Background
Early interest in the cardiovascular effects of polyunsaturated fatty acids (PUFA) emerged from observational studies conducted between the 1950s and 1970s indicating that populations with a high intake of omega-3 PUFA had lower rates of mortality from cardiovascular diseases (CVDs) [
1,
2]. Numerous subsequent randomized controlled trials (RCTs) indicated that dietary substitution of carbohydrates or saturated fatty acids by PUFA had a protective effect on intermediate outcomes, such as a reduction in low-density lipoprotein (LDL)-cholesterol and triglycerides [
3]. However, the hypothesized cardioprotective role for omega-3 and omega-6 PUFA has been challenged by a recent series of Cochrane’s systematic reviews of RCTs of dietary advice or supplementation, which suggest little to no benefit [
4‐
6]. Overall, most RCTs on PUFA intake included in these systematic reviews were at moderate to high risk of bias and there is large uncertainty on the evidence linking PUFA to several cardiovascular outcomes [
4‐
6]. Despite being the gold-standard study design for testing the effect of clinical interventions, in practice RCTs are often limited in statistical power, breadth of outcomes analyzed, and have high risk of bias. Therefore, integrating multiple lines of evidence is key to improve causal inference on the role of PUFAs in CVD aetiology.
Alpha-linolenic acid and linoleic acid are omega-3 and omega-6 PUFAs, respectively, that cannot be produced endogenously by humans and, therefore, need to be obtained from diet. Other omega-3 and omega-6 PUFAs can be produced endogenously through a series of elongation and desaturation reactions [
7]. Circulating PUFA concentration is influenced by environmental and genetic factors [
8‐
13], and the measurement of circulating PUFA can be used as an objective biomarker of PUFA intake, avoiding well-known biases in self-reported assessment of dietary intake. Comparing individuals genetically predisposed to higher or lower circulating PUFA can be used to probe the lifelong effect of circulating PUFA on CVD risk using Mendelian randomization [
14]. Mendelian randomization uses genetic variants associated with biomarkers as instrumental variables to assess their effect on disease aetiology. This approach was developed to improve causal inference by taking advantage of unique properties of genetic variants: (i) germline genotypes remain unchanged throughout one’s life, (ii) the random allocation of parental alleles at meiosis reduces confounding by generating balanced groups and (iii) the unidirectional flow of biological information (from genotype to phenotype) avoids reverse causation [
15‐
17]. In addition to indexing lifelong exposure to a biomarker of interest, genetic variants are subject to relatively little bias due to measurement error [
18].
Previous Mendelian randomization studies have reported conflicting findings regarding the relationship between circulating PUFAs and CVDs risk [
19‐
25]. Overall, shorter (e.g. α-linolenic acid and linoleic acid) and longer (e.g. arachidonic acid) chain PUFA have been associated with lower and higher risk of CVDs, respectively [
19‐
25]. These seemingly contradictory findings have been largely attributed to the inclusion of genetic variants mapping to the
FADS locus in the analyses, which contains the genes
FADS1 and
FADS2 encoding, respectively, the desaturases delta-5 (D5D) and delta-6 (D6D), key enzymes catalyzing rate-limiting steps in PUFA biosynthesis [
22]. Genetic variants modulating the expression/activity of D5D and D6D will lead to changes in shorter- and longer-chain PUFAs in opposite directions, which likely explains such contrasting MR findings of lower vs higher CVD risk being associated with shorter- vs longer-chain PUFAs, respectively. In addition, these
FADS variants are highly pleiotropic and associated with numerous non-fatty acid traits including triglycerides, low-density lipoprotein (LDL)-cholesterol and fasting glucose [
22]. On the one hand, given the well-established link of
FADS1/2 with PUFA biosynthesis, single-nucleotide polymorphisms (SNPs) in the vicinity of the
FADS locus can add to the evidence on the involvement of fatty acids in the development of cardiovascular diseases. On the other hand, the fact that SNPs nearby
FADS1/2 are highly pleiotropic and not specific for individual fatty acids or fatty acid classes complicate inferences on the causal role of circulating fatty acids on CVD risk from Mendelian randomization studies solely/predominantly relying on genetic variants within this locus.
