figure b

Introduction

Metformin is the first-line pharmacological treatment to manage hyperglycaemia in people with type 2 diabetes, and is on the World Health Organization list of essential medicines [1]. Increasing evidence suggests that metformin may differ from other classes of glucose-lowering medications in having superior safety and lower risk of cardiovascular complications [2]. Furthermore, pharmaco-epidemiological studies have suggested that metformin may reduce cardiovascular disease and cancer [3, 4], suggesting the possibility of its use for these diseases. Metformin not only impacts glycaemic traits but also other potentially relevant factors, such as growth differentiation factor 15 (GDF-15) and vascular endothelial growth factors [5]. However, pharmaco-epidemiological studies may be open to immortal time bias and confounding, which may generate spurious protective effects of metformin, in particular for cancer-related studies [6, 7]. To date, relevant randomised controlled trials of metformin in cardiovascular disease are not large enough to be definitive [8], whilst the impact of metformin on cancer has not been evaluated fully in a randomised controlled trial.

Mendelian randomisation studies, which make use of the random allocation of genetic variants at conception, are less susceptible to confounding and time-related biases than other observational studies, and are now increasingly used to infer health effects of medications, such as the use 3-hydroxy-3-methyl-glutaryl-coenzyme A reductase (HMGCR) variants to mimic the health effects of statins [9]. Previous Mendelian randomisation studies have attempted to use genetics to infer health effects of metformin, but they were potentially underpowered and included non-specific instruments [10], or only evaluated health effects of metformin biomarkers instead of metformin itself [11]. To provide more definitive and direct evidence concerning the effects of metformin on cardiovascular disease and cancer risk, we conducted a Mendelian randomisation study using AMP-activated protein kinase (AMPK), the target of metformin, as a proxy of metformin use, in one of the largest prospective cohort studies globally.

Methods

Study design

This is a Mendelian randomisation design, using AMPK, the pharmacological target of metformin, to infer the AMPK pathway-dependent effects of metformin. The study design is depicted in Fig. 1 [12].

Fig. 1
figure 1

Study design of this Mendelian randomisation study (a) and its comparison with a randomised controlled trial (b)

Full size image

Study population

The UK Biobank recruited ~500,000 participants intended to be aged 39–73 years between 2006 and 2010 from 22 recruitment centres across Scotland, Wales and England in the UK. Participants provided biological samples; completed questionnaires, covering self-reported diseases and regular prescription medications; underwent assessments; and had nurse-led interviews. A blood sample for standard haematological tests was collected by venepuncture in ethylenediaminetetraacetic acid tubes, and tested at the central processing laboratory in Stockport, within 24 h of blood collection. HbA1c was measured by high performance liquid chromatography on Bio-Rad Variant II Turbo analysers (Bio-Rad Laboratories, CA, USA). Longitudinal follow-up via record linkage to all health service encounters and death is ongoing. Hospital inpatient data and cancer registries used ICD-9 and ICD-10 codes and death registries used ICD-10 codes. Genotyping was undertaken with two similar arrays, the UK Biobank Lung Exome Variant Evaluation (BiLEVE) Axiom array (49,979 participants) and the UK Biobank Axiom array (438,398 participants). Genotype imputation was based on the reference panel combining the UK10K haplotype and the Haplotype Reference Consortium reference panels. To reduce confounding by latent population structure [13], we restricted the analysis to genetically verified white British participants and further excluded participants with (1) withdrawn consent; (2) sex mismatch (genetic sex differs from reported sex); (3) putative sex chromosomes aneuploidy; (4) poor-quality genotyping (outliers in heterozygosity and missing rate >1.5%); or (5) excessive relatedness (more than ten putative third-degree relatives), as per our previous study [14]. We used genotype and phenotype data from the UK Biobank provided in February 2020.

