Causal impact of genetically-determined fish and fish oil intake on epigenetic age acceleration and related serum markers

With oily fish intake as the exposure, 59 independent single nucleotide polymorphisms (SNPs) were identified as instrumental variables (IVs) (Table 1) for use against each of the four biological clock outcome datasets. 2SMR performed for each of these exposure outcome pairs yielded a directionally consistent effect for the IVW method on all four epigenetic age acceleration measures (Fig. 1). Of these four, only PhenoAge reached statistical significance (PhenoAge PIVW=0.0086).

To further scrutinize the relationship between Oily Fish Intake and PhenoAge, additional MR methods including the simple mode, weighted mode, and weighted median were employed. These three methods also yielded a directionally consistent effect with respect to the IVW method, towards the age-decelerating direction. (Figure 1 S). Further sensitivity tests included the leave-one-out analysis to test for the overpowering influence of any single genetic instrument on the overall estimate, and the MR Egger Intercept to test for horizontal pleiotropy (See heterogeneity p-value in Table 2), to ensure no significant heterogeneity across instruments. In each test, there were no significant confounding effects that influenced this relation (Figure 2 S).

For fish oil supplementation, 4 independent SNPs were used as IVs against all four biological clock outcomes (Table 1). In this case, PhenoAge was not significantly modified (PhenoAge PIVW>0.5) by the exposure, though it remained directionally consistent towards age deceleration. GrimAge, however, did reach statistical significance (GrimAge PIVW=0.037), with the MR-Egger method also showing directional consistency (Fig. 2).

For this fish oil and GrimAge exposure outcome pair, all five MR methods yielded directional consistency (Figure). The leave-one-out analysis did not reveal any individual SNP that overinfluenced the result, and the heterogeneity p-value was less than 0.1, indicating confounding from horizontal pleiotropy is not significant (Figure). 3 S 4 S

To ascertain the established effects of oily fish intake on certain serum lipidemia and inflammatory biomarkers, 2SMR was performed on each exposure outcome pair. There were no statistically significant results between oily fish consumption and the evaluated serum lipid biomarkers, although in the case of C-reactive protein, the result approached statistical significance (PIVW=0.064), with both the IVW and MR Egger methods showing directional consistency (Fig. 3).

In the case of fish oil supplementation, there was a statistically significant relationship towards modifying triglycerides in the expected negative direction (PIVW=0.004), with MR-Egger showing directional consistency (Fig. 4). The fish oil supplementation exposure also produced a significant interaction with hsCRP and HDL via the IVW method (PIVW<0.01), however in each case the estimated effect of the MR Egger method was not directionally consistent, indicating potential confounding by pleiotropy (Figure 5 S).

Further analysis was performed for the fish oil and triglyceride exposure outcome pair. The directional effect of the weighted mode, weighted median, and simple mode MR methods all aligned with the IVW method towards triglyceride reduction. As in the prior comparisons, all of the sensitivity tests including the leave-one-out analysis and Cochran’s Q test did not indicate significant confounding (Figure 6 S, Table 2).

To further validate our results, reverse causation analysis was performed to explore whether there was a significant effect of the outcome on the exposure. In each case (Oily Fish and PhenoAge, Fish Oil and GrimAge, Fish Oil and Triglycerides), there was no evidence for reverse causality, as shown in the following figures (Figure 7 S, 8 S, 9 S, 10 S). All statistically significant results that passed the sensitivity tests are presented in Table 2.

In this 2-sample MR study, we attempt to augment the understanding of the beneficial effects of oily fish consumption on certain measures of healthy aging. Toward this end, we provide evidence for potential causal effects of oily fish and fish oil intake on certain emerging epigenetic measures of biological age. Whole oily fish consumption showed a potential causal relationship with PhenoAge acceleration, whereas supplemental fish oil had evidence for a potential causal relationship with GrimAge acceleration. We also report an overall directionally consistent effect of oily fish consumption on first- and second-generation epigenetic age acceleration measures. In addition, we find supportive evidence for potential causality between fish oil supplementation and triglyceride reduction, though not HDL or LDL, suggesting that this intervention is unlikely to directly modify serum cholesterol levels.

The first and second generation epigenetic clocks estimate biological age using DNA methylation patterns. The former, namely, the Horvath and Hannum clocks, predict chronological age across cells and tissues, but were not originally trained to account for age-related changes in physiological function [10, 11]. However, second-generation clocks, such as PhenoAge and GrimAge, capture intervention effects related to healthspan, physiological function, and mortality risk. This key distinction between the first- and second-generation epigenetic clocks may explain why the dietary interventions examined in the current study exhibited a significant modulating effect only on the latter.

Both PhenoAge and GrimAge were trained on serum biomarkers related to healthspan, such as hsCRP [8, 9, 32]. Indeed, a recent investigation reported that higher diet quality, as measured by the Dietary Approaches to Stop Hypertension (DASH) score, was associated with decelerated epigenetic aging only in the second (and third) generation epigenetic clocks, including GrimAge (p < 0.001) and PhenoAge (p= 0.001) [33]. Hence, it is the latter generation clocks that have proven to be more responsive to dietary and/or other lifestyle interventions [16, 34–36].