Some of these limitations in current Mendelian randomization studies could be addressed by incorporating into the analyses numerous genetic variants modulating circulating fatty acids via different pathways. Consistency of findings across variants from multiple loci would increase confidence in the results and allow the use of Mendelian randomization methods that require multiple independent variants and are more robust to violations of the method’s assumptions. However, the modest size of current genetic association studies (GWAS) on circulating fatty acids (up to ~13,000) have only allowed the discovery of a small number of genetic variants strongly and independently associated with circulating PUFA [
8,
9,
11‐
13].
The aims of this study were (a) conducting the largest GWAS on circulating PUFA to date including a sample of 114,999 individuals and (b) using two-sample Mendelian randomization to investigate the involvement of circulating omega-3 and omega-6 fatty acids on a wide range of cardiovascular disease endpoints in up to 1,153,768 individuals of European ancestry (i.e. coronary artery disease, ischemic stroke, haemorrhagic stroke, heart failure, atrial fibrillation, peripheral arterial disease, aortic aneurysm, venous thromboembolism and aortic valve stenosis).
Discussion
Using the largest-scale genetic association data available for fatty acids from over 114,000 UK Biobank participants, we identified a much larger number of genetic variants strongly and independently associated with circulating PUFA concentration compared to previous GWAS [
8,
9,
11‐
13], which enabled us to conduct key sensitivity analyses such as MR-Egger and multivariable MR, both of which preferably require a large number of independent instruments. As expected, many of these genetic variants mapped to genes involved in lipoprotein-related metabolism. This poses a challenge to Mendelian randomization studies of fatty acids since using SNPs from a single gene region involved in fatty acid metabolism (e.g.
FADS) is likely to be more specific but not amenable to most pleiotropy-robust Mendelian randomization methods, while using SNPs across the genome is likely to introduce non-specificity but allow the use of pleiotropy-robust methods.
When using these multiple genetic variants as instruments for Mendelian randomization, our findings did not confirm the inverse association observed in conventional multivariable regression and provided weak evidence of higher genetically predicted DHA (and total omega-3 fatty acids) concentration being related to higher risk of some cardiovascular endpoints. However, overall, Mendelian randomization findings did not pass our criteria for multiple testing correction and were attenuated when accounting for LDL-cholesterol/apolipoprotein B or excluding a SNP in the vicinity of the FADS locus. Mendelian randomization findings for higher genetically predicted linoleic acid (and total omega-6) concentration were inconsistent across different cardiovascular endpoints and methods and did not confirm the inverse association observed in conventional multivariable regression. There was weak evidence of higher genetically predicted linoleic acid being related to lower risk of ischemic stroke and peripheral artery disease after accounting for LDL-cholesterol/apolipoprotein B. Despite the large increase in the number of instruments in our analyses, there remains considerable imprecision in estimates for the effect of circulating fatty acids on the risk of some cardiovascular disease outcomes. As an example, IVW estimates for the relation of DHA with the risk of aortic aneurysm was 1.11 (95% CI 0.95–1.30), while IVW estimates for the relation of our positive control exposure (LDL-cholesterol) with the same outcome was 1.17 (95% CI 1.02–1.35). This indicates that we cannot confidently rule out the presence of clinically meaningful effects due to the considerable uncertainty in some results.
Previous metanalyses of classical observational studies indicate that higher circulating long-chain omega-3 and omega-6 PUFA are either not associated or associated with lower risk of coronary artery disease and stroke [
61‐
64]. Cochrane recently published a series of systematic reviews of RCTs with the overall conclusion that increasing omega-3, omega-6 or total PUFA intake, via supplementation or diet, has modest to no effect on CVD events or mortality [
4‐
6]. Since then, further large-scale RCTs on long-chain omega-3 PUFA have been published and yielded conflicting results [
65‐
67]. In addition, most Mendelian randomization studies, as well as classical observational studies and RCTs, have focussed on exploring the effect of PUFA on the risk of coronary artery disease and, to a lesser extent, ischemic stroke. The effects of these fatty acids on other types of cardiovascular endpoints, such as heart failure and atrial fibrillation, have been under explored.
