Mendelian Randomization Study for Genetically Predicted Polyunsaturated Fatty Acids Levels on Overall Cancer Risk and Mortality

This article is featured in Highlights of This Issue, p. 1001

Globally, the annual cancer incidence is expected to rise from 14 million to 22 million by 2032 (1). This large increase in cancer case numbers will impose an enormous burden on health care systems. To reduce the impact of the cancer burden, there has been considerable epidemiologic research into which modifiable risk factors cause cancer. However, due to methodologic limitations, drawing definitive conclusions regarding causality is difficult.

Polyunsaturated fatty acids (PUFA) are sourced primarily from the diet and are of potential population health relevance as PUFAs have been shown to reduce the progression of cancer cells through apoptosis, anti-inflammation, and inhibition of angiogenesis (2, 3). As such, the protective effect of PUFAs on cancer is well supported in many animal studies (4–6). PUFAs are categorized into different classes, omega-3 and omega-6, according to their chemical structure. Being a biomarker that can be modified via diet and supplements, PUFAs have gained interest among the research community with previous studies showing an association between PUFA levels and reduced cancer incidence (7–9).

Few observational studies examining PUFA dietary intake and supplementation have provided conflicting results for specific types of cancers (3, 10–14). A large prospective study conducted by Brasky and colleagues showed that long chain omega-3 fatty acids are associated with an increased risk of prostate cancer (15). Conversely, a cohort study in Finland reported that most PUFAs are associated with a reduction in risk of prostate cancer (16). These inconsistent findings may be due to the shortcomings of the observational study design, which can be subject to confounding and reverse causation. In contrast, relatively few studies have been conducted on PUFA levels and cancer mortality, but those that have suggest there may be survival benefits of PUFAs on certain cancer types (17–19). Moreover, there is no study to date evaluating whether PUFAs have an impact on overall cancer risk.

To overcome the limitations of observational study approaches we have utilized Mendelian randomization (MR) to examine the causal relationship between PUFA and all cancers, using data from the UK Biobank cohort (20). MR is analogous to a “natural” randomized controlled trial (21). In MR, genetic variants associated with a risk factor (our exposure of interest) are used as instrumental variables (IV) to infer the true relationship between the risk factor and outcome (21). For a valid inference, three MR assumptions must be satisfied (i) the genetic variants must be associated with the risk factor, (ii) the genetic variants must not be associated with confounders of the risk factor–outcome relationship (the independence assumption), and (iii) the genetic variants must act on the outcome only via their effect on the risk factor (the exclusion restriction assumption; Fig. 1).

Our primary objective was to investigate whether an increase in genetically predicted PUFA levels are associated with overall cancer risk and overall cancer mortality. Second, our aim was to identify whether these associations differed by sex and individual types of cancer.

We included participants from the UK Biobank for our study, which is a prospective, population-based cohort from the United Kingdom. The UK Biobank has collected a wide range of data from 2006 to 2010 for 502,649 participants, age ranging from 37 to 73 years (20). International Classification of Diseases (ICD-10) coding has been used to categorize the illnesses in UK Biobank participants (22). All participants provided informed written consent, the study was approved by the National Research Ethics Service Committee North west – Haydock, and all study procedures were performed in accordance with the World Medical Association Declaration of Helsinki ethical principles for medical research. The UK Biobank data can be accessed via http://www.ukbiobank.ac.uk/.

Affymetrix UK BiLEVE Axiom array or the Affymetrix UK Biobank Axiom array was used for the genotyping (23). A detailed description of the characteristics of the UK Biobank study population and quality control of the generated genetic data can be accessed from the previous publication of Bycroft and colleagues (24). As a brief summary, 805,426 genotyped SNPs were available for 488,377 samples, and this data was imputed to HRC Release 1.1 (25) and UK10K haplotype resources (23).

