Sex-aware genome-wide assessment of de novo variants in autism across coding and noncoding regions

Abstract

Background

Autism spectrum disorder (ASD) shows a consistent sex bias, yet how sex shapes de novo variant (DNV) risk across coding and noncoding sequence remains unclear. I analyzed DNVs in > 41,000 parent–child sequenced trios from three ASD family-based cohorts and compared DNV characteristics and enrichment patterns in males and females. Notably, these trios included individuals with ASD as well as those without ASD. I developed a new sex-aware DNV caller (HAT-FLEX) and thoroughly evaluated each candidate DNV using an additional tool introduced in this study, SNOW, to generate a high-confidence callset.

Results

I identified enrichment of missense and loss-of-function (LOF) DNVs both overall and within known ASD-related genes (i.e., SFARI genes). Gene-specific enrichment analyses revealed twelve genes that were exome-wide significant and specific to males, for significance, including FOXP1, SMAD6, AUTS2, CCDC168, PIEZO1, EML6, ZNF84, IGSF23, OTOG, NHSL1, ADNP, and FREM3 and three genes that were specific to females, for significance, including TAOK1, MECP2, and DDX3X within a variant class. Direct comparisons of DNVs in males and females revealed MECP2 as the only exome-wide significant gene; however, GABBR2 was also trending toward enrichment in the direct males with ASD comparison to females with ASD. Furthermore, probit analyses support a female protective effect and demonstrates that damaging DNVs, particularly LOF variants, including potentially stronger X-linked effects in females, are associated with increased ASD liability, whereas synonymous variants are not. Finally, I analyzed promoters and identified a single significant promoter region (p = 3.8 × 10⁻¹³), associated with the WDR74 gene, with the signal driven by DNVs observed in males with ASD. Surprisingly, the noncoding RNA gene RNU2-2 lies within this significant WDR74 promoter and accounted for most of the DNVs in the region. RNU2-2 DNVs were present in 0.2% of individuals with ASD in comparison to 0.05% of individuals without ASD.

Conclusions

These findings show that ASD DNV risk differs by sex at both the gene and liability levels, supporting a female protective effect while highlighting sex-specific patterns in coding variation. They also identify RNU2-2 as a noncoding contributor to ASD risk, expanding the landscape of ASD-associated variation beyond protein-coding genes.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40246-026-00961-8.

Background

According to the Centers for Disease Control and Prevention, the overall prevalence of autism spectrum disorder (ASD) is one in 31 individuals. Furthermore, ASD exhibits a sex bias with a prevalence in males of one in 20 individuals and in females of one in 70 individuals [1]. ASD is considered to be multifactorial with contributions from both the genome and the environment, which is modeled using a liability distribution. Under Falconer’s threshold model, this indicates that the rare class (females with autism) would require a higher threshold to develop ASD than the more common class (males with autism) [2, 3]. More generally, sex bias in human phenotypes and disease is common and the reasons for this bias are important for understanding of both the underlying mechanisms and ultimately for more tailored future clinical care. Potential mechanisms underlying sex bias include effects from the genome, hormones, and other potential environmental interactions. Regarding the genome, the most obvious culprits are the (1) sex chromosomes, themselves, and the loci which they interact with in the autosomal genome and (2) the genomic loci interacting with hormones. There are several potential downstream effects of this including gene expression differences in males and females. Classical examples of genes related to ASD with effects from the X chromosome are FMR1 repeat expansions in Fragile X Syndrome in males [4] and protein-affecting de novo variants (DNVs) in MECP2 in Rett Syndrome [5]. Fragile X syndrome preferentially affects males and Rett syndrome preferentially affects females due to differences in underlying mechanisms. An example of an effect from the Y chromosome is the significant increase in ASD in individuals with XYY syndrome [6]. A recent example indicating hormonal involvement in ASD has shown a link between ASD and fetal estrogens [7].

With the advent of whole-exome sequencing (WES) and whole-genome sequencing (WGS), the study of rare variants in ASD genome-wide has expanded greatly. This has yielded several studies identifying an excess of rare and de novo variants (DNVs) in females with ASD as compared to males with ASD [8, 9]. In ASD, this phenomenon has taken on the identity of the female protective effect [10] and fits well with the threshold model described above. More generally comparisons of rare variants have been expanded out beyond ASD to a broader set of individuals including those with developmental disorders [11]. This is important; however, it is also relevant to note that developmental disorders more broadly do not exhibit as substantial of a skewing towards males as in ASD [9]. The result of these studies has been the identification of several genes on the X chromosome with enrichment in females with developmental disorders with the most common being DDX3X [12]. This gene is a gene on the X chromosome that escapes X inactivation and thereby is expressed at higher levels in females with ASD than males with ASD presumably providing a type of “buffered” expression.

Since the sex bias in ASD is higher than in general developmental disorders, the focus of this current study is on DNVs in large cohorts of families with ASD. A previous study has examined DNVs in ~ 23,000 trios with ASD and an additional ~ 30,000 individuals with developmental disorders [13]. A recent preprint reports assessment of DNVs in ~ 31,000 individuals with ASD [14], which included DNV calling in a subset of the individuals in the study and the integration of these with existing DNV calls from other cohorts. However, ongoing efforts to sequence more individuals with ASD has generated additional datasets. The most concerted of these efforts have come from the Simons Foundation through their cohorts including the Simons Simplex Collection (SSC) [15] and the Simons Powering Autism Research (SPARK [16]).

In this study, > 41,000 parent–child trios from families with ASD, that were sequenced by either WES or WGS through the SSC or SPARK, are assessed for DNVs in the protein-coding regions of the genome and a subset of > 9800 parent–child trios sequenced by WGS are assessed for DNVs in the noncoding regions of the genome. Importantly, the DNVs generated and analyzed in this current study are not integrated from different DNV callers but rather are from a uniform calling workflow as follows. First, I developed two modern computational tools for calling (i.e., HAT-FLEX) and optimizing DNV calls (i.e., SNOW) from WES and WGS data. Specifically, these sex-aware tools have been developed to handle DNV calling on autosomes, the pseudoautosomal regions (PARs), the non-PAR X chromosome, and the non-PAR Y chromosome. Second, I generated high-quality, sex-aware, DNVs using data from multiple sequencing technologies and characterized DNV patterns across three ASD family cohorts. Third, I evaluated gene-set and gene-specific enrichment of loss-of-function (LOF) and missense DNVs in males and females with or without ASD, including analyses involving the autosomes, the PARs, the non-PAR X chromosome, and the non-PAR Y chromosome. Fourth, I identified the rates of LOF and missense DNVs within genes and through a probit model identified importance of LOF and missense DNVs to ASD. Finally, I tested for enrichment of DNVs in promoter regions in males and females with or without ASD, respectively, leading to further followup that involved a noncoding RNA gene.

Methods

Study cohorts

Three cohorts consisting of families with ASD were accessed from the Simons Foundation Autism Research Initiative (SFARI) Base. These included a total of 41,421 parent–child sequenced trios from the Simons Simplex Collection (SSC) [15] and the Simons Powering Autism Research (SPARK [16] cohorts.

De novo variant calling

I attempted to call DNVs on individuals from all 41,421 trios (Supplementary Table S1). In total, 85 samples failed DNV QC based on outlying DNV counts: < 40 or > 150 DNVs in WGS, or > 8 DNVs in WES. I removed an additional 220 individuals who belonged to monozygotic twin pairs, retaining only one individual per twin pair to avoid duplicate representation. This resulted in a total of 41,201 independent parent–child sequenced trios; of which, 31,268 trios were from WES data (17,114 males with ASD, 4867 females with ASD, 4528 males without ASD, and 4759 females without ASD), and 9848 trios (4808 males with ASD, 1021 females with ASD, 1928 males without ASD, and 2091 females without ASD) were from WGS data. A description of how I called DNVs is below.