Integrating multiple lines of evidence to resolve controversies in research on cardiovascular health effects of fatty acids is essential. However, directly comparing results from different study designs in this context is not straightforward. For illustration, in intervention studies, the effect of PUFA supplementation or diet intake on CVD endpoints is frequently tested over relatively short periods of time due to logistical issues [
4‐
6] and, in some instances, may depend on the overall diet composition [
3]. On the other hand, genetic proxies of circulating PUFA affect their metabolism (not intake) are assumed to have lifelong effects and have pleiotropic effects on lipoprotein-related traits.
Mendelian randomization can provide a valid test of the presence of a causal effect if genetic variants are relevant and valid instruments for the exposure of interest.
Regarding instrument relevance, we have selected independent SNPs strongly associated with circulating PUFA concentration, which explained from 4.8 to 7.9% of phenotypic variance (mean F statistics 109–201) among the UK Biobank participants (discovery sample). In addition, we replicated these associations in an independent dataset [
12] using the same NMR metabolomics platform as the one used in the UK Biobank participants (median sample size 13,516), where SNPs explained 3.1 to 6.9% of phenotypic variance in circulating PUFA. This indicates that bias due to weak instruments is unlikely to be substantial in our analyses, even though bias due to winner’s curse (due to using the UK Biobank to select SNPs and estimate their effect on PUFA) could affect the magnitude of effect estimates. There was little evidence that the selected SNPs impacted PUFA composition in independent datasets with more detailed data on individual PUFA, particularly after removing SNPs nearby the
FADS locus. This might be related to differences in assays, units of analyses or statistical power.
Instrument validity requires that any effect from genetic instrument to the outcome is completely mediated by the exposure of interest. This assumption could be violated in several scenarios, such as in the presence of confounding due to population stratification or horizontal pleiotropy.
Bias by population stratification could result if heterogeneity in genetic ancestry in a given sample was related to different distributions of genetic instrument and outcome. To mitigate that, heterogeneity in genetic ancestry was accounted for when generating genetic association data for cardiovascular outcomes in all data sources by correcting for genomic inflation factor, adjusting for principal components of ancestry or using mixed linear models as detailed in Table S2 (Additional file
1). In addition, the overall consistency of findings across multiple studies provides some reassurance against findings being explained by population stratification.
Horizontal pleiotropy is one of the major threats to the validity of Mendelian randomization studies. By conducting a series of analyses for functional mapping and annotation of fatty acid genetic association results, we showed that genetic variants associated with fatty acids are strongly enriched for genes and pathways involved in lipoprotein metabolism, particularly (triglyceride-rich or high-density) lipoprotein particle remodelling, apolipoprotein binding and reverse cholesterol transport. For illustration, some PUFA SNPs mapped to genes encoding proteins targeted by lipid-lowering drugs, such as
HMGCR (Entrez Gene
3156),
PCSK9 (Entrez Gene
255738) and
CETP (Entrez Gene
1071) [
68,
69]. Considering the pivotal role of lipoprotein metabolism in the aetiology of several cardiovascular diseases, this stresses that the assumption of no horizontal pleiotropy (or the weaker versions of this assumption by MR-Egger) in our analyses is likely implausible. We tried to mitigate that by using multivariable MR to account for total fatty acids, given the lack of specificity of the selected instruments for specific fatty acids, and for triglycerides, LDL-cholesterol or apolipoprotein B, which are key determinants of several CVDs reflecting lipoprotein metabolism. Accounting for LDL-cholesterol/apolipoprotein B revealed a potential direct protective effect of linoleic acid on the risk of ischemic stroke and peripheral artery disease, suggesting that horizontal pleiotropy via LDL-cholesterol might have masked some true underlying protective effect of linoleic acid. In addition, Mendelian randomization findings for the relation between DHA and several CVDs were attenuated when excluding SNPs in the vicinity of the
FADS locus. On the one hand, SNPs regulating
FADS1 and
FADS2 are expected to be more credible instruments given their proximal relation with PUFA biosynthesis. However, these SNPs have been shown to be associated with numerous fatty acids and non-fatty acid traits and may not be valid instruments for circulating DHA [
22]. Results from multivariable regression were often inconsistent with results from MR IVW (e.g. multivariable regression and IVW estimates were in opposite direction for the association between PUFA and CAD risk). While the reason for that is not entirely clear, this is possibly related to pleiotropic genetic instruments since accounting for LDL-cholesterol in MR analyses attenuated those differences.
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.