Causal relationships between genetically determined metabolites and human intelligence: a Mendelian randomization study

Shin et al. reported the most comprehensive exploration of genetic influences on human metabolism so far, by performing a GWAS of non-targeted metabolomics on 7824 healthy adults. [16]. Metabolic profiling was carried out on fasting serum using high-performance liquid chromatography and gas chromatography separation coupled with tandem mass spectrometry. After quality control, 486 metabolites were retained for genetic analysis, among which 309 were chemically identified and could be further assigned to 8 metabolic groups (amino acids, carbohydrates, cofactors and vitamins, energy, lipids, nucleotides, peptides, and xenobiotics), while the other 177 were classified as ‘unknown’. The final genome-wide association analyses were carried out on approximately 2.1 million single nucleotide polymorphisms (SNPs). Full summary statistics for the 486 metabolites can be found at the Metabolomics GWAS Server (http://​metabolomics.​helmholtz-muenchen.​de/​gwas/​).

The foundational principle of MR relies on the existence of valid IVs. A genetic variant is a valid IV if it is (i) significantly associated with the exposure, (ii) independent of confounders, and (iii) associated with the outcome only through the exposure [22]. To identify valid IVs, we first selected the SNPs with significance P < 1 × 10⁻⁵, so as to account for a proportion as large as possible of the variance explained for the corresponding metabolite. We next performed a clumping procedure (linkage disequilibrium threshold of r² < 0.1 within a 500-kb window) to select the independent SNPs using the PLINK software (v1.9). To avoid the negative impact of weak IVs, we further used the proportion of variation explained by each IV (R²) and the F statistics to select SNPs strong enough to be valid IVs. Typically, an F statistic > 10 is considered sufficient for MR analysis [23].

GWAS summary statistics for intelligence were obtained from the study by Savage et al. [10]. Briefly, these authors performed a large GWAS meta-analysis of 269,867 individuals from 14 cohorts of European ancestry. Intelligence was assessed using different neurocognitive tests and the general factor of intelligence (Spearman’s g). Although differences in assessment methods might reduce the power to detect associations in meta-analyses, this approach can at the same time reduce type I errors by removing measurement errors, and therefore identify SNPs with robust associations to the common latent factor underlying intelligence across different methods. Stringent quality control procedures were applied to the summary statistics for each cohort. Association analysis was conducted controlling for covariates of age, sex, genotyping array, socioeconomic status for specific cohort, and twenty European-based ancestry principal components. Finally, a total of 9,295,118 SNPs were included in the meta-analysis.

Primary two-sample MR analyses were performed using the standard inverse-variance weighted (IVW) method. The IVW method provides a consistent estimate of causal effects by combining the ratio estimates of each variant in a fixed-effect meta-analysis model [23]. The P-value was calculated with a standard normal cumulative distribution function on the ratio of the combined causal effect and its standard error. The significance threshold to declare a causal relationship for the IVW-based MR estimate was set, using Bonferroni correction, at P < 1.03 × 10^–4 (= 0.05/486). Associations with P < 0.05, but not reaching the Bonferroni-corrected threshold, were reported as suggestive of association.

The IVW method provides an unbiased estimate under the assumption that all genetic variants are valid IVs. However, this assumption is easily violated, leading to inaccurate estimates, when horizontal pleiotropy occurs (some variants act on the outcome via a different intermediary) [24]. To avoid the effects of widespread horizontal pleiotropy in MR, we further performed sensitivity analyses using three additional MR methods: the MR-Egger method, which provides a consistent causal effect estimate, even when all genetic variants violate the assumptions defining valid IVs, under a weaker assumption (known as the InSIDE [instrument strength independent of direct effect] assumption) [24]; the weighted median method, which introduces a weighted median estimator and provides a more precise estimate than MR-Egger regression without the InSIDE assumption [25]; and the MR-PRESSO method, a newly developed approach which can identify and correct for horizontal pleiotropic outliers in MR [26]. We further used the MR-PRESSO global test as well as the intercept of the MR-Egger regression to test for pleiotropy, and we also evaluated heterogeneity with the I² and the Cochran Q test. Typically, I² > 25% or Cochran Q-derived P < 0.05 were used as indicators of possible horizontal pleiotropy. Analyses were carried out using the packages MendelianRandomization and MR-PRESSO in R (version 3.6.1).

