Drug targets for lipid modification and risk of type 2 diabetes: a cis-Mendelian randomization study

Abstract

Background and aims

Reducing plasma levels of low-density lipoprotein cholesterol (LDL-C) is the cornerstone in the prevention of coronary artery disease (CAD) but may also increase risk of type 2 diabetes (T2D). A comprehensive examination of the genetic evidence of T2D related side-effects of all current lipid-modifying drugs, including those in development, has not yet been performed.

Methods

This cis-Mendelian randomization study used individual level data from the UK Biobank, Lifelines, and publicly available genome-wide association data. We identified loci that are either targeted directly with drugs, or alternatively, targeting their gene products (mRNA and/or protein). Included are, in alphabetical order, the loci ACLY, ANGPTL3, ANGPTL4, APOB, APOC3, CETP, HMGCR, LDLR, LIPG, LPA, MTTP, NPC1L1, and PCSK9. We used cis-genetic instruments weighted for LDL-C, HDL-C, triglycerides, and apolipoproteins as downstream proxies for the drug targets. Main outcomes were prevalent and incident T2D, with CAD as a contrast outcome.

Results

Lipid modification through HMGCR is predicted to reduce CAD risk and increase T2D risk. Modification through targeting APOC3, LDLR, LPA, MTTP, NPC1L1, and PCSK9 is predicted to reduce CAD risk without a change in T2D risk. Modification through ANGPTL4 and CETP is predicted to reduce risk of both CAD and T2D. For ACLY, ANGPTL3, APOB, and LIPG, we found evidence for neither CAD nor T2D.

Conclusions

This study provides genetic evidence for variation in diabetes-related side-effects of different lipid-modifying drugs, with potential relevance for future clinical trials and individual treatment decisions.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12933-026-03209-w.

Introduction

Lipid-lowering drugs are a cornerstone for the treatment of coronary artery disease (CAD) [1–3]. Over the past decade, however, convincing evidence has accumulated that statins, concomitant with reducing low-density lipoprotein cholesterol (LDL-C) levels, perturb glycemic control, thereby increasing the risk of type 2 diabetes (T2D) [4–9]. The effect of statins is at least in part explained by inhibition of 3-hydroxy-3-methylglutaryl-CoA reductase (HMGCR), the target of statins. A recent meta-analysis showed that statins increased the risk of new-onset diabetes [9]. Incident T2D was found to increase by 10% per year with low-intensity statin treatment and by 36% per year with high-intensity statin treatment. In patients with preexisting diabetes, statin treatment worsened glycemic control, again proportional to treatment intensity [9]. Although the benefit for CAD reduction by statin therapy is considered to significantly outweigh the risk of new-onset T2D, such effects of statins are evidently unwanted [9]. A recent analysis among more than 2 million people across 6 continents showed that at a projected age of 50 years the absence of T2D results in a gain in life expectancy of 6.4, and 5.5 years in women and men, respectively [10]. This emphasizes the importance of investigating diabetes-related side-effects of pharmaceutical lipid reduction.

Over the past years various drugs have become available for CAD prevention. In patients at moderately increased risk of CAD, statins are often used in combination with Ezetimibe [11] (target: NPC1L1, Nieman-Pick C1-Like 1) which inhibits intestinal cholesterol uptake. Bempedoic acid (target: ACLY, ATP citrate lyase) like statins also reduces hepatic cholesterol synthesis [12]. Additional LDL-C lowering can be achieved with PCSK9 inhibitors (target: PCSK9, proprotein convertase subtilisin/kexin type 9) which inhibit the degradation of the LDL receptor [13]. In patients with a very high risk of CAD, e.g., those suffering from homozygous familial hypercholesterolemia [14], lipoproteins can be cleared from the circulation through apheresis [15] or through inhibiting the synthesis and secretion of very-low density lipoprotein (VLDL), the precursor of LDL in the circulation. This can be achieved with Mipomersen (target: APOB, apolipoprotein (apo) B), or the small molecule inhibitor Lomitapide (target: MTTP, microsomal triglyceride transfer protein) [16].

In addition, new drugs have been or are being developed which effectively reduce plasma levels of triglycerides. These include Evinacumab [17], Zodasiran [18] and Solbinsiran [19] (target: ANGPTL3, Angiopoietin Like Protein 3) which through reducing ANGPTL3, an inhibitor of lipoprotein lipase (LPL), increase the lipolysis of triglycerides in VLDL and intestinal-derived chylomicrons. Volanesorsen [20], Olezarsen [21], Plozasiran [22] (target: APOC3, apoC-III), and MAR001 [23] (target: ANGPTL4, Angiopoietin Like Protein 4) also increase LPL-mediated triglyceride hydrolysis but through reducing plasma levels of apoC-III and ANGPTL4, respectively.

In the circulation, lipoproteins are also subject to the hydrolysis of phospholipids which is another target for therapy. For example, MEDI5884 (target: LIPG, endothelial lipase) is a monoclonal antibody against endothelial lipase that specifically increases high-density lipoprotein cholesterol (HDL-C) [24]. Obicetrapib (target: CETP, cholesteryl ester transfer protein) [25–30] is developed to reduce apoB-containing lipoproteins LDL as well as lipoprotein(a). Finally, Olpasiran, Lepodisiran, Pelacarsen, Muvalaplin (target: LPA, lipoprotein(a)) are drugs that are developed to specifically lower lipoprotein(a) which is associated with aortic valve disease as well as CAD [31].

The effects of the above drugs on T2D are described for several targets. Randomized controlled trial data and post-hoc analyses have suggested that PCSK9 inhibition could modestly increase T2D risk although data are not entirely consistent [32–37]. Post-hoc analysis on the use of CETP inhibitors in patients at increased risk of CAD have shown that this class of drugs may reduce the risk of new onset diabetes [38] while similar results were obtained with a drug target Mendelian Randomization (MR) study [29]. The use of Ezetimibe does not seem to confer a major risk of new onset diabetes [39, 40]. MR studies furthermore have shown that ANGPTL3 is associated with a reduction in CAD without affecting T2D risk [41, 42] but others have shown that genetic inhibition of ANGPTL3 as well as ANGPTL4 are associated with reductions in both CAD and T2D [43, 44] MR studies into APOC3 have shown that genetic inhibition only reduces CAD [41, 42, 44]. Finally, genetically determined very low lipoprotein(a) levels are reported to be associated with an elevated risk of incident T2D, in particular in subjects on statin treatment [45]. It remains to be determined how robust the effects of these drugs on T2D are, whether these are due to on-target drug effects and how they compare to each other. In addition, studies into the relation between lipid-modulation and T2D are missing for drugs that target ACLY, APOB, LDLR, LIPG, and MTTP.