Prior to analysis, SNPs with minor allele frequency (MAF) < 0.001, and imputation quality score (INFO) < 0.6, were filtered, as well as SNPs which are not present in HRC Release 1.1 reference panel were filtered before the analysis, leaving 14,381,678 SNPs. From this wider UK Biobank imputed dataset the overall cancer risk analysis was restricted to 270,342 cancer-free controls and 46,155 ICD-10 cancer cases, and the overall cancer mortality analysis to 270,342 healthy controls and 6,998 cancer cases following exclusion of related samples and population outliers (Supplementary Table S1). For both risk and survival, data was analyzed using PLINK v2.00 alpha under logistic regression model fitting the first 10 principal components (PC), sex and age as covariates. A total of 23,352 individuals included in the study were diagnosed with cancer before the UK Biobank recruitment date (prevalent cases). A total of 20,865 individuals were diagnosed after the UK Biobank recruitment date (incident cases) and there were no accurate records for the diagnosis date available for 1,938 patients. Usually, in a prospective cohort like this, we should consider including only the incident cases, and not the prevalent cases. However, the introduction of prevalent cases is unlikely to introduce bias as the genetic instruments are inherited and lifelong. The numbers of cases and controls used for each individual cancer are illustrated in Supplementary Tables S1, S2, and S3. The cancer case–control definitions are used as the same as those in Ong and colleagues (26).

The genetic variants used as IVs were selected from two previously reported genome-wide association studies (GWAS) from the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) consortium (27, 28). For each PUFA, up to three genetic instruments were used to identify the causal associations with all cancer risk and mortality (Table 1). All of the identified SNPs operate in the PUFA biosynthetic pathway (29, 30). The chosen SNPs were significantly associated with PUFAs with P < 5 × 10⁻⁸, and within each PUFA were independent of each other (linkage disequilibrium, r² < 0.05). The variance explained for each PUFA as tagged by these SNPs was either moderate (<2.05%) for eicosapentaenoic acid (EPA), α-linolenic acid (ALA), and docosahexaenoic acid (DHA) or large for arachidonic acid (AA), linoleic acid (LA), and docosapentaenoic acid (DPA; Table 1; ref. 31); given our large cancer dataset, these variance explained values are expected to produce either moderately narrow (EPA, ALA, and DHA), or very narrow (AA, DPA, and LA), confidence intervals (CI) on our causal ORs. As a result of competition for enzyme availability within the PUFA biosynthetic pathway there is substantial correlation between the PUFAs, where several SNPs might be implicated in the variation of multiple PUFAs (27). Hence, this should not be misinterpreted as a potential violation of the MR SNP-confounding MR assumption. To ensure that our MR analyses were not biased because of presence of SNP-confounding factor associations, we searched prior MR literature for evidence of potential confounders body mass index (BMI), educational attainment, and height. A schematic diagram summarizing our MR study design is shown in Fig. 1.

First we conducted a GWAS for overall cancer risk (46,155, cases, 270,342 controls), overall cancer mortality (6,998 cases, 270,342 controls), sex-specific stratified analysis, and individual cancer risk for specific cancers (breast, prostate, melanoma, pancreatic, lung, and colorectal) to identify the GWAS significant (5 × 10⁻⁸) SNPs which are associated with the former (sample size of individual cancers are provided in Supplementary Table S1). The quality control procedures are described above.

Using two sample summary data, MR analyses were performed with R 3.3.2. to determine the causal role of PUFA on overall cancer risk and mortality. The six PUFAs considered here are interrelated with each other by enzyme-driven pathways, leading to correlations between the traits (6, 32). Given this, there is overlap in the genetic instruments used for each of the PUFAs (see Table 1). To allow direct comparisons with previous work, we present separate results for each of the six PUFAs.

The significance threshold for all cancer risk and mortality is 0.004 [six PUFAs times two outcomes (risk, mortality) require correcting 0.05 by 12 tests]. Furthermore, we performed a sex-specific analysis to determine if these associations varied by sex (significance threshold for this subanalysis 0.002). As a subsequent analysis, we performed MR evaluating the role of PUFA on susceptibility of individual common cancers in the UK Biobank. We also performed meta-analysis for specific cancers if results from previous studies were available. Individual cancers were categorized according to ICD-10 criteria (Supplementary Tables S1 and S2). Given six individual cancers were considered, we set a significance threshold of 0.004/36 = 0.0001. Due to sample size limitations, we did not repeat the cancer mortality analyses on individual cancers.