For WGS data, family-level genome-wide variant call files (VCFs) were obtained from a previous study [17]. Briefly, these VCFs were generated from genotyping Genomic VCFs (gVCFs) generated by GATK HaplotypeCaller [18] and DeepVariant [19], respectively, with GLnexus [20]. For WES data, I downloaded per-individual gVCFs generated with DeepVariant from SFARI Base (accession: SPARK iWES_v3). For each family, I performed joint-genotyping using GLnexus (release v1.4.1-0-g68e25e5).

For both WGS and WES, I called DNVs using my newly developed tool called HAT-FLEX (version 1.0, https://github.com/TNTurnerLab/HAT-FLEX, full description in Supplementary Methods). Briefly, HAT-FLEX is a Python post-calling computational tool for identifying DNVs from parent–child trios using one or two genome-wide VCFs per trio (e.g., GATK, DeepVariant), with an optional two-caller intersection mode; it supports human and other diploid species, reads trios from a simple family file or standard pedigree file, and applies sex-aware haploid/diploid logic on X/Y including awareness of PARs, so that filtering and parent relevance depend on genomic context. It detects candidate DNVs via sex-aware genotype patterns (reference alleles in parent(s), alternate allele in child), optionally splits multi-allelic sites by allele with normalization, can mask user-provided regions, and then applies layered quality controls including required FORMAT keys, strand-bias checks, depth/GQ thresholds, parent “ALT leak” rules (allele depth and genotype-based), allele-balance thresholds (diploid and haploid), homopolymer handling, and a “haploid unusual” flag for heterozygous-like signals on haploid loci. Outputs include a sorted, normalized trio-only VCF with added INFO/FILTER annotations (e.g., DENOVO, SEXCTX, IRRELPARENT, HAPUNUSUAL, CLUSTER), optional TSV and metrics/manifest summaries, and automatic bgzip compression plus optional tabix indexing (with gzip fallback), along with logging, and verbosity controls.

For WGS data, the parameters of HAT-FLEX included –ped to input a standard pedigree, –caller1_vcf for the DeepVariant family VCF, –caller2_vcf for the GATK HaplotypeCaller VCF, –intersect_mode of allele to intersect the two callers by allele, –normalize_alleles set to true, –par_bed set to build 38 coordinates, –gq_value set to 20, –depth_value set to 10, –haploid_depth_value set to 5, –ab_min set to 0.25, –haploid_ab_min set to 0.85, –require_format_keys set to GT,DP,AD,GQ, –strict_format set to false, –max_homopolymer set to 10, –cluster_window_bp set to 100, and –cluster_min_count set to 2. For WES data, the same parameters were used except for –caller2_vcf and –intersect_mode, which were not used, since only DeepVariant was used in the HAT-FLEX calling on the WES data.

Post-calling of DNVs, I developed a computational tool called SNOW (Second-pass de Novo variant Offspring Workflow, version 0.1.0, https://github.com/tycheleturner/snow, (full description in Supplementary Methods) to further process the DNVs. First, I collapsed DNV calls using snow consolidate, which merges nearby child-only variants into single records when they met strict eligibility criteria. Variants were processed in genomic order per contig, requiring the child to carry an alternate allele while all other samples were homozygous reference, and restricting merging to high-confidence PASS sites annotated with the HAT-FLEX cluster flag. Eligible variants within 100 bp were grouped into clusters; for clusters with multiple variants, I defined the minimal spanning interval, retrieved the reference haplotype, and assembled the alternate haplotype by substituting the alternate allele at each site while disallowing overlaps. I required at least two reads to support the assembled (collapsed) haplotype.

Next, I interrogated alignment data for evidence of the alternate allele at candidate DNVs, using snow parentcheck, which scans parental and child reads around each variant annotated with the HAT-FLEX DENOVO flag. Using the child VCF and the child, father, and mother CRAM files, I counted ALT-supporting reads in each individual. Support was assessed at the haplotype level from aligned query-reference base pairs, with explicit handling of SNVs, MNVs, and INDELs; reads were required to anchor the reference position and, for INDELs, to match the expected inserted or deleted sequence. For variants annotated with SEXCTX = PAR, I also queried the homologous PAR on the partner sex chromosome to capture shared coverage. I required zero high-quality ALT-supporting reads in both parents and at least two high-quality ALT-supporting reads in the child.

Finally, I inferred parent-of-origin for DNVs in trios, using snow phaser, which combines phase-by-transmission generated on a genome-wide trio VCF with read-backed physical phasing from the alignment data of the child. I first identified informative transmission sites genome wide as high-quality biallelic SNVs where the child was heterozygous and all trio genotypes had GQ ≥ 20, excluding non-PAR X/Y in male probands. For each DNV, I collected child reads supporting the ALT allele and phased them to nearby informative SNVs within ± 40 kilobase pairs (kbp) by determining which parental allele those same reads carried. I summed paternal and maternal support across informative sites to assign parent-of-origin by majority vote (ties or insufficient evidence were reported as unknown), and quantified confidence using a binomial test reported as both a 0 to 1 score and a Phred-scaled value. For male probands, non-PAR X and Y calls were assigned deterministically (X assigned maternal, Y assigned paternal).

Benchmarking HAT-FLEX and SNOW

An important check of HAT-FLEX and SNOW is QC of the full DNV callset generated in this study, which I describe later. Here I discuss my use of publicly available WGS alignment and DNV call data from prior studies [21, 22] to further benchmark the methods. Notably, these publicly available datasets were generated from blood-derived DNA rather than lymphoblastoid cell lines (LCLs), which are known to introduce cell line-specific artifacts that can confound DNV calling [23]. Therefore, this is an ideal testing dataset for benchmarking HAT-FLEX and SNOW. Since the source studies reporting DNV calls used the telomere-to-telomere (T2T) reference genome [24], I selected that dataset for benchmarking. I downloaded the DeepVariant-GLnexus callset derived from PacBio data, the DRAGEN callset derived from Illumina data, the PacBio BAM files for each individual, and the T2T reference genome. HAT-FLEX was applied to a single trio comprising child NA12877 and parents NA12889 and NA12890, using the DeepVariant-GLnexus and DRAGEN callsets as caller 1 and caller 2 in HAT-FLEX, respectively. All parameters were kept consistent with those used in the autism cohort analyses. Briefly, variants clustered on the same read were collapsed, additional allele-level validation was performed using the alignment data, and PacBio BAM files were used to phase the DNVs with SNOW.

