For our study, exposure instruments were selected from the previous GWAS on food consumption conducted in United Kingdom Biobank data (up to N = 445,799) (Pirastu et al., 2019) which included 25 traits for which valid instruments were available: consumption of beef, beer, bread, champagne/white wine, cheese, cooked vegetables, decaffeinated coffee, dried fruit, fresh fruit, ground coffee, instant coffee, lamb, nonoily fish, oily fish, pork, poultry, processed meat, red wine, salad, salt, spirits, tea, water adjusted for coffee, and vegetarianism and drink temperature. The paper reporting the results of food consumption GWAS is currently under review and has not been published yet and consequently included in the current study as a preprint. The replication analyses for the food consumption GWAS were conducted on two different cohorts: the EPIC-Norfolk Study and the Fenland Study. The statistical power of the replication cohorts was relatively limited (sample sizes 21,337 and 11,442 individuals, respectively). Despite that, for 82% of the signals a concordant direction of effects was observed and for 32% of the signals the nominal significance was achieved (Pirastu et al., 2019). The performance of the United Kingdom Biobank dietary questionnaire has been previously assessed and shown to reliably rank individuals according to the intake of the measured foods and food groups (Bradbury et al., 2018).

In order to be able to distinguish between the independent effects of single food items and those arising due to the effect of dietary patterns, we defined 15 “Principal Component traits” (PC-traits) by first clustering the single food items using the iCLUST algorithm (Revelle, 1979; Revelle and Zinbarg, 2009). The defining and calculation of PC-traits is not a part of this paper and was done for our previous manuscript—therefore, only a brief description is included here and more detailed information can be found in Pirastu et al. (2019). After clustering the single-food items, we split the resulting tree dendrogram into different layers depending on the items in each cluster and the degree of similarity. For example, Oily-Fish and Non-Oily fish were first grouped in an overall fish consumption variable (Fish-PC1) and then in a more general healthy food measure together with Fruit-PC1 and Vegetables-PC1. Finally, they were all used to estimate a measure of the overall dietary pattern (All-PC1-3). Figure 2 shows the Sankey plot of the relationships between the different defined traits, and Figure 3 represents the loadings of each single food/drink on each of the main PC-traits (a full table of the PC-loadings can be viewed in Supplementary Table S1 and is visualized in Supplementary Figure S1). Higher values of All-PC1 correspond to higher consumption of vegetables, fruit, and fish and lower consumption of meat, coffee, and alcohol. All-PC2 separates foods from drinks (alcohol-containing beverages and coffee) with higher values corresponding to higher consumption of coffee and alcohol and lower consumption of the rest of the foods. Fruit-PC1 corresponds to higher consumption of dried and fresh fruit whereas a higher value on Psychoactive-PC1 corresponds to higher consumption of coffee and alcohol.

The SNP effect estimates for the PC-traits were estimated by applying principal component analysis to the genetic correlation matrix of the items included in each group and by using the resulting rotation matrix to project the estimated effects of the SNPs on each of the items onto the PC space. Standard errors for the PC-trait effects were estimated using the single items SEs and the phenotypic correlation matrix.

As instruments, independent (r2<0.001) SNPs significantly associated with the traits in the exposure GWAS (p-value < 5 × 10–8) were used, for which the effect was not mediated through other confounders or health-related traits. In case of the PC-traits, SNPs were selected if they were associated (p-value <5e-8) with at least one of the items which participated in the trait definitions. Once extracted, each selected SNP was assigned the p-value among the traits of interest from the rotated PC space. Next, the SNPs that were not available in the outcome GWAS were excluded. Finally, we applied LD pruning (r < 0.001) to the selected SNPs. We have previously shown that food frequency GWAS results are strongly affected by educational attainment (as a proxy of socioeconomic status) and by health-related traits such as body mass index, blood pressure, and cholesterol, for which dietary advice is generally given (Pirastu et al., 2019). This results in either bias by indication (where the behavior is determined by health advice or belief, e.g., lower fat consumption in people with high cholesterol) or reporting bias (where people underreport or overreport food consumption due to their health status, e.g., obese people underreport true fat consumption). If the biasing trait is heritable, this leads to spurious results in the GWAS, which can bias the MR results. In order to distinguish which variables are likely directly associated with the food of interest (and not mediated by health conditions), we have previously developed a method called Corrected to Uncorrected ratio filtering (CUR), which is based on the idea that if the SNP is directly associated with food preferences, then its effect should not change when adjusted for education status or health conditions (Pirastu et al., 2019). The following list was considered as potential confounders: body mass index, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, triglycerides, diastolic and systolic blood pressure, T2D, coronary artery disease, Crohn’s disease, ulcerative colitis, and educational attainment. Briefly, the method is composed of three steps. The first step is to conduct multivariable MR using each food trait as the outcome while the exposures are selected through a stepwise procedure. Once the effect of each exposure is obtained for each SNP, we obtain an “expected effect” which represents the overall effect of each SNP on the food trait mediated through the causal exposures. We then obtain a corrected effect which is the difference between the observed effect and the exposure-mediated effect. Finally, we estimate the ratio between the corrected effect the uncorrected observed effect (CUR). When CUR is close to 1, it means that the effect of the SNP on the food trait is entirely direct and thus it is a valid instrument. We have shown elsewhere (Pirastu et al., 2019) that using SNPs with CUR = 1 ± 0.05 maximizes the chances of selecting the correct SNPs. The strength of the instruments was assessed by using F-statistic and is reported in Supplementary Table S2. The SNP effects for the outcomes were obtained from a previous GWAS on 123 plasma metabolites in 24,925 individuals (Kettunen et al., 2016). Most of the metabolites assessed in this GWAS were linked to the lipid profile, and consequently the sample of metabolites used in the current paper is biased toward lipids.

