Background

Extreme prematurity (birth < 28 weeks of gestation) remains one of the leading causes of neonatal morbidity and mortality in the USA [1]. Although survival rates for infants born extremely preterm have improved dramatically in recent decades, children born extremely preterm remain at higher risk for cognitive impairment, with lower average general intelligence and executive function deficit [2,3,4,5,6] and 9-fold higher risk of severe cognitive impairment compared to children born full-term [7,8,9,10,11,12,13]. Adverse neurodevelopmental outcomes, such as cognitive impairment, affect ~ 1 million preterm infants born each year [14] and may persist through adulthood [15,16,17,18]. Although cognitive impairment is not always severe, even mild deficits can have substantial impact, resulting in a spectrum of outcomes from difficulties in school to inability to lead an independent adult life [19]. Specific problems can include deficits in executive function, language, learning and memory, attention, perceptual-motor function, and social cognition [3, 6, 20, 21], which negatively affect well-being [19]. Cognitive impairment has life-long effects on quality of life, with significant familial and social capital costs. Although precise data are limited [22], lifetime costs collectively for children born in 2000 with intellectual disability alone are estimated at $51.2 billion [23].

Despite substantial research efforts to understand neurodevelopment outcomes, we know remarkably little about genetic factors and molecular mechanisms influencing cognitive function in preterm children. Previous genetic studies have discovered hundreds of genetic variants that can predispose children to neurodevelopmental disorders including autism spectrum disorder [24], attention deficit disorder with hyperactivity [25], intellectual disability [26, 27], specific language impairment [28], specific learning disorders [29], and childhood onset schizophrenia [30]. Some genetic studies have evaluated genetic risk factors for neurodevelopmental outcomes for preterm children or children with low birth weight [31,32,33,34,35,36]. MAOA was found to be associated with mental development throughout early childhood among preterm children [31]. A variant rs4074134 of BDNF, and a rare insertion/deletion in the intron region of SLC6A4 were significant predictors of cognitive performance at school age in a study of genetic risk factors for poor cognitive development in children with low birth weight [36]. With preterm infants from a randomized controlled trial (RCT) examining antenatal exposure to corticosteroids, Clark et al. found variants of IL1B, IL4R, and IL6 associated with lower scores on the Bayley’s Scales of Infant Development and developmental delay at age 2 [34], and Costantine et al. [33] found that variants of VIP and GRIN3A were associated with cerebral palsy. A COMT variant was associated with reduced corpus callosum size in adults with history of preterm birth [32].

However, previous studies do not explain the pathways through which these variants or genes might influence the risk of poor cognitive outcomes, and few genome-wide association studies (GWAS) examined the genomic regions associated with cognitive function among children born extremely preterm. Therefore, identifying genetic factors that are associated with children’s cognitive function and understanding related mechanisms are necessary to develop earlier screening assessments and effective precision interventions and understand why some preterm children of the same gestational age do worse than others. To advance along these directions, we utilized samples from the extremely low gestational age newborns (ELGAN) cohort [37], the largest US-based study of children born extremely preterm, to identify genetic factors associated with cognitive impairment at age 10 years. Finally, integrative analysis with brain expression quantitative trait loci (eQTL) and chromatin interactome data was performed to identify potential causal variants and functional genes underlying the GWAS associations.

Methods

Study participants

ELGAN is a multicenter cohort study originally designed to identify exposures increasing risk of structural and functional neurologic disorders in children born extremely preterm [37]. A total of 1506 infants born before the 28th week of gestation and 1249 mothers were enrolled during the years 2002–2004. Study participants were enrolled at 14 hospitals in the United States to achieve a large enough sample size and generalizability. The enrollment and consent procedures were approved by the individual institutional review boards. At the age of 10 years, 889 of the surviving children returned for follow up (ELGAN2, 92% of the 966 who were recruited for this phase of the ELGAN Study) and were assessed for cognition capacity, learning abilities, and impairments in executive function [7]. Of these children, 528 had genotype data available for analysis and thus constitute the sample size of this paper. Table 1 summarizes demographic information for the ELGAN2 cohort and the ELGAN2 subset with genetic data (n = 528) we used in our analysis.