Generating a cross-cohort DNV callset

The DNVs called in the ASD cohorts described above were annotated for their presence in the Human Pangenome Reference Consortium (HPRC) [25, 26]. First, the decomposed VCF from the HPRC was downloaded from https://s3-us-west-2.amazonaws.com/human-pangenomics/pangenomes/freeze/freeze1/minigraph-cactus/hprc-v1.1-mc-grch38/hprc-v1.1-mc-grch38.vcfbub.a100k.wave.vcf.gz. Each DNV was tested for pangenome-graph membership using snow pangenomecheck, which queries each DNV against the pangenome VCF for presence or absence at the site and/or the allele level. DNVs present at the site-level in the pangenome VCF were removed from further analyses. QC of the DNVs was performed using acorn [27] (Supplementary Figure S1, Supplementary Figure S2, Supplementary Figure S3, Supplementary Figure S4, Supplementary Figure S5, Supplementary Figure S6). DNVs were than annotated with Variant Effect Predictor (VEP) [28] for predicted functional consequences. DNVs that are SNVs were summarized as an SBS96 mutational spectrum using deconstructSigs [29]. Single base substitution signatures were visualized by hierarchical clustering heatmaps (individuals × signatures) using heatmap.2 [30] with dendrograms for both rows and columns. For any analyses involving exome data, to ensure comparable callability across samples, analyses were restricted to DNVs in the targeted capture regions. To assess coverage across the capture regions, I used mosdepth [31] to calculate the mean sequencing depth for each individual in each region. I then aggregated the data to identify regions with sufficient coverage: at least 10 × coverage in more than 90% of individuals on the autosomes, at least 10 × coverage in more than 90% of females on the X chromosome, and at least 5 × coverage in 90% of males on each of the X and Y chromosomes.

Gene set enrichment testing

Missense and LOF DNVs were assessed for gene set enrichment using the denovolyzeByClass function in denovolyzeR [32]. The analyses evaluated enrichment across eight gene sets including (1) all genes, (2) non-PAR genes on the X chromosome, (3) non-PAR genes on the Y chromosome, (4) PAR genes, (5) hormone receptor genes from the GtoPdb [33] accessed at https://www.guidetopharmacology.org/GRAC/NHRListForward, (6) genes exhibiting sex-bias in gene expression in the brain [34], (7) genes present in the SFARI Gene database (release 2025 Q4) (https://gene.sfari.org/), and (8) 100 random genes exclusive of the SFARI gene list. For each gene set, enrichment was tested for synonymous DNVs, missense DNVs, and LOF DNVs.

I analyzed four DNV datasets stratified by proband sex (male/female) and affection status (affected/unaffected). After basic cleaning and filtering (including harmonizing gene symbols), each dataset was formatted for denovolyzeR for gene set enrichment tests. A gene set was considered significant if it passed Bonferroni correction for 24 tests (p < 0.002).

Since denovolyzeR does not stratify expected DNV counts by child sex, I adjusted expectations for non-PAR genes on the X and Y chromosomes using a parent-of-origin model with male:female germline mutation-rate ratio (i.e., paternal fraction ). denovolyzeR’s baseline expectation is , where is the number of probands and is the per-gene, per-class probability from denovolyzeR’s internal table; let with males and females. Relative to the autosomal baseline , transmission implies for non-PAR loci: male X and male Y , yielding multipliers and ; female multipliers are and . For mixed-sex groups, and . PAR genes were unadjusted (multiplier ). Implementation-wise, I multiplied the numeric columns of denovolyzeR’s probability table by gene-specific multipliers and supplied the adjusted table to denovolyzeByClass via the probTable argument; genes without a specified multiplier were left unchanged (multiplier = 1.0).

Per-gene enrichment testing

Per-gene DNV enrichment was tested separately within each sex × affection stratum using denovolyzeR (denovolyzeByGene) for LOF, missense, and LOF + missense classes, with non-PAR X/Y expectations adjusted as described above. Genes with a p-value less than 2.5 × 10⁻⁶ were considered exome-wide significant.

Male versus female DNV comparisons

As a complementary analysis, I tested per-gene enrichment using two-sided Fisher’s exact tests on allelic variant counts for DNVs that were missense or LOF. For each gene and chromosomal category (autosomes, PAR, X non-PAR), I constructed four pairwise comparisons: (1) affected females vs affected males, (2) unaffected females vs unaffected males, (3) affected females vs unaffected females, and (4) affected males vs unaffected males. For each comparison, I formed a 2 × 2 table of variant alleles vs non-variant alleles between the two groups, where the group-specific denominator equaled the number of allele copies evaluated given cohort size and chromosome copy number (autosomes/PAR: 2N per group; X non-PAR: 2N in females and N in males). Non-PAR Y genes were not included in between-group comparisons. Genes with a p-value less than 2.5 × 10⁻⁶ were considered exome-wide significant.

Child-level protein-coding DNV analysis using a liability-threshold probit framework

Child affection status was analyzed with child-level probit regression under a latent-liability framework (Additional details in Supplementary Methods). Briefly, DNVs were aggregated by child, mutation class, and genomic region, and grouped as LOF, missense, or synonymous (negative control) variants in autosomal and X non-PAR regions. Expected burden was derived from denovolyzeR gene-level mutation probabilities, with sex-specific adjustment for X non-PAR mutation rates. Primary predictors were observed-minus-expected excess burden terms for damaging and synonymous variant classes. Additive models included damaging burden, synonymous burden, and sex; interaction models additionally tested sex-by-burden effects.

Testing for DNV enrichment in promoter regions

Promoters were tested for enrichment of DNVs using fitDNM [35, 36]. They were defined as 2000 bp upstream of the transcription start site of protein-coding genes from the RefSeq database downloaded from the UCSC Genome Browser [37, 38]. Due to the discovery, in this study, of RNU2-2, in a genome-wide significant promoter I also tested for enrichment of DNVs in exonic coordinates of protein-coding and noncoding genes, respectively (see below).

Assessment of enrichment of DNVs in the exonic regions of coding and noncoding genes

I implemented a custom pipeline (based on previously described approaches [35, 36, 39]) to test for enrichment of DNVs in exonic regions of protein-coding and noncoding genes. Briefly, I (i) estimated gene-level mutability from trinucleotide-based substitution rates using a reference genome and RefSeq gene models, (ii) counted unique DNVs overlapping exons for each gene, and (iii) assessed enrichment by comparing observed counts to expectations under a Poisson model (, with adjusted for sex-chromosome copy number). Enrichment was summarized using one-sided Poisson p-values and observed-to-expected ratios.

Results

Characteristics of high quality DNVs in this study

Throughout the remainder of the paper the following acronyms are used to represent the cohorts. “SSC WGS” refers to whole-genome sequencing data from the Simons Simplex Collection, “SPARK WGS” refers to whole-genome sequencing data from the SPARK cohort, and “SPARK WES” refers to whole-exome sequencing data from the SPARK cohort. For the SSC WGS cohort, I identified 277,345 SNVs, 41,369 INDELS (15,653 deletions, 25,716 insertions), and 2611 MNVs (Fig. 1A, Supplementary Table S2). The mean ± standard deviation of length for the deletions were 4.48 ± 4.88 bp, for insertions was 5.03 ± 5.39 bp, and for MNVs was 14.05 ± 20.52 bp. For the SPARK WGS cohort, I identified 392,037 SNVs, 47,503 INDELS (20,294 deletions, 27,209 insertions), and 3241 MNVs (Fig. 1A, Supplementary Table S3). The mean ± standard deviation of length for the deletions were 4.69 ± 5.16 bp, for insertions was 5.07 ± 5.67 bp, and for MNVs was 12.53 ± 18.87 bp. For the SPARK WES cohort, I identified 46,802 SNVs, 2517 INDELS (1015 deletions, 1502 insertions), and 312 MNVs (Fig. 1A, Supplementary Table S4). The mean ± standard deviation of length for the deletions were 5.21 ± 5.12 bp, for insertions was 4.71 ± 3.64 bp, and for MNVs was 9.25 ± 12.63 bp (Fig. 1A). Please note for any analyses of coding regions across SSC WGS, SPARK WGS, and SPARK WES, I used the subset of WES coordinates (Supplementary Table S5) with high coverage, across most of the samples, as shown in Supplementary Table S6. This resulted in the removal of 4677 DNVs that were in lowly covered exome regions (Supplementary Table S7).

