Decomposing the genetic risk of autism spectrum disorder into discrete molecular subtypes underlie clinical heterogeneity based on transcriptome profile

Abstract

Background

Autism spectrum disorder (ASD) is a highly heterogeneous neurodevelopmental condition with a complex genetic architecture. While over a thousand risk genes have been cataloged, a fundamental challenge remains how this vast genetic landscape translates into diverse clinical manifestations. To address this, we propose a “many-to-few” framework, shifting from traditional “many-to-one” convergence model toward an intrinsic organizational architecture where disparate ASD risk genes funnel into discrete molecular dimensions.

Methods

By leveraging similarity network fusion (SNF) to integrate bulk and single-nucleus RNA sequencing data, we decomposed the 311 ASD reliable risk genes into three stable, spatiotemporally distinct molecular subtypes. These subtypes represent coordinated expression programs that integrate functionally diverse genes, such as those involved in synaptic signaling, mRNA stabilization, and histone modification. Mapping de novo variants from the SPARK cohort onto these subtypes enabled stratification of probands into three genetically defined subgroups (S1–S3) with divergent clinical profiles.

Results

Transcriptomic decomposition partitioned the 311 reliable risk genes into three stable molecular subtypes, Synaptic Signaling (C1), mRNA Stabilization (C2), and Histone Modification (C3), exhibiting divergent spatiotemporal trajectories and cell-type-specific enrichment. Patient subgroups stratified based on these molecular subtypes displayed significant differences in adaptive functioning, core symptoms, and psychiatric comorbidities, while a reference group (S4) lacking de novo variants of these highly reliable ASD risk genes exhibited the most preserved functions. Furthermore, diverging rare and common genetic liability profiles across subgroups, particularly between S1 and S3, provide empirical support for a molecular subtype-based liability threshold model.

Conclusions

Our study establishes a biologically informed framework that links intrinsic molecular subtypes to multidimensional phenotypic constellation, advancing mechanistic insight and offering translational potential for precision stratification and intervention.

Supplementary information

The online version contains supplementary material available at 10.1186/s12967-026-08364-y.

Introduction

Autism spectrum disorder (ASD) is a neurodevelopmental condition characterized by highly clinical heterogeneity and complex genetic architecture [1]. The clinical spectrum is vast, ranging from severe impairments in social communication and restrictive/repetitive behaviors (RRB) to complex psychiatric comorbidities [2–5]. At the genetic level, both rare and common genetic factors contribute to ASD liability. SFARI Gene database have cataloged over a thousand risk genes [6], including hundreds of high-confidence (HC) and syndromic genes (SYN) identified by whole-exome (WES) [7–11] and whole-genome sequencing (WGS) [12–15] studies, as well as limited genes from common variants identified by GWAS [16–18]. However, a fundamental challenge remains how this vast genetic landscape translates into the diverse clinical manifestations observed in patients, a gap that significantly limits our ability to develop precision stratification and intervention strategies.

To date, research has primarily focused on identifying points of mechanistic convergence to explain how disparate risk genes confer a singular diagnosis of ASD [19, 20]. This “many-to-one” paradigm has successfully mapped risk genes onto broad, convergent biological pathways [21]. This convergence is evidenced by static functional annotations indicating that discrete risk gene clusters within core biological pathways such as synaptic signaling and chromatin remodeling [22–24]. Furthermore, dynamic transcriptomics data from the neurotypical brain has revealed convergent co-expression modules across multiple scales as well. Specifically, bulk transcriptomic data has previously revealed aggregate molecular signatures in synaptic and transcriptional regulation pathways [25–29], and single-nucleus RNA sequencing (snRNA-seq) has refined these findings to cell-type-specific level [30–32]. Previous studies have conducted gene enrichment analyses for all identified genes, which, although able to cluster different types of pathways and propose the concept of molecular subtypes, have not formally performed clustering to subgroup ASD genes, nor have they further associated these with phenotypic traits.

Genetic observations have long recognized that ASD is not monolithic but genetically partitioned. On one hand, core symptom domains, such as social communication and restrictive behaviors, often display weak genetic correlations and can dissociate within individual patients [33, 34], underscoring the genetic underpinnings of ASD are partitioned. On the other hand, genomic structural equation modeling (Genomic SEM) has demonstrated that the genetic architecture of ASD can be explicitly decomposed into discrete, independent genetic factors, and each linked to specific biological and clinical profiles [35]. This reveals that genetic risk is organized into identifiable components, providing a structured genetic foundation for the clinical heterogeneity of ASD. Based on this decomposable architecture of genetic risk, we propose that the hundreds of ASD high confidence and syndromic genes may be accurately captured by the “many-to-few” model.

In this “many-to-few” framework, risk genes do not merely converge onto a singular diagnostic endpoint, instead, they funnel into a limited number of molecular dimensions, each independently modulating a specific phenotypic constellation. This hypothesis aligns with the emerging view that genes tightly connected in regulatory networks may exhibit correlated phenotypic profiles [36]. While recent evidence suggests that specific trait combinations harbor distinct genetic risks [37], such investigations are primarily phenotype-driven, defining subgroups by clinical observation first. Consequently, it remains unclear whether the ASD risk genes are governed by an intrinsic organizational architecture, one that originates from molecular programs rather than being defined by outward clinical groupings.

To test this “many-to-few” hypothesis, we adopted a biology-driven approach to decompose the ASD reliable risk genes holistically by leveraging the integrative power of transcriptomics. Specifically, using similarity network fusion (SNF), an integrative method that has identified subtypes when integrating multiple omics datasets in cancer [38], we integrated bulk and single-nucleus RNA sequencing data of ASD to classify 311 highly reliable ASD risk genes into three molecularly connected gene clusters. These clusters represent distinct molecular subtypes and then were further characterized by their predominant biological functions, as well as distinct spatiotemporal trajectories and cell-type enrichments patterns. Subsequently, ASD patients from SPARK cohort were stratified into different subgroups based on which cluster of gene variants they carried. The phenotypic and genetic features of different subgroups were characterized. Our findings prove that the molecular-based “many-to-few” framework can effectively capture and dissect the heterogeneity of ASD.