In the present study, we thus aimed to examine variation in T2D related side-effects between lipid modifying drugs. We used a cis-MR approach to genetically mimic an on-target drug trial [46–48], predicting effects of all current lipid-modulating drugs on T2D (our main outcome), relative to their effect on CAD (as a contrast outcome).

Methods

Study overview and design

We applied cis-MR methodology, in this context also referred to as drug target MR as it leverages cis-genetic variants in and around a drug target locus. An overview of the study design is shown in Fig. 1. In brief, first we identified loci that are either targeted directly with drugs, or alternatively, targeting their gene products (mRNA and/or protein). Included are, in alphabetical order, the loci ACLY, ANGPTL3, ANGPTL4, APOB, APOC3, CETP, HMGCR, LDLR, LIPG, LPA, MTTP, NPC1L1, and PCSK9. Second, we investigated single nucleotide polymorphisms (SNPs) proximal to these loci (i.e. cis-variants 100 kb upstream and 100 kb downstream of these loci). As the majority of GWAS and reference datasets are provided in Ensembl GRCh37 coordinates [49], we used these coordinates rather than those from the GRCh38 assembly to ensure consistency across datasets. We included SNPs associated with plasma lipid biomarkers (HDL-C, LDL-C, and TG). Third, in available GWAS summary data, as well as in individual level, longitudinal cohort data (i.e. UK Biobank and the Lifelines Cohort Study and Biobank, henceforth UKB and LL), we assessed associations of these SNPs with a) the main outcome, T2D; b) the contrast outcome, CAD; and c) intermediate glycemic outcomes (Table 1 for all data sources). Fourth, after quality control (i.e. SNP-lipid and SNP-outcome data harmonization; instrument selection), we performed cis-MR to predict drug target effects, using weights based on HDL-C increase, LDL-C decrease, and TG decrease. Fifth, we performed MR sensitivity analyses to examine robustness against violations of pleiotropy assumptions. Sixth, we conducted colocalization analyses to assess potential confounding through linkage disequilibrium (LD, Fig. 2). Seventh, we assessed and compared the evidence for and between each drug target.

Fig. 1.

Study overview. Abbreviations: Apo, apolipoprotein; CAD, coronary artery disease; GRCh37, Genome Reference Consortium Human Build 37; GWAS, genome-wide association study; HbA1c, glycated haemoglobin; HDL, high density lipoprotein; HOMA-B/ IR, homeostatic model assessment of beta cell function/ insulin resistance; LDL, low density lipoprotein; SNP, single nucleotide polymorphism; T2D, type 2 diabetes; TG, triglycerides

Table 1.

Overview of data sources

Trait	Source	Ancestry	Total N (N cases)	Unit	Covariates	Author	Year	PubmedID
T2D prevalence	DIAMANTE	European subset	933,970 (80,154)	log(OR)	Cohort specific	Mahajan	2022	35551307
T2D prevalence	FinnGen (round 10)	European	400,197 (65,085)	log(OR)	Age, sex, genotyping batch, PC1–10	Kurki	2023	36653562
T2D incidence	UK Biobank	European	456,895 (19,529 incident cases)	log(HR)	Age, age2, sex, PC1–10	Present study	Present study	Present study
T2D incidence	Lifelines	European	57,876 (769 incident cases)	log(HR)	Age, age2, sex, PC1–10	Present study	Present study	Present study
CAD	CARDIoGRAMplusC4D	European	1,165,690 (181,522 cases)	log(OR)	Cohort specific	Aragam	2022	36474045
Random glucose	MAGIC	European subset	459,772	mmol L−1	Age, sex, BMI, time since last meal	Lagou	2023	37679419
HbA1c	MAGIC	European subset	146,806	%	Cohort-specific covariates, PCs	Chen	2021	34059833
Fasting glucose	MAGIC	European subset	200,622	mmol L−1	Cohort-specific covariates, BMI, PCs	Chen	2021	34059833
Two-hour glucose	MAGIC	European subset	63,396	mmol L−1	Cohort-specific covariates, BMI, PCs	Chen	2021	34059833
Fasting insulin	MAGIC	European subset	151,013	pmol L−1	Cohort-specific covariates, BMI, PCs	Chen	2021	34059833
HOMA-IR	MAGIC	European	37,037	log(NA)	Age, age2, sex, geographical covariates, study site	Dupuis	2010	20081858
HOMA-B	MAGIC	European	36,466	log(%)	Age, age2, sex, geographical covariates, study site	Dupuis	2010	20081858
HDL-C	UK Biobank	European	441,016	SD (inverse rank normalization)	Age, sex, genotyping chip, relatedness	Richardson	2020	32203549
LDL-C	UK Biobank	European	441,016	SD (inverse rank normalization)	Age, sex, genotyping chip, relatedness	Richardson	2020	32203549
Triglycerides	UK Biobank	European	441,016	SD (inverse rank normalization)	Age, sex, genotyping chip, relatedness	Richardson	2020	32203549
apoA-I	UK Biobank	European	441,016	SD (inverse rank normalization)	Age, sex, genotyping chip, relatedness	Richardson	2020	32203549
apoB	UK Biobank	European	441,016	SD (inverse rank normalization)	Age, sex, genotyping chip, relatedness	Richardson	2020	32203549

Fig. 2.