Table 1 Participant characteristics of the ELGAN2 subset and ELGAN2 cohort
Full size table

Cognitive function at age 10 years

Cognitive function at age 10 years was assessed with latent profile analysis (LPA) [38], which empirically identifies subgroups of children who share similar profiles on a set of measures. The LPA included 9 cognitive measures including verbal and nonverbal intelligence quotient (IQ) and several measures of executive function (EF). IQ was assessed with the School-Age Differential Ability Scales–II (DAS-II) Verbal and Nonverbal Reasoning scales. EF was assessed with two subtests from the DAS-II and five subtests from the Developmental NEuroPSYchological Assessment-II (NEPSY-II). Working memory was evaluated with the DAS-II Recall of Digits Backwards and Recall Sequential Order test. The NEPSY-II Auditory Attention and Auditory Response Set, Animal Sorting Inhibition, and Inhibition Switching subtests were utilized to examine auditory attention and set switching, concept generation and mental flexibility, and simple inhibition and inhibition shifting, respectively [7]. It has been shown that characterizing cognitive function using measures of executive function in addition to IQ better discriminates the academic performance and educational needs of children born extremely preterm [38]. LPA classifies subjects who share a similar pattern of scores on the measured variables, while maximizing the difference in scoring patterns across distinct profiles [39]. It assigns subjects into a finite number of profiles by identifying the most likely model that describes the heterogeneity of data, which is known as finite mixture models.

To determine the optimal number of profiles, LPA was fit to the data, and Bayesian information criteria (BIC) [40], sample-size-adjusted Bayesian information criteria (SSABIC) [41], and Lo–Mendell–Rubin-adjusted (LMR) likelihood ratio test [42] were used to assess model fit. Children were categorized by their most likely latent profile for further analysis. In this sample, a four-profile model provided the best fit for the data [38]. For our analysis, we used a binary classification that grouped participants into two previously validated distinct profile groups (LPAx) [38, 43]: no or low cognitive impairment and moderate-to-severe cognitive impairment.

Genotype data

Genomic DNA was isolated from umbilical cords and genotyping was performed using Illumina 1 Million Quad (Illumina Inc, San Diego, California). This work was done as part of the candidate gene analysis of severe intraventricular hemorrhage (IVH) in preterm born infants [44], where infants with birth weights 500–1250 g and severe grades IVH and neonates with normal cranial ultrasounds were enrolled prospectively at 24 universities. A subset of ELGAN participants were provided as additional samples along with samples from a few other studies in the IVH study [44].

We performed variant level and sample level quality control (QC) on genotype data. For variant level QC, we excluded variants with call rate < 90% or minor allele frequency (MAF) < 1%. For sample level QC, we excluded samples with missing rate > 10%. These resulted in 700,845 SNPs and 528 samples using plink v.1.90 [45, 46].

Genotype imputation

Starting with the quality controlled (QCed) genotype data, we used the Michigan imputation server [47] for phasing and imputation using TOPMed freeze 5 [48] as the reference panel. Specifically, Eagle [49] was used for phasing and Minimac4 [50, 51] was used for imputation. We performed strand matching by dropping ambiguous (i.e., A/T or C/G) SNPs and by flipping non-ambiguous SNPs that were initially in − strand when compared to alleles in the + strand observed in the TOPMed freeze 5 reference panel. Genotype data was lifted over to genome build hg38. In total, we obtained ~ 34 million well imputed variants, with 5.5 million variants having MAF > 5%, 10.5 million variants having MAF > 1%, and 17.4 million variants having MAF > 0.5%. Well imputed variants were defined by having Rsq > 0.8, where Rsq is the estimated imputation quality metric [50, 51]. To evaluate the imputation accuracy, we randomly selected 5% of the genotyped variants on chromosome 1 and performed genotype imputation with the rest 95% genotyped variants, again with the TOPMed freeze 5 reference panel. The 5% masked genotyped variants were saved for imputation quality evaluation. Specifically, we calculated the squared Pearson correlation between imputed genotypes and true observed genotypes. For the 1956 variants on chromosome 1 tested, the mean squared Pearson correlation was 0.97, which suggests that most variants were well imputed even with a relatively small sample size (Fig. S1, Additional file 1), consistent with what has been reported in the literature [52,53,54].

