation

To ensure valid IVs, the three core assumptions of MR must be satisfied. We derived genetic instruments from the aforementioned GWAS by selecting independent genome-wide significant variants for each exposure of interest (P < 5 × 10⁻⁸). After excluding palindromic SNPs, we performed LD clumping (r² < 0.001) using the European (EUR) reference panel. We screened candidate SNPs for potential pleiotropy using phenotype look-up resources and excluded variants showing genome-wide significant associations with potential confounders or with the outcome (for example, thyroid traits or autoimmune phenotypes) to better satisfy the independence and exclusion-restriction assumptions. We applied Steiger filtering to confirm the causal direction by retaining instruments that explained more variance in the exposure (DM traits) than in the outcome (HT), thereby reducing the risk of reverse causation^[21] (Fig. 1, Assumption 2). We calculated the F statistic for each remaining SNP (F = β²/SE²) and summarized instrument strength across SNPs. We excluded SNPs with weak instrument strength (F < 10)..

2.5. Statistical analysis

All statistical analyses and data visualizations were performed using R version 4.1.2 (https://www.r-project.org/). We conducted the TSMR study using the R packages “Two Sample MR”^[22] and “Mendelian Randomization.”^[23] The inverse variance weighted (IVW) method^[24] was the primary approach used to estimate the association between DM and HT. In addition, MR-Egger and weighted median methods^[25] were applied as complementary analyses.The IVW method provides accurate estimates when the assumption that all included SNPs serve as valid IVs is met.^[26] MR-Egger regression can detect and adjust for pleiotropy, while the weighted median method provides reliable estimates under the assumption that at least 50% of the IVs are valid.^[27] Harmonization of exposure and outcome summary statistics was performed prior to MR analyses. We aligned SNP effect alleles between datasets, flipped the sign of beta estimates when the effect allele was reversed, and removed variants with incompatible alleles. Palindromic SNPs with ambiguous strand orientation were excluded. Only SNPs available in both exposure and outcome datasets after harmonization were retained for downstream analyses.

2.6. Quality control

We used MR-PRESSO to detect outlier instruments and obtain outlier-corrected IVW estimates (implemented in the MRPRESSO R package). A significant MR-PRESSO global test (P < .05) suggests potential horizontal pleiotropy, indicating that the IVW estimates may be biased. We assessed heterogeneity among SNP-specific causal estimates using Cochran’s Q test. A nonsignificant Q statistic (P > .05) was interpreted as limited evidence of heterogeneity.^[28] We performed leave-one-out analyses to identify influential SNPs and assess the robustness of the MR estimates.

3. Results

3.1. LDSC analysis

Twin studies have provided epidemiological evidence for genetic susceptibility to HT.^[29] Concordance for thyroid autoantibodies is higher in monozygotic than in dizygotic twins (80% vs 40%). In addition to twin studies, we used LDSC to estimate the LD score regression intercept, observed-scale SNP heritability (h²), and genetic correlation (rG) between the DM GWAS from UK Biobank and the HT GWAS from IEU OpenGWAS. Figure 2 summarizes the LDSC results for DM and HT. In this analysis, rG provides a genome-wide measure of shared genetic architecture between DM and HT. LDSC identified a significant negative genetic correlation between DM and HT (rG = −0.3285, SE = 0.062, P = 1.20 × 10⁻⁷), with evidence of nonzero SNP heritability (h² P < .005).

Figure 2.

The bar chart displays the mean_chisq, lambda_gc, and intercept values for each trait in the provided data. Each color represents a different measurement metric, enabling a visual comparison of the differences in these three indicators between DM and HT.

To further assess potential causal relationships between genetic liability to DM and HT, we next performed MR analyses. Using genetic variants as IVs, we assessed the causal effects of DM subtypes (T1DM and T2DM) on HT.