We next used GWAS datasets of four other related outcomes to replicate the findings of our MR estimates. The first dataset was obtained from another GWAS of intelligence with 248,482 samples from the UK Biobank [27]. Summary statistics of cognitive performance (n = 257,828) and educational attainment (n = 766,345) were obtained from the study of Lee et al. [28]. Genetic associations with income (n = 286,301) were extracted from the large publicly available Lothian Birth Cohorts of 1921 and 1936 data-sharing resource [29]. Notably, the Davies et al. reported another GWAS for intelligence with a larger sample size, but the summary data for full dataset is not available due to data permissions [30].

Metabolic pathway analysis was carried out using the web-based tool suite MetaboAnalyst 4.0 (https://​www.​metaboanalyst.​ca/​) [31]. For this analysis, we extracted all metabolites showing suggestive associations in the IVW estimates (P_IVW < 0.05). Two libraries of metabolic pathways or metabolite sets were selected for enrichment analysis, namely the Small Molecule Pathway Database (SMPDB, http://​www.​smpdb.​ca) [32] and the Kyoto Encyclopedia of Genes and Genomes (KEGG, https://​www.​kegg.​jp/​) database [33]. P-values < 0.05 were considered statistically significant.

We selected 3–675 independent genetic variants as IVs for each of the 486 metabolites (Additional file 4: Table S1). On average, the IVs explained 4.7% (range 0.8–83.5%) of the variance of their respective metabolic traits. The minimum F statistic used to evaluate the strength of these IVs was 20.33. Using these IVs, IVW identified 16 known metabolites and 16 unknown metabolites that might have causal effects on human intelligence (Fig. 1, Additional file 4: Table S2). Among the 16 known metabolic traits, 5-oxoproline was significantly associated with intelligence after Bonferroni correction (P_IVW = 9.25 × 10^–5). Using 25 SNPs as proxy, we observed a 0.24 increase in the score of the Spearman’s g test for an increase of one standard deviation (SD) in the level of 5-oxoproline (β = 2.10; 95% Confidence interval [CI] 0.12 to 0.35). We also found 15 other metabolites to be suggestive for association, including indolelactate (β = − 0.09; 95% CI − 0.81 to − 0.01, P_IVW = 0.0313), mannitol (β = − 0.03; 95% CI − 0.06 to − 0.01, P_IVW = 0.0223), and 2-oleoylglycerophosphocholine (β = 0.18; 95% CI 0.05 to 0.30, P_IVW = 0.0055).