Mendelian randomization (MR) assumptions. Panel A. Mendelian randomization assumptions. Relevance (IV1): the instrument is strongly related to the exposure. Exchangeability (IV2): there is no confounding between the instrument and outcome (no path exists from confounders to instrument). Exclusion restriction (IV3): SNP effects on the outcome are exclusively mediated through the exposure (i.e. absence of horizontal pleiotropy, no alternative path exists from instrument to outcome). Under these assumptions, associations of SNP-instrumented exposure with an outcome provide unconfounded, causal estimates of the relation between an exposure and an outcome. Panel B. The cis-Mendelian randomization approach to drug target investigation is similar to traditional Mendelian randomization but uses only genetic instruments proximal to the genetic locus (cis-instruments). We used downstream lipid biomarkers as proxies for the drug target, e.g. associations of HMGCR-instrumented LDL-C decrease with T2D provides an estimate of the effect of HMGCR inhibition on T2D. Panel C. Correlation between genetic variants through linkage disequilibrium. If the variant in high linkage disequilibrium has a direct path to the outcome this is a violation of exchangeability, which is potentially the case if the exposure and the outcome do not colocalize (i.e. if they have a low posterior probability of having a shared single causal variant)

Mendelian randomization assumptions

In the population, genetic variation in a trait, e.g. LDL-C, is plausibly randomly distributed, independent of other traits, due to Mendel’s laws. MR exploits genetic variation as a natural experiment: those genetically predisposed towards higher LDL-C can be compared to those predisposed towards lower LDL-C with regards to health outcomes such as T2D. Using genetic variation in LDL-C at a specific drug target locus, i.e. using cis-MR, it is possible to predict the effects of pharmaceutical modification of that drug target.

More technically, MR uses genetic variants, in our case single nucleotide polymorphisms (SNPs), as instrumental variables (IVs). Under Mendel’s laws of inheritance, alleles for a certain trait are randomly assigned at conception (random segregation), independent of other traits (independent assortment). MR yields unconfounded, causal estimates under three key assumptions (Fig. 2). These are: 1) Relevance: the IV is strongly related to the exposure; 2) Exchangeability: there is no confounding between the IV and outcome, and; 3) Exclusion restriction: IV effects on the outcome are exclusively mediated through the exposure. In the summary cis-MR setting of the present study, additional assumptions apply: 4) no sample overlap between exposure and outcome genetic data, and; 5) although derived from two samples, the populations underlying these samples are sufficiently similar [50, 51].

While traditional MR uses variants across the genome, the cis-MR (or drug target MR) approach uses cis-variants, proximal to a genetic locus, to mimic on-target drug effects [52–54]. We used lipid biomarkers as proxies for the drug targets, using the associations of cis-variants with plasma lipid levels as weights; as an example, the association of HMGCR-instrumented LDL-C decrease with T2D represents an estimate of the effect of pharmaceutical HMGCR inhibition on T2D. This method, explained first by Schmidt et al., estimates the total causal effect of a drug target, importantly without the requirement or implication that the biomarker (here LDL-C decrease, HDL-C increase, TG increase) itself causes the outcome [47, 55].

A more detailed discussion on potential violations of the MR assumptions, and how we attempted to mitigate these, can be found in Methods S1.

Exposure data

We extracted SNPs associations from a cross-sectional GWAS on HDL-C, LDL-C, and triglycerides (TG) levels [56], within a 100 kb region of each locus. In the original GWAS, each lipid biomarker was inverse rank normalized to a mean of 0 and variance of 1 for comparability between lipids, lipoproteins, and apolipoproteins; units of SNP estimates are therefore in SD of the normalized lipid biomarkers (Table 1). As secondary biomarkers, we investigated apoA-I and apoB; these are the major protein components of, and thus positive controls to, HDL-C and LDL-C, respectively.

Outcome data

Our main outcome was T2D. We used both individual-level cohort data and population-level GWAS summary data to obtain SNP-outcome estimates for use as outcome data (Table 1). First, we essentially performed a candidate gene study in which we estimated SNP-T2D incidence (T2Di)-associations in longitudinal UKB and LL data using survival analysis methods (details on UKB, LL, and meta-analysis methods in Methods S2). Second, we performed a look up of SNP-T2D associations, at each of the genetic loci, from comprehensive cross-sectional/case–control, European ancestry GWAS summary data on T2D from the DIAMANTE consortium [57]. Third, we extracted SNP-T2D associations from FinnGen (round 10 data) [58] as an independent control outcome dataset, nullifying sample overlap between exposure and outcome data. We examined CAD as a positive control outcome. We used SNP-associations from a cross-sectional/case–control GWAS on CAD from the CARDIoGRAMplusC4D consortium [59] to contrast effects of lipid modification on T2D to known effects on CAD. As secondary outcomes, we examined intermediate glycemic/insulin traits (i.e., random glucose [60], HbA1c, fasting glucose, two-hour glucose, fasting insulin [61], HOMA-IR, and HOMA-B [62]) for which we obtained GWAS summary statistics from the MAGIC consortium.

Data harmonization and instrument selection

For all analyses, we restricted to common variants with a minor allele frequency of > 1%. We harmonized the lipid biomarker data with the T2D outcome data. In case of palindromic SNPs, we inferred the strand based on allele frequency, removing those with ambiguous allele frequencies (minor allele frequency > 0.43). In addition, we removed SNPs with incompatible alleles and non bi-allelic SNPs. From the complete harmonized dataset, SNPs that were significant (p < 1 × 10^–5) and independent (clumping LD r² < 0.001 within 1 Mb) were taken forward as instruments into our cis-MR analysis. For genetic colocalization analysis (both coloc and coloc-SuSIE), we used all available SNPs within 100kB of a certain locus (Fig. 1). Specifically, for UKB + LL T2Di, we excluded SNPs with heterogeneous effects between UKB and LL (heterogeneity p < 0.05). The complete harmonized dataset is presented in Table S1, with the set of clumped instruments presented in Table S2a. We calculated F-statistics for each SNP as indicator of instrument strength.

