In this study, we examined a subset of 1279 participants enrolled in SEED Phase I (SEED 1) with genome-wide genotyping and an adjudicated ASD or typical control outcome classification. SEED is a multisite case–control study based in the United States that was designed to investigate risk factors for ASD, including environmental and genetic factors (Schendel et al., 2012). SEED 1 enrolled children aged 2–5 years, approximately balanced between ASD cases, population-based controls (Peterson et al.), and children with non-ASD developmental delays, and their parents. All children were recruited from the same geographic areas and birth date range (Schendel et al., 2012). SEED ascertained rigorous clinical phenotyping, information on GI symptoms and conditions, as well as biospecimens. A detailed description of SEED recruitment, inclusion/exclusion criteria, clinical phenotyping, and data collection procedures can be found in Schendel et al. (2012). All families provided written consent for participation and institutional review boards at local study sites and the Centers for Disease Control and Prevention (CDC) approved the study.

Detailed information on SEED 1 clinical evaluations and ASD diagnostic criteria are previously published (Wiggins et al., 2015). Briefly, SEED participants were recruited from one of three ascertainment groups: (1) the general population (Peterson et al., 2019); (2) children with a broad array of developmental delays or disorders (DD); and (3) children with ASD. Only ASD cases and POP controls were included in this report. Children were screened with the Social Communication Questionnaire (SCQ) (Rutter et al., 2007). Children who scored 11 or greater, or who had previously been diagnosed with ASD (regardless of SCQ score), completed a full ASD developmental evaluation that included both the Autism Diagnostic Observation Schedule (ADOS) (Anttila et al., 2018) (Gotham et al., 2007) and Autism Diagnostic Interview-Revised (ADI-R) (Lord et al., 1994). All children (including those with SCQ < 11) completed the Mullen Scales of Early Learning (Mullen, 1995) developmental evaluation. Children with a mental age of 24 months or greater were classified as an ASD case if they met ASD diagnostic criteria on the revised ADOS and met the ADI-R diagnostic criteria or any one of three ADI-R relaxed criteria, as approved by the instruments’ authors. Children who had a mental age less than 24 months were classified as ASD if they met the above criteria and a clinician was certain they had ASD via the adapted Ohio State University Autism Rating Scale. Children recruited as POP controls who scored 11 or greater on the SCQ but did not otherwise meet the criteria for ASD classification were retained as POP controls (Schendel et al., 2012) (Wiggins et al., 2015).

GI symptoms were ascertained in SEED using a GI questionnaire designed for the SEED study (Reynolds et al., 2021). Parents reported whether their child had any current “gastrointestinal (bowel) problems on a regular basis”, that is, greater than two times per month. They also reported nine more specific symptoms: vomiting, diarrhea, loose stools, constipation, loose stools alternating with constipation, abdominal pain with meals, abdominal pain relieved by defecation, pain on stooling, and gas.

Genomic DNA was isolated from individuals through blood or buccal samples at the SEED biorepository. Samples were sent to the Genetic Resources Core Facility at Johns Hopkins University for processing on the Illumina Omni1M-Quad array, or to the University of California—San Francisco for processing on the Affymetrix Axiom KP array. Rigorous quality control (QC) metrics were applied to the genotype measurements to ensure data integrity by removing poorly performing samples and genomic markers. Samples with marker call rate <98%, a sex discrepancy between genotype and self-report, evidence of unexpected relatedness (Pi-hat >0.2), or excess hetero- or homozygosity were removed. Single nucleotide polymorphisms (SNPs) that had a call rate <96%, were monomorphic, had minor allele frequency <5%, or that deviated significantly (p < 0.0001) from Hardy–Weinberg equilibrium were also removed. Eigensoft (Price et al., 2006) was used to compute principal components that represent the genetic ancestry of each sample, which were used as covariates in our downstream analyses. Phasing was performed using SHAPEIT (Delaneau et al., 2013) and imputation was performed using IMPUTE2 ( Howie et al., 2009), with 1000 Genome Project, version 3, as the reference panel (Howie et al., 2011).

A total of 1279 unrelated individuals passed QC, completed imputation, had information on GI symptoms, and had a final classification definition as POP or ASD. Three hundred and forty-five of these samples were genotyped on the Axiom array, and 934 were genotyped on the omni array. This included 564 ASD cases, and 715 POP controls.

To create a proxy for genetic susceptibility to GI symptoms, we chose inflammatory bowel disease, including Crohn’s disease and ulcerative colitis as they have a well-documented genetic basis, and similar symptoms as those more frequently observed in ASD cases, such as diarrhea, loose stool, and abdominal pain (Baumgart & Sandborn, 2012; Danese & Fiocchi, 2011). Crohn’s disease and ulcerative colitis are the two types of inflammatory bowel disease (Liu et al., 2015). Large-scale genome-wide association studies have identified genetic variants associated with inflammatory bowel disease (Alonso et al., 2015; Duerr et al., 2006; Festen et al., 2011; Liu et al., 2015), and estimated heritabilities of 0.67 for ulcerative colitis and 0.75 for Crohn’s disease (Chen et al., 2014). Polygenic scores take variation in multiple genes into account and can serve as a measure of underlying aggregate genomic susceptibility (Wray et al., 2007).

We calculated PGSs for Crohn’s disease (CD-PGS), ulcerative colitis (UC-PGS), and inflammatory bowel disease (IBD-PGS) in our SEED analytic sample based on summary statistics from a large genome-wide association study (GWAS) conducted by the International Inflammatory Bowel Disease Genetics Consortium (Liu et al., 2015). After downloading the summary statistics, we applied several filters for QC purposes. We removed SNPs that had minor allele frequency >5%, an imputation information score >0.6, and a p-value for heterogeneity in direction of effect between cohorts >0.05. We excluded duplicate SNPs, insertions, deletions, and palindromic SNPs. In order to account for linkage disequilibrium among SNPs, a two-step clumping procedure was performed in PLINK 1.9 (Chang et al., 2015). We first carried out short rage clumping of SNPs by specifying an r² value > 0.5 within a 250 kb window, and then carried out a longer-range clumping step on the remining SNPs by specifying an r² value > 0.2 within a 5000 kb window. Discovery GWAS sample sizes before and after filtering are listed in Table S1.

