Common genetic variants associated with cognitive performance identified using the proxy-phenotype method

Significance

We identify several common genetic variants associated with cognitive performance using a two-stage approach: we conduct a genome-wide association study of educational attainment to generate a set of candidates, and then we estimate the association of these variants with cognitive performance. In older Americans, we find that these variants are jointly associated with cognitive health. Bioinformatics analyses implicate a set of genes that is associated with a particular neurotransmitter pathway involved in synaptic plasticity, the main cellular mechanism for learning and memory. In addition to the substantive contribution, this work also serves to show a proxy-phenotype approach to discovering common genetic variants that is likely to be useful for many phenotypes of interest to social scientists (such as personality traits).

Abstract

We identify common genetic variants associated with cognitive performance using a two-stage approach, which we call the proxy-phenotype method. First, we conduct a genome-wide association study of educational attainment in a large sample (n = 106,736), which produces a set of 69 education-associated SNPs. Second, using independent samples (n = 24,189), we measure the association of these education-associated SNPs with cognitive performance. Three SNPs (rs1487441, rs7923609, and rs2721173) are significantly associated with cognitive performance after correction for multiple hypothesis testing. In an independent sample of older Americans (n = 8,652), we also show that a polygenic score derived from the education-associated SNPs is associated with memory and absence of dementia. Convergent evidence from a set of bioinformatics analyses implicates four specific genes (KNCMA1, NRXN1, POU2F3, and SCRT). All of these genes are associated with a particular neurotransmitter pathway involved in synaptic plasticity, the main cellular mechanism for learning and memory.

Twin and family studies have shown that at least a moderate share of variation in most facets of cognitive performance (i.e., performance by healthy individuals on cognitive tests) is associated with genetic factors (1, 2). However, despite considerable interest and effort, research to date has largely failed to identify common genetic variants associated with cognitive performance phenotypes (3–5) with the exception of APOE, which predicts cognitive decline in older individuals (6–8). Existing studies have relied on one of two research strategies. The first strategy is a candidate gene design, in which researchers test a small number of genetic variants for association with the phenotype of interest, typically based on hypotheses derived from the known biological functions of the candidate genes. The candidate gene associations that have been reported with cognitive performance (9), however, fail to replicate when larger samples are used (3). The second research strategy is a genome-wide association study (GWAS), in which researchers atheoretically test hundreds of thousands of SNPs for association with the phenotype and apply a threshold for genome-wide statistical significance—typically 5 × 10⁻⁸—to account for multiple hypothesis testing. For physical and medical phenotypes, GWASs have identified many novel associations that replicate (10). GWASs on cognitive performance, however, have not yet identified any genome-wide significant associations (4, 5).

Here, we apply an alternative two-stage research strategy, which we call the proxy-phenotype method. In the first stage, we conduct a GWAS on a proxy phenotype to identify a relatively small set of SNPs that are associated with the proxy phenotype. In the second stage, these SNPs serve as candidates that are tested in independent samples for association with the phenotype of interest at a significance threshold corrected for the number of proxy-associated SNPs. In the study reported here, our phenotype of interest is cognitive performance, for which we use Spearman’s measure of general cognitive ability (usually abbreviated as g; it is the general factor measured by a battery of diverse cognitive tests) (4). Our proxy phenotype is educational attainment measured by self-reported years of schooling.

Rietveld et al. (11) had suggested the strategy of using SNPs associated with educational attainment as “empirically-based candidate genes” for association with cognitive performance (11); here, we conduct that analysis and further develop the methodology for doing so. SI Appendix contains our formal framework, which builds on that in ref. 11, as well as power calculations under a range of assumptions. According to the framework, educational attainment is a good proxy phenotype for cognitive performance, because cognitive performance is strongly genetically influenced and causally affects educational attainment; also, much larger samples are available for GWASs on educational attainment. The high genetic correlation (estimated to be roughly 0.65 or higher) (12–14) between the two traits does not have straightforward implications for the statistical power to identify specific SNPs influencing cognitive performance. It does, however, imply that a polygenic score associated with educational attainment will be associated with cognitive performance; thus, it may be viewed as providing an additional suggestive justification for the approach to identifying specific SNPs.

Results

In our first stage, we conducted a GWAS of educational attainment in a pooled Education Sample of 106,736 individuals. We used the same data, analysis protocol, and quantitative years of schooling measure as in ref. 11, except that we omit cohorts with high-quality measures of cognitive performance; we, instead, include these cohorts in the subsequent Cognitive Performance Sample. We chose our inclusion threshold of P < 10⁻⁵ for selecting candidate SNPs based on ex ante power calculations, with a goal to maximize the number of true positives among the candidates (SI Appendix). Pruning for linkage disequilibrium the 927 SNPs that reach this threshold resulted in 69 approximately independent SNPs (SI Appendix).