Materials and methods

We utilized the bulk and single-nucleus transcriptome data from ASD brains to find molecular patterns by similarity network fusion (SNF) after performing quality control and normalization on the transcriptome separately. We obtained stable gene clusters based on genes from the SFARI database (Fig. 1A–B), and then performed two sensitivity analyses to validate the robustness and specificity of clustering results (Fig. 1C). The functional, clinical, and genetic features of the gene clusters were then profiled (Fig. 1D).

Fig. 1.

Overview of the analysis workflow. (A) Data preparation and preprocessing of transcriptomic datasets. (B) Gene clustering via SNFtool and obtain the gene clusters with stable identity. (C) Implementation of two sensitivity analyses to evaluate the robustness and specificity of ASD clustering results. (D) Functional, phenotypic, and genetic characterization of the resulting gene clusters

ASD genes for clustering analysis

SFARI Gene database (https://gene.sfari.org/database/human-gene/) cataloged more than 1000 genes, but the strength of evidence for some genes is insufficient. To more accurately assess potential genetic causes, SFARI Gene database established a set of criteria that rank genes into one of four categories (including high confidence, strong candidate, suggestive evidence and syndromic genes). Thus, to define a more reliable candidate set, we first extracted high confidence (HC, N = 232) genes which clearly implicated in ASD, defined by the presence of at least three de novo likely-gene-disrupting mutations being reported in the literature. In addition, since syndromic autism is one important category in autism spectrum disorder and its related genes were mostly definitive, the syndromic genes (SYN, N = 288) from SFARI Gene were also included in our analysis, which results in 405 unique genes in total accounting for overlap (Table S1). By focusing on these HC and SYN categories, we ensured that our analysis was centered on high-risk genes with compelling causal evidence.

Bulk RNA-seq data of ASD brain

Bulk transcriptomic data were obtained from Gandal et al. (bulk_gandal), comprising 725 cortical tissue samples from 49 ASD patients and 54 controls (ages 2–68 years, across 11 cortical regions) [39]. Processed data were downloaded from the PsychENCODE Knowledge Portal. Counts were normalized to fragments per kilobase of transcript per million mapped reads (FPKM), lowly expressed genes (mean FPKM < 0.5) were removed, and values were converted to transcripts per million (TPM) after regressing out sequencing depth. Technical covariate (sequencing batch) was regressed out from the expression data, while biological variables (brain region, sex, age) were retained. Principal component analysis (PCA) was performed on the adjusted data, and visualization of the resulting component plots confirmed that clustering was not driven by technical covariates (Figure S1A). To capture the transcriptomic profile of ASD patients, only the brain samples from ASD patients were used for the following analysis. After intersecting with SYN and HC genes, we obtained a matrix of 397 genes × 384 ASD samples (Table S2). A subset of control transcriptomes were extracted (a matrix of 397 genes × 341 control samples) for negative control as well.

Single nuclei RNA-seq data of ASD brain

The snRNA-seq data (sn_Velmeshev) comprising 41 postmortem cortical samples (15 ASD, 16 controls, ages 4–22 years) were downloaded from Velmeshev et al. (https://autism.cells.ucsc.edu) [40]. Cells were filtered (500 < nFeature_RNA <8,000, 500 < nCount_RNA <40,000, mitochondrial/ribosomal fraction < 5%), normalized using log-transformation and scaled to 10,000 counts. Cell types (N = 17) were annotated according to the original metadata. Covariates such as sequencing batch were regressed out, while biological variables (brain region, sex, age) were retained. Uniform Manifold Approximation and Projection (UMAP) confirmed that clustering was not driven by batch effects (Figure S2A-B). Pseudo-bulk profiles were generated by aggregating counts within each cell type from ASD patients. After intersection with SYN and HC genes, we obtained a matrix of 392 genes × 17 cell types (Table S2). The abbreviations for each cell type in Velmeshev et al. were present in Table S3. Another pseudo-bulk analysis based on control transcriptomes were extracted (a matrix of 392 genes × 17 cell types) for negative control as well.

Similarity network fusion to identify different gene clusters

We utilized Similarity Network Fusion (SNF) method [38] to identify gene clusters by integrating several gene expression matrices. This approach offers a distinct advantage over other data integration methods, as it clusters genes in a manner that is not influenced by varying normalization techniques across different data types. More importantly, SNF computes and integrates gene similarity networks obtained from each of their data types separately, taking advantage of the complementary information between genes in the data, which could be well utilized to integrate different omics data [41]. We used the expression matrices from both bulk RNA-seq and snRNA-seq data of ASD brains, which have been normalized, for the SNF analysis. Specifically, for each normalized dataset, it first constructed a separate N_genes × N_genes similarity (affinity) matrix using a scaled exponential similarity kernel. These affinity networks were then fused through an iterative process to capture complementary biological signals across scales. The similarity network fusion was implemented using the SNFtool package, adhering to the standard hyperparameter settings (neighbor size K = 20 and iterations T = 15) as recommended in the foundational SNF framework. Following the original methodological rationales, we chose K = 20 to prioritize reliable local neighborhood structures and effectively suppress technical noise from distant, lower-weight edges. The fusion process was set to T = 15 iterations to ensure that the interchanging diffusion process reached empirical convergence into a stable, unified similarity network, as demonstrated across various genomic datasets in the original study. The function ‘OptimumNumberOfClustersGivenGraph’, in SNFtool could provide the optimal number of clusters by examining the mathematical structure of the fused similarity matrix via Eigen-gap Heuristic and rotation cost. The SNF analysis was also applied in bulk RNA-seq and snRNA-seq data of neurotypical brain data.