In this analysis, we do observe that PhenoAge and GrimAge acceleration are significantly modified in the direction of rejuvenation for either whole food-based and/or supplemental exposures to fish oils. This result appears to be in line with ample evidence that points to the general cardio-protective effects of omega-3 PUFA [37–39], which is in alignment with reports that CVD accelerates aging as measured by the two second-generation epigenetic clocks [40].

Multiple reports suggest that oily fish and/or supplemental fish oil intake impart health benefits via improved cardiovascular health and systemic inflammation levels, though, aside from its effects on triglycerides, the effects of this intervention on serum lipid and inflammatory markers appear to be less convincing [41–43]. Systematic reviews that assessed the impact of fish or fish oil consumption on vascular risk factors found that it was associated with a reduction in plasma triglyceride levels, as well as an increase in HDL cholesterol levels [25, 44]. Another meta-analysis of 191,558 individuals reported that higher fish intake was linked to lower triglyceride levels, however no associations were found with serum HDL, whereas LDL exhibited a somewhat paradoxical increasing trend with increased consumption of fish [41].

With respect to the effects of oily fish intake on systemic inflammation, the ATTICA study reported that individuals who consumed more than 300 g of whole fish per week had significantly lower levels of all measured inflammatory markers including hsCRP, while supplemental fish oil’s effects on established clinical inflammatory biomarkers such as hsCRP were less consistent [42, 43, 45]. For example, Yang et al. noted a reduction in hsCRP induced by cod liver oil in gestational diabetics, but the effect was null in type 2 diabetics [42, 43].

We note that our 2SMR-based analysis of oily fish intake and fish oil supplementation aligns with the above-mentioned effects of these interventions on serum biomarkers, where fish oil supplementation was found to only causally reduce serum triglycerides. Interestingly, oily fish consumption did not modify this biomarker in our investigation, which may be a property of the wholefood based exposure not supplying enough omega-3 PUFAs to induce an intervention effect. When these fatty acids are consumed in quantities above 3 g daily, a significant decrease in serum triglycerides is observed, with a potential mechanism being a reduction in the production of VLDL-TG in the liver, a key contributor to plasma triglyceride levels [23].

The absence of a similar effect from oily fish consumption in our study might therefore be attributed to wholefood-based exposure not providing sufficient omega-3 PUFAs to elicit this intervention effect. Finally, we did observe a mild hsCRP-reducing effect of oily fish intake, which is consistent with the literature [45] and may be explained by a general anti-inflammatory effect of various components of fish, including the antioxidants selenium and astaxanthin, which have been associated with attenuated levels of this serum and other serum proxies of systemic inflammation [46–49].

Although we selected genetic instruments that are robustly associated with our exposures of interest, which should minimize the probability of a biased causal inference, we wanted to ascertain that the individual genetic variants are indeed related to dietary behaviors. Toward this end, we note that several of the variants within oily fish and fish oil intake exposures contribute to the variance within several nutrient-seeking habits, the most well-characterized of these being rs1421085 within the FTO locus. Multiple studies have implicated this and other FTO variants in nutrient-seeking behaviors, both directly and via gene-diet interaction [50–52]. For example, rs7206790, which is in linkage disequilibrium (D’ = 0.84) with the genetic instrument used in the current study, was found to be associated with EPA and DHA consumption within the CHARGE consortium at near genome-wide significant level [50]. Moreover, rs1421085 has also previously been indicated in dietary protein-seeking, independent of its effects on adiposity [52].

Additionally, recent UK Biobank analyses have revealed strong interactions of this causal variant with dietary variation [51]. Another variant, rs11859365, lies within the CDH13 gene, which harbors a quantitative trait locus (QTL) of the metabolically protective adipokine, adiponectin [53]. To this point, rs11859365 has, at genome-wide significant level, been linked to liking and/or consumption of specific foods, including fish, herring, salmon, mackerel, sardines, seafood, pork, capers, and savory foods [54]. Additionally, our PheWAS has revealed that the SNP rs2533273 within the fish oil supplement exposure is also associated with other dietary behaviors, including dried and fresh fruit intake, vitamin supplementation, poultry, pork, tea and alcohol intake. In conjunction with our sensitivity analyses, the above observations provide evidence that fish consumption exposures examined in the context of the current study are unlikely to violate the above-mentioned MR assumptions that must hold to enable causal inference.

This study has several limitations. To minimize confounding by population stratification, our results are based on European-ancestry populations, and may not generalize to more ethnically diverse populations due to different genetic backgrounds and dietary habits. Although Mendelian randomization (MR) methods offer robust evidence, the assumptions might not always hold, especially with unrecognized pleiotropy or unaccounted pathways. The use of novel second-generation epigenetic age measures may not fully capture all biological aging aspects. Therefore, while our results suggest potential causal relationships, they must be cautiously interpreted, and their clinical relevance is still to be determined.