In our second stage, we tested these 69 education-associated SNPs for association with cognitive performance in the Cognitive Performance Sample, which comprises 24,189 genotyped subjects from 11 cohorts (SI Appendix). The specific cognitive tests differ across cohorts, but the cognitive performance measure in every cohort is calculated as Spearman’s g (SI Appendix); previous research has found that g values from different test batteries are highly correlated, especially if the batteries have many tests or if the test is specifically constructed to measure g (15–17). We tested each SNP individually for association with cognitive performance using ordinary least squares, controlling for sex, age, and (depending on the cohort) at least four principal components of the genome-wide data (to reduce confounding from population stratification). At the cohort level, the analyses were conducted according to a prespecified plan that we preregistered on the Open Science Framework (https://osf.io/z7fe2/). The cohorts’ results were then meta-analyzed using an inverse-variance weighting scheme. Two independent teams of analysts cross-checked and verified the results.

To confirm that the education-based first stage identifies reasonable candidate SNPs for cognitive performance, Fig. 1 plots the standardized regression coefficients from the regression of years of schooling on the education-associated SNPs in the Education Sample (with the reference allele chosen to ensure that the coefficient is positive) against the standardized coefficients from the second-stage regression of cognitive performance on the SNPs in the Cognitive Performance Sample. The direction of the effect coincides in 53 of 69 cases (two-sided binomial test, P = 9.10 × 10⁻⁶), indicating that this context is a good context for applying the proxy-phenotype method. We were surprised that the correlation between the effect size on educational attainment and the effect size on cognitive performance is negative (ρ = −0.25, P = 0.03), although not significantly after dropping a possible outlier (the bottommost point of the figure; ρ = −0.14, P = 0.26). If the population correlation is truly negative, within our theoretical framework, it suggests that SNPs that affect cognitive performance more strongly tend to affect other factors that matter for educational attainment (such as personality traits) less strongly and vice versa (SI Appendix).

Fig. 1.

The relationship between standardized coefficients from the first-stage regression of years of schooling on the education-associated SNPs in the Education Sample (x axis) and standardized coefficients from the second-stage regression of cognitive performance on these SNPs in the Cognitive Performance Sample (y axis). The reference allele is chosen such that the coefficient on years of schooling is positive. Each point represents 1 of the 69 education-associated SNPs. (The cloud of points is bounded away from zero effect on years of schooling, because the criterion for including an SNP was reaching P < 10⁻⁵ in the GWAS on years of schooling in the Education Sample.) Because the SD of years of schooling is ∼3, a coefficient of 0.03—a typical size for a years of schooling standardized coefficient (before correcting for the winner’s curse)—means that each reference allele is associated with an increase of 0.03 × 3 ∼ 0.09 y of educational attainment. In conventional IQ units that have an SD of 15, a standardized regression coefficient on cognitive performance of 0.03 corresponds to ∼0.45 IQ points.

To provide a benchmark for evaluating our list of education-associated candidate SNPs, we generated (through a prespecified algorithm) a list of theory-based candidate SNPs for cognitive performance drawn from published findings in the candidate gene literature (SI Appendix). (This list does not include the SNPs comprising the APOE haplotype, because these SNPs were not available in the cohort GWAS results.) After applying the same pruning procedure as for the education-associated SNPs, our list of theory-based SNPs contains 24 independent SNPs, of which only one is in a genomic region close to an education-associated SNP. Fig. 2 overlays Q–Q plots for the theory-based and education-associated candidates. The education-associated candidates taken altogether are more strongly associated with cognitive performance than would be expected by chance (z = 5.98, P = 1.12 × 10⁻⁹). Whereas a visual inspection of the plot suggests that the theory-based candidates exhibit some association with cognitive performance, we cannot reject the null hypothesis for any SNP individually or all of them taken together (z = 1.19, P = 0.12).

Fig. 2.

A Q–Q plot for a regression of cognitive performance on the education-associated SNPs (black circles) with a 95% confidence interval around the null hypothesis (gray shaded region) and a Q–Q plot for a regression of cognitive performance on the theory-based SNPs (red circles) with a 95% confidence interval around the null hypothesis (orange shaded region).

The top three education-associated SNPs—rs1487441, rs7923609, and rs2721173—show clear separation from the others in Fig. 2 and are significantly associated with cognitive performance after Bonferroni correction for multiple hypothesis testing (Table 1). Consistent with the negative correlation in Fig. 1, these SNPs are different from the three SNPs that reached genome-wide significance for association with educational attainment in the analyses in ref. 11. After adjusting the estimated effect sizes of the SNPs (each R² ∼ 0.0006) for the winner’s curse, we estimate each as R² ∼ 0.0002 (SI Appendix), or in terms of coefficient magnitude, each additional reference allele for each SNP is associated with an ∼0.02 SD increase in cognitive performance [or 0.3 points on the typical intelligence quotient (IQ) scale]. R² ∼ 0.0002 is about the same as the R² value for the known SNP associations with educational attainment (11) but far smaller than the largest effect sizes for complex physical traits, such as height (R² ∼ 0.004) and body mass index (R² ∼ 0.003) (18, 19).