3.2. MR analysis of DM and HT

Instrument selection and key MR results assessing the causal effect of DM on HT in the UK Biobank discovery dataset and the FinnGen validation dataset are summarized in Table 2 and Figure 3A. In UK Biobank, 67 independent SNPs were retained as instruments (mean F = 75.40, all F-statistics > 10, range: 29.965–898.776) (Table S1, Supplemental Digital Content 1). The primary IVW analysis supported a positive causal effect of DM on HT (OR = 7.26, 95% CI 1.95–27.11; P = .00317). Sensitivity estimators were less precise and not statistically significant: MR-Egger OR = 0.55 (95% CI 0.03–11.67; P = .704),and weighted median OR = 1.35 (95% CI 0.23–7.84; P = .738). Cochran’s Q test suggested heterogeneity (P < .05); therefore, we applied MR-PRESSO, which produced a similar corrected estimate (OR = 6.56, 95% CI 1.92–22.40; P < .05). In FinnGen, 54 SNPs were retained (mean F-statistic = 79.84, all F-statistics > 10; range 29.933–586.347) (Table S2, Supplemental Digital Content 2), and IVW again indicated an increased risk of HT associated with genetically predicted DM (OR = 1.08, 95% CI 1.02–1.14; P = .00480). Consistent but nonsignificant results were observed with alternative methods: MR-Egger OR = 1.02 (95% CI 0.93–1.12; P = .696), and weighted median OR = 1.07 (95% CI 0.98–1.16; P = .125). No heterogeneity was detected in FinnGen (Q P > .05); thus, MR-PRESSO was not performed. Scatter plots further showed concordant positive slopes across methods (Fig. 4A, 4B and 4C). Across both datasets, leave-one-out analyses indicated that sequential removal of any single SNP did not materially change the IVW estimates (Fig. 5A, 5B and 5C), supporting the robustness of the association between DM and HT.

Table 2.

Summary of instrument selection and Mendelian randomization results for the causal effect of diabetes mellitus on Hashimoto thyroiditis in UK Biobank and FinnGen.

Dataset	GWAS ID	IVs retained (n)	Mean F (range)	IVW OR (95% CI), P	Heterogeneity (Q P)	MR-PRESSO OR (95% CI), P
UK Biobank (DM)	ukb-b-10753	67	75.40 (29.965–898.776)	7.26 (1.95–27.11), P < .05	P < .05	6.56 (1.92–22.40), P < .05
FinnGen (DM)	finn-b-E4_DIABETES	54	79.84 (29.933–586.347)	1.08 (1.02–1.14), P < .05	P > .05	-
FinnGen (T2DM)	finngen_R10_T2D	170	68.87 (29.762–1004.781)	1.10 (1.05–1.16), P < .05	P < .05	1.09 (1.06–1.17), P < .05

Figure 3.

Results of MR analysis. (A) MR analysis of DM and HT; (B) MR analysis of different types of DM and HT.

Figure 4.

Scatter plot of MR analysis. (A) Scatter plot of MR analysis between UKB DM datasets and HT; (B) Scatter plot of MR analysis between Finn DM datasets and HT; (C) Scatter plot of MR analysis between T2DM datasets and HT.

Figure 5.

Leave-one-out analysis. (A) Leave-one-out analysis of MR analysis between UKB DM datasets and HT; (B) Leave-one-out analysis of MR analysis between Finn DM datasets and HT; (C) Leave-one-out analysis of MR analysis between T2DM datasets and HT.

Notably, the estimated ORs differed in magnitude between UK Biobank and FinnGen. This does not necessarily indicate conflicting biological effects: for a binary exposure, two-sample MR using GWAS summary statistics typically yields a causal estimate on the underlying liability (log-odds) scale, and the scaling of SNP–exposure associations can vary across cohorts and analysis pipelines. In our study, the UK Biobank estimate was also less precise with substantial heterogeneity and a wider confidence interval, whereas FinnGen provided a more precise estimate. Therefore, we emphasize the direction and consistency of evidence across datasets and sensitivity analyses.

3.3. MR analysis of type 1 and type 2 diabetes with HT

Building on the overall evidence for a causal association between DM and HT, we performed subtype-stratified analyses to assess whether the association differed between T1DM and T2DM.

We excluded SNPs in LD or with palindromic alleles from the FinnGen T2DM dataset and removed seven variants associated with potential confounders (rs12440774, rs12507026, rs12534035, rs340882, rs4325676, rs62448618 and rs703972). In total, 170 SNPs were retained as IVs for MR analyses (Table S3, Supplemental Digital Content 3). All instruments were strong (mean F-statistic = 68.87; all F-statistics > 10), indicating no bias from weak instrument associations. IVW analyses supported a positive association between T2DM and HT risk (OR = 1.10, 95% CI = 1.05–1.16, P < .05), as shown in Figs. 3B and 4C. Cochran’s Q test indicated heterogeneity (P < .05), whereas the MR-Egger intercept provided no evidence of horizontal pleiotropy (P > .05). To address heterogeneity, we applied MR-PRESSO to detect outliers and obtain outlier-corrected IVW estimates. Outlier-corrected estimates remained positive (OR = 1.09, 95% CI = 1.06–1.17, P < .05). Leave-one-out analyses showed that sequential exclusion of individual SNPs did not materially alter the IVW estimates, indicating that no single SNP drove the overall association (Fig. 5C). Complementary analyses were directionally consistent (weighted median: OR = 1.03, 95% CI = 0.95–1.12, P = .495; MR-Egger: OR = 1.01, 95% CI = 0.90–1.12, P = .921). In contrast, two-sample MR analyses using FinnGen T1DM summary statistics showed no evidence for a causal effect of T1DM on HT (OR = 1.04, 95% CI = 0.97–1.12, P > .05).