Fig. 1.

Overall Features of DNV Callset Generated in this Study. Three cohorts are shown including Simons Simplex Collection whole-genome sequencing (SSC WGS), Simons Powering Autism Research (SPARK) WGS, and SPARK whole-exome sequencing (SPARK WES). A Shown are the number of DNVs per individual (expect ~ 75 in WGS and ~ 2 in WES), the number of variants by type, the Ti/Tv ratio (expect ~ 2 in WGS and ~ 2.8 in WES, and the mean ± standard deviation of length for various DNV types. B Shown is information on DNVs that could be phased for parent-of-origin chromosome. Inside each bar is the count of DNVs, on the y-axis are the percent phased to the paternal chromosome, on the x-axis is the category (i.e., genomic region). For autosomal, X chromosome in females, and PAR, it is expected that ~ 75% are phased to the paternal chromosome of origin. For the X chromosome in males, 0% are expected to be phased to the paternal chromosome. For the Y chromosome in males, 100% are expected to be phased to the paternal chromosome

Several metrics were considered for determination of high quality DNVs (Fig. 1A). First, the number of DNVs per individual was calculated and was 76.82 ± 14.27 in SSC WGS, 78.16 ± 15.94 in SPARK WGS, and 1.59 ± 1.30 in SPARK WES. Second the Ti/Tv ratio was calculated and was 2.25 in SSC WGS, 2.27 in SPARK WGS, and 2.81 in SPARK WES. Genome-wide the percent of DNVs that could be phased to the parent-of-origin chromosome was 37.7% in SSC WGS, 36.7% in SPARK WGS, and 12.0% in SPARK WES. Third, phased variants were considered in 5 different categories including the autosomes, the X chromosome in females, the X chromosome in males, the Y chromosome in males, and the pseudoautosomal regions (PAR). All DNVs in males on the X chromosome implicitly phase to the maternal chromosome and on the Y chromosome phase to the paternal chromosome (Fig. 1B). For the autosomes, the percent phased to the paternal chromosome was 76.6%, 77.8%, and 79.1% in SSC WGS, SPARK WGS, and SPARK WES, respectively (Fig. 1B). For the X chromosome in females, the percent phased to the paternal chromosome was 79.5%, 80.0%, and 73.9% in SSC WGS, SPARK WGS, and SPARK WES, respectively (Fig. 1B). For the PARs, the percent phased to the paternal chromosome was 78.9%, 80.5%, and 53.8% in SSC WGS, SPARK WGS, and SPARK WES, respectively (Fig. 1B). The lower percent in SPARK WES is likely due to a small DNV count of 13 that could be phased in that category.

WGS data were used to assess additional characteristics of the DNVs. Overall, there was an increase in DNV counts with parental age at birth (Fig. 2A, B). A linear model of DNV counts and paternal and maternal ages at birth, respectively, were assessed and identified an increase of 1.4 [1.3,1.5] DNVs per year for paternal age and 0.5 [0.4,0.6] DNVs per year for maternal age in SSC WGS with an adjusted r² for the model of 0.45. For SPARK WGS, these values were an increase of 1.5 [1.4,1.6] DNVs per year for paternal age and 0.5 [0.4,0.6] DNVs per year for maternal age with an adjusted r² for the model of 0.54. For the combined dataset, these values were an increase of 1.5 [1.4,1.5] DNVs per year for paternal age and 0.5 [0.4,0.5] DNVs per year for maternal age with an adjusted r² for the model of 0.50. Overall, there was a shift towards older fathers compared to mothers in this study (Fig. 2C). Next, single base substitution signature analyses were performed on the DNV data. The top contributing signatures were Signatures 1A, 1B, 5, and 16 (Fig. 2D, Supplementary Table S8). This is consistent with prior findings in DNV datasets [40].

Fig. 2.

DNVs Increase with Parental Age at Birth and Show Expected Clock-Like Mutation Signatures. A Hexagonal binning plot of DNV counts per individual versus the age of the father at birth. B Hexagonal binning plot of DNV counts per individual versus the age of the mother at birth. C Density plot showing the ages of father and mother at birth. Fathers in this study are, generally, slightly older than mothers. D Heatmap of mutation signatures generated using deconstructSigs. Each row is an individual, and each column is a signature. The color corresponds to the fraction of that signature contributing to the individual signature profile. The top signatures are the expected clock-like signatures including Signature 1A, Signature 1B, Signature 5, and Signature 16

QC metrics on the DNV dataset show that HAT-FLEX and SNOW generate high-quality DNVs. To further assess the quality of these tools I pursued benchmarking of HAT-FLEX and SNOW on publicly available, high-quality WGS data, and found that both methods perform well (Supplementary Table S9). In one trio with NA12877 as the child, the previous publication [22] reported 72 DNVs identified across three sequencing technologies: PacBio, Illumina, and ONT. Since HAT-FLEX uses at most two callers, I used the publicly available genome-wide PacBio and Illumina VCFs as the two input callsets for intersection-level calling with HAT-FLEX and the PacBio alignment files as input for SNOW. Consequently, some DNVs reported in the previous study were not represented in the HAT-FLEX/SNOW input data because they were supported only by ONT. Of the 72 DNVs reported in the previous paper, HAT-FLEX/SNOW identified 54. Among the 18 DNVs not recovered, four were flagged as LowGQ by HAT-FLEX, three were absent from the publicly available Illumina genome-wide VCF, two were absent from both the publicly available Illumina and PacBio genome-wide VCFs, and two were genotyped as 0/0 in all family members in one publicly available genome-wide VCF but appeared de novo in the other publicly available genome-wide VCF. The remaining seven cases reflected a range of discordances between datasets: one variant was inherited in the publicly available PacBio genome-wide VCF, one was inherited in the publicly available PacBio genome-wide VCF but absent from the publicly available Illumina genome-wide VCF, one was inherited in the publicly available Illumina genome-wide VCF but appeared de novo in the publicly available PacBio genome-wide VCF, one was missing from the publicly available Illumina genome-wide VCF, one had a missing child genotype (./.) in the publicly available PacBio genome-wide VCF, and one appeared de novo because of a complex allele-matching issue compounded by LowGQ in the publicly available genome-wide VCF. Overall, these discrepancies are driven primarily by unavailability of high-quality information to call DNVs in the intersection of the publicly available genome-wide Illumina and PacBio VCFs. None of the ones missed by HAT-FLEX/SNOW were high-quality DNVs at the intersection of Illumina and PacBio data. In addition, HAT-FLEX/SNOW identified 4 DNVs that were not reported in the previous publication. All four appeared high confidence upon visual inspection of the reads. Two of these differences may be explained by the use of a different Y chromosome reference in the previous study [22]. To compare phasing performance on the 54 shared DNVs, I found that SNOW assigned parent-of-origin for all but one variant; the single exception failed phasing because of SNOW phaser’s genotype quality requirement for phase-by-transmission variants. By contrast, the previous publication phased only 46 of these variants, and the parent-of-origin assignments were identical for all variants phased by both SNOW and the previous publication [22]. Overall, SNOW phased 7 additional variants relative to the previous publication (Supplementary Table S10). In total HAT-FLEX/SNOW contained 57 DNVs as two of the DNVs were collapsed to one DNV using snow consolidate (Supplementary Table S11). HAT-FLEX and SNOW demonstrated robust accuracy on high-quality benchmark WGS data, recovering the great majority of evaluable published DNVs, identifying additional high-confidence events not reported previously, and achieving fully concordant but more complete parent-of-origin phasing.