Gene cluster stability analysis

To assess the stability of clustering outcomes, we extracted random subsets of samples with varying sample size from the bulk_Gandal datasets to run SNF analysis with the single- nucleus data of sn_Velmeshev. This procedure was repeated 1,000 times. Five proportions of bulk data, 10, 30, 50, 70, and 90%, were tested, resulting in a total of 5,000 fused matrices. The consistency of clustering results across all 5,000 subsampled trials was measured using the Kappa statistic. At the same time, we identified the genes with stable cluster identity across 5,000 subsampled SNF trials that reproduced the three-cluster solution. Genes were defined as having stable cluster identity if they were assigned to the same cluster in ≥95% of high-concordance trials (Kappa ≥ 0.8).

Sensitivity analysis

Two sensitivity analyses were conducted in this study to evaluate the robustness and specificity of the gene clustering results. The first sensitivity analysis assessed the reproducibility by substituting the snRNA-seq data (sn_Velmeshev) with another snRNA-seq data (sn_Wamsley), focusing on the 311 genes with stable identity. The dataset sn_Wamsley was obtained from Wamsley et al., comprising 32 cortical samples (32 ASD, ages 2–60 years) [42]. Cells were filtered (500 < nFeature_RNA <6,000; 500 < nCount_RNA <40,000; mitochondrial/ribosomal fraction < 5%), normalized using log-transformation and scaled to 10,000 counts. Cell types (N = 19) were re-annotated using canonical markers provided by the original paper. Covariates such as sequencing batch were regressed out, while biological variables (brain region, sex, age) were retained. UMAP confirmed that clustering was not driven by batch effects (Figure S3A-B). Pseudo-bulk profiles were generated using only ASD samples, yielding a matrix of 396 genes × 19 cell types (Table S3) after intersection with SYN and HC genes (Table S2). The processed data were downloaded for general research use according to the following requirements for data access (https://psychencode.synapse.org/DataAccess).

We performed the second sensitivity analysis by applying SNF method to control samples to confirm the observed three-cluster pattern of the 311 genes with stable identity is specific to ASD pathology and not intrinsic to general brain gene expression profiles. The bulk RNA-seq data containing 13 cortical regions from neurotypical donors were obtained from the GTEx project (version 10) (https://gtexportal.org). TPM data of these 13 cortical regions were available. After intersection with SYN and HC genes, we obtained 14 matrices of 311 genes × N samples (N = 181–3234) for 13 cortical regions and the merged transcriptome data of the 13 brain regions (Table S2). The single-cell transcriptomic data from neurotypical brain tissue were obtained from Trevino et al. (GEO accession GSE162170) [43]. The dataset includes 31,304 single cells grouped into 22 cell types (Table S3). Cells were filtered (500 < nFeature_RNA <8,000; 500 < nCount_RNA <40,000; mitochondrial/ribosomal fraction < 5%). Covariates such as sequencing batch were regressed out, while biological variables were retained. UMAP confirmed that clustering was not driven by batch effects (Figure S4A-B). Pseudo-bulk profiles were generated, yielding a matrix of 396 genes × 22 cell types after intersection with SYN and HC genes (Table S2). The consistency of clustering results was assessed using the kappa statistic, providing a quantitative measure of agreement across clustering outcomes.

GO biological process enrichment analysis for three gene clusters

Gene Ontology (GO) enrichment analysis was conducted using rrvgo, a Bioconductor package designed to facilitate the biological interpretation of large sets of GO terms. Functional enrichment analysis was carried out for each gene cluster individually. GO biological process (GO-BP) terms were retained if they met the following criteria, Bonferroni adjusted p-value < 0.05 and gene set size between 3 and 200 [44].

Temporal/Spatial enrichment analysis for three gene clusters

Transcriptomic data (RNA-seq) from 16 human brain regions were obtained from the Allen Institute BrainSpan Atlas, which comprise samples from 57 post-mortem brains of neurotypical donors. Both sexes are represented, and the developmental stages range from 8 post-conceptual weeks (pcw) to 40 years of age [45, 46]. Given the high proportion of missing values in the original BrainSpan dataset (approximately 52%), we utilized an imputed version of the dataset provided by Pei et al. [47], available at (https://github.com/bsml320/BrainSpan) [47]. For temporal expression analysis, gene expression values were first adjusted for sex and brain region, followed by z-score normalization. Smoothed developmental trajectories were then generated using these standardized values. Spatial enrichment analysis was performed using lmFit and eBayes functions from limma package (version 3.48.1). A multivariable linear model was fitted for each gene to estimate the effects of brain region on gene expression, with sex and age as covariates. Genes were then ranked in descending order based on |t-statistics|. The top 5% of genes were selected to construct region-specific gene sets. Within this subset, genes with positive t-statistics were classified as significantly upregulated, while those with negative t-statistics were classified as significantly downregulated for each brain region. Fisher’s exact test was applied to assess the enrichment of the three gene clusters across these gene sets, yielding odds ratios and p-values for each region. Enrichment was considered significant if the Benjamini–Hochberg (BH) p-value was < 0.05 and the odds ratio (OR) exceeded 1.

Cell type enrichment analysis for three gene clusters

Cell type enrichment analysis was performed using single-nucleus transcriptomic data from controls (sn_Trevino), processed as described above. Enrichment scores for each gene cluster were calculated using the AUCell package (version 1.20.1, https://github.com/aertslab/AUCell). AUCell evaluates whether a subset of input genes are enriched within the expressed genes of individual cells by computing the Area Under the Curve (AUC) for ranked gene expression. This ranking-based approach is independent of expression units and normalization procedures. AUC score distributions across all cells were used to assess the relative expression of each gene signature.