4. Discussion

We combined LDSC and two-sample MR in discovery and validation analyses to examine genetic correlation and potential causal effects of DM subtypes on HT. The results were consistent with a causal effect of T2DM on increased HT risk. These findings support thyroid-function monitoring in individuals with DM and improve understanding of the genetic architecture and pathogenesis of HT, with potential implications for therapeutic development.

HT is an autoimmune endocrine disorder characterized by thyroid inflammation and elevated circulating thyroid autoantibodies.^[30] HT is a leading cause of hypothyroidism in regions with sufficient dietary iodine intake. In early stages, patients may be asymptomatic and show positivity for anti–thyroid peroxidase antibodies (TPOAb); later, hypothyroidism may develop.^[31,32] The incidence of HT ranges from 0.3 to 1.5 cases per 1000 person-years,^[33]with particularly high prevalence among women. More than 10% of women test positive for thyroid antibodies, and the prevalence of clinically overt disease exceeds 2%.^[34] Previous studies suggest that environmental exposures and genetic susceptibility may trigger immune responses against thyroid antigens, contributing to HT development.^[35–37]

As a polygenic disease, DM may share genetic risk factors with thyroid disorders, and the two conditions often co-occur clinically.^[38] Evidence suggests that autoimmune processes in T1DM may extend beyond pancreatic islets and contribute to HT risk.^[39–41] T1DM is characterized by autoimmune destruction of insulin-producing β-cells and is often associated with other autoimmune diseases, consistent with autoimmune polyendocrine syndromes.^[42]Studies report that 15–30% of individuals with T1DM develop autoimmune thyroid disease, including HT or Graves’ disease.^[43] Similarly, analyses of human leukocyte antigen (HLA) genes in families with autoimmune diabetes and thyroiditis have revealed shared susceptibility between T1DM and autoimmune thyroid diseases (AITD), including HT.^[44] Evidence also suggests an association and potential interaction between T2DM and HT.^[45] Parlițeanu and Cheța (2018) reported associations between clinical features of T2DM and HT, suggesting possible interaction, particularly through metabolic pathways. For example, studies in hospitalized endocrine and diabetes cohorts report that a substantial proportion of individuals with HT have disordered carbohydrate metabolism, often presenting as T2DM with increased insulin resistance.^[46]

Prior observational evidence on the DM–HT association is heterogeneous and varies by diabetes subtype. In T1DM, systematic reviews and meta-analyses consistently report an increased prevalence of thyroid autoimmunity and thyroid dysfunction compared with the general population, supporting the concept of autoimmune clustering driven by shared immune susceptibility.^[47,48] In contrast, studies in T2DM show more variable results: several clinic-based and population studies report higher rates of thyroid dysfunction and thyroid autoantibodies in T2DM than in controls, often enriched among women, individuals with longer diabetes duration, poorer glycemic control, or visceral obesity,^[49,50] whereas other cohorts report comparable prevalence to the general population or attenuation after accounting for metabolic confounders.^[50] Notably, a two-sided analysis suggested an asymmetric relationship in which AITD was more prevalent in T2DM cohorts, but T2DM was not necessarily more prevalent among patients with thyroid disease, highlighting possible surveillance bias and shared risk factors rather than symmetric causality.^[51] These consistent (T1DM) versus heterogeneous (T2DM) observational patterns motivated our genetic approach, and our MR results help reconcile prior discrepancies by reducing confounding and reverse causality.

A Mendelian randomization study evaluated causal relationships among thyroid-stimulating hormone (TSH), free thyroxine (fT4), deiodinases, insulin resistance, and T2DM, highlighting complex links between thyroid function and glucose metabolism.^[52]

Our LDSC analysis identified a significant genetic correlation between DM and HT, supporting further investigation of shared biological mechanisms. Building on this, two-sample MR analyses supported a causal effect of T2DM on increased HT risk, whereas there was no evidence for a causal effect of T1DM on HT. Together, these results suggest that T2DM and HT may share biological pathways that link metabolic dysfunction to thyroid autoimmunity.