We identified SNPs within 100 kb from each locus and extracted or estimated SNP associations with exposure lipid biomarkers, CAD, and T2D/glycemic outcome traits. In total, we harmonized 468,584 pairs of SNP-lipid and SNP-outcome associations (Table S1). After clumping (i.e., removing all SNPs that are correlated with the strongest signal at a threshold of LD r² < 0.001, within 1 Mb from that signal) and p-value filtering (SNP-lipid p < 10^–5), 1305 SNP-lipid and SNP-outcome pairs remained. Instrument strength was sufficient (F-statistic range 19.6 to 3325, Table S2a-b).

cis-Mendelian randomization analysis

Before cis-MR analysis, we oriented the SNP-lipid biomarker effect directions towards the canonical drug effect, i.e. HDL-C increase, LDL-C decrease, and TG decrease. This was done so that the SNPs weights and thus the resulting MR estimates can be interpreted as the predicted drug effect (e.g. HMGCR-instrumented LDL-C decrease mimics HMGCR inhibition). We calculated the Wald ratio to obtain single SNP causal estimates (SNP-outcome effect divided by SNP-exposure effect, with standard errors derived using the delta method). When more than one independent SNP was available, we pooled Wald ratios using a random effect, inverse variance weighted meta-analysis (MR-IVW). MR-IVW is powerful but assumes the absence of horizontal pleiotropy, or that any horizontal pleiotropy is balanced. We therefore performed sensitivity analyses in case ≥ 3 SNPs were available; Cochran’s Q to assess heterogeneity as a first sign of horizontal pleiotropy; Egger intercept to find evidence for unbalanced, directional horizontal pleiotropy; and MR weighted median, MR simple mode, MR weighted mode, and MR Egger, which are less powerful than MR-IVW but robust to a range of violations of exclusion restriction. All analyses were performed using the TwoSampleMR library version 0.5.8 [63], with R version 4.2.2 software. As a sensitivity analysis, we reran MR including only SNPs with a strict, more traditional GWAS p-value cutoff of 5 × 10^–8, trading off statistical power for instrument strength. Next to this traditional approach, using the MendelianRandomization library version 0.6 [64], we performed a generalized least squares approach to MR that allows for correlated SNPs and thus the inclusion of additional instruments, trading off instrument validity for statistical power. In contrast to the main analysis that strictly allows only uncorrelated SNPs, we here allowed a small degree of correlation between SNPs (lenient clumping threshold, p < 1 × 10^–5, LD r² < 0.2 within 1 Mb, Table S2b). We in addition performed MR using weights for lipoprotein(a) (exclusively for LPA) from a recent GWAS [65] in UKB, and weights for gene expression in liver tissue (for each locus) taken from GTEx v11 liver tissue data (https://gtexportal.org/home/) [66]. To account for the multiple testing burden, we applied a two-sided threshold of p < 0.001 to indicate statistical significance for MR results. This study adhered to the STROBE-MR recommendations for reporting MR studies [67, 68] (Methods S2).

Genetic colocalization analysis

MR estimates may be confounded when a SNP in high LD with the instrument is the actual causal variant for the outcome (Fig. 2). This is thought to be less likely in case two traits colocalize (i.e. share a causal variant) at the genetic locus under investigation [69]. Therefore, using all available harmonized SNPs in each region, and associations of these SNPs with both exposure and outcome, we performed bivariate genetic colocalization analysis under the single shared causal variant hypothesis with the coloc.abf() function in the coloc R library version 5.2.2 [70]. We applied default, skeptical priors. We assessed the following hypotheses: h0: neither trait is associated, h1: only trait one (lipid exposure) is associated; h2: only trait two is associated (T2D or CAD outcome); h3: both traits are associated, but with distinct causal variants; h4: both traits share a common causal variant. A posterior probability of h4 > 0.80 was considered evidence of colocalization, whereas 0.50 < h4 ≤ 0.80 was considered suggestive. We built regional association plots with the geni.plots R package (https://github.com/jrs95/geni.plots) to visually present SNP associations and (co)localization. LD correlation matrices were obtained using LDLink [71], using the 1000G v3 European subset [72] as reference. Whereas colocalization was performed using pairwise deletion (i.e., using all available SNPs between two traits under investigation), we built regional association plots using listwise deletion (i.e., using only SNPs available for all traits) for more intuitive, simultaneous comparison across all traits.

For interpretation: if there is evidence for colocalization (high h4), it is more likely that MR assumptions hold (Fig. 2 panel B) and we assign more confidence to the MR results. If there is evidence for distinct causal variants (high h3), this suggests that the MR assumption of exchangeability is violated because of potential confounding through LD (Fig. 2 panel C) and thus we assign less confidence to such MR results.

Our primary colocalization analyses were performed under the single causal variant assumption. This assumption is violated in case there are multiple causal variants, as may be the case with multiple MR instruments per locus. Thus, as a sensitivity analysis, we performed coloc Sum of Single Effects (SuSiE) Regression, a fine-mapping informed approach that allows for multiple causal variants. We used the coloc_susie() function implemented in the coloc R package (version 5.2.3), using default priors, considering evidence to be suggestive in case of 0.50 < h4 ≤ 0.80 and strong in case of h4 > 0.80 [73].

Results

cis-Mendelian randomization and genetic colocalization results

In this study, we included all loci that are targeted directly with drugs, or with drugs that target their gene products. The loci that encompass these 13 lipid genes are listed in alphabetical order. For reference, all main results (i.e., MR, MR sensitivity analysis, genetic colocalization analysis, apoA-I and apoB) are presented together in Table S3. For ease of interpretation, the MR results presented in Figs. 3–4 and Table S3, are oriented towards canonical drug effects (i.e. per SD HDL-C increase, LDL-C decrease, and TG decrease). The supplemental information includes Fig. S1 which provides MR estimates of drug effects on intermediate glycemic/insulin traits. Probabilities for each of the 5 genetic colocalization hypotheses are visually presented in Fig. S3. We show regional association plots to visualize colocalization (or lack thereof) between lipid exposures and the main outcomes in Fig. S4. MR analyses after a stricter SNP selection threshold (p < 5 × 10^–8) are shown in Fig. S5. The results of coloc-SuSiE (credible sets, h0 to h4) are shown in Table S4, and we explore concordance between the coloc-SuSiE and our primary coloc method in Table S8 and Results S2. MR analyses using GTEx liver expression are shown in Table S9 and discussed in Results S1.

Fig. 3.

