Introduction

Reading is a multifaceted ability requiring the integration of diverse cognitive skills supported by extensive neural circuitries. Ultimately, it involves transforming arbitrary strings of visual symbols into meaningful sounds, words, sentences, and paragraphs1. Learning to read is a dynamic and cumulative process that starts at birth and continues across childhood, during which children are exposed informally to reading and then progress to formal reading education2. Although normative patterns of development are described in widely accepted theories of reading development3,4, considerable variability exists in the long-term pathways of reading development5,6,7,8,9 due to genetic and environmental influences10. As reading skills have wide ranging implications for daily life and later academic attainment, economic success and participation as active citizens11, understanding the etiological factors contributing to reading skills and identifying the sequence underlying these processes is important.

Reading skills are highly heritable. A recent meta-analysis estimated that genetic contributions (around 46–68%) were largely responsible for individual differences in reading-related skills12. Since the 1980s, molecular genetic studies identified promising genetic risk loci and candidate genes associated with reading and reading-related phenotypes13, albeit with limited replicability given their apparently polygenic nature. In the last 10 years, large genome-wide association studies (GWASs) have enabled the calculation of pan-genomic propensity scores for a variety of complex phenotypes. These polygenic scores (PGSs) incorporate hundreds of thousands of single nucleotide polymorphisms (SNPs) spread across the genome, each weighted based on its relative contributions to predicting a specific outcome, and then aggregated into an individual index of genetic susceptibility14.

GWASs have yielded new insights about the molecular basis of the heritability of reading and reading-related traits15,16,17,18,19,20,21,22,23. However, the impact of most of these studies is constrained by their modest sample sizes17,18,19,20,21,22,23, as well as a lack of direct reading measures15 and of replication. Recently, Eising and colleagues16 conducted a multicohort GWAS of 13,633 to 33,959 5-to-26-year-old participants assessed on some reading-related traits, which revealed a robust SNP heritability accounting for 13–26% of traits variability and a shared genetic component underlying word/nonword reading, spelling and phonological awareness (PA). Moreover, there is evidence that educational attainment (i.e., the highest level of education completed by a person) and cognitive ability share genetic variability with reading skills24,25. This leaves open a question regarding the specificity of the genetic association between the reading PGS and reading skills. Indeed, it could simply reflect the shared genetic variation with educational attainment and/or cognitive ability, without being specific to the neurocognitive mechanisms of reading.