Genome-wide association analysis

We used EPACTS 3.3.0 [55] for single variant association testing. To account for relatedness among samples, we used the EMMAX (Efficient Mixed Model Association eXpedited) test [56], which is an efficient implementation of mixed model association accounting for sample structure including population structure and hidden relatedness. Biallelic SNPs with MAF > 2% (did not account for relatedness) and Rsq > 0.8 were included in the analysis. In total, 8,535,130 variants were included in the association analysis. For the 528 samples who had genotype and covariates data available, we inferred kinship matrix using EPACTS and top 10 principal components (PCs) from the genotype data using PLINK. We performed the association test on the outcome LPAx, a binary outcome that classifies children into no or low cognitive impairment and moderate-severe cognitive impairment groups. The covariates for the single variant association analysis included gestational age, maternal education, maternal race, sex of the infant, and top 10 PCs. We performed sensitivity analysis for variants showing suggestive signals by further adjusting interaction terms between covariates.

Results

Association analysis results

We conducted a GWAS on LPAx of 528 samples from the ELGAN2 cohort. We identified two genome-wide significant loci from the 8,535,130 variants tested: STX18 and TEAD4, which are located on chromosome 4 and chromosome 12, respectively (Fig. 1). The index SNPs are rs79453226 (MAF = 0.036) and rs11829294 (MAF = 0.145) at the STX18 and TEAD4 loci, respectively. The genomic inflation factor λ, which measures the inflation in the test statistics and is calculated as the ratio of the median of the empirically observed distribution of the test statistics to the expected median (median of a chi-square distribution with one degree of freedom), is 1.038, which suggests no significant inflation of test statistics or excess false positive rate (Fig. 2). Table 2 shows genome-wide significant variants, and suggestive variants with p values less than 1e-6. For this set of suggestive variants, we performed additional association analysis by including interaction terms between non-genetic covariates (gestational age, maternal education, maternal race, and sex of the infant). We evaluated each interaction term separately, and we found that the two genome-wide significant loci remained significant and most suggestive loci had similar significance levels as in the original model (Fig. S2, Additional file 1), indicating that our top loci are robust to the further adjustment of interactions between covariates.

Fig. 1
figure 1

Manhattan plot. The Manhattan plot visualizes the association of SNPs along the genome with the LPAx trait. X-axis represents genomic location and y-axis represents -log10(p value). Each dot represents a SNP tested. SNPs above the red horizontal line, which marks the 5 × 10−8 are considered genome-wide significant. This plot was generated using the R package karyoploteR [57]. NCBI build 38

Full size image
Fig. 2
figure 2

QQ plot. A quantile-quantile (Q-Q) plot is used to characterize the extent to which the observed distribution of the test statistics follows the expected null distribution. This plot was generated using the R package qqman [58]

Full size image
Table 2 Significant and suggestive association results for LPAx
Full size table

Figure 3 shows LocusZoom [59] plots for the two genome-wide significant loci, with linkage disequilibrium (LD) from TOP-LD [60], calculated using TOPMed European and African participants. We can see that in the European population, the lead variant rs11829294 in the TEAD4 region has a number of LD tags (e.g., 21 variants with r2 ≥ 0.8) and some of them had highly significant p values; by contrast, the lead variant rs79453226 in the STX18 region has fewer LD tags (2 variants with r2 ≥ 0.8) that showed suggestive association (Fig. 3a). In the African population, the lead variant rs11829294 in the TEAD4 region has only 2 LD tags, which did not have significant or suggestive association, and the lead variant rs79453226 in the STX18 region has no LD tag with r2 ≥ 0.8 (Fig. 3b).