Table 1.

SNPs significantly associated with cognitive performance after Bonferroni correction (full results are in SI Appendix, Table S4)

SNP	Chromosome	Base pair position	Nearest gene	Effective allele	Allele frequency	Years of schooling (Education Sample)		Cognitive performance (Cognitive Performance Sample)
						Standardized coefficient	P value	Standardized coefficient	P value
rs1487441	6	98660615	LOC100129158	A	0.473	0.026	1.78 × 10−9	0.036	1.24 × 10−4
rs7923609	10	64803828	JMJD1C	A	0.521	−0.021	1.06 × 10−6	−0.034	2.58 × 10−4
rs2721173	8	145715237	LRRC14	T	0.473	−0.020	8.61 × 10−6	−0.034	2.88 × 10−4

The chromosome and base pair position are from the National Center for Biotechnology Information genome annotation (build 36), and the nearest gene is from the SCAN database (www.scandb.org/newinterface/about.html). Allele frequency refers to the Cognitive Performance Sample.

Power calculations that we report in SI Appendix help shed light on why the proxy-phenotype method succeeded in identifying SNPs, whereas GWASs to date on cognitive performance have not. A GWAS in our Cognitive Performance Sample of n = 24,189—which is larger than the largest GWASs (n = 17,989 in ref. 4 and n = 3,511 in ref. 5)—would have had power of 0.06% to identify any given SNP with an association that has R² = 0.0002. In contrast, our proxy-phenotype approach had power of 12%. Given this power and the rather stringent significance threshold (0.05/69 ∼ 0.00072), Bayesian calculations using reasonable assumptions regarding priors suggest that the posterior probabilities that these three SNPs are associated with cognitive performance are high (SI Appendix).

Turning from specific SNPs to the set of all 69 education-associated SNPs, we assess the explanatory power of a linear polygenic score that aggregates their coefficients (SI Appendix). In pooled results from four family-based cohorts (4,463 individuals in total), we find that the score is significantly associated with cognitive performance (P = 8.17 × 10⁻⁴), with R² ranging approximately from 0.2% to 0.4% across samples. Using only within-family variation, the pooled coefficient has the same sign but is smaller and has a larger SE (P = 0.36). Thus, we cannot rule out that some of the score’s explanatory power is because of population stratification, although even without stratification, the nonsignificance of the within-family coefficient is not surprising given the low power of this test (SI Appendix).

Next, we explore whether educational attainment might serve as a proxy phenotype for cognitive health phenotypes (as opposed to cognitive performance in the normal range). Our sample comprises 8,652 European descent individuals over the age of 50 y from the Health and Retirement Study (HRS) (SI Appendix). We confirm that, for 60 of 69 SNPs available in the HRS data, the direction of the effects on educational attainment generally coincides with the direction of the effects on the two cognitive health phenotypes that we study: total word recall, which is a test for memory problems (two-sided binomial test, P = 0.0067) and total mental status, which is a battery that screens for early signs of dementia (P = 0.0775). We obtain the weights for a polygenic score by conducting a de novo meta-GWAS analysis of educational attainment just as in the first stage described above, but this time, we exclude the HRS from the Education Sample.

Fig. 3 shows that the score is associated with both of the cognitive health phenotypes. The strength of the protective effect is approximately constant across age categories from age 50–80 y and becomes weaker for total word recall after age 80 y. These associations are essentially unaffected when we control for up to 20 principal components of the genome-wide data, suggesting that the associations are not driven by population stratification (20). The R² values of these associations range roughly from 0.2% to 0.4% (similar magnitudes as in the analysis of cognitive performance in the family-based cohorts). When we control for years of schooling, the estimated effect of the score falls roughly in one-half but remains statistically significant (SI Appendix). The score is not associated with cognitive decline (i.e., the change in a cognitive phenotype across longitudinal survey waves), except for total word recall after age 80 y.

Fig. 3.

Coefficients from regression of standardized cognitive phenotype (total word recall or total mental status) on standardized polygenic score within age category controlling for sex and clustering SEs by individual (details in SI Appendix, section 14). Error bars show ±1 SE.