While these discoveries provide avenues to explore the biological underpinnings of reading skills, the links between genetic predisposition and reading proficiency remain unclear. Testing possible endophenotypes (EPs) as mediators is argued as an effective approach to unravel the etiology of complex traits26,27. EPs reflect basic neurophysiological, neuroanatomical, biochemical, endocrine, cognitive, or neuropsychological processes, presumed to more closely link genes with complex traits26,27,28. At the cognitive level, reading acquisition relies upon the rapid and accurate integration of a vast circuit of neural structures and connections subserving several neurocognitive skills such as phonology, morphology, syntax, and semantics, as well as visual and orthographic processes, working memory, attention, motor movements, and higher-level comprehension and cognition1,11,29,30,31. Phonological awareness (i.e., the ability to identify and manipulate the sounds of spoken words) is considered a foundational skill for reading development as it is essential to establishing and subsequently automatizing letter-sound correspondences which consistently predict and support reading performance across languages30,31,32, and it significantly and substantially mediates the relationship between neurites’ density and orientation and reading skills in most tracts spanning the dorsal and ventral reading pathways33. Attending to and maintaining phonological units in a phonological memory loop are among the best predictors of reading skills34. Multiple memory systems are involved in remembering the meaning of words and comprehending text: working memory is a reliable predictor of reading comprehension35 and explains a significant proportion of variance in children’s reading comprehension skills36, whereas short-term memory is involved in storing phonological stimuli (e.g., phonological memory—PM)36. The ability to rapidly name linguistic and non-linguistic stimuli (e.g., rapid automatized naming - RAN) outlines the processes of connecting and automatizing whole sequences of letters and words with their linguistic information, regardless of writing systems29. Extensive research has provided reliable findings supporting RAN as one of the best longitudinal predictors of reading fluency across all orthographies29. The ability to process auditory stimuli in rapid sequences (e.g., rapid auditory processing—RAP) allows for fine sound-speech discrimination, which is critical for linguistic and phonological development and in turn supports the adequate development and acquisition of reading skills37,38. Visual and auditory attention is involved in the basic letter-sound correspondences during letter string processing as it is crucially involved in parsing and identifying relevant sub-lexical orthographic units by efficient attentional processing on each letter39,40,41,42. Several longitudinal studies have shown that visual and auditory attention in pre-readers significantly predicted future reading proficiency in primary school43,44,45,46. Taken together, these findings support previously addressed reading-related neurocognitive EPs as reliable EPs for reading skills29,30,31,32,36,37,42,47,48,49. Previous works showed that genetic influences were largely responsible for individual differences in some of these reading-related neurocognitive traits (i.e., PA, PM, RAP, and attentional components)12. Molecular genetic studies reported significant association between reading-related genetic risk loci and candidate genes and individual variations in all the above-described reading-related neurocognitive EPs13,50,51,52, identifying shared genetic underpinnings between reading skills and reading-related EPs. Perhaps more interestingly, some of previously addressed reading-related neurocognitive EPs (i.e., PA, visual motion processing and RAP) have been shown to act as mediators in the association between candidate gene and behavior in both clinical53 and general population54 samples. By emphasizing the complex relationships between genes, EPs and behavior, these findings highlight the importance of adopting an integrative approach to study the etiology of complex traits such as reading skills via EPs26,28,55.

Accordingly, in an age-homogenous sample of French-speaking Canadian school-aged twins simultaneously assessed for reading skills and reading-related EPs (N = 328 from 208 twin pairs), the present study tested (a) the direct and indirect associations between a PGS based on Eising and colleagues16 (hereafter, Reading-PGS) and reading decoding and comprehension skills via multiple reading-related EPs (i.e., PM; PA; RAN; RBTP; and, RAP); (b) the specificity of these direct and indirect associations by controlling for PGSs for educational attainment (EduYears 4—EY4-PGS)56 and cognitive ability (IQ3-PGS)57; and, (c) the direct and indirect associations between Reading-PGS and reading decoding and comprehension skills via a theory-driven sequence among mediators. Although the majority of the current PGSs account for a small proportion of the phenotypic variance of complex traits, dissecting such a small effect using structural equation modeling (SEM) could provide valuable insights on the complex and multifactorial nature of reading skills as it allows to test relationships among variables according to specific theoretical frameworks, providing additional, more comprehensive, meaningful theoretical insights58,59,60.

Results

Bivariate correlations

Bivariate Pearson correlations show that (1) all EPs were moderately correlated with both Decoding and Comprehension reading skills, whereby better performances in EPs correlated with better performances on both reading tasks. EPs explained around 3–17% of the phenotypic variance in both Decoding (range from 4.0% for RAN to 16.8% for PA) and Comprehension (range from 3.1% for RAP to 13.0% for RAN); and (2) the Reading-PGS was modestly associated with PA (explained variance of 2.2%), RBTP (explained variance of 3.5%), Decoding (explained variance of 1.8%) and Comprehension (explained variance of 3.2%), with higher PGS values significantly predicting better performance in all skills (Table 1).

Table 1 Correlations among reading-PGS, EPs, decoding, and comprehension in the total sample (n = 328)
Full size table

The multiple-mediator model

The multiple-mediator model tested the role of all the EPs simultaneously (Fig. 1).

Fig. 1: The multiple-mediators model.
figure 1

Reading-PGS Reading-PolyGenic Score as estimated by the summary statistics provided by (19); PM phonological memory, PA phonological awareness, RAN rapid automatized naming, RBTP rapid bimodal temporal processing, RAP rapid auditory processing. As the sample included children nested in families (i.e., twins), to control for the degree of kinship, we considered relatedness as a clustering variable upon the SNPs’ effects. Moreover, school grade level was used as an estimated covariable in the model (Table 1). We controlled for the covariance among the EPs and, separately, between Decoding and Comprehension (Table 1).