Fig. 3
figure 3

Locus zoom plots for the two genome-wide significant loci. Colors represent linkage disequilibrium r2 values calculated from TOPMed individuals with the lead SNP in each plot. a Locus zoom plots with linkage disequilibrium r2 values calculated from TOPMed European ancestry individuals. b Locus zoom plots with linkage disequilibrium r2 values calculated from TOPMed African ancestry individuals. NCBI build 38

Full size image

Epigenetic functional annotations

To further investigate the two loci identified for potential mechanisms, we examined several functional annotation metrics, including the CADD phred score [61] and the fathmm MKL score [62]. CADD phred score measures the deleteriousness of variants and is computed as -10*log10(rank/total). A CADD phred score of ≥ 10 indicates that the variant is predicted to be among the 10% most deleterious variants in the human genome, a score of ≥ 20 indicates among the 1% most deleterious. The fathmm MKL score predicts the functional consequences of variants where values above 0.5 are generally considered deleterious, and values below 0.5 neutral or benign. We also looked at the Genehancer feature [63] and the genes predicted by Genehancer. Table 3 shows functional annotations for variants that passed the suggestive p value threshold (p value < 1e-6). We observed that variants rs9424366, rs79946490, rs58545250, and rs17031018 were among the top 10% most deleterious in the human genome, and variant rs16913588 was predicted to be deleterious (with a fathmm MKL score of 0.97). Several variants were assigned by Genehancer as falling into enhancer regions with target genesTSPAN9, ITPR1, and CLIC4. These results provide evidence that some of the variants might have deleterious effects that are relevant to neurocognitive development in preterm children and suggest additional genes that might be functionally related.

Table 3 Epigenetic functional annotations for selected genome-wide significant and suggestive associations
Full size table

Chromatin interactions

We examined chromatin conformation data for additional functional implications based on physical contacts from Hi-C and alike technologies. Figure 4 shows virtual 4C plots generated by HUGIn2 [64] for the top two loci in adult cortex and fetal cortex Hi-C data [65]. HUGIn2 is a web-based viewer of genome-wide chromatin conformation data to explore chromatin spatial organization across multiple human cell lines and primary tissues. HUGIn2 can additionally incorporate data from multiple sources including genetic variants, chromatin organization features (e.g., topologically associating domains (TADs) [66], frequently interacting regions (FIREs) [67]), gene expression, and epigenetic annotations. For our purpose, we examined ± 500 kb regions around each locus. Significant chromatin interactions between the putative regulatory regions (harboring some GWAS variant(s)) and promoters of genes suggest the likely causal or effector genes regulated by the GWAS variant(s). The results were consistent between adult cortex and fetal cortex. The variant rs79453226 at the STX18 locus was linked to the promoter regions of several genes, including STX18 and NSG1 (Fig. 4a), and the variant rs12322215 at the TEAD4 locus was linked to FKBP4, FOXM1, RHNO1, TULP3, TSPAN9, and PRMT8 (Fig. 4b).

Fig. 4
figure 4

Virtual 4C plots. Centered at a rs79453226 b rs12322215 in adult cortex and fetal cortex. The bin containing the anchor position is indicated as a thick grey vertical bar. Different genes or regions can be highlighted in yellow. On the top is gene expression data with gene locations. Each gene is indicated by an arrow pointing the direction of transcription. The start site is indicated by the tail of the arrow. Each gene is labeled by its common name and highlighted in red indicating the expression level: the deeper the red color the higher the expression. On the bottom is the chromatin interaction Hi-C data that is plotted as a virtual 4C plot with the given anchor position. The black line shows the observed counts, the red line shows the expected counts, and the blue line shows the -log10(p value). The range of the -log10(p value) is plotted on the y-axis on the right while the range of the count data is shown on the left. The x-axis is the genomic location in Mb. NCBI build 37

Full size image

Overlapping with brain eQTL

Next, we investigated whether we could find any brain eQTL signals among the top variants. We examined all variants with LD r2 ≥ 0.6 with variants that passed the suggestive p value threshold (p value < 1e-6) using LD calculated from TOPMed European ancestry samples. Table 4 shows variants overlapped with commonMind eQTL [68] with FDR < 5%. Multiple brain eQTLs for PRMT8 on chromosome 12 in LD with the index SNP rs11829294 were identified.