Single SNP MR estimates were obtained using the Wald-type ratio estimator (33). For PUFA traits with multiple SNP instruments, the inverse variance weighted (IVW) model was used to derive a combined causal estimate (34). In the IVW method, the ratio of SNP-outcome (cancer) effect estimate and the SNP-risk factor (PUFA) effect estimate of each genetic variant is combined into a single-effect estimate by weighting the each genetic variant according to its variance (21).

The MR results for overall cancer risk and mortality are shown in Fig. 2. The results are expressed in terms of a “large change” of 1 standard deviation (SD) in each PUFA. A change of 1.9 units (1 SD) in AA, which is equivalent to moving a person from 16th percentile to the median, showed no association with overall cancer risk (OR, 1.02; 95% CI, 1.0–1.03) or cancer mortality (OR, 1.02; 95% CI, 0.98–1.06; Fig. 2). As for AA, all other all PUFAs results were not significantly associated (all P > 0.004) with overall cancer risk or mortality (Fig. 2).

Overall cancer risk and mortality for a 1 SD increase in AA in men was OR = 1.03 (95% CI, 1.01–1.06; P = 0.014, i.e., not significant after correction for multiple testing) and OR = 1.05 (95% CI, 0.97–1.13), respectively (Fig. 2), as they are interconverted and compete for the same conversion enzymes (35). AA levels were correlated with the other PUFAs (some positively correlated and some negatively; ref. 36) and they showed broadly consistent results, albeit with wider CIs (Fig. 2). For example, a 1 SD increase in LA was associated with lower mortality (OR, 0.88; 95% CI, 0.78–1.00), and there is a negative correlation between the effects of the SNP instruments on the levels of AA and LA (27, 28). ALA was associated with lower risk in men (OR, 0.92; 95% CI, 0.86–0.98; P = 0.015), which is not significant after correction for multiple testing (Fig. 2).

For women, overall cancer risk and mortality for AA was OR = 1.01 (95% CI, 0.98–1.03) and OR = 0.99 (95% CI, 0.85–1.15), respectively (Fig. 2). The other PUFAs yielded similar nonsignificant results (Fig. 2). The mortality results for ALA were nonsignificant with a relatively wide CIs (men, OR = 0.86, 95% CI = 0.74–1.01; women, OR = 0.99, 95% CI = 0.93–1.05; Fig. 2).

The causal association between each PUFA and risk of selected individual cancers is illustrated in Fig. 3. We did not perform sex-stratified individual cancer risk analyses due to low statistical power. We found no association between AA and breast cancer risk (OR, 0.98; 95% CI, 0.94–1.01), prostate cancer risk (OR, 1.02; 95% CI, 0.98–1.06), colorectal cancer risk (OR, 1.05; 95% CI, 0.99–1.11), melanoma risk (OR, 1.04; 95% CI, 0.98–1.11), or lung cancer risk (OR, 1.07; 95% CI, 1.0–1.16).

We performed fixed effect meta-analysis on colorectal cancer combining our estimates with those from May-Wilson and colleagues and the meta-analyzed result showed increased risk for AA and reduced risk for LA (AA, OR = 1.05, 95% CI = 1.03–1.07; LA, OR = 0.95, 95% CI = 0.93–0.97; ref. 37; Supplementary Table S4). Furthermore, we conducted a meta-analysis for prostate cancer risk of our UK Biobank estimates with a MR study by Khankari and colleagues (31). The meta-analysis OR for AA for 1 SD change was nonsignificant (OR, 1.01; 95% CI, 0.99–1.03; Supplementary Table S5). Similarly, for melanoma, we obtained more precise estimates of OR by combining our study with a previously published study by Liyanage and colleagues (AA, meta-analysis OR, 1.03; 95% CI, 1.00–1.07; ref. 38; Supplementary Table S6).