Full size image

This model provided a good fit to the data (χ2(2) = 2.940, p = 0.230; RMSEA = 0.038, 90% CI = 0.000–0.123; CFI = 0.997; SRMR = 0.012), accounting for 28.4% and 27.0% of the variance in Decoding and Comprehension, respectively (all fitted parameters are reported in Table 2 and Supplementary Table 1). Using 5000 bootstrapping replicates and a bias-corrected 95% CI, we found a significant total indirect effect from Reading-PGS to Decoding (β = 0.090, SE = 0.031; 95% CI = 0.028/0.150) and to Comprehension (β = 0.079, SE = 0.029; 95% CI = 0.019/0.140). Within these total indirect effects, three specific indirect effects were significant, involving PA and RBTP as mediators (Table 2).

Table 2 Specific indirect effects from the reading-PGS to decoding and comprehension via EPs
Full size table

Specifically, PA mediated the relationship between Reading-PGS and Decoding and Comprehension (specific indirect effect = 0.051 [95% CI = 0.014 / 0.090] and specific indirect effect = 0.036 [95% CI = 0.008 / 0.069], respectively; Table 2), whereas RBTP mediated only the relationship between Reading-PGS and Decoding (specific indirect effect = 0.034 [95% CI = 0.012 / 0.059]; Table 2). Inspection of beta scores revealed that the specific indirect effects along all relationships were positive whereby higher Reading-PGS predicted better skills in PA and RBTP which, in turn, predicted better skills in Decoding and Comprehension (Fig. 2).

Fig. 2: Indirect effects from reading-PGS to decoding and comprehension via EPs.
figure 2

Reading-PG Reading-PolyGenic Score as estimated by the summary statistics provided by Eising and colleagues16; PA phonological awareness, RBTP rapid bimodal temporal processing. 95% CIs are reported in squared brackets. Non-significant paths are indicated by a dotted line.

Full size image

Power analyses were conducted using MedPower (https://davidakenny.shinyapps.io/MedPower/) to determine the statistical power reached given a sample size equal to 328 subjects and alpha set at 0.05. The analyses were modeled for each of the three significant specific indirect effects we found (Fig. 2). Under these assumptions, the statistical power was above 80% (range between 82.2 and 83.6%), suggesting that our sample size was able to provide reliable findings. To confirm the reliability of the hypothesis-driven models we tested (Fig. 1), we ran alternative models in which we flipped the order among the observed variables (see Supplementary Results).

We tested the specificity of the direct and indirect relationships by controlling for EY4-PGS and IQ3-PGS (Supplementary Table 2—Supplementary Fig. 1). The inclusion of these PGSs lead to a good fit of the data (χ2(1) = 1.998, p = 0.158; RMSEA = 0.055, 90% CI = 0.000–0.169; CFI = 0.997; SRMR = 0.013; all fitted parameters are reported in Supplementary Table 3 and Supplementary Table 4). The total indirect effect from Reading-PGS to Decoding showed a trend toward significance (β = 0.057, SE = 0.030; 90% CI = 0.005/0.106), with minimal changes in the percent of explained variance (28.1%) and no changes to the other observed indirect paths (Supplementary Table 4—Supplementary Fig. 2).