As the main MR method, we used the inverse-variance weighted method, with a random-effect standard error if the heterogeneity p-value was less than 0.05/123. One of the problems of MR is when SNPs are associated with the outcome through causal paths which do not pass through the exposure of interest, also referred to as horizontal pleiotropy. In order to remove the SNPs with the highest heterogeneity (and thus likely are subject to horizontal pleiotropy), we used the method called MR-Radial (Bowden et al., 2018). All analyses were thus run on the instruments selected with this method. As sensitivity analyses, we used MR-Median (Bowden et al., 2016), MR-RAPS (Zhao et al., 2019), and MR-Egger (Bowden et al., 2015). These methods have been thoroughly described elsewhere and are all sensitive to the breaking of different MR assumptions. When only one instrument was available, the Wald ratio method was used. Finally, we defined as significant the food–metabolite relationships where Storey’s q-value (Storey, 2003) was less than 0.05. The analyses were run using R version 3.6.1 (R Core Team, 2019).

Out of the 413 significant results, 19 were based on the Wald ratio and 394 were based on the inverse-variance weighted method. For the latter, we used MR-Median, MR-RAPS, and MR-Egger as sensitivity analyses to detect potential violation of assumptions. We found that the sensitivity analyses broadly agreed with the results of the main analysis with the exception of one case (Alcohol-PC2 on “Description of average fatty acid chain length, not actual carbon number”), where the direction of the effect estimate from MR-Median was opposite to the one from the main analysis, and thus this result cannot be considered reliable. In 16 cases (mostly when Alcohol-PC2 was the exposure), the betas from MR-median were smaller (>50% effect difference) than the ones in the main analysis, which indicates that in these cases the effect sizes might be inflated. For some food–metabolite pairs, we detected some indication of heterogeneity. This was mostly related to PC-traits which reflect the effects of many different exposures, and it is thus not unexpected.

Most significant relationships were found with All-PC1 (higher value corresponds to lower meat/coffee/alcohol/salt and higher fruit/salad consumption), higher consumption of coffee (Coffee-PC1), and consumption of alcoholic beverages and coffee (Psychoactive-PC1 and -PC2). The majority of the effects of these traits were on the same metabolites (VLDL, IDL, and LDL related) and with opposite directions: namely, showing negative correlations for All-PC1 and positive correlations for Coffee-PC1 and Psychoactive-PC1. These effects were very similar to each other in size, with betas ranging from −0.502 to −0.322 for All-PC1, from 0.327 to 0.436 for Coffee-PC1, and from 0.425 to 0.778 for Psychoactive-PC1. The latter is likely partly due to the strong correlation between these lipid measurements.

All-PC1 had significant associations with 41 of the measured 123 metabolites, which was, as expected, the highest number of associations we saw for any of the tested foods or PCs. Interestingly, these effects were mostly related to atherogenic lipoproteins—the components of VLDL, IDL, and LDL of various sizes, and ApoB—whereas surprisingly there were no notable effects on any of the components of high-density lipoprotein (HDL) particles or other metabolites measured (for example, omega-3 or omega-6 fatty acids). Additionally, All-PC1 had significant effects on serum total cholesterol and serum total triglycerides. When a significant potentially causal relationship was found for a metabolite, the associations of this metabolite with each dietary item were examined. Often, the effects on PC-traits were driven only by a subset of dietary items. For example, in the case of total lipids in IDL, although we see a clear effect of overall diet (All-PC1), it seems that this is driven primarily by the components of Fruit-PC1 and Psychoactive-PC1 while the remaining foods do not seem to play any role. This observation may explain some of the significant heterogeneity we detected.