Participants carrying the De novo variants in three gene clusters

De novo variant curation was based on processed data from Fu et al. [11], who applied Hail’s de_novo () function to identify candidate de novo variants of SPARK cohort while accounting for population allele frequencies. Candidate variants were further filtered using the following criteria, (1) population frequency in gnomAD and within the dataset, (2) ‘ExcessHet’ filter, (3) allele balance and parent/child depth ratio, (4) variant quality score log-odds (VQSLOD), and (5) excessive number of de novo candidate variants per sample. Detailed procedures were described in the original manuscript [11]. Here, only European individuals were included (N = 4,282) in our genetic analysis [11], where 14.6% (625/4,282) probands carrying de novo mutations or single-nucleotide variants (SNVs) in the 405 ASD risk genes. Then, we further filtered these probands who had both qualified phenotype and genotype data after imputation, applied from SFARI Base (https://base.sfari.org) [48]. In total, 4,187 probands remained, in which 620 probands carried the de novo mutations or SNVs in the 405 ASD risk genes and 3,567 probands did not. Probands were stratified into three subgroups (S1, S2, S3), based on the cluster assignment of the mutated gene. A reference group (Subgroup 4, S4; N = 200) was established by random sampling from 3,567 probands lacking de novo mutations in highly reliable ASD risk genes, representing those without de novo mutations or SNVs in known, highly reliable ASD genes. This sample size was chosen to balance statistical power and group parity for pairwise comparisons.

Phenotypic data processing and Subgroup comparison

Phenotypic measures included assessments of repetitive and restrictive behaviors (Repetitive Behavior Scale-Revised, RBS-R), social communication (Social Communication Questionnaire, SCQ), and adaptive functioning (Vineland Adaptive Behavior Scales, Third Edition, VABS-III). Additional information on comorbidities was obtained from the Child Behavior Checklist (CBCL). Data quality was evaluated according to the 2025 release notes. Samples were excluded if any item was flagged (e.g., variables with ‘xx_flag = 1’, for example, diagnosis_flag = 1). For the Vineland scale, it mainly includes four domains (communication, daily living skills, socialization and motor domain) and each of them has two or three subdomains, where the variable ‘xx_est’ (xx represents name of each subdomain) in each subdomain indicates the proportion of items estimated by caregivers. If ‘xx_est’ was estimated no less than 25% in two or more subdomains, the Vineland scores were deemed invalid and the corresponding samples were removed. The values coded as ‘888’, ‘998’, or ‘−9999’ were treated as missing values, and rows with incomplete data were removed using the complete.cases () function.

The differences in clinical phenotypes across the four subgroups were assessed using covariate-adjusted models. For continuous variables, residuals from linear models adjusting for age at evaluation and sex were tested for normality. If normally distributed, ANOVA was performed on the adjusted residuals, followed by pairwise t-tests with BH correction. For non-normally distributed residuals, Kruskal–Wallis tests were applied, followed by pairwise Wilcoxon rank-sum tests with BH correction. For categorical variables, likelihood ratio tests from logistic or multinomial regression models were used. Variables that were inherently age-standardized were adjusted for sex only (Figure S5). Detailed sample sizes (N), missing data rates, and the specific statistical tests applied for each phenotypic measure are comprehensively reported in Table S4. For the phenotypic analyses, we employed the Benjamini-Hochberg (BH) procedure to control the False Discovery Rate (FDR). To maintain maximum statistical stringency, we performed a global correction across the total pool of all pairwise subgroup comparisons from all clinical measures assessed.

Genotyping, quality control and imputation

To compare the genetic burden of four subgroups, the genotype data were processed using standard quality control procedures. Individuals were excluded if they exhibited genotyping call rates < 95%, sex discrepancies, or excess heterozygosity (>3 standard deviations from the sample mean). Variants were retained if they met the following criteria, minor allele frequency (MAF) >1%, genotyping rate > 95%, and Hardy–Weinberg equilibrium p > 1 × 10^{− 6}. After quality filtering, a total of 414,124 high-confidence SNPs were retained for downstream analyses, including genetic relatedness inference, principal component calculation, and genotype imputation.

Genotype imputation was performed using the TOPMed Imputation Server, based on reference panels from the NHLBI TOPMed Program [49, 50], following phasing with Eagle (version 2.5). Genomic coordinates after imputation were converted from GRCh38/hg38 to GRCh37/hg19 using UCSC liftOver. For all downstream analyses, we retained imputed variants with MAF > 0.1% and imputation quality score (R²) >0.3.

Polygenic risk score analyses and group comparison

Polygenic risk scores (PRS) for ASD were calculated using PRS-CS, a Bayesian regression approach that applies continuous shrinkage priors to single nucleotide polymorphism (SNP) effect sizes [51]. For ASD, we obtained harmonized GWAS summary statistics [18] and used the UK Biobank European reference panel for linkage disequilibrium estimation. The analysis was performed chromosome-wise using a global shrinkage parameter phi = 1e-2, with chromosome-specific posterior SNP effect sizes subsequently concatenated into genome-wide weight files. Individual PRS was calculated using PLINK’s score function by summing the product of allele counts and posterior effect sizes across autosomal SNPs.

To assess group differences in PRS while controlling for potential confounders, we employed linear mixed-effects models that incorporated multiple covariates. The model included sex and age as covariates, along with the top ten principal components derived from PC-AiR method implemented in the GENESIS R package (version 2.22.2) to control for population stratification [52]. The full model was specified as, PRS ~ group + sex + age + PC1 + PC2 +  …  + PC10.

Results

Gene Clustering

ASD risk genes were curated from the SFARI Gene database, including both high-confidence (HC) and syndromic (SYN) genes (N = 405). After intersection with transcriptomic data, 385 genes were retained for analysis. We applied Similarity Network Fusion (SNF) to decompose the genetic landscape by integrating two distinct ASD transcriptomic matrices, one from bulk brain tissue (bulk_Gandal) and one from single-nucleus RNA-seq (sn_Velmeshev) (Figure S6A-B). This analysis identified three gene clusters (Table S5), with 125, 98, and 162 genes in Clusters 1, 2, and 3, respectively (Fig. 2A). In contrast, applying the same integrative approach to control transcriptomes yielded a markedly different clustering structure (Fig. 2B–C), underscoring the specificity of the observed patterns to ASD.