For the first assumption (strong instrument assumption) to be true, it is required that the genetic variants are robustly associated with the exposure of interest. We have selected our IVs from the largest GWAS performed so far for the PUFAs (27, 28). All SNPs were genome-wide significant (P <5 × 10⁻⁸); this is a much more stringent threshold than the traditional F statistic >10 used in conventional MR analyses (39) and confirms our instruments fulfil the first MR assumption. Most IVs explain a good proportion of variance of the PUFAs, which further validates that these instruments are suitable for MR analysis (See Table 1).

With regard to the evidence from MR for causality, spurious findings can occur due to confounding arising from the differences in the underlying genetic architecture of subpopulations (40). Our MR study uses IVs selected from the previous GWAS that adequately controlled the population stratification (27, 28). Overall cancer risk, all cancer mortality, and individual cancer GWAS analysis were limited to defined European populations using principal component analysis, with principal components included as covariates to adjust for any potential residual stratification (26).

The presence of pleiotropic associations of genetic instrument variable may introduce bias into the MR framework if the genetic instrument is associated with the outcome through non-PUFA pathways (potentially violating independence assumption and exclusion restriction assumption). We examined previously published results to assess if our selected SNP instruments had pleiotropic effects on selected potential confounders (height, BMI, and educational attainment; refs. 31, 37, 38). The result of the pleiotropy assessment can be found in Supplementary Tables S7–S9. Various PUFA SNPs, that is, rs174547 and rs10740118 were shown to be associated with height and BMI. Whenever applicable (i.e., for the case of LA, because LA was the only PUFA which has three genetic variants), leave-one-out MR analyses excluding these potentially pleiotropic variant revealed that our results on LA were not meaningfully different although as expected our estimates had wider CIs (Supplementary Figs. S1 and S2).

Previous studies have examined the role that PUFAs may play in cancer prevention (2, 3, 5, 11, 41). To date, our study is the first study to evaluate the relationship between multiple PUFAs and overall cancer risk and mortality through MR approach. We did not find any evidence for associations between genetically determined PUFA levels and overall cancer risk or cancer mortality in our study. When stratifying by sex, we found a suggestive protective effect with ALA in males only, but this did not remain significant after multiple testing corrections. Our results provide no support for an association between PUFA levels and overall cancer risk and mortality. Because even lifelong changes in PUFA levels (as indexed by inherited genetic variation) have little or no effect on risk or mortality, our results suggest that dietary supplementation over a shorter period is unlikely to affect cancer diagnosis or death rates.

The sample distribution for our overall cancer risk study comprises approximately equal proportions of prevalent and incident cases. This might have influenced the findings on aggressive forms of cancer in the population (e.g., lung cancer, pancreatic cancer) as they are likely to have been under-sampled in the UK Biobank cohort. However, given that our individual cancer analyses reveal at most weak associations, it is unlikely that our overall cancer MR estimates were heavily biased by under-sampling of these cancers.

An observational study conducted in Japan reported that consumption of fish ≥3 times a week (fish is rich in PUFAs; ref. 42), reduced the lung cancer risk by 4-fold (RR, 0.23; 95% CI, 0.10–0.54; ref. 43). Meta-analysis of eight prospective epidemiologic studies identified that high PUFA intake had no effect on lung cancer risk (RR, 0.91; 95% CI, 0.78–1.06; ref. 10). Our results suggested that there is no association between lung cancer and PUFAs (for 1 SD increase in AA, OR, 1.07; 95% CI, 1.0–1.16; Fig. 3).

The evidence from previous observational studies for the association between PUFAs and breast cancer suggested there were protective effects (13, 44). However, we found no association between genetically determined PUFA concentrations and breast cancer (for 1 SD change in AA, OR, 0.98; 95% CI, 0.95–1.01; Fig. 3).

