Multivariate genome-wide covariance analyses of literacy, language and working memory skills reveal distinct etiologies

The study investigates how reading-related skills—literacy (reading fluency and spelling), phonological awareness, oral language (listening comprehension), and phonological working memory—are interrelated at the genetic level. Building on frameworks like the Simple View of Reading, which distinguishes decoding from language comprehension, and theories implicating working memory in reading, prior work has shown moderate to strong heritability for these traits and substantial genetic overlap (generalist genes). However, most multivariate insights come from twin studies with debated assumptions. This study aims to map multivariate genome-wide covariance structures in a large, population-based cohort of unrelated youth, to identify shared and trait-specific genetic influences across domains and to compare genetic and residual (non-genetic) covariance patterns, thereby informing etiological mechanisms and modifiability of these skills.

Prior research documents strong to moderate phenotypic interrelations among language, literacy, and working memory across development, with sizeable heritability for reading (decoding, fluency, comprehension), spelling, phonological awareness, language comprehension, and non-word repetition. Twin studies indicate considerable genetic overlap (pleiotropy) consistent with generalist genes affecting learning abilities, and developmental genetic stability for reading and language. Yet, genetic overlap differs by domain: oral language correlates strongly genetically with reading comprehension but only moderately with reading fluency, aligning with the Simple View of Reading’s distinction between comprehension and decoding. Working memory (often assessed by non-word repetition) is related to decoding and vocabulary development and shows genetic links to reading. Despite these findings, comprehensive multivariate genetic covariance structures integrating multiple reading measures, language, and PWM in unrelated individuals remain underexplored.

Design and cohort: Population-based sample from the Avon Longitudinal Study of Parents and Children (ALSPAC), UK. Unrelated youths aged 7–13 years with genome-wide genotyping and dense phenotyping. After QC, up to N≈6453 had relevant phenotypes and genetic data. Measures: Reading fluency (non-word reading speed and accuracy; word reading speed and accuracy; passage reading speed and accuracy), spelling accuracy (ages 7 and 9), phonemic awareness (Auditory Analysis Test, age 7), listening comprehension (WOLD, age 8), and non-word repetition (CNRep, age 8). Psychometric properties are reported, with scores adjusted for age (except already standardized), sex, and 2 genetic PCs; residuals were rank-transformed. Genotyping and QC: Illumina HumanHap550; post-QC 465,740 SNPs; unrelated European-ancestry individuals (pairwise relatedness <0.05). SNP-based heritability and bivariate genetic correlations were also estimated via GCTA GREML as sensitivity checks. Analytic strategy: Genetic-relationship-matrix structural equation modeling (GSEM) fitted to GRMs from SNP data to decompose phenotypic variance into additive genetic (A) and residual (E) components. Three multivariate model forms: (i) saturated Cholesky decomposition, (ii) Independent Pathway (IP) with common and trait-specific A and E, and (iii) a hybrid Independent Pathway/Cholesky (IPC: IP for genetic part, Cholesky for residual part). Model fit compared via LRTs, AIC, and BIC. Two-stage modeling due to computational constraints: - Stage 1 (single-domain): Modeled six reading fluency measures to identify proxy measures capturing shared and specific genetic variance; and two spelling measures (ages 7 and 9) to select a spelling proxy. • Best-fitting for reading was IPC, indicating a strong shared genetic factor across all six measures plus some measure-specific genetic factors; proxy selection: passage reading accuracy (age 9) for strongest common factor loading, and word reading speed (age 13) for strongest specific genetic factor. • Spelling (Cholesky): Genetic factors at age 7 captured ~94% of age-9 spelling genetic variance; selected age-7 spelling as proxy. - Stage 2 (multi-domain): Two separate five-trait models using either passage reading accuracy (age 9) or word reading speed (age 13) as the reading proxy, alongside spelling (age 7), phonemic awareness (age 7), listening comprehension (age 8), and non-word repetition (age 8). Best-fitting models were IPC in both datasets (N=6453 for passage subset; N=6383 for word-speed subset). Sensitivity: GSEM-based SNP-h² and rg were consistent with univariate/bivariate GREML; factor loadings evaluated via Wald tests; multiple-testing considerations via MatSpD noted though not directly applied to joint models.