To interpret these subtype-specific MR findings, it is important to distinguish immune-driven autoimmunity from metabolically driven inflammation. Our results suggest that genetic liability to T2DM (but not T1DM) is associated with a higher risk of HT, supporting a pathway whereby insulin resistance and related metabolic disturbances promote thyroid autoimmunity. T2DM is characterized by chronic low-grade systemic inflammation that is closely linked to visceral adiposity and insulin resistance, with increased circulating pro-inflammatory mediators (e.g., IL-6, TNF-alpha, and CRP) and adipokine dysregulation.^[53,54] These inflammatory cues can enhance antigen presentation and skew T-cell responses toward Th1/Th17 phenotypes, which are central to HT immunopathology. Leptin, an adipokine elevated in obesity and metabolic disease, has been implicated in HT by promoting Th17-related immune activity in thyroid tissue.^[55,56] Hyperglycemia-related oxidative stress and advanced glycation end products (AGEs) are also increased in HT and may further amplify inflammatory signaling and tissue injury, providing a plausible mechanistic bridge between dysglycemia/insulin resistance and thyroid autoimmunity.^[57]

By contrast, although T1DM and HT frequently co-occur clinically, much of this clustering is explained by shared immune-genetic susceptibility (notably HLA class II and other immune-regulatory loci) and autoimmune clustering, rather than a straightforward unidirectional causal effect of T1DM liability on HT.^[58] Consistent with this interpretation, our MR analysis using FinnGen T1DM instruments showed no evidence of a direct causal effect on HT. Because many strong T1DM variants lie in the HLA region, which is also implicated in thyroid autoimmunity, part of the epidemiologic association may reflect horizontal pleiotropy (shared autoimmunity) rather than mediation through the diabetic phenotype. In addition, the markedly smaller number of T1DM cases compared with T2DM in available GWAS reduces power to detect modest causal effects. Recent bidirectional MR work has reported asymmetric causal signals between T1DM and autoimmune thyroiditis, underscoring the complexity of shared autoimmunity and the need for further large-scale studies.^[59]

Our findings highlight the importance of thyroid-function monitoring in individuals with T2DM, particularly at diagnosis and during treatment. Clinicians should remain alert to changes in thyroid function when managing individuals with DM, particularly in those with risk factors for HT.

However, our study has certain limitations. Although we utilized large datasets and advanced statistical methods, further biological research is needed to elucidate the specific interaction mechanisms between T2DM and HT. Another limitation is potential participant/sample overlap between the exposure and outcome GWAS datasets (e.g., shared contributing studies or biobanks), which can induce correlation between SNP–exposure and SNP–outcome association estimates. In two-sample MR, such overlap can shift weak-instrument bias away from the null toward the confounded observational association and may inflate type I error as the overlap increases. Although we sought to minimize this risk by using largely independent data sources where possible and by applying strong instruments (all F-statistics > 10), we could not quantify the exact degree of overlap based on publicly available summary statistics. Future research should focus on these potential mechanisms and how this knowledge can be used to improve the clinical management and treatment of patients with DM and HT. In addition, because our exposure and outcome GWAS were mainly based on European-ancestry populations, the generalizability of our findings to other ancestries may be limited. Population differences in allele frequencies and linkage disequilibrium, as well as residual population stratification, could influence the MR estimates. Future multi-ancestry and ancestry-specific MR studies are needed to validate the transferability of our results. We also note that for binary exposures such as DM, MR effect estimates based on GWAS summary statistics are interpreted on an underlying liability scale and may not be directly comparable in magnitude across cohorts that differ in phenotype definition or scaling; thus, we focus on consistent direction and robustness across analyses.

In conclusion, our study provides genetic evidence that liability to T2DM increases the risk of Hashimoto thyroiditis, whereas no direct causal effect is observed for T1DM. Clinically, our findings support proactive thyroid risk awareness and periodic thyroid function assessment in individuals with T2DM, and encourage coordinated metabolic-endocrine management. Future multi-ancestry and mechanistic studies are warranted to refine risk stratification and identify actionable pathways.

Acknowledgments

We are grateful to the UK Biobank, the FinnGen database, and all participants in our study.

Author contributions

Formal analysis: Yanan Song, Yingchun He.

Investigation: Yanan Song, Yingchun He, Lingling Qin.

Methodology: Yanan Song.

Project administration: Yanan Song, Yingchun He, Tonghua Liu.

Supervision: Yanan Song, Tonghua Liu.

Writing – original draft: Yanan Song, Yingchun He.

Writing – review & editing: Yanan Song, Tonghua Liu.

Data curation: Yingchun He, Lingling Qin.

Funding acquisition: Yingchun He, Tonghua Liu.

Software: Yingchun He.

Conceptualization: Lingling Qin, Tonghua Liu.

Validation: Lingling Qin.