Fig. 2.

Results of gene clustering and sensitivity analysis. (A-B) Gene-to-gene similarity matrices derived from the integrated expression profiles of 385 genes across brains transcriptomes, combining bulk RNA-seq (bulk_gandal) and scRNA-seq (sn_velmeshev) datasets for similarity network fusion (SNF) analysis. Clusters are visualized in bluescale and arranged according to subtypes identified through spectral clustering of the fused gene network. (A) shows the clustering results for ASD cases (C1: 125, C2: 98, C3: 162), while (B) depicts the corresponding results for controls (C1: 106, C2: 13, C3: 80, C4, 186). (C) Alluvial diagrams illustrate gene cluster mapping between cases and controls. (D) summary statistics of clustering results across different subsampling proportions of bulk_gandal (10%, 30%, 50%, 70%, 90%). (E) Distribution of overall Kappa scores across 1,000 iterations under 50% subsampling; the red dashed line indicates the mean Kappa score. (F) Distribution of cluster assignment frequencies for 385 ASD risk genes across subsampled SNF trials that reproduced the three-cluster with high concordance (Kappa ≥ 0.8). Finally, it obtained 311 genes with stable identity, cluster 1, 89 genes; cluster 2, 81 genes; cluster 3, 141 genes. (G–I) Validation of clustering robustness and specificity. (G) Alluvial diagrams illustrate cluster mapping of 311 stable genes, confirming reproducibility after replacing sn_wamsley with sn_ Velmeshev. (H) Kappa scores show poor agreement for various non-ASD brain transcriptome from GTEx datasets and sn_trevino. (I) Alluvial diagrams between original clustering results and SNF trials from the 13 merged brain regions of GTEx and sn_trevino

Identification of genes with stable cluster identity

To evaluate the robustness of transcriptomics-guided decomposition, we performed random subsampling of the bulk brain transcriptomic data at five proportions (10, 30, 50, 70, and 90%), with 1,000 iterations for each proportion. These 5,000 subsets were individually fused with the single-nucleus ASD brain dataset (sn_Velmeshev) and subjected to SNF analysis. In total, 4,916 trials consistently yielded three clusters, with clustering stability increasing alongside sample size (Fig. 2D). Consistency was quantified using the Kappa statistic, which demonstrated strong agreement with the original clustering (mean Kappa > 0.8, p < 0.05 across all proportions; 4,702 trails had Kappa > 0.8; Fig. 2E, Figure S7A–D). Although the overall decomposition structure was robust, individual gene assignments were not uniformly stable. To identify the core constituents of each molecular dimension, we defined a gene as having a stable identity if it was consistently assigned to the same cluster in ≥95% of high-concordance trials (Kappa ≥ 0.8) (Fig. 2F). Applying this criterion yielded a core set of 311 highly reliable ASD risk genes and 74 unstable genes. These stable members, including 89 genes in Cluster 1, 81 in Cluster 2, and 141 in Cluster 3, form the molecular foundation of our three molecular subtypes (Table S5). Moreover, the proportions of HC and SYN did not differ significantly among clusters (p = 0.55) (Figure S8A).

Validation with alternative datasets and specificity assessment

We next assessed the robustness of the decomposition results by repeating the original SNF analysis via replacing sn_Velmeshev with the latest single-nucleus ASD brain dataset from Wamsley et al. (sn_Wamsley). This independent validation successfully regrouped the 311 core genes into three stable clusters with a strong Kappa score of 0.74 (Fig. 2G), reaffirming the overall stability of the decomposition framework.

To rigorously evaluate the specificity of our clustering to ASD pathology, we conducted additional analyses using transcriptomic data from neurotypical population. We sequentially integrated each of the 14 bulk brain transcriptomic datasets from GTEx with single-nucleus data of the general population (sn_Trevino). Across these 14 integration trials, only seven outcomes yielded three clusters. However, consistency analysis revealed poor agreement with our ASD-derived framework, with Kappa scores ranging from −0.04 to 0.19, indicating minimal concordance (Fig. 2H–I, Figure S9A–M). The remaining seven integration outcomes failed to reproduce the three-cluster structure entirely, yielding only two clusters.

Functional characterization of stable gene clusters

To explore the biological landscapes captured by each cluster, we performed functional enrichment analysis on the 311 stable genes. The results revealed that each cluster represents a biologically coherent and molecularly distinct biological functions (Fig. 3A, Table S6). Cluster 1 was enriched for synaptic signaling processes, including regulation of postsynaptic membrane potential and chemical synaptic transmission. Cluster 2 was predominantly associated with mRNA stabilization pathways, such as regulation of CRD-mediated mRNA stabilization and negative regulation of nuclear-transcribed mRNA decay. Cluster 3 showed primary enrichment in histone modification processes, including histone methylation and acetylation (Fig. 3B). Despite these distinct primary functions, the biological profiles of the clusters were not mutually exclusive. For example, several functional pathways were distributed across multiple clusters, particularly pathways of excitatory postsynaptic potential between Clusters 1 and 3, and pathways of protein amino acid acetylation between Clusters 2 and 3 (Fig. 3C). This pattern suggests that while the clusters define discrete functional dimensions biologically, they possess overlapping functions that may facilitate coordinated biological regulation.

Fig. 3.