Some strikingly clear patterns can be noted, when examining the results that are depicted on Figure 4—for example, looking at the significant positive effects that Psychoactive-PC1 has on the components of IDL particles, large LDL particles, and serum total cholesterol. In the case of these results, both subcategories—coffee and alcohol—clearly contribute to the overall effect of Psychoactive-PC1. More importantly, all the four subcategories of alcohol show independently significant effects. All the significant effects that Psychoactive-PC1 has on the components of IDL particles and large LDL particles are accompanied by a contrary effect from Fruit-PC1, which shows significant negative effects. The latter indicates that higher consumption of fruits has a lowering effect on several IDL and large LDL components. In these cases, the effects of Fruit-PC1 are with the same direction as the effects of All-PC1 (“low meat/coffee/alcohol/salt + high fruit/salad”). Furthermore, this association pattern is also partly followed by the components of medium LDL particles, but in this case, when looking into the components of Psychoactive-PC1, alcohol seems to play a less important role compared to coffee.

In most of the cases, alcohol and coffee both contributed to the overall effect of Psychoactive-PC1, and we cannot distinguish between the effects of the components (although alcohol seems to play a slightly larger role in the case of IDL particles). Therefore, in most of the cases Alcohol-PC1, Coffee-PC1, and the items comprising these behave in the same way. Nevertheless, there is a group of clear and notable counterexamples: the components of medium VLDL and small VLDL particles. In these cases, Coffee-PC1 shows clear positive effects and all the betas of all the coffee subgroups agree with the direction of these effects, whereas Alcohol-PC1 and all of its subgroups show no clear effects and do not seem to be associated with the components of medium and small VLDL particles. Another interesting example that behaves differently from other items in the group of Psychoactive-PC1 is beer, which has significant positive effects on glucose (β = 0.43, 95% CI: 0.15; 0.70), concentration of very large HDL particles (β = 0.47, 95% CI: 0.16; 0.77), free cholesterol in very large HDL (β = 0.45, 95% CI: 0.16; 0.75), phospholipids in very large HDL (β = 0.46, 95% CI: 0.16; 0.76), total cholesterol in large HDL (β = 0.53, 95% CI: 0.23; 0.83), free cholesterol in large HDL (β = 0.57, 95% CI: 0.27; 0.86), and 22:6 docosahexaenoic acid (DHA, a subgroup of omega-3 fatty acids; β = 0.84, 95% CI: 0.32; 1.37). The latter is an example, where beer is clearly the odd-one-out compared to the effects of other alcohol subgroups, indicating that the effect comes from other ingredients in beer rather than alcohol.

The effects of Meat-PC1 do not have much contribution from beef, processed meat, and poultry and are driven mostly by pork (in the case of polyunsaturated fatty acids than other 18:2, and average number of double bonds in a fatty acid chain) or lamb (in the case of ratio of bisallylic groups to total fatty acids) or both (ratio of bisallylic groups to double bonds, DHA, and average number of methylene groups per double bond). Furthermore, it is noteworthy that lamb as a separate item had significant negative effects on several components of medium, large, very large, and largest VLDL particles, namely, on total cholesterol, cholesterol esters, free cholesterol, and phospholipids in medium and large VLDL; total lipids and triglycerides in very large and largest VLDL; phospholipids in very large VLDL; concentration of small and large VLDL particles; and mean diameter for VLDL particles. Thus, lamb has significant negative effects on the components of those lipoproteins that are largest and with lowest density, whereas showing no notable effects on the components of any of the lipoproteins that are smaller and with higher density than small VLDL particles.

As expected, Fish-PC1 had significant effects on omega-3 fatty acids and DHA, and these effects were clearly driven by oily fish. Furthermore, looking at Figure 5, the results regarding creatinine notably stand out, namely, there is a lowering effect of Psychoactive-PC1 on creatinine, which is clearly driven by coffee (which in turn shows significant negative effect on creatinine). Surprisingly, vegetarianism and bread share largely the structure of effects—they both have significant effects on omega-6 fatty acids (β = 3.73, 95% CI: 1.75; 5.70; β = 0.67 95% CI: 0.30; 1.04, respectively), 18:2 linoleic acid (LA, a subgroup of omega-6 fatty acids; β = 3.83, 95% CI: 1.86; 5.80; β = 0.55 95% CI: 0.18; 0.91), phosphatidylcholine and other cholines (β = 3.27, 95% CI: 1.31; 5.24; β = 0.53 95% CI: 0.17; 0.89), and total phosphoglycerides (β = 3.19, 95% CI: 1.23; 5.16; β = 0.63 95% CI: 0.25; 1.00). Of note, the results of vegetarianism are based only on one instrument and the method of the Wald ratio was used.