Finally, we used 14 (of 69) education-associated SNPs that are nominally significantly associated with cognitive performance (P < 0.05) to explore possible biological pathways in a set of bioinformatic analyses (SI Appendix); 2 of 14 SNPs are in gene deserts, but the other 12 SNPs are in close vicinity to at least one gene predicted (based on its expression profile) to be involved in the nervous system (SI Appendix). Among the most promising genes across these loci are KNCMA1, NRXN1, POU2F3, and SCRT, all of which are predicted to be involved in a glutamate neurotransmission pathway [labeled in REACTOME as “unblocking of NMDA receptor, glutamate binding, and activation” (21)] that is involved in synaptic plasticity, a cellular mechanism for learning and memory. Using different methods (but some overlapping data), this same pathway has previously been implicated in human cognitive performance (21).

Discussion

This paper makes two contributions. The first contribution is that we show that the proxy-phenotype method generates positive findings in a domain in which neither candidate gene nor GWAS approaches have so far made substantial progress. Similar approaches have sometimes been used in prior work (e.g., to find rare structural variants associated with cognition) (22), and there is existing work focused on the related idea of increasing statistical power in GWAS by analyzing correlated phenotypes jointly (23, 24).

We propose that the proxy-phenotype method, if systematically applied in social science genetics, could be a useful complement to traditional gene discovery methods (such as GWAS) in cases where it affords greater statistical power. In this case, it does so, because (i) much larger genotyped samples are available for educational attainment than cognitive performance and (ii) some genetic variants are likely to be associated with educational attainment because of their more direct, stronger relationships with cognitive performance. For the same reasons, educational attainment might similarly serve as a proxy phenotype for personality traits, such as persistence and self-control. In other contexts, the proxy-phenotype method may be better powered for different reasons. For example, for behavioral phenotypes with substantial measurement error—such as smoking, drinking, exercise, or eating habits—the proxy phenotype could be a medical outcome associated with the behavior (e.g., pulmonary disease for smoking or cirrhosis for alcohol consumption). We also note that, although our analysis plan specified that cohorts look up a relatively small set of education-associated SNPs in their existing GWAS results on cognitive performance, researchers with access to full GWAS results on the phenotype of interest could implement a more powerful version of the proxy-phenotype method. For example, first-stage results on the proxy phenotype could inform priors that are updated using GWAS results on the phenotype of interest.

We caution that the proxy-phenotype method (like theory-based candidate SNP approaches) could generate an unacceptably high rate of false positives if it were applied when underpowered and if results were reported selectively. To minimize this danger, we propose a set of best practices that proxy-phenotype studies should follow: researchers should (i) conduct power calculations ex ante to justify the use of the method for a particular phenotype of interest and report these calculations in the supplemental information, (ii) circulate an analysis plan to all cohorts before conducting any analysis and register the plan in a public repository, (iii) commit to publishing all findings from the study, including null results, and (iv) conduct Bayesian calculations of the credibility of any findings. We followed these procedures in this paper. Although replication of findings in an independent cohort would be ideal, we anticipate that it will often be infeasible given the unavailability of genotyped samples that may motivate the proxy-phenotype approach in the first place.

The second contribution of this paper is to identify common genetic variants associated with cognitive phenotypes. Knowing the three significant SNPs is not useful for predicting any particular individual’s cognitive performance because the effect sizes are far too small, but it does enable follow-up research (e.g., pinpointing the causal variants and then conducting KO experiments in animals) that may ultimately shed light on biological pathways underlying cognitive variation. The polygenic scores constructed from our results may prove useful for studying gene–environment interactions. In future work, the magnitude of explained variance will increase as researchers gain access to datasets with even larger first-stage samples. Our results suggest that such scores hold promise for eventually identifying individuals whose cognitive health at older ages is at greatest risk, which could allow for appropriate preparation and (if possible) preventative intervention.

Materials and Methods

The first stage of our two-stage procedure consisted of conducting a GWAS meta-analysis on years of schooling in a pooled Education Sample (n = 106,736) using the same analysis plan as in the work by Rietveld et al. (11) and the same cohorts, except for omitting the individuals who we include in the second stage. To obtain our set of education-associated SNPs, we selected all SNPs with P value < 10⁻⁵ from the first-stage meta-analysis results and then pruned for linkage disequilibrium. The second stage of our two-stage procedure consisted of conducting a meta-analysis of these 69 SNPs on high-quality measures of cognitive performance in the independent Cognitive Performance Sample, which included 11 cohorts (n = 24,189). We constructed linear polygenic scores from the meta-analysis results of the second stage and tested them for association with cognitive performance in four family-based cohorts (pooled n = 4,463), with the meta-analysis sample excluding the respective validation sample. Analyses on cognitive health phenotypes were conducted in an independent cohort of older Americans, the HRS, using the two measures that are available in more than one wave in that sample: total word recall and total mental status (n = 8,652 and n = 8,539, respectively). We tested the association between these two phenotypes and a linear polygenic score that was constructed using the coefficient estimates from the GWAS meta-analysis of years of schooling (as in the first stage, excluding only HRS; n = 98,110). SI Appendix includes all of the details on the samples and methods.