Gene-based enrichments reveal specific categories

I performed gene set enrichment on the combined WGS data and the combined WGS and WES data, respectively. For exome region analyses, testing was restricted to regions of high coverage across most samples (Supplementary Table S6). For both analyses, four different groups were assessed including males with ASD (n = 4808 individuals in WGS, n = 21,922 individuals in the combined WGS/WES set), females with ASD (n = 1021 individuals in WGS, n = 5888 individuals in the combined WGS/WES set), males without ASD (n = 1928 individuals in WGS, n = 6456 individuals in the combined WGS/WES set), and females without ASD (n = 2091 individuals in WGS, n = 6850 individuals in the combined WGS/WES set). The testing in each group was across eight different gene sets (all genes, non-PAR X chromosome genes (Supplementary Table S12), non-PAR Y chromosome genes (Supplementary Table S13), PAR genes (Supplementary Table S14), hormone receptor genes (Supplementary Table S15), genes with sex-bias in expression in the brain (Supplementary Table S16), SFARI genes (Supplementary Table S17), and random genes (Supplementary Table S18) and three variant types (synonymous, missense, LOF).

In the WGS dataset (Fig. 3A), no group had significance for synonymous variants in any category. In males with ASD, there was a significant enrichment of missense DNVs in the all gene set (p = 6 × 10⁻¹⁰, enrichment = 1.11), the LOF DNVs in the all gene set (p = 2 × 10⁻¹⁰, enrichment = 1.32), missense DNVs in the SFARI gene set (p = 3 × 10⁻³⁹, enrichment = 1.77), and LOF DNVs in the SFARI gene set (p = 4 × 10⁻⁶⁹, enrichment = 4.42). In females with ASD, there was a significant enrichment of missense DNVs in the all gene set (p = 3 × 10⁻⁵, enrichment = 1.16), LOF DNVs in the all gene set (p = 1 × 10⁻⁸, enrichment = 1.65), LOF DNVs in the non-PAR X chromosome genes (p = 1 × 10⁻⁵, enrichment = 4.4), missense DNVs in SFARI genes (p = 2 × 10⁻¹⁷, enrichment = 2.1), and LOF DNVs in SFARI genes (p = 5 × 10⁻⁴⁴, enrichment = 7.65). In unaffected males, there was an enrichment of missense DNVs in SFARI genes (p = 0.0005, enrichment = 1.29). In unaffected females, there was an enrichment of missense DNVs in the non-PAR X chromosome genes (p = 9 × 10⁻⁸, enrichment = 1.91).

Fig. 3.

Gene Set Enrichment Testing. Shown are gene set enrichment analyses generated using denovolyzeR for four different settings (males with autism, females with autism, males without autism, females without autism). Eight different gene sets are considered in the analyses and within each gene set analysis three different variant types are considered including synonymous, missense and LOF. A Results shown for the combined WGS data and B results shown for the full combined WGS and WES data. Multi-test correction significant p-values are labeled and have a *

In the combined WGS/WES dataset (Fig. 3B), no group had significance for synonymous variants in any category. In males with ASD, there was a significant enrichment of missense DNVs in the SFARI gene set (p = 4.10 × 10⁻⁶⁴, enrichment = 1.45). There was also a significant enrichment of LOF DNVs in the SFARI gene set (p = 1.9 × 10⁻¹¹⁶, enrichment = 2.89). In females with ASD, there was a significant enrichment of missense DNVs in the all gene set (p = 2.44 × 10⁻⁵, enrichment = 1.07), the non-PAR X chromosome genes (p = 1.31 × 10⁻¹⁰, enrichment = 1.63), and SFARI genes (p = 5.82 × 10⁻³⁷, enrichment = 1.65). There was a significant enrichment of LOF DNVs in the all gene set (p = 2.94 × 10⁻⁵, enrichment = 1.18, the non-PAR X chromosome genes (p = 1.93 × 10⁻⁵, enrichment = 2.17), the genes with sex-bias in expression in the brain (p = 0.0008, enrichment = 3.48), and SFARI genes (p = 1.45 × 10⁻⁷⁸, enrichment = 4.19). In unaffected males, there was an enrichment of missense DNVs in SFARI genes (p = 3.41 × 10⁻⁵, enrichment = 1.19). In unaffected females, there was an enrichment of missense DNVs in the non-PAR X chromosome genes (p = 0.0001, enrichment = 1.33) and the SFARI genes (p = 0.0004, enrichment = 1.15).

Gene-based enrichment of DNVs

I first performed denovolyzeR analyses for LOF, missense, and combined LOF and missense DNVs in the full set of individuals with ASD (n = 27,810 individuals) (Supplementary Table S19, Supplementary Figure S7). The 24 genes significant in the LOF category were SCN2A, CHD8, ADNP, SHANK3, GRIN2B, PTEN, MIB1, CHD2, FOXP1, POGZ, SMAD6, MARK2, SYNGAP1, GIGYF1, BRSK2, KDM6B, KDM5B, DLG4, AUTS2, ASH1L, TANC2, ARHGEF9, CSNK2B, and ANKRD11. The 22 genes significant in the missense category were CCDC168, MECP2, PTEN, SCN2A, MUC19, PIEZO1, TBL1XR1, ZNF469, EML6, GREB1L, SLC6A1, OTOG, KDM5B, R3HCC1, CACNA1I, ACTG1, DDX3X, KCTD1, FREM3, PPP2R5D, FOXP2, and ZNF84. The 32 genes significant in the combined LOF and missense category were SCN2A, CCDC168, PTEN, MECP2, ADNP, CHD8, FOXP1, KDM5B, CHD2, SHANK3, GRIN2B, MUC19, DDX3X, PIEZO1, MIB1, ZNF469, TBL1XR1, SLC6A1, SMAD6, OTOG, TLK2, GREB1L, EML6, POGZ, GIGYF1, R3HCC1, FREM3, PPP2R5D, BRSK2, ACTG1, DNMT3A, and KCTD1. Several of these genes have come up in other studies of ASD (e.g., SCN2A [41], CHD8 [42]).

Next, I performed the same analyses in males with ASD (n = 21,922 individuals) (Supplementary Table S20, Fig. 4A–C). I list the significant genes in each category and label in parentheses after the gene as follows (NSAF) if there was not even nominal significance in females with ASD for the same category. The 17 genes significant in the LOF category were CHD8, SCN2A, SHANK3, FOXP1 (NSAF), SMAD6 (NSAF), PTEN, SYNGAP1, CHD2, MARK2, GIGYF1, ADNP, AUTS2 (NSAF), KDM6B, MIB1, POGZ, CCDC168 (NSAF), and ASH1L. The 11 genes significant in the missense category were CCDC168, PIEZO1 (NSAF), MUC19, OTOG (NSAF), EML6 (NSAF), ZNF469, PTEN, ZNF84 (NSAF), SCN2A, IGSF23 (NSAF), and NHSL1 (NSAF). The 18 genes significant in the combined LOF and missense category were CCDC168, SCN2A, PTEN, FOXP1, CHD8, MUC19, PIEZO1 (NSAF), CHD2, SMAD6 (NSAF), KDM5B, ZNF469, OTOG (NSAF), ADNP (NSAF), EML6 (NSAF), SLC6A1, SHANK3, FREM3 (NSAF), and MIB1. Overall, there were 12 unique genes that met exome-wide significant enrichment in males that were not nominally significant for the same variant class in females. These included FOXP1, SMAD6, AUTS2, CCDC168, PIEZO1, EML6, ZNF84, IGSF23, OTOG, NHSL1, ADNP, and FREM3. To determine if the females with autism data was underpowered to detect these genes, I ran a test to see if increasing the sample size to the size of the males with autism dataset would recover any of these genes in the females with autism. Only one gene (ADNP) was projected to no longer be an NSAF gene in that scenario (Supplementary Table S21).