Previous observational studies of prostate cancer and PUFAs have yielded inconclusive findings (45–52). A previous MR study conducted by Khankari and colleagues showed little evidence that PUFAs are associated with prostate cancer (AA: OR = 1.01, 95% CI = 0.99–1.03; LA: OR = 1.0, 95% CI = 0.98–1.02; ref. 31). Our results were consistent with the above study suggesting a null association between PUFAs and prostate cancer risk (AA: OR = 1.02, 95% CI = 0.98–1.06; LA: OR = 0.96, 95% CI = 0.88–1.11). A meta-analysis of our study with the above mentioned study further reduced the CI on the null estimates (AA: OR = 1.01, 95% CI = 0.99–1.03; LA: OR = 0.99, 95% CI = 0.98–1.02; Supplementary Table S5).

Our study estimated an OR for a 1 SD change in LA and AA on colorectal cancer risk of OR = 0.94 (95% CI, 0.78–1.12) and OR = 1.05 (95% CI, 0.99–1.05), respectively (Fig. 3). When we meta-analyzed our study with another larger MR study (ref. 37; 9,254 cases and 18,386 controls of European ancestry), we found a similar result; increased LA decreases the risk (OR, 0.95; 95% CI, 0.93–0.97) whereas increased AA increases the risk (OR, 1.05; 95% CI, 1.03–1.07). EPA, DPA, and DHA showed no associations in either study (Supplementary Table S4). These MR results show the same direction of effect as published observational studies; increased AA (comparing the upper and lower quartiles of intake) has been shown to be associated with an increase in the risk of colorectal cancer (OR, 2.03; 95% CI, 1.16–3.54; P = 0.001) among males (53). A recent case–control association study revealed similar outcomes for AA (OR, 2.18; 95% CI, 1.16–4.08; ref. 54).

A recent meta-analysis of observational studies reported that a high dietary intake of PUFAs are protective for pancreatic cancer (RR, 0.87; 95% CI, 0.75–1.0; ref. 55). Our study found no evidence for an association between genetically determined PUFA levels and risk of pancreatic cancer (AA: OR, 0.94; 95% CI, 0.73–1.24; Fig. 3).

Observational study findings are inconclusive regarding the association of PUFAs and melanoma risk (56, 57). Our meta-analyzed results combining the estimates from Liyanage and colleagues (38) suggested that there is no association between PUFAs and melanoma after the adjustment for multiple testing (AA: OR, 1.03; 95% CI, 1.00–1.07; Supplementary Table S6).

Our study had several strengths. Provided that the MR assumptions are met, the MR method provides a more robust inference on causality free from the traditional caveats of observational studies. As genetic variants are randomized at meiosis, MR estimates are less likely to be biased because of confounding and reverse causation. Furthermore, the majority of observational studies examining the effect of PUFAs on cancer risk were based on food frequency questionnaires, which are self-reported, and hence highly susceptible to recall bias and measurement errors (58). In contrast, our PUFA SNP instruments were previously shown to encode biomarkers which are relevant to the PUFA metabolic pathway (30), providing an alternative angle to evaluate the relationship of PUFA and cancers via genetic predictors of PUFAs that are not affected by bias from self-report.

Our study was well powered to detect moderate effect sizes and hence enable precise estimates on the cancer ORs with tight CIs. Here, our statistical power comes from the large size of the UK Biobank cohort (N > 300,000), and that in general, our instruments explain a high proportion of PUFA level phenotypic variance (r² for AA = 33.1%, LA = 8.3%). The variance explained for the PUFAs here is much higher than in many conventional MR studies where the instrument is commonly r² ∼ 1%–3%. Furthermore, the UK Biobank participants in our study consist of an ethnically homogenous population, which allowed us to draw robust inferences. Although the ethnic homogeneity of the UK Biobank may suggest that these findings are not readily generalizable to other populations, it can also be inferred that population stratification is unlikely to be biasing the results given such artefacts usually inflate chances of type I error while our findings were predominantly null. All the reported cancer cases and cancer-related deaths were confirmed with the UK national cancer and death registries, ensuring that we had accurate data.