According to our results, PA and RBTP seem to mediate the effects from Reading-PGS to Decoding and Comprehension. As there is theoretical evidence that multimodal sensory processing of transient and dynamic stimuli enables the establishment and automatization of letter–sound correspondences necessary to reading decoding41,42,61,62,63,64,65,66 which, in turn, supports proficiency in reading comprehension36,67,68,69, we tested three putative sequential associations from Reading-PGS to reading skills to formally test this hypothesis by following a step-by-step approach. Firstly, we examined whether PA underpins reading decoding which, in turn, supports reading comprehension by testing a sequential multiple mediation over-identified model (Supplementary Fig. 3). This model provided a good fit to the data (χ2(1) = 2.464, p = 0.117; RMSEA = 0.067, 90% CI = 0.000–0.177; CFI = 0.991; SRMR = 0.026) and accounted for 33.60% of the variance in Comprehension. We found a significant total indirect effect from Reading-PGS to Comprehension via PA and Decoding (β = 0.052, SE = 0.022; 95% CI = 0.011/0.097), whereby a higher Reading-PGS predicted better skills in phonological awareness which in turn predicted proficiency in reading decoding, followed by higher levels of reading comprehension (specific indirect effect = 0.031 [95% CI = 0.006 / 0.058]). Secondly, we assessed whether RBTP enables the establishment and automatization of PA skills necessary to reading decoding by testing a sequential multiple mediation (Supplementary Fig. 4). This model accounted for 22.6% of the variance in Decoding. We found a significant total indirect effect from Reading-PGS to Decoding via RBTP and PA (β = 0.098, SE = 0.027; 95% CI = 0.044/0.152), indicating that higher Reading-PGS predicted more rapid temporal processing leading up to better PA skills and reading decoding (specific indirect effect = 0.010 [95% CI = 0.001 / 0.027]). Finally, we formally evaluated the theoretical evidence showing that RBTP enables the establishment and automatization of PA skills necessary to reading decoding which, in turn, supports reading comprehension by testing a sequential multiple mediation over-identified model (Fig. 3).

Fig. 3: The serial mediation model from Reading-PGS to Comprehension via RBTP, PA, and Decoding.
figure 3

Reading-PGS Reading-PolyGenic Score as estimated by the summary statistics provided by Eising and colleagues16, RBTP rapid bimodal temporal processing, PA phonological awareness. 95% CIs are reported in squared brackets. Non-significant paths are indicated by a dotted line.

Full size image

This model provided a good fit to the data (χ2(2) = 2.538, p = 0.281; RMSEA = 0.029, 90% CI = 0.000–0.117; CFI = 0.997; SRMR = 0.018) and accounted for 33.9% of the variance of Comprehension. We found a significant total indirect effect from Reading-PGS to Comprehension via RBTP, PA and Decoding (β = 0.071, SE = 0.022; 95% CI = 0.028/0.115), suggesting that children with a higher Reading-PGS are more efficient in rapid temporal processing, which is linked to better skills in PA, which lead to higher scores in reading decoding and comprehension (specific indirect effect = 0.005 [95% CI = 0.001 / 0.013]). To confirm the reliability of the sequential associations found in this hypothesis-driven model, we ran alternative models in which we flipped the order among the observed variables (see Supplementary Results).

Discussion

This study examined the association between Eising and colleagues16 word reading PGS and reading-related skills in a French-speaking Canadian sample and tested the mediating role exerted by cognitive endophenotypes within these associations. We also examined the specificity of these direct and indirect effects after controlling for EY4-PGS and IQ3-PGS, and the specific sequential associations anchored in one of the most reliable theoretical (i.e., multimodal sensory processing of transient and dynamic stimuli) underlying reading acquisition.

While this study is not the first to document the predictive value of Eising and colleagues16 PGS, it is, to the best of our knowledge, a first formal test of the specificity of the direct and indirect associations between Reading-PGS and reading skills, showing that the effect of the reading PGS is marginally inflated or induced by its genetic covariation with educational attainment and general cognitive function. In addition, this study represents a first formal test of the sequential association from genes to reading skills via EPs, contributing to the growing literature on the neurogenetic influences on reading development27,28. Consistent with the multiple deficits model underlying the liability of complex traits26,70,71, the genetic variation represents the first step in a multiple-layered liability that ultimately leads to behavioral phenotypes55. Accordingly, our multiple-mediation model testing the associations between the Reading-PGS and reading-related skills via multiple reading-related neurocognitive EPs explained approximatively a third of the variance in reading decoding and comprehension. Within this total effect, we found small, although significant, indirect paths from the Reading-PGS to reading decoding and comprehension skills via phonological awareness and rapid bimodal temporal processing. These results are consistent with previous evidence implicating both phonological awareness and basic multimodal processing speed as reliable EPs of reading skills30,32,47,49, and further identify them as targets to uncover the association between genes and reading skills. Furthermore, results of sequential mediation models indicate that these EPs might act in a sequential order that offers a more nuanced representation of the association between genes and reading.