Fig. 4.

Gene-Based Enrichment Analyses. Shown in A to C are gene set enrichments analyses generated using denovolyzeR for two different settings (males with autism, females with autism) and shown in Figure D are results from Fisher’s exact tests for four different comparison settings. A Manhattan plots of denovolyzeR results for LOF DNVs, B Manhattan plots of denovolyzeR results for missense DNVs, and C Manhattan plots of denovolyzeR results for combined LOF and missense DNVs. D Fisher’s exact test results with heatmap colors corresponding to the − log10(p-value). * Indicates exome-wide significant. For A to C the boxes indicate the gene is not even nominal significant in that category in the other sex

Next, I performed the same analyses in females with ASD (n = 5888 individuals) (Supplementary Table S22, Fig. 4A–C). I list the significant genes in each category and label in parentheses after the gene as follows (NSAM) if there was not even nominal significance in males with ASD for the same category. The 11 genes significant in the LOF category were SCN2A, ADNP, GRIN2B, CHD8, SHANK3, CSNK2B, MIB1, TAOK1 (NSAM), BRSK2, POGZ, and ARHGEF9. The 5 genes significant in the missense category were MECP2 (NSAM), PTEN, SCN2A, TBL1XR1, and DDX3X (NSAM). The 10 genes significant in the combined LOF and missense category were SCN2A, MECP2 (NSAM), PTEN, ADNP, DDX3X (NSAM), GRIN2B, TBL1XR1, SHANK3, CHD8, and KDM5B. Overall, there were three genes that met exome-wide significant enrichment in females that were not nominally significant for the same variant class in males. These included TAOK1, MECP2, and DDX3X.

Next, I performed the same analyses in all unaffected individuals (n = 13,306 individuals) (Supplementary Table S23) (Supplementary Figure S8). No genes were significant in the LOF category. The six genes significant in the missense category were MUC19, HEATR5A, WDFY4, DNAH14, STARD9, and ZNF69. The five genes significant in the combined LOF and missense category were MUC19, HEATR5A, WDFY4, DNAH14, and STARD9. I than performed the analyses in unaffected males (n = 6456 individuals) (Supplementary Table S24) (Supplementary Figure S9). No genes were significant in the LOF category. The one gene significant in the missense and the combined LOF and missense category was STARD9. Next, I performed the same analyses in unaffected females with ASD (n = 6850 individuals) (Supplementary Table S25) (Supplementary Figure S9). No genes were significant in the LOF category. The four genes significant in the missense category were MUC19, HEATR5A, WDFY4, and DNAH14. The three genes significant in the combined LOF and missense category were MUC19, HEATR5A, and WDFY4.

Finally, I performed Fisher’s exact test analyses (Supplementary Figure S10, Supplementary Figure S11, Supplementary Figure S12, Supplementary Figure S13) in four different comparison groups including males with ASD versus males without ASD (Supplementary Table S26), females with ASD versus females without ASD (Supplementary Table S27), males with ASD versus females with ASD (Supplementary Table S28), and males without ASD versus females without ASD (Supplementary Table S29) (Fig. 4D). Only two genes were exome-wide significant, and they were SCN2A in the females with ASD versus females without ASD comparison and MECP2 in the males with ASD versus females with ASD comparison. Though not significant, DDX3X met expectation of trending towards enrichment in females with ASD, specifically. One intriguing gene was GABBR2 that trended towards significance in females in only the comparison of males with ASD versus females with ASD. I also examined overall burden and found an excess of LOF and missense DNVs in the comparison of males with ASD to males without ASD (p = 0.005) on the autosomes and in the comparison of females with ASD to females without ASD (p = 0.0007) on the autosomes. Females with ASD exhibited an enrichment of LOF and missense DNVS on the X chromosome in comparison to females without ASD (p = 0.006). (Supplementary Table S30).

Rates of alleles per gene

To give estimates for rates of DNV LOF and missense alleles per gene, I estimated them using the data and a Poisson model. I performed these estimates on the autosomes (Fig. 5A, Supplementary Table S31), PAR (Fig. 5B, Supplementary Table S32), non-PAR X Chromosome (Fig. 5C, Supplementary Table S33), and non-PAR Y Chromosome (Fig. 5D, Supplementary Table S34). In Fig. 5, I show the top genes by statistical significance and there are a few genes which show clear rate differences between males and females. These include the autosomal genes SCN2A (rate in males with ASD is 5.93012 [3.87375, 8.68899] per 10,000 alleles, rate in females with ASD is 19.5312 [12.3811, 29.3065] per 10,000 alleles), GRIN2B (rate in males with ASD is 2.28081 [1.09374, 4.1945] per 10,000 alleles, rate in females with ASD is 10.1902 [5.26543, 17.8003] per 10,000 alleles), ADNP (rate in males with ASD is 2.5089 [1.25243, 4.48911] per 10,000 alleles, rate in females with ASD is 9.34103 [4.66301, 16.7137] per 10,000 alleles), and CHD8 (rate in males with ASD is 4.78971 [2.96491, 7.32158] per 10,000 alleles, rate in females with ASD is 8.49185 [4.07217, 15.6168] per 10,000 alleles) and the X chromosome genes MECP2 (rate in males with ASD is 0.456163 [0.011549, 2.54158] per 10,000 alleles, rate in females with ASD is 13.587 [7.76612, 22.0644] per 10,000 alleles) and DDX3X (rate in males with ASD is 0 [0, 1.68273] per 10,000 alleles, rate in females with ASD is 8.49185 [4.07217, 15.6168] per 10,000 alleles).

Fig. 5.

LOF and Missense DNV Rates Per 10,000 Alleles. Shown are the top genes based on statistical enrichment of LOF and missense DNVs on the A autosomes, B PAR, C non-PAR X chromosome, and D non-PAR Y chromosome. Rates are calculated in four categories: males with autism, females with autism, males without autism, and females without autism

Female protective effect evident in probit model of damaging de novo variant burden

