Multi-polygenic score prediction of mathematics, reading, and language abilities independent of general cognitive ability

The study explores the heritability of specific cognitive abilities (SCAs) independent of general cognitive ability (g). For over a century, research has shown substantial correlations between all cognitive abilities, with shared variance termed 'g'. The Cattell-Horn-Carroll (CHC) model illustrates this hierarchical structure, placing 'g' at the top, followed by broad SCAs (e.g., reading, math), and finally specific cognitive measures. Family, twin, and adoption studies consistently demonstrate substantial heritability for both g and SCAs. Previous research, however, has primarily focused on SCAs without controlling for g, leading to predictions confounded by g's influence. While the genetic correlations between SCAs are substantial, they are not perfect, indicating a unique genetic component to each SCA beyond its overlap with g. Studies correcting for g (yielding SCA.g) show surprisingly high heritability, similar to uncorrected SCAs. This high heritability, both corrected and uncorrected, makes SCAs good targets for genome-wide association (GWA) studies to identify single nucleotide polymorphisms (SNPs) associated with these abilities and aggregate them into polygenic scores (PGSs) for prediction. However, the strong genetic correlation between SCAs and g necessitates GWA studies of SCA.g to develop truly specific genetic profiles. Existing GWA studies have generally had small sample sizes, limiting their ability to detect significant SNPs and build powerful predictive PGSs. This study uses a larger sample and a multi-PGS approach to address these limitations, aiming to maximize the prediction of mathematics, reading, and language abilities, both corrected and uncorrected for g.

The literature review extensively cites previous research on the heritability of general and specific cognitive abilities, highlighting the consistent finding of substantial heritability in both domains. Studies using twin designs and adoption studies support this. The review then addresses the challenges in disentangling the genetic influences of specific cognitive abilities from the pervasive influence of general cognitive ability (g). It mentions previous attempts to isolate the unique genetic component of SCAs using statistical corrections for g but notes limitations in sample size and the types of cognitive abilities examined. The authors discuss the potential of polygenic scores (PGSs) derived from genome-wide association studies (GWAS) to predict SCAs, while emphasizing the limitations of previous GWAS studies, particularly the small sample sizes leading to limited power in identifying significant SNPs and building effective predictive models. The review concludes by justifying the current study's approach of using a multi-PGS strategy, combining multiple existing PGSs to potentially enhance the predictive power of genomic models for SCAs, especially when accounting for the influence of g.

The study utilized data from the Twins Early Development Study (TEDS), a longitudinal twin study with over 16,000 twin pairs. The sample was restricted to participants without serious medical conditions or missing background information, with same- and opposite-sex dizygotic twins combined. At age 12, twins completed a battery of 14 cognitive assessments, yielding composite scores for reading, mathematics, language abilities, and g. SCA.g measures were created by regressing the g factor from the respective SCA scores and using the standardized residuals. Twin analyses, using maximum-likelihood model-fitting in OpenMx, estimated additive genetic (A), shared environmental (C), and non-shared environmental (E) influences on SCA and SCA.g. Genomic analyses calculated SNP-based heritabilities using GCTA, incorporating family data. PGSs were constructed using LDpred2-auto from existing GWAS summary statistics; those with smaller sample sizes (under 10,000) or including TEDS participants were excluded. A multi-PGS approach employed elastic net penalized regression models with out-of-sample comparisons to predict SCA and SCA.g, using an 80/20 training/hold-out split and 10-fold cross-validation. Standard multiple regression analyses were conducted for comparison. Analyses were conducted to test for sex and zygosity differences, with measures corrected for age and sex to avoid inflation of correlations.

Twin analyses revealed substantial heritability for both SCA (average 53%) and SCA.g (average 40%). Although SCA.g heritability was lower, the difference wasn't statistically significant. The heritability estimates followed a consistent pattern across SCAs, both corrected and uncorrected. SNP heritabilities mirrored the twin heritability findings, with higher values for SCA (35%) compared to SCA.g (26%). The multi-PGS models significantly predicted all three SCAs and SCA.gs, but the predictive power was substantially higher for uncorrected SCAs (average 11.1% variance explained) than for SCA.g (average 4.4% variance explained). The most predictive PGS were consistently those from GWA studies of highly general traits such as educational attainment and general intelligence, while those from more specific SCA measures showed more modest independent predictive power. The multi-PGS approach only marginally outperformed the single best-predicting PGS for individual SCAs. The EA4 PGS for educational attainment consistently exhibited the strongest prediction for SCAs, and generally demonstrated stronger predictions for SCA.g compared to other PGSs. Analysis revealed that reading ability was less g-loaded than mathematics and language ability, as indicated by both the lower decrease in heritability after correcting for g and the slightly stronger prediction of g-corrected reading by multi-PGS. However, genetic correlation analyses from previous work demonstrated the high correlation between reading and g, which partially mitigated this finding.

The findings support the substantial heritability of SCAs independent of g. The difference in heritability between SCA and SCA.g, which is more pronounced in PGS prediction than in twin or SNP heritability, likely stems from the influence of highly g-loaded PGS derived from studies of general traits like educational attainment and intelligence. Despite this, the multi-PGS models were able to explain a portion of the variance in SCA.g, indicating potential for more precise prediction with more specific PGSs derived from larger GWAS focused on SCA.g. The results highlight the need for large-scale GWA studies targeting SCA.g to create more powerful predictive PGSs, which may aid in developing interventions to address cognitive strengths and weaknesses from an early age. The study also notes differences in g-loading across the specific cognitive domains; Reading, in particular, appears to be less influenced by g, indicating potential differences in the genetic architecture underlying various cognitive abilities.

This study provides further evidence for the substantial heritability of specific cognitive abilities independent of general cognitive ability. While multi-PGS prediction of SCA.g was lower than that of uncorrected SCAs, it was still significant, suggesting the potential for improved prediction with larger GWAS focusing on SCA.g. Future research should prioritize large-scale GWA studies of SCA.g and explore the development of brief, cost-effective cognitive assessments to facilitate these studies. Utilizing self-report measures, GWAS-by-subtraction methods and analysis of existing GWAS summary statistics may also be fruitful avenues for future research.

The study acknowledges limitations inherent to the twin method, PGS analyses (focus on additive effects of common SNPs), and the predominantly white UK sample, limiting generalizability to other populations. The reliance on existing PGSs, many derived from studies with smaller sample sizes, could also have limited the predictive power of the multi-PGS analyses.