According to the phonological theory of reading (dis)ability, there is an association between phonological awareness and reading-related skills across languages30,31,32,36,68,72,73,74,75. Although phonological processing has been consistently reported as the best predictor of reading ability30,72,73,74,75, it has also been argued that basic visual or/and auditory attentional shifting with its noise-exclusion mechanism61, may be etiologically relevant for reading acquisition and development39,40,41,42,62,63,64,65,66,76. An accurate multimodal sensory processing of transient and dynamic stimuli, requiring precise and rapid timing across distributed brain networks in which perceptual neural noise exclusion is fundamental63,77, enables the consolidation of mental representations of phonemes and the establishment and automatization of letter–sound correspondences65,74,78 that are necessary to accurate and fluent reading decoding30,72,73,75. Possibly, a well-paced attentional shifting facilitates the processing of sensory input chunks, thereby decreasing the neural noise and improving the cortical representation of speech components essential for reading acquisition and accounting for the processing of rapid stimuli, such as phonemes and syllables63,64,65,66,78,79,80,81. Longitudinal studies have shown that a sluggish multimodal processing speed in pre-readers predicted future reading impairments in primary school43,44,45,46 and that training programs targeting attentional control improved both reading decoding and comprehension in typical readers82,83,84,85,86,87,88, pre-readers at risk of language-based learning disability89, and children with language-based learning disability90,91,92,93. Together, our multiple mediation models demonstrated that the indirect effects of Reading-PGS upon reading skills were accounted for by sequence from sensory processing to phonological skills. Empirical studies should therefore thrive to uncover the sequence by which genetic contributions to the development of reading skills are phenotypically expressed during childhood.

Our findings need to be considered within the limits of the study. First, although PGSs have the potential to be used for stratified intervention94, our findings may have been constrained by the limited predictive power of the PGSs. Second, Reading-PGS was calculated using summary statistics provided by a multicohort GWAS of 13,633 to 33,959 children, adolescents, and young adults aged 5 to 26 years16. Although this could have limited the ability to capture child-specific effects within our 8-year-old sample, we nevertheless found significant direct and indirect effects of the Reading-PGS on reading skills. Third, participants who provided biological samples for genotyping came from higher socioeconomic backgrounds, and we only included children with European ancestry to minimize the risk or errors due to population stratification, limiting the generalization of results to other populations. Fourth, although we cannot determine causal influences among measures over time because of the cross-sectional nature of the study, we provided a first formal test of the sequential association from genes to reading skills via cognitive skills. However, longitudinal studies investigating the role of PGSs are needed to explain the robustness of these relations over time. Fifth, although the EPs we used provide very sensitive measures boosting the statistical power to detect the presumed association28, our sample size remains relatively small. Nonetheless, power analyses yielded good estimates, complemented by the bootstrap resampling methods to test the indirect effects95,96,97. Replications in independent, larger datasets are needed. Sixth, due to the nature of the study and to logistic constraints, we were not able to test and include in this study other putative reading-related EPs (e.g., language-related skills). Future studies could extend the current model to include language-related skills.

As already outlined by Carrion-Castillo and colleagues98, the current findings point towards a genes-neurocognitive processes-reading skills sequence. By revealing insights into a putative sequence from genes to reading skills, these findings add to a growing body of literature showing that trait-specific PGSs are needed in order to achieve a better prediction of complex phenotypes99 and thereby facilitate the early identification of children with genetic susceptibility for reading (dis)ability. Within this objective, EPs need to be viewed as potentially valuable markers for both genetic mapping of complex traits and helping reshape classical nosological systems and diagnostic categories27,55.

Methods

Data and codes of all the analyses are available from the first author upon reasonable request.