Based on prevalence estimates in ASD, I calculated the thresholds on the liability distribution for males and females, respectively. With an ASD prevalence of one in 20 in males and one in 70 in females these yielded thresholds of 1.645 in males and 2.189 in females (Fig. 6A). This is consistent with a female protective effect in which females require greater latent liability to cross the affected threshold. For the probit model using data generated in this study, final individual counts considered were 21,922 affected males, 5888 affected females, 6456 unaffected males, and 6850 unaffected females. In the unified additive model (Fig. 6B), damaging de novo burden was positively associated with liability, with the strongest effects seen for autosomal LOF variants (β = 0.218, 95% CI 0.172 to 0.264, p = 1.82 × 10⁻²⁰) and X non-PAR LOF variants (β = 0.317, 95% CI 0.044 to 0.591, p = 0.023), while smaller but still significant effects were observed for autosomal missense variants (β = 0.035, 95% CI 0.019 to 0.051, p = 1.16 × 10⁻⁵) and X non-PAR missense variants (β = 0.119, 95% CI 0.006 to 0.231, p = 0.039). In contrast, synonymous burden showed no evidence of association for either autosomes (β = 0.010, 95% CI − 0.015 to 0.035, p = 0.45) or X non-PAR regions (β = 0.051, 95% CI − 0.138 to 0.239, p = 0.60). Female sex was associated with a markedly lower baseline liability in the same model (β = -0.847, 95% CI − 0.874 to − 0.820). In the sex-stratified interaction model (Fig. 6C), autosomal LOF effects were present in both males (β = 0.195, 95% CI 0.137 to 0.253, p = 5.58 × 10⁻¹¹) and females (β = 0.257, 95% CI 0.182 to 0.331, p = 1.54 × 10⁻¹¹), and autosomal missense effects were similarly modest in males (β = 0.034, 95% CI 0.015 to 0.054, p = 6.37 × 10⁻⁴) and females (β = 0.037, 95% CI 0.011 to 0.064, p = 0.005). For X non-PAR burden, the estimates were larger and statistically supported in females for both LOF (β = 0.343, 95% CI 0.037 to 0.649, p = 0.028) and missense variants (β = 0.138, 95% CI 0.012 to 0.263, p = 0.031), whereas the corresponding male estimates were smaller and imprecise, with confidence intervals crossing zero for X non-PAR LOF (β = 0.207, 95% CI − 0.388 to 0.803, p = 0.50) and X non-PAR missense (β = 0.043, 95% CI − 0.208 to 0.295, p = 0.74). Overall, Fig. 6B and C together indicate that the pooled liability signal is driven primarily by damaging, especially LOF, burden, with sex-stratified analyses suggesting that X-linked effects may be more pronounced in females than in males.

Fig. 6.

Burden Assessments. A Shown are the liability distribution and male and female thresholds calculated based on ASD prevalence in males and females, respectively. B Shown is the uniform additive model (damaging burden + synonymous burden + sex). C Shown is the model with sex-specific burden effects (damaging burden + sex + burden × sex)

Promoter-based testing of DNVs reveals enrichment of DNVs in RNU2-2

I evaluated all RefSeq gene promoters, in the WGS data only, across six groups: all individuals with ASD (Supplementary Table S35, Supplementary Table S36) (Supplementary Figure S14), all individuals without ASD (Supplementary Table S37, Supplementary Table S38) (Supplementary Figure S15), males with ASD (Fig. 7A, Supplementary Table S39, Supplementary Table S40), females with ASD (Supplementary Table S41, Supplementary Table S42) (Supplementary Figure S14), males without ASD (Supplementary Table S43, Supplementary Table S44) (Supplementary Figure S15), and unaffected females without ASD (Supplementary Table S45, Supplementary Table S46) (Supplementary Figure S15) for enrichment of DNVs. A single promoter region, for two separate transcripts of WDR74 (chr11:62840210-62842210 and chr11:62840156-62842156) reached genome-wide significance for DNVs (p = 3.8 × 10⁻¹³), and this signal was specific to males with ASD (Fig. 7A). Follow-up analyses revealed that most DNVs within this promoter clustered in the noncoding RNA gene RNU2-2, which overlaps a subregion of the promoter (Fig. 7B). I therefore tested RNU2-2 directly and observed significant enrichment (p < 2.2 × 10⁻¹⁶) in males with ASD (Fig. 7C). I further looked for DNVs throughout the other data and found that DNVs in RNU2-2 are present in 0.19% of males with ASD and 0.20% of females with ASD (Fig. 7D, Supplementary Figure S16). Overall, they are observed in 0.2% of individuals with ASD. They are also present in 0.10% of unaffected males and were not observed in any unaffected females (Fig. 7D, Supplementary Figure S16). Overall, they are in 0.05% of individuals without ASD.

Fig. 7.

Promoter-Based Testing Reveals Enrichment of DNVs in the RNU2-2 gene, which resides in the promoter of WDR74. A Manhattan plot of − log10(p-value) versus chromosome in males with autism. Each point represents a promoter. The top promoter is genome-wide significant and is a promoter of the WDR74 gene. B Careful examination of DNVs in the promoter reveals localization of them to the RNU2-2 noncoding RNA gene within the promoter region. C Shown are the DNVs, in RNU2-2, in males with autism (navy), females with autism (red), and males without autism (green). Two of the DNVs are recurrent in unrelated individuals with autism. D Percent of individuals that contain a DNV in RNU2-2

Testing for enrichment of DNVs in exons of coding and noncoding genes

After discovery of RNU2-2, I decided to test for enrichment of DNVs, generally, in the exonic regions of all isoforms of protein-coding and noncoding genes (n = 178,356 isoforms). I evaluated all RefSeq protein-coding and noncoding genes (Supplementary Table S47), in the WGS data only, across six groups: all individuals with ASD (Supplementary Table 48, Supplementary Table S49, Fig. 8A), all individuals without ASD (Supplementary Table S50, Supplementary Table S51, Supplementary Figure S17), males with ASD (Supplementary Table S52, Supplementary Table S53, Fig. 8B), females with ASD (Supplementary Table S54, Supplementary Table S55, Fig. 8C), males without ASD (Supplementary Table S56, Supplementary Table S57, Supplementary Figure S17), and unaffected females without ASD (Supplementary Table S58, Supplementary Table S59, Supplementary Figure S17) for enrichment of DNVs. Overall, the top gene in this analysis was RNU2-2 (Supplementary Table S60) and it contained DNVs in nine males with ASD, two females with ASD, and two males without ASD. Assessment of its expression in the Brainspan Atlas of the Developing Human Brain (https://www.brainspan.org) revealed cortical expression of this gene (Supplementary Figure S18).

Fig. 8.

Enrichment of Exonic DNVs in Coding and Noncoding Genes. A Testing in all individuals with autism, B testing in males with autism, and C testing in females with autism

Discussion

In this study, I investigated DNVs in ASD, with a particular focus on differences by sex and diagnostic status. ASD shows a sex bias, with males affected more frequently than females. Previous investigations of DNVs related to sex bias in ASD have often included individuals with developmental disorders more broadly, whereas this current study focuses on the primary ASD phenotype. Using data from three ASD family cohorts, I generated uniform high-quality DNV callsets for individuals with and without ASD. These callsets enabled a range of complementary analyses, including: 1) testing enrichment across variant and gene-class categories; 2) identifying genes reaching exome-wide significance in males with ASD, females with ASD, males without ASD, and females without ASD; 3) directly comparing DNV enrichment between males and females across multiple comparison groups; 4) estimating rates of LOF and missense DNV alleles, 5) examining burden of LOF and missense DNVs, and 6) evaluating the enrichment of noncoding variants. Importantly, I performed analyses that were aware of sex chromosomes including PARs.

The foundation of this study is in the generation of a high-quality DNV callset. This was done by developing a sex-aware DNV caller (HAT-FLEX) and another tool for further optimizing DNV calls (SNOW). Rather than filtering DNVs using genomic annotations (e.g., repeat regions), I used intensive assessment of each candidate DNV, through SNOW, to produce a robust, high confidence callset. This opens up new regions of the genome to investigation in the context of ASD. After generating the high-confidence DNV callset, I tested enrichment across predefined gene sets stratified by DNV class. As expected, synonymous DNVs showed no significant gene-set enrichments. Consistent with prior observations, I observed enrichment of missense and LOF DNVs in individuals with ASD. Additional enrichments were detected in females, including gene sets on the non-PAR regions of the X chromosome and genes with sex-biased expression in the brain. Several gene sets are relatively small (e.g., hormone receptor genes); although they showed large enrichment, they did not remain significant after multiple-testing correction. These signals may warrant followup in larger cohorts.