Table 4 Variants overlapped with commonMind eQTL
Full size table

Overlapping with selective sweeps

We also examined whether the identified top variants overlapped with selective sweeps detected by S/HIC [69]. Table 5 shows multiple variants in LD (r2 ≥ 0.6) with the top variants overlapped with selective sweeps: all except one on chromosome 12 at the TEAD4 locus and one in an intergenic locus on chromosome 2. We found that some variants at the TEAD4 locus were located in soft sweep regions, where selection on standing variation produced qualitatively different skews in LD and allele frequencies.

Table 5 Variants overlapped with selective sweeps
Full size table

Discussion

Cognitive impairment is highly prevalent among children born extremely preterm. Yet, limited evidence is available on the genetic factors that may contribute to this kind of impairment. In this study, we aimed to identify genetic factors that are associated with children’s cognitive function and understand related genetic mechanisms by utilizing samples from the ELGAN cohort. Leveraging an LPA-derived phenotype and genetics data, we identified two genome-wide significant loci in our genome-wide association analysis for LPAx (a data-derived cognitive impairment outcome): TEAD4 (rs11829294, p value = 2.40e-8) and STX18 (rs79453226, p value = 1.91e-8).

We utilized chromatin conformation data from multiple human cell lines and primary tissues to see whether there are significant chromatin interactions between the two genome-wide significant loci and their neighboring regions. In adult cortex and fetal cortex, we found that variant rs12322215 (p value = 1.08e-07) in high LD with rs11829294 (r2 = 0.883) is linked to promoter regions of a few genes including TSPAN9 and PRMT8 (Fig. 4). Furthermore, the association at the TEAD4 locus rs11829294 and a few other variants that showed suggestive significance at the same locus were assigned by Genehancer as falling into the enhancer region of TSPAN9 (Table 3). We also observed TSPAN9 is highly expressed in both adult cortex and fetal cortex but not in hippocampus, and we did not observe similar chromatin interactions in hippocampus (Fig. 4). TSPAN9 is located at chr12:3,077,355-3,286,564 (GRCh38/hg38) and is one of tetraspanins, a superfamily of glycoproteins that function as “organizers” of cell membranes by recruiting other receptors and signaling proteins into tetraspanin-enriched microdomains and induce normal platelet activation [70]. Those proteins mediate signal transduction events that play a role in the regulation of cell development, activation, growth, and motility. TSPAN9 is highly expressed in normal brain tissues, including cerebellum and cerebellar hemisphere [71]. These pieces of evidence suggest the potential regulatory role of rs11829294 and its LD buddies on the TSPAN9 gene that could impact cognitive development among children born extremely preterm.

We also performed integrative analysis with brain eQTL to identify potential functional genes underlying the genome-wide significant association. A few brain eQTL for PRMT8 were found to be in high LD with rs11829294 (Table 4). PRMT8 is a member of the protein arginine N-methyltransferase (PRMT) family, which mediates protein arginine methylation, a common post-translational modification that has been implicated in signal transduction, RNA processing, transcriptional regulation, and DNA repair [72]. PRMT8 was found to be associated with the plasma membrane and has a tissue-specific expression in brain. Specifically, it is highly expressed in nucleus accumbens (basal ganglia), putamen (basal ganglia), cortex, caudate (basal ganglia), frontal cortex (Brodmann area 9), and anterior cingulate cortex (Brodmann area 24) [71]. It was also identified as a tissue-restricted enzyme responsible for proper asymmetric dimethylarginine (ADMA) level in postmitotic neurons where PRMT8-dependent arginine methylation is required for neuroprotection against age-related increased of cellular stress [73]. Moreover, PRMT8 in human embryonic stem cells (hESCs) plays an important role not only in maintaining pluripotency but also in controlling mesodermal differentiation [74]. Along with the evidence that variant rs12322215 is linked to the promoter region of PRMT8, we conclude that PRMT8 is another biologically plausible gene regulated by eQTL at the TEAD4 locus that could have potential effects on cognitive impairment among preterm children.

Additionally, we found that multiple variants in LD with rs11829294 overlapped with selective sweeps detected by S/HIC (Table 5). There is evidence that soft sweeps are widespread and account for the vast majority of recent human adaptation, and positive selection may often proceed via “soft sweeps” acting on mutations already present within a population. Furthermore, linked positive selection affects patterns of variation across much of the genome, and may increase the frequencies of deleterious mutations [69]. Therefore, these variants in soft sweep regions provide additional evidence for the associations found at the TEAD4 locus being biologically plausible and potentially causal.