Sample

Participants were part of an ongoing longitudinal population-based birth cohort study of twins born between 1995 and 1998 in the Montreal area, the Quebec Newborn Twin Study-QNTS100. For the present study, only second and third graders attending French schools (i.e., French-speaking twins) were preserved. We therefore included 328 (87 MZ and 241 DZ; males, N = 159) 8-year-old participants (M = 8.26 years, SD = 1.59 years) from 208 twin pairs (one child per MZ pairs) for whom measurements of reading skills, reading-related EPs and genotypic information was available. Procedures were approved by the ethics committees of Université Laval and Ste-Justine Hospital Research Center in Montreal. Informed written consent was obtained from parents and participants expressed their agreement to participate.

Measures

Reading skills and reading-related EPs were assessed by trained research assistants during a visit at the Ste-Justine Hospital Research Center. Descriptive statistics for reading skills and EPs are reported in Supplementary Table 6.

Le Test d’Habiletés en Lecture-THAL101 assessed the decoding of graphemes (Decoding) and comprehension of simple paragraphs (Comprehension). For Decoding, a phoneme from a word shown on screen was heard on the computer and the child had to say if the phoneme appeared in a comparison word presented on the screen as fast as possible. The 50 items were rated as pass (1) or fail (0). The task was interrupted after five consecutive failed items or within a set of ten. The total score is the sum of correct items. For Comprehension, children silently read short paragraphs (maximum four sentences) from which words were deleted. Participants had to select from a set of two or four grammatically correct alternatives the word that best filled the deletion. The 40 items were timed and the test interrupted after three consecutive failed items or two items exceeding the 30-second time limit. Each item was rated as pass (1) or fail (0). For both Decoding and Comprehension, when the response time was faster than the Z-scores of the same-age normative sample, a time component bonus of up 2 points per items was awarded. We derived a total score as the sum of correct items and time bonuses. Normative data were collected on 1457 French-speaking students from the first year of primary school to the first year of the middle school in different provinces in Quebec. Good psychometric properties have been reported for children attending Grade2 and Grade3 (test-retest reliability above 0.80 for both Decoding and Comprehension, and internal consistency equal to 0.93 and 0.98 for Decoding and Comprehension, respectively)101.

The PA, PM, RBTP, RAP and RAN measures are described in detail elsewhere102. Briefly,

  • PA was measured using a phonological deletion task (Test d’analyse auditive en français)103 adapted from the Auditory Analysis Test104. Children were asked to repeat a word that they just heard (Please, say “orteille”) and then to repeat the same word without a sound segment in it (Please, say the word again but without the /t/). Twenty-four items were administered to children in order of increasing difficulty. The total score was the sum of correct answers. A good internal consistency was reported (alpha between 0.68 and 0.93)103;

  • PM was assessed using the Repetition of Nonsense Words subtest of the NEPSY105. The subtest is composed of 13 items with an increasing number of syllables children had to repeat. Children had to pronounce each syllable of pseudo-words as fast as possible. A maximum of three minutes was allowed for the test. The total score is the sum of correctly repeated syllables. Good reliability (.77) and validity indicators have been reported105;

  • RBTP was assessed using a task inspired by the Auditory Repetition Task106 in which children had to say which of two stimuli, a light or a sound came first. The visual stimulus was a flash produced by a small circular Light-Emitting Diode placed three feet from the child. The auditory was a 5 ms 1-kHz sound; inter-stimulus intervals (ISIs) of 60, 180, 300, and 420 ms were applied, and the stimuli and the ISIs were randomly distributed through 32 trials. The total score was the sum of correct answers;

  • RAP was assessed using a task inspired by the Seashore Measure of Musical Talents107 in which children had to identify which of two consecutive 1-kHz sounds (a 800 ms standard and 700, 675, or 650 ms comparators) was longer. The order of presentation of the two stimuli was randomly distributed and an ISI of 500 ms was applied. The total score was the sum of correct answers;

  • RAN was measured with the non-alphanumeric serial Rapid Automatized Naming task of 50 items of color108. Children had to name the colors from left to right as fast as possible. Good reliability (0.92) and validity (0.90) was reported108,109. Time (seconds) was retained for analyses.