Strength and direction of drug effects predicted by cis-Mendelian randomization (MR). Each panel represents a genetic locus. Colored tiles represent statistical strength (-log₁₀(p-value)) and direction (color) of the cis-MR estimate. Red color indicates positive effects: drug target inhibition is predicted to increase outcome risk, while blue color indicates negative effects: drug target inhibition is predicted to reduce outcome risk, and finally white color indicates neutral effects: drug target inhibition does not affect outcome. MR p-values < 10^–10 are truncated. Tiles remain grey in case of unavailability of instruments. Circles represent results from genetic colocalization under the single causal variant assumption. Larger size indicates higher posterior probability (0–1) of a single shared causal variant (hypothesis 4, h4) between exposure and outcome, thus stronger support for the MR estimate. Lipid exposures represent the lipids that were used as weights, oriented towards canonical drug effect (i.e. HDL increase, LDL decrease, TG decrease), for the analysis. In this analysis, we used independent cis-instruments (p-value < 10^–5, clumping LD r² < 0.001 with LD calculated over 1 Mb, from SNPs within 100 kb of the genetic locus)

Fig. 4.

Predicted drug effect estimates from cis-Mendelian randomization, presented as odds ratio (OR) per SD in lipid weights with 95%CI for main binary outcome (type 2 diabetes, T2D, and coronary artery disease, CAD). We used independent cis-instruments (p-value < 10^–5, clumping LD r² < 0.001 within 1 Mb, within 100 kb of the genetic locus). MR estimates are based on the Wald ratio (SNP outcome divided by SNP exposure effect, in case of a single instrument) or random-effects inverse variance weighted meta-analysis of Wald ratios (in case of multiple instruments). Lipid exposures represent the lipids that were used as weights, with estimates oriented towards canonical drug effect (i.e. HDL increase, LDL decrease, TG decrease)

In summary, cis-MR predicts protective effects on T2D of drugs that target ANGPTL4 and CETP, and deleterious effects of drugs that target HMGCR. The evidence was strongest for ANGPTL4, with consistent effects across three T2D datasets (Mahajan, FinnGen, and UKB + LL longitudinal data), and supported by colocalization (high posterior probability of hypothesis 4: a shared single causal variant, meaning no evidence of confounding through LD). We found no suggestive effects of other drug targets. We examined CAD as a positive control outcome: cis-MR predicts that drugs for all targets reduce CAD risk, except for ACLY, ANGPTL3, APOB, and LIPG. Our main aim was to investigate effects of lipid modifying drugs on T2D. We therefore zoom in to the loci with suggestive effects on that outcome, i.e., HMGCR, ANGPTL4, and CETP, respectively, below, interrogating the data further (i.e. assessing sensitivity analyses examining instrument validity). In addition, we more closely examine APOC3—given its predicted deleterious effect on continuous HbA1c, and MTTP- given its predicted protective effect on random glucose. A detailed description of the results of the remaining 8 genes can be found in Results S1.

HMGCR

At the HMGCR region, we identified instruments for HDL-C, LDL-C and TG (Table S1). We first looked closely at LDL-C as the foremost biological and clinical readout for the use of HMGCR inhibitors. As anticipated, HMGCR-instrumented LDL-C decrease associated with reduced CAD risk (OR = 0.63 per SD LDL-C decrease, 95%CI: 0.54 to 0.74, p = 1.3 × 10^–8) with evident colocalization given the overlapping association signals (Fig. S4c) and strong colocalization probabilities (h4 = 1.00, Figs. 3, S3, Table S3). Using LDL-C weights, HMGCR-instrumented LDL-C decrease associated with increased T2D risk (Mahajan T2D: OR = 1.55 per SD LDL-C decrease, 95%CI: 1.25 to 1.93, p = 7.2 × 10^–5) while colocalization analysis suggests distinct causal variants for LDL-C and T2D at HMGCR (posterior probability of distinct causal variants h3 = 0.97; shared single causal variant h4 = 0.02) suggesting violation of exchangeability (i.e., likely confounding through correlated variants). The MR estimate was corroborated in independent FinnGen data (OR = 1.55 per SD LDL-C decrease, 95%CI: 1.26 to 1.91, p = 3.7 × 10^–5), with suggestive colocalization (h4 = 0.72), and similar but nonsignificant association with T2D incidence (UKB + LL, HR = 1.26 per SD LDL-C decrease, 95%CI: 0.91 to 1.73, p = 0.16, h4 < 0.01).

Using HDL-C weights, HMGCR-instrumented HDL-C increase surprisingly associated with CAD risk increase, (IVW OR = 3.64 per SD HDL-C increase, 95%CI 1.74 to 7.57, p = 5.6 × 10^–4) although colocalization analyses suggested distinct causal variants (h3 = 0.93, h4 = 0.05) and thus potential confounding through LD. We found evidence for reduced T2D risk (Mahajan OR = 0.06 per SD HDL-C increase, 95%CI: 0.02 to 0.16, p = 9.3 × 10^–9) with strong colocalization (h4 = 0.99). This was corroborated in FinnGen (OR = 0.13 per SD HDL-C increase, 95%CI 0.05 to 0.35, p = 4.2 × 10^–5, h4 = 0.78) and suggestively in T2D incidence data (HR = 0.16 per SD HDL-C increase, 95%CI: 0.04 to 0.73, p = 0.02, albeit without colocalization h4 = 0.02). With weights based on TG decrease, no associations with CAD or T2D were found.

Associations of HMGCR-instrumented lipid levels with intermediate glycemic outcomes were directionally consistent with associations with T2D (Figs. 3, S1, Table S3), but generally inconclusive with only borderline nominally significant effects and poor evidence of colocalization. No MR sensitivity analyses were possible due to a limited number of IVs (< 3 SNPs). In MR analyses where we used multiple correlated instruments, effect estimates were generally consistent with those of the main MR analysis (Table S3). Here, we were able to examine potential directional horizontal pleiotropy and generally found none (for MR Egger with correlated instruments, Cochran’s Q p > 0.05, Table S3), again supporting the main findings. Applying a stricter SNP p-value threshold did not result in different IV sets and thus resulted in similar estimates (Figs. S2, S5). Summarizing, these results predict that HMGCR inhibition reduces CAD and concomitantly increases T2D risk.

