Main findings
Based on large-scale metabolomics data in UKB, we identified a wide range of metabolites causally affected by smoking, including an inflammatory biomarker, fatty acids, and different lipids subclasses. Most of these metabolites were associated with type 2 diabetes, and when we aggregated them into a smoking-related metabolic signature, this signature mediated more than one third of the association between smoking and type 2 diabetes. The metabolic signature was confirmed by external validation, and it also mediated part of the smoking-diabetes association in the TwinGene study. We also observed additive interaction between the metabolic signature and genetic susceptibility to type 2 diabetes or insulin resistance. This implies that smoking increases the risk of type 2 diabetes in part through effects on the metabolome and suggests that such effects are more pronounced in individuals with genetic risk factors for diabetes.
Comparison with previous studies
Our finding that smoking affects a variety of metabolites confirms those of previous, smaller observational studies [
4‐
13] (eTable
12) and a one-sample MR-study [
15]. The observation that most of them also associate with incidence of type 2 diabetes is consistent with previous prospective studies [
19,
40,
41]. For free fatty acids, we observed a general pattern where smoking seemed to decrease the degrees of unsaturation, leading to higher levels of saturated (SFA) and monounsaturated fatty acids (MUFA), and lower percentages of polyunsaturated ones such as DHA, omega-6 fatty acids, and omega-3 fatty acids. Previous studies find that smoking is positively associated with triglycerides and LDL cholesterol [
42], and inversely associated with HDL cholesterol [
43] and HDL particle sizes. We extend these observations by showing that smoking was associated with larger VLDL particle sizes, higher levels of VLDL regardless different particle sizes, higher levels of triglycerides regardless of types/particle sizes of lipoproteins, lower levels of all forms of cholesterols (total cholesterol, free cholesterol, or esterified cholesterol) in HDL regardless of particle sizes, and lower levels of different forms of cholesterols in IDL. We and others [
5,
12,
44] observed associations between smoking and some amino acids but since our MR analyses did not confirm most of them, they are probably not causal. Importantly, our findings in former smokers indicated that most smoking-related metabolic changes are reversible after smoking cessation. The exact biological pathways linking smoking to metabolic changes warrant deeper exploration. Notably, the effect of smoking on gut microbiota has been observed [
45] and gut microbiota is an important determinant of metabolite levels [
14,
46].
Our study is the first to quantify the overall mediation role of smoking-related alterations of the metabolome in the association between smoking and type 2 diabetes. We did this by integrating multiple metabolites influenced by smoking into one metabolic signature. Our findings suggest that variation in this signature explains 38·3% of the excess risk of type 2 diabetes conferred by smoking. More than half of the smoking-diabetes association was not mediated by the smoking-related metabolic signature, indicating that other pathophysiological consequences of smoking play a role in diabetes development. Such mechanisms may include direct adverse effects of smoking on pancreatic tissue and β-cell function [
47]. Part of the effects of smoking on type 2 diabetes may also be mediated by metabolites not measured in the NMR platform which primarily targeted lipid-related metabolites. This remains to be investigated. Nevertheless, non-biological factors such as misclassification of smoking status and measurement errors of the metabolites may also affect the estimated proportion and therefore the exact proportion should be interpreted with caution.
Diabetes is a heterogenous metabolic disorder typically caused by the combination of insulin resistance and insulin deficiency [
48]. Several of the smoking associated metabolites are known to be associated with insulin resistance. As an example, glycoprotein acetyls is an inflammatory biomarker and inflammation is an important promotor of insulin resistance [
49]. Inflammation and insulin resistance may exacerbate each other [
18]. Triglycerides, HDL cholesterol, and free fatty acids are also closely linked to insulin resistance [
17,
50]. Metabolites can regulate insulin sensitivity directly by modulating components of the insulin signaling pathway, indirectly by altering the flux of substrates through multiple metabolic pathways such as lipogenesis and lipid oxidation and protein synthesis, and though post-translational modification of proteins [
17]. The mediating role of the metabolome in the association between smoking and type 2 diabetes thus seems to involve insulin resistance-related pathways. This is consistent with experimental studies showing that smoking causes insulin resistance [
16]. Interestingly, there was additive interaction between the metabolic signature and GRS-IR but not between smoking and GRS-IR. Given the close relationship between smoking-related metabolites and insulin resistance, it is possible that GRS-IR interacts specifically with the metabolic alterations caused by smoking (the metabolic signature) and not with other effects of smoking. The interaction between metabolic signature and GR-IR may reflect synergistic effects of inherited and acquired insulin resistance in the development of type 2 diabetes. Regarding GRS-T2D, we observed additive interaction with both smoking and the metabolic signature. For smoking, previous studies either did not investigate interaction on the additive scale [
20] or did not find evidence of additive interaction [
51,
52], which might be due to limited statistical power. GRS-T2D primarily captures β-cell function and insulin secretion [
53,
54] while the metabolic signature mainly encompassed indicators of insulin resistance. Therefore, our findings indicated that individuals with inherited tendency towards dysfunctional insulin secretion may be more susceptible to adverse effects of acquired insulin resistance. Of note, interaction was detected solely on the additive scale and not on both additive and multiplicative scales, which is considered the strongest form of interaction. Further studies on the topic are clearly warranted.
Strengths and limitations
The strengths of this study include the use of large-scale metabolomics data and the integration of both observational and MR analyses which allowed us to identify metabolites influenced by smoking. We also had access to genomic information and could, for the first time, investigate if smoking induced alterations of the metabolome interacts with genetic susceptibility on diabetes incidence. We used elastic net regression to derive an overall smoking-related metabolic signature for mediation analysis. Such a model works well for data with high collinearity and has been used by previous metabolomics studies [
32,
55]. The robustness and generalizability of the metabolic signature was supported by external validation in TwinGene and internal cross-validation in UKB. A further strength was the ability to adjust for potential confounding from a wide range of lifestyle factors including diet. Both smoking status and metabolite levels may change during the follow-up, but this is most likely to underestimate the associations of smoking and the metabolite signature with diabetes. Furthermore, we confirmed the metabolites identified at baseline by MR analyses and many of the smoking-related metabolites were also associated with current smoking at repeat assessment. MR analyses assume no pleiotropy. We accounted for this issue by applying the MR pleiotropy residual sum and outlier approach (MR-PRESSO) estimator to detect and correct for potential pleiotropy. This study was conducted in people of primarily European origin, and it remains to be explored if our findings are generalizable to non-European populations.