Functional and expression characterization of three gene clusters. (A) Venn diagram of significantly enriched GO-BP terms for each cluster. (B) Top ten GO-BP terms enriched in each gene cluster (bonferroni-corrected). (C) Shared GO-BP terms across the three clusters. (D) Developmental expression trajectories of the three clusters across BrainSpan stages. Expression patterns of cluster 1 (red), cluster 2 (green), and cluster 3 (blue) across human brain development. Solid lines represent mean expression; shaded areas indicate ±1.5 SE. (E) Regional expression patterns of each gene cluster across brain regions. Odds ratios and p-values were computed by comparing each region to all others. Positive t-statistics indicate upregulation, while negative t-statistics indicate downregulation in the corresponding region. (F-H) UMAP plots showing cell-type enrichment based on AUCell score. The AUC score reflects the relative activity of each gene cluster in individual cells, with higher scores indicating stronger enrichment

Spatiotemporal expression profiles of gene clusters

To determine the developmental and anatomical specificity of the identified molecular subtypes, we examined the spatiotemporal expression profiles of the three gene clusters. As shown in Fig. 3D, Cluster 1 exhibited a biphasic trajectory, with elevated expression during late fetal and early postnatal stages, followed by a marked decline throughout childhood. Cluster 2 showed relatively stable expression across development, with a modest elevation before infancy and a gradual decrease thereafter. Cluster 3 maintained a steady expression level during prenatal stages, followed by a slow postnatal decline that persisted into adulthood. Notably, the overall expression levels of Cluster 3 genes were generally lower than those of the other two clusters across most developmental stages. Spatial enrichment analysis, which identified genes with significantly high or low expression in specific brain regions, revealed that Cluster 1 genes were significantly depleted in the amygdala (AMY) compared to other brain regions (Fig. 3E).

Cell-type enrichment analysis revealed further divergence among these molecular subtypes. Cluster 1 were predominantly enriched in excitatory glutamatergic neurons and inhibitory interneurons, and were significantly depleted in non-neuronal populations such as early radial glia and cycling progenitors (Fig. 3F). In contrast, Cluster 2 were highly expressed in non-neuronal cell types, including early and late radial glia and multipotent glial progenitor cells (Fig. 3G), while being depleted in glutamatergic neurons. Cluster 3 showed moderate enrichment in both glutamatergic neurons and interneurons, despite their overall lower expression across brain cell types (Fig. 3H). These distinct spatiotemporal and cellular signatures suggest that the three molecular subtypes act at different biological scales and developmental windows to modulate ASD risk.

Phenotypic profiles of ASD subgroups stratified by Gene Cluster

We hypothesized that the molecular subtypes identified in this study may contribute to the phenotypic heterogeneity observed among individuals with ASD. To evaluate this, we examined the relationship between ASD phenotypes and molecular subtypes by identifying de novo mutations in probands from the SPARK cohort.

Overview of the population

There were 4,187 European probands from SPARK database carrying de novo mutation. Among them, 620 carried de novo mutations or SNVs in the 405 ASD risk genes from our initial set, while 3,567 carried mutations in genes outside this set. We further focused on the 311 ASD risk genes with stable cluster identity, in which 480 individuals harbored qualifying de novo mutations. The majority (95.8%, 460/480) carried mutations in only one gene, while 14 individuals had mutations spanning two clusters. To ensure subgroup specificity and eliminate confounding from mixed molecular signatures, 14 individuals with mutations spanning multiple clusters were excluded from further stratification. The remaining 466 probands were stratified into three genetic subgroups based on the cluster assignment of the mutated genes, 123 in Subgroup 1 (S1), 63 in Subgroup 2 (S2), and 280 in Subgroup 3 (S3) (Table S7). The number of mutations per gene ranged from 1 to 11, with a median of 2 (Figure S10A–B). Additionally, a comparable sample (N = 200) was randomly drawn from 3,576 probands to form the reference group (Subgroup 4, S4), representing those without highly reliable rare de novo risk. Pairwise comparisons of clinical phenotypes among the four subgroups revealed statistically significant differences across multiple developmental and behavioral domains.

Developmental trajectory and adaptive function

Significant group differences were observed in early developmental trajectory and real-world functional competence. S4 consistently exhibited the most pronounced preservation of adaptive skills, reporting the highest median composite scores across all major Vineland-3 domains. This trend was uniformly observed not only in the ABC composite but also across its three constituent subdomains, communication, daily living skills, and socialization, as well as the motor score. In sharp contrast, Subgroup 1 (Synaptic Signaling) demonstrated the lowest median scores across all Vineland domains, significantly lower than all other subgroups (P_adj < 0.01 for all contrasts), reflecting the most substantial and pervasive functional impairment. Subgroup 2 (mRNA Stabilization) and Subgroup 3 (Histone Modification) occupied an intermediate position, though S4 scores were consistently and significantly higher than S2 and S3 scores for major domains (Fig. 4A–B, Figure S11A–C).

Fig. 4.

Clinical phenotypic and genetic profiles of individuals carrying distinct gene clusters. (A–N) phenotypic profiles of ASD subgroups stratified by gene cluster membership. (A–B) Vineland-3 composite and motor standard scores across subgroups. (C–F) developmental milestones and diagnostic timing. (G–H) core ASD symptom severity (RBS-R, SCQ). (I–L) Co-occurring behavioral and psychiatric features (CBCL subscales). (M–N) functional of language and cognitive across subgroups. “language_age_level” and “cog_age_level” refer to the language and cognitive ability that probands have compared to his/her age. “above_age”, “at_age”, “slight_below_age” and “signif_below_age” denotes the function of probands is higher, average, slightly lower and significantly lower than the peers with the same age separately. (O) Summary radar plots illustrating functional and symptom domain rankings for each subgroup. Radar plots illustrate the standardized ranking of phenotypic domains across the four subgroups and rankings are identical if the differences between subgroups are not statistically significant. The distance of each wedge extending from the center to the outer edge represents its ranking from 4 to 1. Consequently, a longer physical length indicates a higher ranking (closer to the periphery), symbolizing superior performance in functional domains (left hemisphere) or increased severity in symptom domains (right hemisphere). Left hemisphere represents the function domains including adaptive function (adapt.), developmental milestones (Dev.); cognitive function (cog.), language function (Lang.). Right hemisphere represents the symptom domains including social communication (soc.), repetitive behaviors (RBBs.), internalizing symptoms (int.), externalizing symptoms (ext.). (P–S) genetic architecture of ASD subgroups. (P–Q) functional annotation of de novo variants showing subgroup differences in missense constraint (MPC) and variant class (MisB). (R) Proportion of individuals carrying protein truncating variants across subgroups. (S) Polygenic risk score (PRS) distribution across ASD subgroups. Adjusted PRS values are shown across three percentile bins. Significance levels are indicated as follows. p < 0.05, *BH-adjusted p < 0.05, **0.001 < BH-adjusted p < 0.05, *** BH-adjusted p < 0.001