ANGPTL4

We identified instruments for TG and HDL-C (and apoA-I; Table S1-S2) at the ANGPTL4 region (Fig. 3). Using TG as weights, we found ANGPTL4-instrumented TG decrease associated with reduced CAD risk (OR = 0.45 per SD TG decrease, 95%CI: 0.38–0.52, p = 1.2 × 10^–24, h4 = 1.00) as well as reduced T2D risk (Mahajan T2D OR = 0.66 per SD TG decrease, 95%CI: 0.54–0.80, p = 2.2 × 10^–5) with a strongly overlapping regional association pattern between TG and T2D (Figure S4) and colocalization probability (h4 = 0.98, Fig. 3, Table S3, Figure S3) and similar estimates for FinnGen T2D (OR = 0.70 per SD TG decrease, 95%CI: 0.59–0.83, p = 4.3 × 10^–5 per SD TG decrease, h4 = 0.94) and UKB + LL T2Di (HR = 0.59 per SD TG decrease, 95%CI: 0.43—0.83, p = 1.95 × 10^–3, h4 = 0.44, Fig. 4, Table S3).

Weighting by HDL-C increase showed strong consistency with the TG-weighted results with essentially similar effect estimates of drug target effects. These results were corroborated by colocalization, and MR analyses with apoA-I. In MR analyses using an IV-set based on a stricter p-value threshold (5 × 10^–8), we observed highly similar effect estimates. Using correlated instruments, we again found highly similar effect estimates, with MR sensitivity analyses yielding no evidence of heterogeneity (Cochran’s Q p > 0.05) or directional pleiotropy (Egger-intercept p > 0.05), supporting our main findings (Table S3). Summarizing, ANGPTL4 inhibition is predicted to reduce both CAD and T2D risk, both when using TG weights and HDL-C weights, supported by strong colocalization (Fig. 3).

CETP

We identified instruments for all lipid biomarkers (Table S2) at the CETP region. MR predicted protective effects of CETP inhibition on CAD using weights for LDL-C and TG decrease. Surprisingly, using HDL-C weights, CETP-instrumented HDL increase only suggestively associated with reduced CAD risk despite strong colocalization (OR = 0.89 per SD HDL-C increase, 95%CI: 0.79 to 1.01, p = 0.07, h4 = 0.95). Allowing for correlation between SNPs, more instruments could be included; these analyses yielded similar effect estimates but much stronger p-values (e.g. with IVW using correlated SNPs: OR = 0.87 per SD HDL-C increase, 95%CI: 0.84 to 0.90, p = 1.38 × 10^–13, Table S3). Sensitivity analyses did not suggest heterogeneity or directional pleiotropy. All lipid biomarkers and CAD strongly colocalized at CETP (all h4 > 0.94, Figs. 4, S1, S4) with overlapping regional association patterns (Figure S4).

CETP-instrumented HDL-C increase, LDL-C decrease, and TG decrease were not associated with T2D except suggestively in FinnGen (e.g. OR = 0.49 per SD LDL-C decrease, 95%CI: 0.32–0.75, p = 0.0012, similar per SD TG decrease), though with only modest evidence for colocalization (h4 = 0.44–0.47).

There was no evidence for unbalanced, directional horizontal pleiotropy from MR sensitivity analyses. Analyses on apoA-I and apoB yielded consistent results with those of HDL-C and LDL-C, respectively. Using an IV subset based on a stricter p-value threshold (5 × 10^–8), resulted in essentially the same results, except for CAD for which we found slightly stronger associations (Figs. S2, S5, Table S3). Summarizing, these results predict that CETP inhibition reduces CAD risk and tentatively T2D risk.

APOC3

IVs for all lipid biomarkers were found at APOC3 (Table S1-S2). We found strong evidence of protective effects on CAD across all lipid biomarker weights except for apoA-I (e.g. OR = 0.80 per SD TG decrease, 95%CI: 0.74 to 0.87, p = 1.49 × 10^–9). This was supported by sensitivity analyses (stricter IV p-value threshold, MR sensitivity analyses, MR with correlated instruments), and strong colocalization probabilities (h4 = 1.00), Table S3, Figs. S3-S4). APOC3-instrumented lipid levels associated with increased HbA1c (B = 0.037 ± 0.0097 per SD TG decrease, p = 1.3 × 10^–4) supported by colocalization (h4 = 0.95) and sensitivity analyses, for all lipid biomarkers, again except for apoA-I. However, associations with T2D did not survive multiple testing adjustment and were not consistent across the T2D datasets. Coloc-SuSiE disagreed with our primary colocalization result (coloc-SuSiE max h4 = 0.13), favoring distinct causal variants (h3 = 0.64–0.81) (Table S4, Table S9). Summarizing, these results predict that APOC3 inhibition reduces CAD risk, and concomitantly increases HbA1c with unclear effects on T2D risk.

MTTP

IVs for all lipid biomarkers were found at MTTP (Table S1-S2). Using weights for LDL-C decrease and TG decrease, we found associations with reduced CAD risk (OR = 0.29 per SD TG decrease, 95%CI:0.14 to 0.57, p = 4.1 × 10^–4), with modest support from colocalization (h4 = 0.67). There were too few instruments to assess risk of pleiotropy using sensitivity analyses. There was some evidence of protective effects on T2D but inconsistently across outcome data sets and without support from colocalization. Using MTTP-instrumented LDL-C decrease, we found suggestive evidence of reduced random glucose (B = -0.093 ± 0.026 per SD LDL-C decrease, p = 3.2 × 10^–4, h4 = 0.72) (Table S3). Summarizing, MTTP inhibition is genetically predicted to reduce CAD risk and tentatively reduce random glucose.

Discussion

Here, we aimed to predict T2D risk conferred by currently used lipid-modifying drugs, and those in development, by investigating associations of genetically instrumented lipid levels with T2D risk. Our study confirms previously established associations of HMGCR-instrumented lipid levels with CAD and opposite effects on T2D, suggesting that lipid modification through HMGCR reduces CAD risk but concomitantly increases T2D risk. LDLR, LPA, NPC1L1, and PCSK9-instrumented lipid levels were found to be associated with CAD, but not with T2D, suggesting that lipid modification through these targets reduces CAD risk without elevating T2D risk. We found little evidence for associations of ACLY, ANGPTL3, APOB and LIPG-instrumented lipid levels on either T2D or CAD. Finally, the directionally concordant associations of ANGPTL4 and CETP instrumented lipid levels with T2D and CAD suggest that lipid modification through ANGPTL4 and CETP simultaneously reduces CAD and T2D risk. In both cases, however, these results were not corroborated by anticipated changes in glycemic outcome traits. Modification through APOC3 is predicted to reduce CAD risk with a concomitant increase HbA1c, with unclear effects on T2D risk. Modification through MTTP tentatively reduces CAD risk and random glucose, with unclear effects on T2D risk.