Drawing causal inferences using 2SMR requires strong adherence to 3 critical assumptions [31]. First, the chosen IVs - in this case, SNPs associated with oily fish intake—must show a robust association with the exposure of interest. Second, the IVs should not be directly associated with the outcome or through any confounding variables. Lastly, the IVs must influence the outcome exclusively through the exposure (Fig. 5). If the 3 above mentioned assumptions hold, MR-estimate effects of exposure on outcomes are not likely to be significantly affected by reverse causation or confounding [55].

The exposure datasets for oily fish intake (ukb-b-2209; N=460,443) and fish oil consumption (ukb-b-11075; N=461,384) originate from the MRC-IEU consortium, and are composed of both males and females of European ancestry. These datasets are publicly available, and no ethical approval was required. Only those instruments with genome-wide significance (p<5e-8) were used as IVs. GWAS summary statistics from the Integrative Epidemiology Unit (ieu) database were used as the blood biomarker exposures. Specifically HDL cholesterol (id:ieu-b-4843; N=37,120), LDL cholesterol (id:ieu-b-4846 N=70,814), triglycerides (id:ieu-b-4850; N=78,700), and C-reactive protein (id:ieu-b-35, N=204,402) from the Within Family GWAS Consortium were used. In each dataset, all, participants were of European descent to minimize population stratification. Additionally, all exposure/outcome datasets had undergone adjustment for principal components to adjust for population structure as described in Bycroft et al [56]. Epigenetic datasets, including GrimAge, Hannum, PhenoAge and intrinsic epigenetic age acceleration (IEAA), were derived from McCartney et al., the largest epigenetic aging GWAS to date, with samples sizes ranging from 34, 449 to 34, 467 individuals [57]. Figure 6 outlines the entire methods process.

We pruned the IVs for linkage disequilibrium (LD), which removed SNPs exhibiting genome-wide significance for a given exposure (p<5e-8) and fell within a 10, 000kb window with a correlation of r2>0.001. These exposure-associated SNPs, were then utilized as IVs in the MR analysis. To ensure consistency between alleles across the exposure and outcome datasets, palindromic and ambiguous SNPs were removed from the data. Additionally, the F-statistic was used to confirm the strong correlation of the IV with the exposure, such that IVs with F<10 were removed from the analysis. This value was calculated using the formula below:

where R2 is the proportion of explained variability from the IVs of the exposure dataset, n is the sample size, and k is the number of IVs.

The exploration of potential causal relationships for each exposure-outcome pair involved two primary MR methods, including Inverse-variance weighting (IVW) and MR Egger [31, 55]. The multiplicative random-effects IVW method is a weighted regression of instrument-outcome effects on instrument-exposure effects with the intercept is set to zero. This method generates a causal estimate of the exposure trait on the outcome traits by regressing the, for example, SNP-oily fish intake trait association on the SNP-epigenetic age measure association, weighted by the inverse of the SNP-epigenetic age measure association, while constraining the intercept of this regression to zero. However, this constraint can result in unbalanced horizontal pleiotropy whereby the instruments influence the outcome through causal pathways distinct from that through the exposure (which would violate the second assumption described above). Such unbalanced horizontal pleiotropy may not represent a true association between the exposure and the outcome, exaggerating or attenuating the effect estimate from the IVW method. However, unbalanced horizontal pleiotropy can be readily assessed by the MR Egger method (via the MR Egger intercept), which provides a valid MR causal estimate that is adjusted for the presence of such directional pleiotropy, albeit at the cost of statistical efficiency. Consistency across effect sizes from these MR methods was evaluated, and the IVW results were deemed significant with a p-value <0.05.

We further assessed pleiotropy explicitly for each exposure–outcome pair, retaining only those results with a pleiotropy p-value >0.1. Leave-one-out was then used to to ensure that no single variant excessively influenced the overall effect estimate. We also conducted a reverse causation test to exclude the possibility that the outcome might drive changes in the exposure. Finally, we applied the MR Steiger test to verify that each instrument explained more variance in the exposure than in the outcome, validating the assumed direction of causality [31].

Furthermore, the variant effect predictor (VEP) was applied to each IV of the oily fish intake dataset to determine the implicated genes, and performed a PheWAS across the IVs to ascertain relationships with nutrient-seeking behaviors [70] (see https://gwas.mrcieu.ac.uk/phewas/#info) [58].

The TwoSampleMR package in R was used for the 2SMR analysis, with additional R software packages being utilized for data formatting (see data availability statement) [59]. For each exposure-outcome pair, the IVW method was considered significant with a p-value <0.05.

Not applicable to this study.

Not applicable.

None to Declare.

The 2S-MR analysis was performed using publicly available datasets via the TwoSampleMR R package.

Additional code available upon request.