- SNP-based heritability: Across reading fluency, spelling, phonemic awareness, listening comprehension, and non-word repetition, GCTA SNP-h² ranged ~0.30–0.50 (SE ~0.06–0.09). - Stage 1 (reading fluency structure): • Strong shared genetic factor across all six reading measures; near-perfect genetic correlations among reading measures (e.g., passage reading accuracy rg with others ~0.88–0.97). • Additional measure-specific genetic influences for some measures (largest for word reading speed); despite small phenotypic variance explained, they accounted for notable fractions of SNP-h² for specific tests. • Proxies chosen: passage reading accuracy (highest common factor loading) and word reading speed (highest specific factor). - Stage 2 (multi-domain IPC models): • Overarching shared genetic factor across literacy, phonemic awareness, listening comprehension, and PWM. • In the passage-reading subset, the shared factor explained phenotypic variance of approx. 27% (spelling), 31% (phonemic awareness), 45% (passage reading), 13% (listening comprehension), 14% (non-word repetition), corresponding to 91%, 98%, 97%, 44%, and 53% of each trait’s SNP-h², respectively. • Trait-specific genetic factors were substantial for listening comprehension (explaining ~17% phenotypic; 56% of SNP-h²) and non-word repetition (~12% phenotypic; 47% of SNP-h²), indicating domain-specific genetic influences. • In the word-reading-speed subset, the shared factor captured ~56% of word-speed SNP-h², with ~44% trait-specific; genetic correlations of word speed with other traits remained moderate to high (e.g., rg≈0.52 with listening comprehension; 0.50 with non-word repetition; 0.72 with spelling; 0.74 with phonemic awareness). - Genetic vs residual covariance: • Strong genetic correlations across literacy and phonological awareness were accompanied by modest to strong residual correlations among literacy traits (even across 6-year age gaps). • Marked discordance for oral language vs literacy/phonological awareness: moderate-to-strong genetic correlations contrasted with near-zero residual correlations (e.g., listening comprehension vs passage reading residual r≈0.08; vs spelling r≈0.02; vs phonemic awareness r≈0.11), implying that most phenotypic covariance with literacy is genetic. • Bivariate SNP-heritabilities often approached 1 (within 95% CI), indicating phenotypic covariation largely driven by genetic covariance. - Overall, findings support widespread pleiotropy (“generalist genes”) plus clear evidence for trait-specific genetic architecture, particularly for oral language and PWM.

The study addressed the research question by quantifying and modeling multivariate genetic covariance across literacy, phonological awareness, oral language, and PWM in unrelated youths. The identification of a robust core genetic factor spanning these domains supports the concept of generalist genes underpinning shared cognitive processes relevant to decoding and related skills. At the same time, the presence of sizable trait-specific genetic components for listening comprehension and PWM, and for word reading speed as a reading proxy, indicates meaningful domain- or measure-specific etiologies beyond the shared architecture. Discordance between genetic and residual covariance for oral language vs literacy/phonological awareness suggests different etiological mechanisms: while genetic factors link language to literacy, shared environmental or other non-genetic influences that correlate literacy/phonological awareness measures do not similarly link to oral language. This implies different modifiability profiles across domains, consistent with theoretical models distinguishing decoding from language comprehension. These insights have implications for educational strategies and interventions, highlighting that improvements in literacy-related skills may not translate equivalently to oral language without targeted approaches.

This work maps multivariate genome-wide covariance among literacy, phonological awareness, oral language, and PWM, revealing a pervasive pleiotropic genetic factor complemented by substantial trait-specific genetic influences, particularly for listening comprehension and non-word repetition. It also uncovers distinct genetic vs residual covariance patterns, especially between oral language and literacy/phonological awareness, pointing to divergent etiological mechanisms and differential modifiability across domains. Future research should: (i) replicate in independent cohorts with comparable phenotyping; (ii) extend to longitudinal, measurement-invariant models to chart developmental changes; (iii) explore potential population-level confounders (e.g., assortative mating, dynastic effects) using family-based designs; and (iv) examine how educational/schooling contexts and targeted interventions impact domain-specific vs shared components.

- Developmental scope: Measures were assessed primarily in mid-childhood to early adolescence; lack of repeated longitudinal measures across multiple ages limits inferences about developmental changes in genetic architecture. - Population phenomena: Assortative mating and dynastic effects may inflate SNP heritability and genetic correlations even among unrelated individuals, potentially biasing estimates of shared genetic variance. - Replication: Few independent cohorts have comparable comprehensive measures across literacy, phonological awareness, language, and PWM, limiting direct replication. - Computational constraints: Necessitated a two-stage proxy selection approach rather than modeling all measures simultaneously, though sensitivity analyses support robustness. - Residual components: Residual variance includes untagged genetic effects, environment, and error; disentangling these subcomponents is beyond scope.