Smoking, use of smokeless tobacco, HLA genotypes and incidence of latent autoimmune diabetes in adults

Aims/hypotheses Smoking and use of smokeless tobacco (snus) are associated with an increased risk of type 2 diabetes. We investigated whether smoking and snus use increase the risk of latent autoimmune diabetes in adults (LADA) and elucidated potential interaction with HLA high-risk genotypes. Methods Analyses were based on Swedish case–control data (collected 2010–2019) with incident cases of LADA (n=593) and type 2 diabetes (n=2038), and 3036 controls, and Norwegian prospective data (collected 1984–2019) with incident cases of LADA (n=245) and type 2 diabetes (n=3726) during 1,696,503 person-years of follow-up. Pooled RRs with 95% CIs were estimated for smoking, and ORs for snus use (case–control data only). The interaction was assessed by attributable proportion (AP) due to interaction. A two-sample Mendelian randomisation (MR) study on smoking and LADA/type 2 diabetes was conducted based on summary statistics from genome-wide association studies. Results Smoking (RRpooled 1.30 [95% CI 1.06, 1.59] for current vs never) and snus use (OR 1.97 [95% CI 1.20, 3.24] for ≥15 box-years vs never use) were associated with an increased risk of LADA. Corresponding estimates for type 2 diabetes were 1.38 (95% CI 1.28, 1.49) and 1.92 (95% CI 1.27, 2.90), respectively. There was interaction between smoking and HLA high-risk genotypes (AP 0.27 [95% CI 0.01, 0.53]) in relation to LADA. The positive association between smoking and LADA/type 2 diabetes was confirmed by the MR study. Conclusions/interpretation Our findings suggest that tobacco use increases the risk of LADA and that smoking acts synergistically with genetic susceptibility in the promotion of LADA. Data availability Analysis codes are shared through GitHub (https://github.com/jeseds/Smoking-use-of-smokeless-tobacco-HLA-genotypes-and-incidence-of-LADA). Graphical abstract Supplementary Information The online version contains peer-reviewed but unedited supplementary material available at 10.1007/s00125-022-05763-w.


Two-sample Mendelian randomisation (MR) study
A typical two-sample MR study uses uncorrelated genetic variants as instrumental variables to proxy modifiable exposures and provides a novel opportunity for causal inference [1,2]. Such a study is based on summary statistics (β and standard errors) for the associations between genetic instruments and the exposure, and between genetic instruments and the outcome [3] from genome-wide association studies (GWAS).
Valid genetic instruments in MR studies affect the outcome only through the exposure, and the genetic instruments should not affect the outcome directly or through confounders of the exposure-outcome association [4].

GWAS of LADA
There is only one GWAS of LADA hitherto [7], which was conducted in 2634 LADA cases versus 5947 population controls of European ancestry. LADA was defined according to age of diabetes onset, presence of diabetes-related autoimmune antibodies and the lack of insulin dependence within the first six months or one year of diagnosis. Summary statistics for the association between the 250 SNPs and LADA were extracted from this GWAS.

GWAS of type 2 diabetes
The DIAbetes Genetics Replication And Meta-analysis Consortium (DIAGRAM) conducted a GWAS of type 2 diabetes including 26,676 cases and 132,532 controls of European ancestry [8]. Summary statistics for the association between SNPs and type 2 diabetes were extracted from this GWAS. One of the 250 SNPs was unavailable in this GWAS and only 249 SNPs were used as instrumental variables for smoking when assessing its association with type 2 diabetes.

Genetic instruments for smoking
The GWAS and Sequencing Consortium of Alcohol and Nicotine use (GSCAN) conducted a GWAS in up to 1.2 million European individuals and identified 259 sentinel single nucleotide polymorphisms (SNPs) for smoking initiation (ever smoking) [5]. A total of 250 out of the 259 SNPs were available in the GWAS of LADA (described below) and were used as instrumental variables for smoking in this MR study. These SNPs explained approximately 4% of the variance in smoking. Three (rs6011779, rs11783093, and rs12027999) of the 250 SNPs are closely related to CHRN genes, which encode subunits of neuronal nicotinic acetylcholine receptors (nAChRs) [6]. There was no SNP located near or in linkage disequilibrium (LD) with genes of the human leucocyte antigen (HLA) among the 250 SNPs used as instrumental variables. Summary statistics for the marginal associations between the 250 SNPs and smoking were extracted from the GWAS dataset generated by GSCAN based on 632,802 individuals (the 23andMe cohort was excluded from the released GWAS dataset).

ESM Discussion about MR results
Causal inferences are more plausible when there are consistent findings across different MR methods. In our study, different MR estimators (except for MR-Egger) estimated the same direction and similar magnitude of association for LADA as well as type 2 diabetes, indicating the robustness of our findings. I 2 GX statistics in the MR-Egger analysis were significantly less than 1, meaning that the MR-Egger causal effect estimates were biased, as reflected by the much smaller ORs estimated by MR-Egger than ORs estimated by other MR estimators. Moreover, MR-Egger often suffers from low statistical power, which was reflected by the wider 95% CIs estimated from MR-Egger than 95% CIs estimated by other MR estimators ( Figure  1). A low I 2 GX statistic means that either the NOME (NO Measurement Error) assumption or the InSIDE (Instrument Strength Independent of Direct Effect) assumption that MR-Egger relies on is violated [15]. The violation of NOME assumption indicates "weak instrument bias", which will bias the risk estimate to null in MR-Egger analysis [15]. "Weak instrument bias" can also be a problem for other MR estimators such as IVW. However, different from MR-Egger, the strength of instruments in other MR estimators relies on F statistics and not on I 2 GX. An F statistic of >10 is generally regarded as evidence of strong instrument strength [2] and the F statistics for 249 out of the 250 SNPs used in the present MR analysis were >30 (ESM Table  1). The InSIDE assumption is violated if pleiotropic effects act via a confounder [11]. In addition to MR-Egger, other MR methods such as MR-PRESSO also need to satisfy the InSIDE assumption [12]. However, we have excluded SNPs associated with any other trait at genomewide significance level from conservative analysis 1 and positive associations of smoking with LADA and type 2 diabetes still existed.
The p values for heterogeneity tests in different MR estimators indicated the presence of heterogeneity. The existence of heterogeneity, however, does not necessarily imply the existence of pleiotropy. Heterogeneity can also be due to different biological mechanisms linking different SNPs to the exposure and there is no violation of MR assumptions in this scenario [3]. Heterogeneity can also exist when the overall pleiotropic effects from different SNPs happen to cancel out ("balanced pleiotropy") [16]. The causal estimates from different MR methods are still unbiased in the case of balanced pleiotropy. The outlier test for MR-PRESSO in fact detected no outlier for the smoking-LADA association (Figure 1). MR-PRESSO detected two outliers for the association between smoking and type 2 diabetes, but the outlier-corrected estimates are similar to the ORs estimated by IVW. Therefore, horizontal pleiotropy does not seem to severely bias the causal estimates in our study.  Table 3. Odds ratios (OR) with 95% confidence intervals (CI) for combinations of tobacco use and HLA genotypes in the risk of LADA, and attributable proportions due to interaction (AP) with 95% CI (ESTRID). ESM Table 9. Hazard ratios (HR) with 95% confidence intervals (CI) for combinations of smoking and HLA genotypes in the risk of LADA, and attributable proportions due to interaction (AP) with 95% CI (HUNT).