Many of our significant results were related to All-PC1 (“low meat/coffee/alcohol/salt + high fruit/salad”), which seems to affect mostly LDL, VLDL, IDL, and related subclasses. The fact that the effect sizes were very similar can be partly explained by the strong correlations between the different metabolites explained in the Results section. Another explanation might be the effect that All-PC1 has on ApoB. Namely, higher values on All-PC1 (corresponding to lower meat/coffee/alcohol/salt and higher fruit/salad) have a lowering effect on ApoB. The latter is the protein part of LDL, IDL, and VLDL, specifically one ApoB per each lipoprotein, thus reflecting the total amount of atherogenic lipoprotein particles (Elovson et al., 1988). This is important in the context of health risks, because it shows that ApoB, being a component of all atherogenic lipoprotein particles, might reflect the actual CVD risk better than the amount of cholesterol in any lipoprotein particle type on their own (Leslie, 2017). Ference et al. (2019) came to the conclusion that the risk of CVD posed by LDL particles is more determined by the concentration of LDL particles measured by ApoB compared to the mass of cholesterol or triglycerides in LDL particles. Furthermore, a recent study using multivariable MR found that ApoB has a robust elevating effect on CVD risk, while the effects of LDL-C, triglycerides, and HDL-C on CVD risk were not significant after accounting for ApoB (Richardson et al., 2020). Our results confirm the possible role of ApoB as a substantial factor in the relationship between diet and health and thus can prove useful to investigate when the causal effects of dietary patterns on risk of CVD are of interest. Nevertheless, despite the important role of ApoB, there is still a wide body of literature showing other features of atherogenic lipoproteins affecting health risks and these should not be neglected when interpreting the findings.

One of the most notable patterns in our results was the one regarding IDL and LDL particles, where higher consumption of coffee and alcohol had elevating effects, whereas higher consumption of fruits and higher value on All-PC1 had lowering effects. Higher levels of atherogenic particles and their components is part of a less desirable blood lipoprotein profile due to the elevated risk for several cardiovascular diseases (Carmena et al., 2004; Ference et al., 2017). For example, IDL-C has been shown to be associated with the degree and frequency of CAD independent of LDL-C (Tatami et al., 1981). Furthermore, different properties of LDL, like size, amount of cholesterol esters and cholesterol, and fatty acid composition are all considered to be aspects of its CVD-causing capability (Lada and Rudel, 2004). The clear pattern of effects we saw on IDL and LDL indicates that these metabolites are likely affected by dietary habits. Since these metabolites are largely shown to affect cardiovascular health, and Mendelian randomization results indicate possible causal pathways, there is considerable scope for further investigation of these results. Furthermore, our results indicate that the harmful effects alcohol and coffee have on cardiovascular health are at least partly mediated by IDL and LDL lipoproteins.

There was a surprising positive effect of beer on DHA, which has been shown to have a cardioprotective effect, as concluded in a recent large meta-analysis of randomized control trials (Bernasconi et al., 2020). This effect on DHA was in the opposite direction compared to other alcoholic beverages, indicating that the beneficial effect is likely not due to alcohol itself but some other ingredient in beer. We propose that the beer–DHA relationship is worth further investigation, and it would be especially interesting to compare regular beer with nonalcoholic beer to detect whether the potential beneficial effect remains. Furthermore, even if beer had some beneficial effect, this is unlikely to counterbalance the negative effect of alcohol raising IDL and LDL and their very well-established effect in predisposing to cardiovascular disease.

Many observational and epidemiological studies have pointed to the possibility of moderate alcohol intake to have a protective effect from cardiovascular diseases (Ronksley et al., 2011) and beneficial effects on cardiovascular biomarkers (Brien et al., 2011). The positive effects of alcohol on the lipid profile have mostly been associated with raised HDL-C levels (Brien et al., 2011). Although our results do show such a trend (with only beer having a significant effect on HDL levels), there have been conflicting results in other studies on the actual benefit of higher HDL-C levels protecting from developing cardiovascular diseases (Briel et al., 2009). Looking at the possible beneficial effect from raised HDL, one cannot underestimate the negative impact of raised IDL and LDL levels on health due to alcohol consumption. Furthermore, the reduced risk of cardiovascular disease in the case of light and moderate alcohol consumption compared with abstinence and heavy drinking, or in other words the U-shaped association of alcohol and CVD risk, found in many epidemiological studies has been found to be originating from other factors such as abstaining from alcohol due to poor health (Shaper et al., 1988) or might be resulting from reverse causation (Millwood et al., 2019). Further, alcohol has been shown to have harmful effects on health regardless of the quantities consumed (Holmes et al., 2014). Thus, the beneficial effects of alcohol on health are debatable and our results do not indicate that alcohol itself would have any beneficial effects.