Analyses of gene-specific enrichments by sex and diagnostic category, were insightful. Few genes met enrichment in any category in individuals without ASD as one would expect. Several genes were enriched in individuals with ASD and by performing sex-stratified analyses I could pinpoint genes that were enriched only in one sex for the DNV class being tested. These genes included genes previously implicated in ASD and some newly implicated genes. Those specific to males included FOXP1, SMAD6, AUTS2, CCDC168, PIEZO1, EML6, ZNF84, IGSF23, OTOG, NHSL1, ADNP, and FREM3 and to females included TAOK1, MECP2, and DDX3X. Both MECP2 and DDX3X are well-established genes with known enrichment in females. Some genes of interest from these lists include the newly implicated PIEZO1, which is a mechanosensitive ion channel and is important in the brain [43, 44] and TAOK1, which is involved in a neurodevelopmental disorder [45] wherein it has recently been reported that endocrinopathies are also common [46]. One intriguing gene in the comparison of males with ASD to females with ASD was the gene GABBR2 with enrichment of DNVs in females. While not exome-wide significant, this is a gene that may be of interest in future studies as it is the only one that was not an X chromosome gene that was close to significance. The neurodevelopmental disorder associated with GABBR2 variants has been documented to resemble Rett syndrome [47] (typically caused by variants in the X-linked gene MECP2). Additionally, GABBR2 variants have been reported in association with Rett-like phenotypes in individuals with classical Rett syndrome caused by MECP2 variants [48].

The probit analyses provide an additional way to interpret my findings within a latent-liability framework. In this setting, the results suggest that the observed sex difference is unlikely to be explained simply by differences in mutation rate. Even after normalizing de novo burden using observed-minus-expected counts, including sex-specific expectations for X-linked variants, damaging variation remained associated with affection status, indicating that mutational opportunity alone does not fully account for the pattern we observed. The absence of a similar association for synonymous variants further supports the specificity of this signal and makes a broad technical artifact or general ascertainment effect less likely, although diagnostic bias cannot be completely excluded on the basis of genetic data alone. At the same time, the substantially lower baseline for females in the probit model, together with damaging-burden effects that were comparable to or even stronger than those in males, is most consistent with a female protective effect interpretation. On the liability scale, this implies that females may need to accumulate a greater overall burden before crossing the threshold for diagnosis, offering a coherent explanation for the sex disparity seen in our cohort.

In the final part of this study, I focused on promoters as they have previously been implicated in bulk as important for ASD [49, 50]. Herein, I queried each promoter for significant excess of DNVs and identified the WDR74 promoter as significant specifically in males with ASD. By inspection of the promoter in the UCSC Genome Browser [38], I found that the RNU2-2 noncoding RNA gene was located within this promoter and nearly all the DNVs were within it. This led me to hypothesize that the signal may be from this gene, and I found that RNU2-2 was also genome-wide significant for excess DNVs. RNU2-2 DNVs are present in 0.2% of males with ASD. I further found that DNVs occur in females with ASD as well and the overall rate is also at 0.2% of individuals with ASD. There were two males without ASD that also had a DNV leading to an overall rate of 0.1% in unaffected males. Assessment of females without ASD identified no DNVs in this region. Overall, 0.05% of individuals without ASD have DNVs in this noncoding gene. These results indicate that RNU2-2 may be an important gene related to ASD in addition to the severe neurodevelopmental disorder for which it is already associated with in the literature [51]. Further investigation of DNVs in the exonic regions of protein-coding and noncoding genes identifies RNU2-2 as the top gene for enrichment of DNVs in ASD. More broadly spliceosomal RNA variants have been implicated in neurodevelopmental disorders and represent an emerging class of noncoding risk signals. Among the spliceosomal RNA genes linked to neurodevelopmental disorders are RNU4-2 [52], RNU2-2 [51], RNU4ATAC [53], and RNU5B-1 [54]. In the present study, only RNU2-2 showed significant enrichment for DNVs. RNU2-2 encodes a U2 snRNA expressed in the cortex early in development in timepoints relevant to autism (e.g., the 18–20 post-conception week window and the 1- to 4-year-old postnatal ages windows) (Supplementary Figure S18). Disruption of splicing in the developing cortex could preferentially affect long, multi-exonic genes by altering exon usage. Given that ASD-associated genes are often long and contain numerous exons [55], this mechanism may increase transcriptional variability and contribute to phenotypic heterogeneity in ASD.

Conclusions

Overall, this study establishes sex as an important modifier of the contribution of DNVs to ASD risk and provides further support for a female protective effect, with females appearing to require a greater burden of damaging variation to reach comparable ASD liability. The findings also suggest that ASD risk is shaped, in part, by sex-specific patterns at the gene level, indicating that important signals may be obscured when males with ASD and females with ASD are analyzed together. In addition to these coding-sequence differences, the identification of RNU2-2 within the WDR74 promoter implicates noncoding variation in ASD susceptibility and broadens the spectrum of ASD-associated variation. These results are important to the field because they support a more complete model of ASD genetic architecture that incorporates both sex-dependent effects and contributions from noncoding sequence. Moreover, HAT-FLEX and SNOW increase the broader utility of this work by providing a sex-aware, high-confidence framework for DNV discovery and evaluation that can support future family-based sequencing studies. Collectively, this study deepens our understanding of the genetic basis of ASD and its pronounced sex bias, while offering insights that may inform and support broader research across the genomics community.

Abbreviations

AD

Allelic depth

ALT

Alternate

ASD

Autism spectrum disorder

CRAM

Compressed reference-oriented alignment map

DENOVO

De novo

DNV

De novo variant

DP

Depth

GQ

Genotype quality

GT

Genotype

gVCF

Genomic variant call file

HAPUNUSUAL

Haploid unusual

HPRC

Human pangenome reference consortium

INDEL

Insertion or deletion

IRRELPARENT

Irrelevant parent (in relation to sex chromosome transmission)

LOF

Loss-of-function

MNV

Multi-nucleotide variant

NSAF

Not even nominal significance in females with ASD

NSAM

Not even nominal significance in males with ASD

PAR

Pseudoautosomal region

probTable

Probability table

SBS96

Substitution mutation signatures (96 type classification)

SEXCTX

Sex chromosome context

SFARI

Simons Foundation Autism Research Initiative

SNV

Single-nucleotide variant

SPARK

Simons Powering Autism Research

SSC

Simons Simplex Collection

TSV

Tab-separated value

VCF

Variant call file

WES

Whole-exome sequencing

WGS

Whole-genome sequencing

Author contributions

TNT conceptualized the study, designed the methodology, developed the software, performed formal analysis and investigation, acquired the funding, and wrote the paper.

Funding

This work was supported by grants from the National Institutes of Health (R01MH126933) and the Simons Foundation (734069).

Data availability

Approved researchers can obtain the SPARK and SSC population datasets described in this study by applying at (https://base.sfari.org). All other relevant data are included within the manuscript. The code for HAT-FLEX is available at (https://github.com/TNTurnerLab/HAT-FLEX) and the code for SNOW is available at (https://github.com/tycheleturner/snow).

Declarations

Ethics approval and consent to participate

The Washington University in St. Louis Institutional Review Board determined this study as "Not Human Subjects Research." For reference, the WashU IRB number for this study is 202103260. The statement regarding their determination was as follows: “Because you are not receiving identifiable data or specimen and have signed an agreement with the original holder which prohibits re-identification, this activity is not considered to meet federal definitions under the jurisdiction of an IRB and therefore falls outside the purview of the HRPO.” All research was conducted in accordance with the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.