AMPK genetic score

We created a weighted AMPK genetic score to mimic the effects of AMPK activation by metformin use based on the strength of the association of genetic variants in the relevant gene regions with HbA1c in the Meta-Analyses of Glucose and Insulin-related traits Consortium (MAGIC), a genome-wide association study (GWAS) of HbA1c, with validation in the UK Biobank. Specifically, we selected genetic variants within 1 megabase pairs downstream and upstream of each of the PRKAA1, PRKAA2, PRKAB1, PRKAB2, PRKAG1, PRKAG2 and PRKAG3 genes that encode AMPK subunits [15]. We selected low-linkage disequilibrium (r2 < 0.3) variants associated with HbA1c at a nominal level of statistical significance (p ≤ 0.05) in MAGIC, restricted to people of European ancestry to minimise population stratification (n = 123,665) [16]. We then validated the associations in the UK Biobank (using multivariable linear regression, adjusted for sex, age at recruitment, genotyping array and the first 20 principal components of genetic ancestry) and only retained variants also reaching statistical significance (p ≤ 0.05) in the UK Biobank, which were used to construct the AMPK score. Electronic supplementary material (ESM) Table 1 and ESM Fig. 1 show the details regarding the 44 variants used to construct the AMPK score. A weighted AMPK score was calculated for each participant by summing the number of HbA1c-lowering alleles that a participant inherited at each variant included in the AMPK score, weighted by the effect of that variant on HbA1c measured in percentage (as estimated in MAGIC) [16]. This score was then considered as above and below the median to mimic metformin use and non-use. We also considered AMPK score in quartiles to assess whether our finding was robust to the way we consider AMPK scores.

Sensitivity analysis

As a sensitivity analysis, we used a stringent variant selection criterion by using only variants associated with HbA1c at genome-wide significance (p ≤ 5 × 10−8) in both MAGIC and UK Biobank and not in linkage disequilibrium with the other variants (r2< 0.01), which gave rs2732480 (ESM Table 1). rs2732480 was associated with lower HbA1c in both MAGIC (p = 2 × 10−9) and UK Biobank (p = 1.07 × 10−142). The effect allele was associated with lower HbA1c percentage (β −0.012; 95% CI −0.016, −0.008).

Study outcomes

The primary outcomes were coronary artery disease and overall cancer. The secondary outcomes were stroke and three main cancers, i.e. breast cancer, colorectal cancer and prostate cancer. Each disease outcome was defined based on self-reported medical conditions at baseline, or subsequent primary and secondary diagnoses of hospital episodes (ICD-9 and ICD-10), or cancer diagnosis (ICD-9 and ICD-10), or underlying and contributing causes of death (ICD-10).

Positive control outcomes

We included type 2 diabetes and HbA1c as positive control outcomes given that these are the expected effects of metformin use. Type 2 diabetes was ascertained using a validated algorithm [17]. Specifically, the criteria included (1) self-reported type 2 diabetes at baseline; (2) indication of type 2 diabetes based on diagnostic codes (ICD-9250 and ICD-10 E11); (3) diabetes medications (metformin, sodium-glucose cotransporter 2 inhibitors, glucagon-like peptide 1 receptor agonists, dipeptidyl peptidase 4 inhibitors, sulfonylureas and thiazolidinediones); and (4) hyperglycaemic blood result (either HbA1c ≥6.5% or 48 mmol/mol, or random glucose ≥11.1 mmol/l). The algorithmic definitions are described in ESM Table 2. HbA1c was measured in mmol/mol (International Federation of Clinical Chemistry [IFCC] unit), and was converted to percentage (National Glycohemoglobin Standardization Program [NGSP] unit) using the equation: NGSP = (0.09148 × IFCC) + 2.152 [18].

External validation

To validate our findings from the UK Biobank, we conducted an external validation study for the outcomes using summary statistics for type 2 diabetes (12,171 cases and 56,862 control participants) from the DIAbetes Genetics Replication And Meta-analysis (DIAGRAM) consortium [19], for coronary artery disease (60,801 cases and 123,504 control participants) from the Coronary ARtery DIsease Genome wide Replication and Meta-analysis plus The Coronary Artery Disease Genetics (CARDIoGRAMplusC4D) consortium [20], for stroke (40,585 cases of stroke and 406,111 control participants) from the MEGASTROKE consortium [21], for breast cancer (122,977 cases and 105,974 control participants) from the Breast Cancer Association Consortium (BCAC) [22] and for prostate cancer (79,148 cases and 61,106 control participants) from the Prostate Cancer Association Group to Investigate Cancer Associated Alterations in the Genome (PRACTICAL) consortium [23]. All participants were of predominantly European ancestry and non-overlapping with the participants in the UK Biobank to avoid bias due to population structure and any potential bias due to participant overlap for a weak instrument [24].

Statistical analysis

To assess the assumption of independence of the genetic instruments (AMPK groups) from potential confounders, we assessed the association of AMPK groups with confounders (age at recruitment, BMI, smoking status, alcohol drinking status, education level, Townsend deprivation index) using χ2 tests or ANOVA. To demonstrate that AMPK had the expected effect on HbA1c, we assessed the differences in HbA1c and random glucose between each group using ANOVA. We assessed the association of AMPK categories with HbA1c using multivariable linear regression. We assessed the association of AMPK categories with risk of type 2 diabetes, cardiovascular diseases and cancers using multivariable logistic regression. All regression analyses were adjusted for sex (if relevant), age at recruitment, genotyping array and the first 20 principal components of genetic ancestry. As per previous studies, we also assessed the impact of genetically predicted reduction in HbA1c (%) instrumented by AMPK variants on risk of type 2 diabetes, coronary artery disease and overall cancer [25].

