Causal effects of the metabolites on intelligence
We selected 3–675 independent genetic variants as IVs for each of the 486 metabolites (Supplementary Table 1). 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 (Figure 1, Supplementary Table 2). Among the 16 known metabolic traits, 5-oxoproline was significantly associated with intelligence after Bonferroni correction (PIVW = 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, PIVW= 0·0313), mannitol (β = -0·03; 95% CI: -0·06 to -0·01, PIVW= 0·0223), and 2-oleoylglycerophosphocholine (β = 0·18; 95% CI: 0·05 to 0·30, PIVW= 0·0055).
Sensitivity analysis
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 (PMR-Egger = 0·0001, PWeighted median = 6·29×10-6, PMR-PRESSO = 0·0007), and no horizontal pleiotropy was observed (PIntercept = 0.09, PGlobal test = 0·06, I2 = 25%, PHeterogeneity = 0.13). Two other metabolites also showed robust associations with intelligence, namely dihomo-linoleate (20:2n6) (PMR-Egger = 0·0494, PWeighted median = 0·0236, PMR-PRESSO = 0·0293, PGlobal test = 0.16) and p-acetamidophenylglucuronide (PMR-Egger = 0·0075, PWeighted median = 0·0060, PMR-PRESSO = 0·0454, PGlobal test = 0·0611), and there were no evidence of horizontal pleiotropy (PIntercept =0.24, PGlobal test = 0.17, I2 = 0%, PHeterogeneity = 0.96 for dihomo-linoleate (20:2n6) and PIntercept =0.06, PGlobal test = 0·06, I2 = 17%, PHeterogeneity = 0.13 for p-acetamidophenylglucuronide; Table 1). Funnel plots appeared generally symmetrical for all the three metabolites, also suggesting no evidence for horizontal pleiotropy (Supplementary Figure 1). Dihomo-linoleate (20:2n6) showed a negative association with intelligence (βIVW = -0·14; 95% CI: -0·25 to -0·04), while the association between p-acetamidophenylglucuronide and intelligence was positive (βIVW = 0·01; 95% CI: 0·00 to 0·01). The causal association between 5-oxoproline and human intelligence is shown on Figure 2, while the associations for dihomo-linoleate (20:2n6) and p-acetamidophenylglucuronide with intelligence are represented on Figure 3. Notably, the very small effect size for p-acetamidophenylglucuronide on intelligence might limit its potential utility as a biomarker.
Associations with other relevant outcomes
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, PIVW= 0·0087) to the result of initial MR estimates, and the causal associations were robust when different methods were performed (PWeighted median = 0·0003, PMR-Egger = 0·0035). The results also showed that 5-oxoproline was significantly associated with cognitive performance (PIVW= 0·0001, PWeighted median = 1·44×10-6, PMR-Egger = 0·0009). However, no evidences for association were found between 5-oxoproline and educational attainment (PIVW = 0·5595, PWeighted median = 0·3417, PMR-Egger = 0·4611), as well as income (PIVW = 0·7854, PWeighted median = 0·4287, PMR-Egger = 0·6178). Besides, the effects of dihomo-linoleate (20:2n6) and p-acetamidophenylglucuronide on intelligence were also significant in the replication stage (Supplementary Figure 2; Supplementary Figure 3).
Genetic basis for the causal associations
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 Supplementary Table 3 and Supplementary Table 4.
Metabolic pathway analysis
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.