Our study had a few limitations. First, the phenotypic variance tagged by SNP instruments for some specific PUFA traits were low (i.e., DHA = 0.65%, ALA = 1.03%). However, this is unlikely to have affected the statistical power for our MR analyses as the sample size(s) in our study were large enough to overcome the low variance tagged by these SNPs and provide precise estimates, as shown by the narrow CIs on our estimates (Figs. 2 and 3). Second, some of our SNP instruments are pleiotropically associated with several potential confounders. We investigated the potential for this to adversely affect our conclusions in a previous MR paper using the same SNPs. Height is an established all cancer risk factor (26) and if our PUFA SNP instruments affect height, there is the potential for these SNPs to affect all cancer risk through a horizontal pleiotropy pathway (i.e., not via PUFA). However, when we examined this previously (38), our selected SNPs either affect height via PUFA (vertical pleiotropy) or have such a tiny effect on height (Supplementary Table S7) that it was very unlikely that our conclusions would be altered. Similarly for the other potential confounders, BMI and educational attainment, our selected SNPs have only a tiny effect on the trait (Supplementary Tables S8 and S9), limiting the effect these could have on our causal estimates.

In some MR studies, many SNP instruments are required to explain sufficient variance in the exposure for adequate power; in such studies statistical robustness approaches are typically applied. Robustness approaches include leave-one-out analysis (where single SNPs are sequential dropped and results are recomputed) and regression-based methods to infer the effect of outliers (34, 59, 60). For the PUFAs considered here, a small number of SNPs (between 1 and 3) explain substantial trait variance, with the selected SNPs all falling within the well understood PUFA biosynthetic pathway (61). Given the very low number of instrument used for each PUFA trait (n < 4), we are unable to apply these statistical robustness approaches to every PUFA. However, we have conducted leave-one-out analysis whenever possible.

Our results showed no association between PUFAs and overall cancer risk and mortality for a large change in genetically predicted PUFA levels (1 SD change correspond to a change of 16th percentile to the median in PUFA levels). Supported by approaches used in previous studies, we assumed a linear relationship between PUFAs and cancer risk, and disregard the need to consider potential nonlinear associations (31, 37). Our MR inference cannot rule out predominantly null association(s) in our study, being a product of unmodeled nonlinear nonzero associations, which experienced a dilution or “cancelling out” effect; although this explanation is extremely unlikely. Furthermore, our sample sizes were not large enough to improve precision of our estimates on individual cancer mortality, and other stratified analyses.

There is an extensive collection of studies on the purported beneficial effects of PUFAs on cancer risk and progression. However, our study showed no evidence for an association between genetically predicted increased PUFA levels on overall cancer risk and mortality, and our CIs were able to rule out all but very small effects (e.g., OR, 1.02; 95% CI, 1.00–1.03 per 1 SD change in AA). However, there was suggestive evidence that increasing ALA levels may be associated with a reduction in overall cancer risk in men, while increasing LA and reducing AA levels were identified to be protective for colorectal cancer risk. Future studies are warranted to replicate these findings on individual cancers.

No potential conflicts of interest were disclosed.

Conception and design: U.E. Liyanage, J.-S. Ong, M.H. Law, S. MacGregor

Development of methodology: U.E. Liyanage

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J. An, M.H. Law, S. MacGregor

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): U.E. Liyanage, J.-S. Ong, J. An, M.H. Law, S. MacGregor

Writing, review, and/or revision of the manuscript: U.E. Liyanage, J.-S. Ong, J. An, P. Gharahkhani, M.H. Law, S. MacGregor

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): U.E. Liyanage, J.-S. Ong

Study supervision: M.H. Law, S. MacGregor

This work was conducted using the UK Biobank Resource (application number 25331). We thank Scott Wood and John Pearson from QIMR Berghofer for IT support. This work was supported by a project grant (1123248) from the Australian National Health and Medical Research Council (NHMRC). J.-S. Ong and U.E. Liyanage received scholarship support from the University of Queensland and QIMR Berghofer Medical Research Institute. S. MacGregor was supported by a fellowship from the Australian Research Council.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.