Ideally, we would evaluate the associations of our PGSs with each GI disorder in our sample to evaluate an optimal p-value threshold for inclusion of SNPs in a GI-PGS calculation. Our analytic sample of young children has only one UC and no CD cases, as expected in a childhood sample, so we could not set a p-value threshold for PGS prediction of these disorders in our sample empirically. Instead, we included SNPs in the three PGSs whose discovery findings fell below p < 0.001, as suggested by a previous simulation study (Zhang et al., 2018) to maximize prediction. The total number of SNPs included in the PGS calculated for ulcerative colitis was 615, for Crohn’s disease was 740, and for inflammatory bowel disease was 826. Genetic scoring was carried out in PLINK 1.9 using the—score command, which computes a score for each individual by summing across the number of risk alleles for each SNP, weighted by the discovery effect size. We standardized each PGS by centering to the mean and rescaling so a 1-unit increase is equal to 1 SD to aid in the interpretation of the association effect sizes.

Density plots for each PGS by ASD status were visually inspected to assess normality and identify outliers. To formally test the associations of each GI-PGS with ASD case status, we employed logistic regression. All modeling included the first 10 principal components for ancestry. In addition, we performed logistic regression stratifying by European ancestry status to reduce population stratification. Each regression coefficient represents the change in log odds of ASD for every one standard deviation unit increase in the PGS. p-Values <0.05 were considered statistically significant.

Using logistic regression, we assessed the strength of association between the three GI-PGS and each of nine GI symptoms in POP controls and ASD cases separately. The results among POP control children could help determine which symptoms could be explained by the variants captured by the GI-PGS, uninfluenced by the potential behavioral causes of GI symptoms in ASD cases. In Model 1, each of nine GI symptoms served as the dependent variable, and were regressed onto each of the three GI-PGS, for a total of 27 models which were restricted to controls: Model 1 : E GI symptom among controls = α + β1GI − PGS + C ancestry principal components 1−10 + ε.

In Model 2, we performed the same analysis in ASD cases to give insight on whether the PGS has a different effect across ASD cases and controls and to inform the modeling below: Model 2 : E GI symptom among cases = α + β1GI − PGS + C ancestry principal components 1−10 + ε.

In Model 3, we include a cross-product term to evaluate the statistical interaction between ASD status and GI-PGS to formally test for effect modification of the PGS effect on GI symptoms by ASD case and control status. Outcomes for Model 3 included any GI symptom, and specific GI symptoms that were significantly predicted by the GI-PGS at a multiple testing threshold of p < 0.016: In addition, we model the GI-PGS as a binary covariate, with a “high GI risk” group that had GI-PGS in the top tertile, and a “low GI risk” group that had GI-PGS in the bottom two tertiles.

Children with ASD were significantly more likely to have any GI symptoms, and all nine specific GI symptoms, compared to POP controls (Table 1). A larger percentage of ASD cases were males and European ancestry was significantly higher among controls (70.0%) than ASD cases (53.7%) (Table 1).

The three GI disorder PGSs appeared normally distributed within both the ASD cases and the control samples, and we did not identify any outliers (Figure S1). Unadjusted mean differences between ASD cases and POP controls were observed for each PGS (Table 1). For example, the CD-PGS was significantly (p = 0.019) lower in the ASD group (mean = −0.08) compared to the control group (mean = 0.05). However, these associations were not statistically significant after adjusting for genetic principal components or after stratifying by European ancestry (Table 2).

Among POP control children, each GI-PGS was associated with increased odds of having any GI symptom, adjusting for ancestry, but this only reached statistical significance for the UC-PGS (p = 0.05) (Figure 1a). When considering nine specific GI symptoms, control children with an increased GI-PGS showed higher odds of constipation, diarrhea, gas, loose stools, and loose stools alternating with constipation, with loose stools alternating with constipation reaching statistical significance for CD-PGS and diarrhea reaching statistical significance for IBD-PGS and UC-PGS (Figure 1b).

Among controls, Diarrhea was significantly predicted by the UC-PGS (p = 0.015) at a multiple testing threshold of p < 0.016 (Figure 1b), suggesting that this PGS captured variation in diarrhea in our POP control sample. No other GI symptoms were significantly associated with other PGS at the multiple testing threshold.

Among ASD cases, none of the GI-PGS were significantly associated with increased odds of having any GI symptom (Figure 2a) nor with any specific symptoms (Figure 2b).

Given the differences in ASD outcome-stratified results for the UC-PGS, we tested for GI effect modification for UC-PGS risk group by ASD status. Children were included in the “high GI risk” group if their UC-PGS was in the top tertile of all children included in the analysis. We assessed effect modification on any GI symptom and on diarrhea because these traits were well captured by the UC-PGS in our POP control sample. We find a difference in the effect of UC-PGS on any GI symptom risk by ASD status (cross-product term for interaction aOR = 0.68, 95% CI = 0.37–1.27, p = 0.22). This relationship reached statistical significance within a European ancestry subset (aOR = 0.42, 95%CI = 0.19–0.88, p = 0.02). Similarly, we also found UC-PGS effect modification on diarrhea risk by ASD among a European ancestry subset (cross-product term for interaction aOR = 0.68, 95% CI = 0.11–3.52, p = 0.66).