We first discuss our findings for ANGPTL3, ANGPTL4 and APOC3 as these genes all encode proteins that inhibit LPL, the enzyme responsible for hydrolysis of triglycerides in apoB-containing lipoproteins. While LPL has been extensively studied for its pivotal role in triglyceride metabolism, a recent MR analysis shows that increased LPL activity is not only associated with reduced triglycerides but—important for the current study—also by a reduced risk of CAD and T2D [74]. Drugs that target ANGPTL3, ANGPTL4, and APOC3 are all developed to correct dyslipidemia and to ultimately reduce the risk of CAD. Based on the shared biological function of these inhibitors, it can be expected that the MR outcomes would be similar, but this is not the case. As already indicated above, we find no evidence for ANGPTL3 to affect CAD or T2D. Similar outcomes have been reported in a drug-target MR study by Richardson [41] and a two-sample MR and interaction MR study by Gagnon [74]. In another drug-target two-sample MR study, Landfors et al. [44] likewise reported no effect of ANGPTL3 on CAD. However, the latter investigators also showed—using exome sequencing data – that ANGPTL3 protein-truncating variants do protect against CAD. The latter results are in line with studies showing that rare ANGPTL3 variants are associated with reduced CAD risk [75–77]. In fact, these studies supported the development of ANGPTL3 inhibitors to lower the risk of atherosclerosis for which preliminary evidence exists [78, 79]. Collectively, these inconsistent results illustrate that MR evidence should not be interpreted in isolation of trial data and rare-variant analyses.

In line with previous studies [41, 44, 74], our investigation suggests that ANGPTL4 inhibition reduces the risk of both CAD and T2D. This result is almost entirely attributable to carriers of a coding variant (p.E40K) in all cohorts studied. This variant is associated with improved glucose homeostasis, lower triglycerides, and a reduced risk of T2D as well as CAD [80, 81]. More recently p.E40K was shown to impair the potential of ANGPTL4 to inhibit LPL and can thus be classified as a mutation [82]. Finally, in line with earlier studies [41, 42, 44], our study shows that genetic APOC3 inhibition reduces the risk of CAD but not T2D. The differential effects of genetic inhibition of ANGPTL3, ANGPTL4 and APOC3 likely relate to their modes of action that can extend beyond LPL inhibition. ANGPTL3 for example does not only inhibit LPL but also endothelial lipase [83–85] which together results in reductions of HDL-C, LDL-C, and TG. ANGPTL4 primarily acts through inhibiting LPL which results in a decrease in TG and an increase in HDL-C [86–88]. ApoC-III inhibits LPL but in addition to other actions also interferes with hepatic lipoprotein clearance [89, 90] which, in combination, results in a reduction of HDL-C, LDL-C and an increase of HDL.

Our study supports the development of ANGPTL4 inhibitors to simultaneously decrease CAD and T2D risk. So far, two approaches are described. Mouse studies that tested antisense oligonucleotides against ANGPTL4 have shown promising reductions in cholesterol, triglyceride and glucose [91] as well as atherosclerosis [92] but after a first study in humans, further clinical development of this compound was halted. In addition, a monoclonal antibody against ANGPTL4 was recently shown to be safe and efficacious to lower triglycerides and remnant cholesterol tested in a phase 1b/2a clinical trial [23] but it is too early to discuss effects on T2D. Next to ANGPTL4, our study also provides evidence that genetic CETP inhibition reduces the risk of CAD as well as T2D albeit in only the FinnGen cohort and with modest evidence of co-localization. This result is concordant with a recent MR study in which on-target CETP inhibition was predicted to reduce the risk of CAD and diabetes [29]. Similarly, the outcomes of pharmacological CETP inhibition collectively show that there is a consistent reduction in the risk of new-onset diabetes [38]. The development of CETP inhibitors has been cumbersome due to off-target effects and futility observed in cardiovascular outcomes trials with two other compounds [93, 94] leaving only one CETP inhibitor, Obicetrapib, in development [30]. Further study is needed to further elucidate the mechanism through which genetic ANGPTL4 and CETP inhibition affect the risk of T2D. When concentrating on plasma lipid levels as exposures, genetic ANGPTL4 inhibition operates through increasing HDL-C and decreasing TG, while genetic CETP inhibition on top also reduces LDL-C. Possibly, the effects are in both cases associated with marked effects on HDL-C highlighted by modest to strong colocalization for both ANGPTL4 and CETP in our study.

The fundamental assumption underlying MR is that of gene-environment equivalence [51, 95]. In the present study, we assume that genetic variation in circulating lipid levels has effects on downstream traits (i.e. CAD, T2D) that are equivalent to those of environmental variation (e.g. pharmaceutical modification) of lipid levels. This assumption is in our view met as cis-variants for a drug target locus plausibly affect the same pathways as pharmaceutical modification of that same target. Quantitatively the assumption might be violated as MR estimates are thought to represent lifetime effects of exposure, whereas pharmaceutical effects are dose, timing, duration dependent, and potentially reversible [96, 97], though this could also be considered an advantage of MR over trials.

Some caution is warranted in interpreting our results. First, perhaps redundantly, absence of evidence (i.e., non-significant results, low colocalization probabilities) does not equal evidence of absence. Each drug target in our study was evaluated using locus-specific instruments, possibly inducing variation in power. Second, it has been argued that in drug target MR, SNP weighting for downstream biomarkers, in our case HDL-C, LDL-C, and TG, in fact tests the total effect of drug target inhibition [47, 55]; using the results of each individual lipid to infer the relevant biological mechanism is not necessarily appropriate. Multivariable cis-MR, in which direct effects of each biomarker are estimated independent of each other, may be a solution. This however would require a -currently unavailable- multitude of powerful cis-instruments per biomarker. Furthermore, sensitivity coloc-SuSiE analysis was only successful in a minority of our analyses. Coloc-SuSiE failed when it was not able to identify credible sets, likely due to modest SNP p-values for some traits. In Results S2, we list those gene-biomarker-outcome trios to we assign the most confidence due to converging colocalization evidence, as well as those in which some caution is warranted due to failed coloc-SuSiE or disagreement between coloc-SuSiE and coloc. Nevertheless, colocalization methods have been observed to disagree, even in experiments that are designed to favor colocalization [98]. Finally, we base our predictions of drug effects on CAD and T2D outcomes on statistical significance of the cis-MR results, and on support by sensitivity analyses, not on effect estimate size. Whether these genetically predicted drug effects will be clinically meaningful remains to be determined following evaluation e.g. through clinical trials.