For the external validation, we performed a standard Mendelian randomisation analysis. We obtained the summary statistics of each variant included in the AMPK score on risk of type 2 diabetes, coronary artery disease, stroke, breast cancer and prostate cancer as reported by each consortium. We obtained the Wald ratio for each variant (the ratio of the genetic association with outcome to the genetic association of exposure), and then combined them using weighted generalised linear regression in an inverse variance-weighted manner and accounted for the correlation between variants [26]. The correlations between variants were obtained in 503 participants of European ancestry from the 1000 Genomes Project (Phase 3). Since variants are from multiple gene regions that may have different mechanisms of effect, a random-effects model was used [26]. We aligned the effect allele of each variant to the HbA1c decreasing allele. We used the Cochran’s Q statistic to assess heterogeneity of the Wald ratios [27], where high heterogeneity may indicate the presence of invalid genetic variants [28].

Exploring the association of HbA1c with cardiovascular disease and cancer risk using Mendelian randomisation

To preclude the possibility that the observed effects of AMPK activation, a target of metformin, are due to lowering HbA1c, we also assessed the association of genetically predicted lower HbA1c on cardiovascular disease and cancer risk. As previously, we obtained 38 independent genetic variants strongly related to HbA1c (p ≤ 5 × 10−8) from MAGIC (ESM Table 3), and applied them to the relevant outcomes in the UK Biobank using inverse variance-weighted, MR-Egger and a weighted median method [14].

The AMPK score was generated using PLINK 2.0 (https://www.cog-genomics.org/plink/2.0/) [29]. Mendelian randomisation analyses were performed with the MendelianRandomisation package version 0.4.2 [30] and all analyses were performed using R software, version 3.5.1 (R Foundation for Statistical Computing, Vienna, Austria) [31]. A two-tailed p value less than 0.05 was considered statistically significant.

Ethical approval

The UK Biobank received ethical approval from the North West Multi-centre Research Ethics Committee (11/NW/0382), and all participants provided written, informed consent. No ethics approval was acquired for the analyses using summary statistics. The contributing studies to the consortium received ethical approval from their specific institutional review boards, and written, informed consent was obtained from all participants.

Results

Participant characteristics

A total of 391,199 participants were included in the main analysis (mean age, 56.9 years; 54.1% women). For type 2 diabetes, there were 26,690 cases. For cardiovascular disease, there were 38,098 cases of coronary artery disease and 11,358 cases of stroke. For cancer, there were 80,941 cases of overall cancer, 9251 cases of breast cancer, 5861 cases of colorectal cancer and 8970 cases of prostate cancer. In Table 1, HbA1c, glucose, insulin therapy users and metformin therapy users were significantly lower in the high AMPK group than in the low group. No other significant differences in baseline characteristics between the two groups were found.

Table 1 Baseline characteristics of the participants in the UK Biobank
Full size table

Association of AMPK score with glycaemic traits and type 2 diabetes

Compared with participants with a low AMPK score (below median favouring higher HbA1c), participants with a high AMPK score (above median) had lower HbA1c (β −0.032%; 95% CI −0.035, −0.028; p = 2.34 × 10−64) and lower random blood glucose (β −0.013 mmol/l; 95% CI −0.022, −0.005, p = 1.09 × 10−3). High AMPK score (above median) was also associated with a decreased risk of type 2 diabetes (OR 0.96; 95% CI 0.94, 0.99; p = 4.16 × 10−3), as shown in Fig. 2a. AMPK quartiles were associated with a stepwise decrease in HbA1c (quartile 2, −0.011%; 95% CI −0.016, −0.006; p = 2.70 × 10−5; quartile 3, −0.028%; 95% CI −0.033, −0.022; p = 2.02 × 10−25; and quartile 4, −0.047%; 95% CI −0.052, −0.042; p = 4.67 × 10−70), and a corresponding stepwise decrease in the risk of type 2 diabetes (p for trend = 4.18 × 10−3, Fig. 2a). Genetically predicted reduction in HbA1c (%) instrumented by AMPK variants was associated with a 61% decrease in the risk of type 2 diabetes (OR 0.39 per % reduction; 95% CI 0.20, 0.78; p = 7.69 × 10−3) (Fig. 3).