Further analysis of key developmental milestone markers corroborated this functional hierarchy. For example, S1 demonstrated significantly delayed attainment of motor and language milestones, such as independent walking age and combined phrases age, compared to S4. Conversely, S4 attained all tested developmental milestones earlier than the three genetic subgroups (Fig. 4C–D, Figure S11D–I). Moreover, S1 reported the earlier age at which caregivers first expressed developmental concern (age_onset) compared to S4 (P_adj < 0.001) and earlier age of clinical diagnosis (diagnosis_age) than S2 and S3 (P_adj = 0.003 and P_adj = 0.029, respectively) as well (Fig. 4E–F).

Core ASD symptoms and Co-occurring features

Severity profiles for the core dimensions of ASD and co-occurring features demonstrated a complex dissociation across the subgroups. In assessing repetitive and restrictive behaviors (RBS-R overall score), S2 exhibited a significantly higher median score than S3 (P_adj = 0.013). Though S4 reported the highest median score, no pairwise contrast involving S4 reached the significance threshold. Conversely, S3 exhibited the lowest RBS-R score among all four groups. Despite the overall lack of significant RBS-R differences involving S4, S4 displayed the lowest severity of social and communication impairment (Social Communication Questionnaire, SCQ). In contrast, S1 and S2 demonstrated comparable levels of socialization skills, both of which were significantly poorer than those observed in S3 and S4. This pattern highlights a pervasive impact in core social function within the S1 and S2 (Fig. 4G–H).

Analysis of co-occurring behavioral and psychiatric features using the Child Behavior Checklist (CBCL) further revealed subgroup-specific profiles. Both S4 and S1 presented lower overall burden of behavioral problems in total CBCL T-scores than S2 and S3 (Figure S11J). However, S4 tended to be characterized by a significant internalizing burden, with a significantly higher median Internalizing Problems T-score than S1 (P_adj = 0.002), represented as the highest median score on the anxiety disorder (AD) (P_adj < 0.001) (Fig. 4I–J). S2 was primarily characterized by a prominent externalizing burden, with the median score on the conduct disorder (CD) and ADHD Subscale significantly elevated compared to S4 (Figure S11K-L). Notably, S1’s median score on the ADHD Subscale was comparable to S2 (P_adj = 0.40), indicating a specific high-risk domain even within this lower-burden behavioral problems group. S3 demonstrated the broadest and highest pattern of behavioral dysregulation, reporting relative higher median scores on both internalizing and externalizing dimensions, marked by a highly significant elevation in oppositional defiant disorder (ODD), CD and AD (Fig. 4J–L, Figure S11K). Furthermore, assessments of language and cognitive functioning aligned well with these phenotypic profiles. The proportions of individuals performing “significantly below age level” or “slightly below age level” were highest in S1 for both language and cognitive domains. In contrast, S4 had the highest proportions of individuals performing “at age level” or “above age level”. S2 and S3 showed intermediate profiles between these two extremes (Fig. 4M–N). Additionally, we calculated Cohen’s d to quantify the effect sizes of these phenotypic differences (Figure S12A). The overview of the phenotypic profile among four subgroups were presented in Fig. 4O.

To evaluate the robustness of these phenotypic profiles, we performed several sensitivity analyses. First, we conducted covariate-adjusted models by incorporating variant type (such as protein-truncating variants vs. missense variants) as an additional covariate. Second, after identifying SCN2A and SHANK3 as the most recurrent risk genes in our cohort (each with N = 11) (Figure S13), we performed independent sensitivity tests by separately excluding carriers of these two genes. These analyses confirmed the remarkable stability of the subgroup-specific clinical signatures. Even after controlling for variant type or excluding top recurrent drivers, the overall phenotypic profiles for S1, S2, and S3 remained highly consistent (Figure S12B-D). Notably, while the reduction in sample size during gene-exclusion analyses led to expected minor fluctuations in statistical significance for certain measures, particularly those near the primary significance threshold, the effect sizes (Cohen’s d) remained stable in both direction and magnitude. This preservation of clinical gradients indicates that the identified phenotypic constellations are a robust property of the molecular subtypes, independent of specific mutation types or recurrent genes.

Divergent genetic patterns underlying ASD phenotypic heterogeneity

We hypothesized that the phenotypic constellations observed across ASD subgroups, including differences in behavioral and developmental co-occurring behavioral features, may reflect distinct underlying genetic architectures involving both rare and common variants. To test this, we characterized the genetic profiles of individuals stratified by gene cluster membership.

As shown in Fig. 4P–R, we checked the functional annotation of the de novo variants and found there were significant differences in the proportion of the missense variants across clusters. The scores of missense badness, PolyPhen-2, and constraint MPC, variants with MPC > 2 (hereafter referred to as MisB, considered highly deleterious) were lower in S3 and S4 compared to S1 and S2, indicating that mutations in S3 and S4 genes tend to be less functionally disruptive. This pattern was further supported by categorical analysis that Subgroups 1 and 2 contained higher proportion of individuals carrying MisB variants. The comparison of protein truncating variants (PTV) showed the genetic subgroups generally had higher proportion of PTV than S4.