For rs79453226, we found that it is linked to promoter regions of STX18 and NSG1 (Fig. 4). We did not find as much evidence for the STX18 locus supporting the significant association as for the TEAD4 locus. By examining LD r2 values calculated from TOPMed, we observed that rs11829294 and rs79453226 have different LD structures at their loci (Fig. 3). Specifically, rs11829294 has a number of LD buddies with r2 ≥ 0.8 showing suggestive association with LPAx in the European population. In contrast, rs79453226 has fewer LD buddies and is not in high LD with any of the suggestive variants. Therefore, rs11829294 is more likely to tag effects from causal variants than rs79453226.

With association results and other information considered, we did not have direct evidence indicating TEAD4 and STX18 as causal genes. However, there is evidence that TEAD4 and STX18 are related to placental development and brain respectively.TEAD4 is a member of the TEAD transcription factor family, which is best known for transcriptional output of the Hippo signaling pathway and has been implicated in processes such as development, cell growth and proliferation, tissue homeostasis, and regeneration [75]. TEADs have been found to be evolutionarily conserved, and have been shown to play important roles in various biological processes and human disease [76, 77]. Mouse knockout studies showed that TEAD4 is specifically required for embryo implantation and trophectoderm lineage determination [78, 79], which play important roles in placental development. TEAD4 null mice are embryonic lethal due to failure in embryo implantation; however, disruption of TEAD4 after embryo implantation results in normal development [78, 79]. TEADs seem to have important biological functions, but studies thoroughly characterizing TEAD function and regulation are lacking. In the future, we can utilize genome-wide DNA methylation, mRNA, and miRNA data from the placenta to study this gene more closely. The gene STX18 encodes a member of the syntaxin family of soluble N-ethylmaleimide-sensitive factor attachment protein receptors (SNAREs) which is part of a membrane tethering complex that includes other SNAREs and several peripheral membrane proteins, and is involved in vesicular transport between the endoplasmic reticulum (ER) and the Golgi complex [80]. It has also been shown that STX18 is important for the organization of two ER subdomains, smooth/rough ER membranes and ER exit sites by mediating the fusion of retrograde membrane carriers with the ER membrane [81]. Knockdown of STX18 caused a global change in ER membrane architecture, leading to the segregation of the smooth and rough ER. Moreover, the organization of ER exit sites was markedly changed concomitantly with dispersion of the ER-Golgi intermediate compartment and the Golgi complex. Variants in STX18 were previously found to be associated with brain volume measurement and neuroimaging measurement [82, 83].

One limitation of our analysis is that our results may not be generalizable to children who are not extreme premature. Another issue is the small sample size, although we were able to impute most variants well (Fig. S1, Additional file), it limits the statistical power of the association analysis. The few genome-wide significant single variant associations we found, and the non-statistically significant heritability estimate also suggest the need for better powered analyses (Additional file 1). It is also possible that variants included in our analyses are in low or moderate LD with true causal variants which are rare and cannot be well-imputed in the ELGAN2 cohort. While ELGAN2 is the largest cohort with genotype and long-term cognitive assessment for extremely preterm children currently available in the USA, in the future we hope to study a larger population with longitudinal data of cognitive function, to investigate whether there are genetic variants that interact with perinatal and neonatal immune factors to increase risk for development of trajectories of impaired cognitive function.

Conclusions

In this work, we present an innovative computational approach that combines LPA with multi-faceted genomic analysis to investigate potential genetic risk factors underlying cognitive impairment among children born extremely preterm. Our association analysis identified two genome-wide significant loci: TEAD4 at rs11829294 and STX18 at rs79453226. Further genomic analysis suggests that rs11829294 and its LD buddies have potential regulatory roles on likely functionally relevant genes TSPAN9 and PMRT8. This study provides new mechanistic insight into neurocognitive function among children born extremely preterm by performing an imputation-based GWAS with subsequent prioritization of causal variants and effector genes.