Fig. 2
figure 2

Association of AMPK score with risk of type 2 diabetes (a), coronary heart disease (b) and overall cancer (c) in the UK Biobank. Boxes represent ORs and lines represent 95% CIs

Full size image
Fig. 3
figure 3

The impact of genetically predicted reduction in HbA1c (%) instrumented by AMPK variants on risk of type 2 diabetes, coronary artery disease and overall cancer in the UK Biobank. Boxes represent ORs and lines represent 95% CIs

Full size image

Association of AMPK score with cardiovascular diseases

High AMPK score (above median) was associated with a 3% lower risk of coronary artery disease (OR 0.97; 95% CI 0.95, 0.99; p = 5.69 × 10−3; Fig. 2b), but not stroke (ESM Fig. 2a). AMPK quartile was associated with a stepwise decrease in the risk of coronary artery disease (p for trend = 4.37 × 10−3; Fig. 2b). Genetically predicted reduction in HbA1c (%) instrumented by AMPK variants was associated with a 53% decrease in the risk of coronary artery disease (OR 0.47 per % reduction; 95% CI 0.26, 0.84; p = 0.01) (Fig. 3).

Association of AMPK score with cancer

High AMPK score (above median) was associated with lower risk of overall cancer (OR 0.98; 95% CI 0.96, 1.00; p = 0.01 and p for trend = 4.04 × 10−3; Fig. 2c), but not with prostate cancer, breast cancer or colorectal cancer (ESM Fig. 2b–d). AMPK quartile was associated with prostate cancer (quartile 2, OR 0.91; 95% CI 0.85, 0.96; p = 1.61 × 10−3; quartile 3, OR 0.93; 95% CI 0.88, 0.99; p = 0.02), although the dose–response was unclear (ESM Fig. 2b). Genetically predicted reduction in HbA1c (%) instrumented by AMPK variants was associated with a 44% decrease in the risk of overall cancer (OR 0.56 per % reduction; 95% CI 0.36, 0.85; p = 7.23 × 10−3) (Fig. 3).

Sensitivity analysis by using a more stringent variant selection criterion

A 1% reduction in HbA1c instrumented by rs2732480 was associated with a decreased risk of type 2 diabetes (OR 0.11; 95% CI 0.02, 0.50; p = 4.08 × 10−3) and coronary artery disease (OR 0.22; 95% CI 0.06, 0.81; p = 0.02). The direction with overall cancer was consistent with the main analysis but with wider CI (OR 0.45; 95% CI 0.17, 1.14; p = 0.09) (ESM Table 4).

External validation

In external replication analyses, genetically predicted lower HbA1c instrumented by AMPK variants was associated with decreased risk of type 2 diabetes (OR 0.11 per % reduction; 95% CI 0.04, 0.35; p = 1.78 × 10−4) and coronary artery disease (OR 0.48 per % reduction; 95 CI 0.33, 0.72; p = 2.89 × 10−4), but not with stroke, breast cancer or prostate cancer (ESM Table 5 and ESM Fig. 3a–e). The Q statistic suggested possible heterogeneity for the associations with type 2 diabetes, coronary artery disease, stroke and breast cancer.

Association of HbA1c with cardiovascular disease and cancer risk in the UK Biobank using Mendelian randomisation

ESM Table 6 shows that genetically predicted higher HbA1c was associated with higher risk of coronary artery disease (OR 1.41; 95% CI 1.03, 1.93; p = 0.03), and possibly with lower risk of overall cancer (OR 0.84; 95% CI 0.70, 1.01; p = 0.07), but not with stroke, or any cancer subtype.

Discussion

To the best of our knowledge, this is one of the first Mendelian randomisation studies to ascertain the effects of metformin, based on AMPK variants, on cardiovascular diseases and cancer. Using a design more robust to immortal time biases and confounding, our study is consistent with previous pharmaco-epidemiological studies suggesting that metformin use may reduce coronary artery disease and overall cancer risk. We added some genetic evidence that the putative cancer-protective effect of metformin via AMPK pathways is unlikely to be associated with glycaemic control.