Genotyping, quality control, and imputation

A first round of DNA collection (saliva and blood) was run on a subsample of twins and their parents when children were 8 years old (N = 581, including 136 MZ twins and 407 parents)100. When participants were 19 years old, a second round of DNA collection was performed (N = 328 including 38 MZ twins)100. Concerning MZ pairs, one child per pair was selected based on the maximum availability of phenotypic data, or otherwise randomly. Genotyping was done by Genome Quebec in 2016. Genome-wide genotype data were generated using semi-custom chip based on Illumina® PsychArray-24v1.1 with an additional 790 SNPs selected based on prior knowledge of regions of interest for psychological, cognitive or social phenotypes. The genetic data was processed using Illumina’s Genome Studio platform, following the manufacturer’s guidelines.

Quality control was executed using PLINK v1.90110. SNPs with call rates <98% or a minor allele frequency (MAF) < 1% were removed, as well as participants with <98% genotyping rate. Additional participants were excluded because of sex mismatch, genetic duplicates, cryptic relatives (pi-hat≥12.5), or deviation from the mean autosomal heterozygosity ( > 4 SD). A multi-dimensional scaling analysis of genome-wide genotype data was used to identify and discard participants who did not cluster with most of the other participants ( > 4 SD) to avoid a type I error related to the population structure. Genotype phasing was conducted using SHAPEIT v2 (r837)111. Imputation of SNPs was carried out using IMPUTE2 v2.3.2112 in 5-megabasepair chunks with 500 kilobase buffers and the 1000 Genomes Phase 3 was used as the reference panel. After imputation, we removed variants with a MAF < 1%, an HWE test p < 1 × 10–6, and an INFO metric<0.8. We then converted probabilities to best-guess genotypes using PLINK with the default 0.1 threshold110. The genotypes were subsequently used for the computation of PGSs.

Polygenic score calculation

Reading-PGS was calculated using PRS-CS113. The contribution of each SNP was weighted by the effect size of each allele as reported in the original GWAS on word reading within the mixed ancestry sample16; the weighted contribution of each SNP was then added to estimate the genetic susceptibility for each subject. The global shrinkage parameter phi was fixed to 0.01113. The same method was applied to estimate PGS-EY4 and PGS-IQ3. All the PGSs were corrected for population stratification using twenty principal components of genetic relatedness matrix, and standardized residuals were used.

Statistical analyses

IBM SPSS Statistics for Windows, version 26.0114 and Mplus software package version 8.5115 were used respectively for descriptive\correlation analyses and SEM mediation modeling. As we analyzed direct and indirect relationships between observed variables in a hypothesis-driven model, we ran path analysis which represents a subset of SEM116,117,118,119,120. Moreover, as at the cognitive level several neurocognitive and sensory traits have been established as reliable EPs for reading skills, we concurrently modeled all paths and covariances among EPs in the implemented SEM mediation modeling (Fig. 1).

This method provides more accurate, robust and powerful estimations of mediation paths when weighed against traditional tests, enables the comparison of the relative importance of each path and controls for type I error121. As RAN was not normally distributed (Supplementary Table 6), we winsorized the data to minimize the influence of outliers in the SEM mediation modeling. Indirect effects were examined using the 5,000 bootstrap technique to assess non-normality in the product coefficient122 and 95% confidence intervals were used to identify significant mediation pathways123. The chi-square statistic, standardized root mean square residual (SRMR ≤ 0.08), root mean square error of approximation (RMSEA ≤ 0.08), and comparative fit index (CFI ≥ 0.95 indicating adequate fit) were used to assess model fit121. As the sample included children from the same families (DZ twins), we accounted for the relatedness using a clustering variable to control for this non-independence when estimating the contributions of Reading-PGS (Fig. 1). According to its correlation with reading-related skills and EPs (Table 1), we controlled for school grade level in all models.