Table 1 shows the results of the sensitivity analyses for the 16 IVW-identified known metabolites. The causal relationship between 5-oxoproline and intelligence was robust when additional MR methods were applied (P_MR-Egger = 0.0001, P_{Weighted median} = 6.29 × 10^–6, P_MR-PRESSO = 0.0007), and no horizontal pleiotropy was observed (P_Intercept = 0.09, P_{Global test} = 0.06, I² = 25%, P_{Heterogeneity} = 0.13). Two other metabolites also showed robust associations with intelligence, namely dihomo-linoleate (20:2n6) (P_MR-Egger = 0.0494, P_{Weighted median} = 0.0236, P_MR-PRESSO = 0.0293, P_{Global test} = 0.16) and p-acetamidophenylglucuronide (P_MR-Egger = 0.0075, P_{Weighted median} = 0.0060, P_MR-PRESSO = 0.0454, P_{Global test} = 0.0611), and there were no evidence of horizontal pleiotropy (P_Intercept = 0.24, P_{Global test} = 0.17, I² = 0%, P_{Heterogeneity} = 0.96 for dihomo-linoleate (20:2n6) and P_Intercept =0.06, P_{Global test} = 0.06, I² = 17%, P_{Heterogeneity} = 0.13 for p-acetamidophenylglucuronide; Table 1). Funnel plots appeared generally symmetrical for all the three metabolites, also suggesting no evidence for horizontal pleiotropy (Additional file 1: Fig. S1). Dihomo-linoleate (20:2n6) showed a negative association with intelligence (β_IVW = − 0.14; 95% CI − 0.25 to − 0.04), while the association between p-acetamidophenylglucuronide and intelligence was positive (β_IVW = 0.01; 95% CI 0.00 to 0.01). The causal association between 5-oxoproline and human intelligence is shown on Fig. 2, while the associations for dihomo-linoleate (20:2n6) and p-acetamidophenylglucuronide with intelligence are represented on Fig. 3. Notably, the very small effect size for p-acetamidophenylglucuronide on intelligence might limit its potential utility as a biomarker.

We next repeated the main findings using summary statistics from other data sources. Figure 4 showed the results of causal effects of 5-oxoproline on human intelligence from another data source, cognitive performance, educational attainment, and income. The effect of genetically determined 5-oxoproline on intelligence (Replication) was similar (β = 0.17; 95% CI 0.04 to 0.30, P_IVW = 0.0087) to the result of initial MR estimates, and the causal associations were robust when different methods were performed (P_{Weighted median} = 0.0003, P_MR-Egger = 0.0035). The results also showed that 5-oxoproline was significantly associated with cognitive performance (P_IVW = 0.0001, P_{Weighted median} = 1.44 × 10^–6, P_MR-Egger = 0.0009). However, no evidences for association were found between 5-oxoproline and educational attainment (P_IVW = 0.5595, P_{Weighted median} = 0.3417, P_MR-Egger = 0.4611), as well as income (P_IVW = 0.7854, P_{Weighted median} = 0.4287, P_MR-Egger = 0.6178). Besides, the effects of dihomo-linoleate (20:2n6) and p-acetamidophenylglucuronide on intelligence were also significant in the replication stage (Additional file 2: Fig. S2; Additional file 3: Fig. S3).

We further investigated the genetic variants that affected both metabolite levels and intelligence. Table 2 shows the 25 SNPs used as IV of 5-oxoproline. Among them, rs11986602 showed the most significant association with 5-oxoproline (β = − 0.0620; SE = 0.0029, P = 6.29 × 10^–104). Notably, it also showed a strong association signal with intelligence (β = − 0.0196; SE = 0.0044, P = 9.53 × 10^–6). Moreover, this SNP had the largest effect sizes on both 5-oxoproline and intelligence, suggesting that the related genetic locus might provide valuable information on the biological mechanisms of intelligence, and that 5-oxoproline might be an important functional intermediate to understand the biological process through which genetics affects intelligence. The IVs for dihomo-linoleate (20:2n6) and p-acetamidophenylglucuronide are shown in Additional file 4: Tables S3 and S4.

Table 3 shows the results of the metabolic pathway analysis. Based on the 16 known metabolites identified by the IVW method, we detected only one significant metabolic pathway associated with intelligence, namely Alpha linolenic acid and linoleic acid metabolism (P = 0.0062). Two metabolites identified by IVW, docosapentaenoate (n3 DPA; 22:5n3) and linolenate (18:3n3 or 6), are involved in Alpha linolenic acid and linoleic acid metabolism according to the SMPDB database. Importantly, many of the metabolites found by our analysis have not been assigned to any metabolic pathway currently recorded in the SMPDB or KEGG databases. Extensive further research will be needed to explore whether these metabolites are involved in biological processes relevant to differences in human intelligence.