We saw a significant increasing effect from Coffee-PC1 on several VLDL, IDL, and LDL lipoprotein subclasses and their components and in addition a negative effect on the mean HDL diameter and cholesterol esters in large HDL. Effects on other HDL parameters were not significant, but there is a notable trend toward a negative correlation. Overall, a higher value on Coffee-PC1 results in a more unhealthy lipid profile, raising ApoB, serum total cholesterol, VLDL, IDL, and LDL levels, and their constituents. There is consistent evidence in the literature of coffee raising LDL-C and total cholesterol. Poole et al. (2017) showed in their meta-analysis of randomized clinical trials (RCTs) that coffee has elevating effects on total and LDL cholesterol, but no clear effect on HDL cholesterol. Furthermore, coffee consumption has been associated with elevated ApoB levels (Williams et al., 1985; Periti et al., 1990; Cornelis and van Dam, 2020), but the causal link has not been established. Interestingly, Cornelis and van Dam (2020) observed higher HDL-C associated with higher coffee consumption, which is not in line with previous RCTs showing no significant effect and contrary to our results showing a non-significant HDL-C lowering effect. Thus, the effect of coffee on HDL-C and other HDL-related metabolites remains unclear.

We are not aware of any studies showing the effects of coffee on IDL and VLDL particles or their constituents, for which we observed consistent elevating effects. Raised VLDL levels might be due to cafestol, a common ingredient in coffee, having an effect of increased VLDL particle assembly rate on the liver (de Roos et al., 2001). Furthermore, we are the first ones to observe the effect of coffee on such a wide range of lipoprotein subclasses, allowing us to see the consistency of coffee’s effects on atherogenic lipoproteins. Looking at these results in the context of health outcomes, there is considerable scope for further investigation. In the case of coffee, the results regarding cardiovascular health are controversial and mostly show either beneficial or neutral effects (Chrysant, 2015) or that moderate consumption is unlikely to have adverse effects (Rebello and van Dam, 2013). On the other hand, raised ApoB levels indicate that the overall atherogenic particle amount is higher, which, as mentioned before, is a good predictor of CVD risk. Overall, we did not see any beneficial effect of coffee on lipid profile in our results; in fact, coffee turned out to have a negative effect on a larger variety of atherogenic lipoproteins than other dietary items. Hence, the total effect of coffee on health remains still unclear, but we propose that the potential harmful effect of coffee on health might be mediated by ApoB and thus via VLDL, IDL, and LDL. Furthermore, these results suggest that coffee consumption should be limited in people at risk of cardiovascular diseases.

Our study has several limitations: for some items, only a few SNPs were available to be used as an instrument in MR; self-reported dietary data are a difficult trait to investigate since it encompasses several biases—we aimed to mitigate this issue by using the corrected-to-uncorrected ratio (Pirastu et al., 2019); MR analysis can suffer from horizontal pleiotropy—we tried to mitigate this issue by using sensitivity analyses; item heterogeneity (the same dietary trait can incorporate items with varying nutritional values); sample homogeneity whereby our analyses are conducted on the European population and might not be fully generalizable to the population of the whole world; and when analyzing dietary patterns, we are limited to the ones arising from the data and cannot therefore fully account for the context of different diets. Furthermore, despite using the corrected-to-uncorrected ratio to identify SNPs that are not influenced by confounders, we cannot rule out the possibility that there were additional confounders that were not included in the model or that some might not have been captured properly (for example, when using educational attainment as a proxy for socioeconomic status). Nevertheless, our study has several strengths: to our best knowledge, we are the first ones to perform MR analysis between dietary items and blood metabolites with such a large amount of dietary SNP instruments available; we found several distinct patterns that shed more light on how dietary changes might affect cardiovascular health; and we found multiple interesting associations that are worth further investigation via feeding studies or randomized trials. We investigated the effects on metabolites profiled with NMR spectroscopy, which encompass mostly lipid profiles. Future studies with proteomics and metabolite data from mass spectrometry might give more detailed insight about the mechanisms by which food affects health.