A protective effect of metformin on cardiovascular health was observed in small randomised controlled trials using surrogate outcomes [32], the UK Prospective Diabetes Study post-trial analysis [33] and a recent meta-analysis [8], which were consistent with our findings. Although a genetically predicted reduction in HbA1c is protective against coronary artery disease [14], it is apparent that metformin’s protective effect is not solely due to its improvement in glycaemic profile given that these benefits are not clearly observed for all other classes of glucose-lowering medications [8], such as sulfonylureas and insulin [2, 34]. Metformin increases GDF-15, a stress-responsive cytokine which suppresses appetite and promotes weight loss [35], hence providing a potential mechanistic pathway by which metformin reduces cardiovascular disease risk. However, changes in GDF-15 were not clearly associated with coronary artery disease risk based on our previous Mendelian randomisation study [11]. On the contrary, sulfonylureas and insulin may lead to cardiotoxicity via weight gain, hypoglycaemia [34] or alteration of hormone levels [36].

The relation of metformin use with cancer risk is more controversial given the concern over immortal time bias [6]. Our study, where the start of ‘exposure’ is at birth, effectively removes this bias. As such, our study adds by showing that immortal time bias alone may not have explained the inverse relation of metformin use with cancer risk. Given that previous studies generally have ruled out the causal role of glycaemic traits in cancer risk [10, 37], possible mechanisms underlying the anti-cancer property of metformin are likely via pre-cursors of glycaemic traits or of glycaemic-independent pathways [38]. People without growth hormone appear to be protected against both diabetes and cancer [39]. This may suggest a possible pathway via growth hormone or the closely related insulin like growth factor-1 [40]. Glycaemic-independent pathways may include inhibition of tumour-mesothelial cell interaction by suppressing hypoxia-inducible factor 1α and TGF-β signalling [41], immune-mediated responses via metabolic reprogramming of tumour-specific T cells [42] and GDF-15 overexpressing fibroblasts promoting the growth of tumour xenografts [43]. The examination of these potential mechanisms can be explored in further studies. Big data approaches, such as metabolomics, may also be warranted to better understand the full spectrum of effects of metformin, and hence to help identify the main pathways in which metformin confers the additional benefits on cardiovascular disease and cancer [44].

Although our study is more robust to confounding and immortal time bias than previous observational studies, there are limitations. First, whilst our study suggested that AMPK activation by metformin may protect against coronary artery disease, and possibly cancer, the estimates from this study cannot be used directly to infer the health impact of metformin given the differences in exposure time, where randomised controlled trials often consider short-term pharmacological treatment in contrast to the effects of lifelong exposures estimated by Mendelian randomisation [12]. Moreover, our study using AMPK variants may only predict the effect of metformin which acts on the AMPK activation pathways, and metformin may also have AMPK-independent pathways that could be explored in additional studies to fully capture the overall effect of metformin on cardiovascular disease and cancer [45]. Second, we used a lower threshold than genome-wide statistical significance to select AMPK variants as a proxy of metformin use to maximise total prediction of AMPK function by the genetic score. We reduced the possibility of false positives by cross-checking the variants’ associations with HbA1c in two independent studies. We also repeated the analysis with stringent variant selection criteria which gave a consistent conclusion. However, this may compromise the generalisability of the genetic score in other studies [46]. Third, we cannot rule out selection bias resulting from the recruitment of generally healthier participants and survivors in the UK Biobank, which may bias the estimate towards the null. We also cannot rule out selection bias from competing risk before recruitment for diseases which share risk factors with other diseases that typically occur at younger ages, which could have biased estimates for stroke and prostate cancer to the null. Forth, the Q statistic suggested possible heterogeneity in some analyses. These heterogeneities may imply that multiple gene regions encoding subunits of AMPK may have different mechanisms of influencing the outcomes, and should be explored in future studies [28]. Fifth, given the pleiotropic effects of metformin and its association with multiple non-glycaemic makers [5], it would be difficult to identify a suitable negative control outcome. Nevertheless, we also assessed the impact of HbA1c on these outcomes and found that HbA1c unlikely explained all of the observed associations related to AMPK. Lastly, we could not exclude the possibility that metformin may reduce subtypes of cancer as the number of cases was not large enough for adequate statistical power, although the directions of effect for some cancer subtypes are similar to overall cancer. Few AMPK genetic variants were available for prostate cancer and breast cancer in the consortia and we were unable to create an overall genetic score in the associated analyses to increase statistical power. Together with possible selection biases embedded in these GWAS [47], these factors might explain the discrepancies between the estimates from the UK Biobank and from external consortia. Further investigations in large consortia on specific cancers may help verify the potential anti-cancer property of metformin.

Conclusion

This Mendelian randomisation study provides some genetic evidence that AMPK activation by metformin may reduce coronary artery disease risk, and possibly overall cancer risk. Whether metformin can be repurposed for coronary artery disease and cancer should be explored in large randomised controlled trials.