Distinct features of the present study are the comprehensive and systematic assessment of drug targets, for the first time including all targets that are currently established or in development. We used genetic instruments to predict drug target effects, reducing bias due to confounding. A wide range of sensitivity analyses were applied to further assess robustness against horizontal pleiotropy and confounding through LD. We performed comprehensive analyses on T2Di obtained from meta-analyzed, individual-level longitudinal cohort data from UKB and LL, as well as from large-scale cross-sectional GWAS data on T2D and a wide range of glycemic outcome traits. Several limitations must be mentioned. First, this study was carried out in populations of European ancestry; portability to populations of other ancestry is uncertain. Second, we used weights for downstream biomarkers e.g., HDL-C increase and LDL-C reduction, to proxy for drug target effects, which some have argued is preferred in MR of drug targets as it more closely represents the mode of action [54]. Others however argue that exploring traits more proximal to gene function, e.g., protein (using pQTLs) and transcripts (using eQTLs), reduces risk of horizontal pleiotropy [53]. Future large-scale tissue-specific pQTL and/or eQTL data, currently not widely available, will facilitate further exploration of this issue, and may provide further validation of our results. Third, available GWAS on intermediate glycemic traits were generally limited in sample size, limiting power for our analyses. Fourth, we could not use genetic rare variation (i.e., minor allele frequencies well below 0.01) for our analysis, as this is not captured well by common GWAS genotyping chips and cannot be reliably imputed. Rare variants include loss-of-function variants and deleterious missense variants that likely have stronger phenotypic effects than the common variants used in the present study. The analyses of such rare variants in whole-exome/whole-genome sequencing data, which are beyond the scope of the present study, may yield additional insights [28]. Fifth, we did not investigate drug-drug interaction in T2D risk. Though also beyond the scope of the present study, such analyses would be informative given that different classes of lipid modifying drugs are often administered together, e.g. statins (target: HMGCR) and Ezetimibe (target: NPC1L1). Finally, although we performed extensive sensitivity analyses (i.e., pleiotropy-robust MR, genetic colocalization, and multiple outcome datasets) to mitigate potential violations of MR assumptions, these assumptions ultimately remain untestable. Randomized controlled trials may provide more conclusive evidence.

In conclusion, this cis-MR study genetically predicts that HMGCR lipid modification reduces CAD risk and increases T2D risk, consistent with known T2D effects of statins. Drugs targeting APOC3, LDLR, LPA, MTTP, NPC1L1, and PCSK9 are genetically predicted to reduce CAD risk, without affecting T2D risk. For ACLY, ANGPTL3, APOB, and LIPG, we found no evidence of an effect on either CAD or T2D risk. ANGPTL4 and CETP lipid modification are genetically predicted to reduce both CAD and T2D risk, with the most consistent evidence for ANGPTL4. These comprehensive findings, considering lipid targets of established drugs and a palette of novel drugs, underscore genetic evidence for variation in diabetes-related side-effects of different lipid-modifying drugs, with potential relevance for future clinical trials and personalized treatment decisions.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

CAD

Coronary artery disease

GWAS

Genome-wide association study

HDL-C

High density lipoprotein cholesterol

IV

Instrumental variable

LD

Linkage disequilibrium

LDL-C

Low density lipoprotein cholesterol

LL

Lifelines Cohort and Biobank

MR

Mendelian randomization

SNP

Single nucleotide polymorphism

T2D

Type 2 diabetes

TG

Triglycerides

UKB

UK Biobank

Author contributions

CHLT, HS, RFPD, JAK conceived the study. CHLT designed the study. CHLT, ZC and RDT performed data analysis; CHLT, ZC, RDT, LL, RFPD, JAK, JJPK drafted the manuscript. All authors revised the manuscript and contributed intellectual content.

Funding

The work was carried out with unrestricted funding from NewAmsterdam Pharma, who are developing Obicetrapib, a CETP inhibiting agent (Educational Grant 112346 awarded to HS).

Data availability

UKB individual-level data can be obtained through a data access application available (https://www.ukbiobank.ac.uk/). Researchers can apply to use the Lifelines data used in this study. More information about how to request Lifelines data and the conditions of use can be found on their website (https://www.lifelines-biobank.com/researchers/working-with-us). Summary level data from the UKB + LL meta-analyses on type 2 diabetes incidence for each of the loci are available in Table S6. Publicly available GWAS data can be obtained from the IEU Open GWAS Project (https://gwas.mrcieu.ac.uk/). GWAS data on type 2 diabetes can be obtained from the DIAGRAM consortium (https://diagram-consortium.org/index.htm) and FinnGen r10 data from the Finngen consortium page (https://www.finngen.fi/en/access_results). GWAS data on coronary artery disease can be obtained from the CARDIoGRAMplusC4D consortium page (http://www.cardiogramplusc4d.org/). GWAS data on random glucose can be obtained from the MAGIC consortium page (http://magicinvestigators.org/). All harmonized SNP-trait summary data that is required to replicate our Mendelian randomization and genetic colocalization analyses is available in Table S1. Analysis code can be shared upon reasonable request.

Declarations

Ethics approval and consent to participate

UK Biobank received ethical approval from the National Information Governance Board for Health and Social Care and the National Health Service North West Centre for Research Ethics Committee (Ref: 21/NW/0157). All participants provided informed consent at recruitment to the study for their data to be used for health-related research that was in the public interest. The Lifelines protocol was approved by the UMCG Medical ethical committee under number 2007/152. All participants signed an informed consent form. This study also used aggregated GWAS data, the underlying cohorts of which received ethical approval. We refer to the original GWAS for details (Table 1).

Competing interests

The work was carried out with unrestricted funding from NewAmsterdam Pharma, who are developing Obicetrapib, a CETP inhibiting agent. MD, JJPK are employed by New Amsterdam Pharma, as chief development officer, and chief scientific officer, respectively.