In addition, we assessed PRS of ASD across subgroups to evaluate the contribution of common variant burden. While a global comparison of the full continuous PRS showed a directional trend consistent with our percentile-based findings, with S3 exhibiting a higher mean adjusted PRS than S1 (Figure S14), these differences did not reach statistical significance across the entire cohort spectrum. Instead, to enhance the sensitivity for detecting subgroup divergence at the risk extremes, we implemented a decile-based stratification approach, a paradigm frequently employed in large-scale ASD genetic studies. This analysis revealed a significant difference across the common variant risk spectrum, underscoring that Subgroup 3 consistently carried a higher common variant burden than other subgroups. This pattern was most pronounced at the extremes of the PRS distribution. Specifically, in the lowest PRS percentile, the adjusted PRS of S3 was significantly higher than that of either Subgroup 1 or the Subgroup 4. This relative elevation of common variant risk in S3 was maintained in the highest PRS percentile, where S3’s adjusted PRS was again significantly higher than that of the S1 (Fig. 4S).

Discussion

This study demonstrates that transcriptomics-guided decomposition provides a robust framework for parsing the genetic heterogeneity of ASD holistically. Our decomposition revealed three stable gene clusters corresponding to three molecular subtypes, Synaptic Signaling (C1), mRNA Stabilization (C2), and Histone Modification (C3). These molecular subtypes exhibit distinct developmental trajectories and cellular enrichment patterns, yet display overlap in core pathways such as excitatory postsynaptic potential and histone acetylation. By stratifying a large clinical cohort based on these molecular subtypes, we demonstrated that disturbances to these discrete molecular subtypes independently modulate separable phenotypic constellations. This empirical evidence validates the “many-to-few” model proposed in this study, providing a mechanistic link that anchors clinical diversity in the structured molecular architecture of ASD risk.

In preparing for the transcriptomics-guided decomposition, we focused on high-confidence (HC) and syndromic (SYN) ASD genes from the SFARI database. Both categories represent the most reliable and penetrant genetic risk with established relevance to ASD, with genetic carriers consistently exhibiting ASD features despite differences in risk gene classification. Indeed, with overlapping phenotypes and increasing recognition of shared biology, the distinction between syndromic and non-syndromic genes remains contentious. Moreover, traditional boundaries between gene categories are increasingly blurred, as many risk factors are documented as both SYN and HC genes within the SFARI Gene database [24, 53, 54]. This overlapping classification underscores a shared biological relevance that transcends categorical labels. Importantly, the comparable proportions of HC and SYN genes across clusters suggest that these two categories of genes may interact or converge within shared biological functions, reinforcing the validity of our decomposition approach.

To capture these functions with maximize disease relevance, we restricted analysis to transcriptomic profiles derived from ASD patients. This case-only framework is supported by evidence that case-derived data, different from controls, more effectively capture ASD-specific molecular signatures compared to neurotypical controls, which may lack the perturbed biological context necessary to reveal disease-relevant gene interactions [55, 56]. By leveraging cross-modal transcriptomic integration within this disease-perturbed state, we captured dynamic expression patterns at both spatiotemporal and cell-type scales that might otherwise remain fragmented. By preserving variation across diverse biological contexts, we captured the broad and persistent signals of genetic risk that span different biological state such as developmental stages and brain regions, ensuring that the identified molecularly connected gene clusters reflect robust pathological signatures. Our comparative analyses using neurotypical brain transcriptomes confirmed that the three-cluster organization was not well replicated in controls, underscoring that the subtype structure we identified reflects ASD-specific pathology rather than general brain biology.

In assessing the stability of the transcriptome-based decomposition, we used random subsampling to assess cluster stability. This procedure revealed a subset of genes (N = 74) that frequently reassigned among clusters. Although these genes were involved in two pathways including learning and regulation of synapse structural plasticity, exhibited a unique sensitivity to the subsampling procedure. The shifting affiliations of these genes likely indicate that their regulatory patterns are highly sensitive to spatiotemporal or cellular heterogeneity in different subsampling of the bulk transcriptomes. Unlike the 311 stable genes that maintain stable co-expression signatures, this subset may represent a dynamic biological background whose molecular associations fluctuate across different neurobiological states. Consequently, by prioritizing the stable 311-gene core for downstream analysis, we focused on the most constitutive pathogenic programs that remain robust despite the inherent biological variance in heterogeneous brain data.

Following the transcriptomics-based decomposition, our analysis revealed that the identified stable gene clusters presented distinct expression patterns, corroborating previous results that disrupted ASD risk genes could propagate dysregulation through downstream networks and converge on specific gene patterns to drive pathology [57–59]. Our results both replicate and extend previous classifications of ASD risk genes. Large-scale exome sequencing studies have cataloged ASD risk genes into broad functional categories, most notably neuronal communication (NC) and gene expression regulation (GER) [7, 10]. Consistent with this framework, we found that our three clusters are enriched for synaptic signaling, mRNA stabilization, and histone modification, processes that map onto the NC/GER dichotomy. However, previous studies characterized pathway-level enrichment without explicitly grouping genes. In contrast, our transcriptomics-guided decomposition directly clusters genes based on their co-expression patterns, yielding three discrete gene sets. This allows us to identify both the predominant biological functions enriched in each cluster and the relationships among them. We found that while each cluster aligns with known biological domains, NC or GER, they are not mutually exclusive. Cluster 1 is predominantly enriched in NC genes such as those involved in synaptic signaling, but it also includes GER genes involved in RNA processing and activity-dependent regulation. Cluster 2 integrates GER genes related to mRNA stabilization with a subset of NC genes, forming a hybrid module preferentially expressed in glial lineages. Cluster 3 is dominated by GER genes involved in histone modification, yet it also contains NC genes including NLGN3, and SHANK3. These NC genes, while classically annotated for synaptic function, are known to be regulated by epigenetic mechanisms during neurodevelopment [60–63]. This combination of cluster-specific preferences and cross-cluster functional composition suggests that different subtypes may be driven by perturbations in distinct core mechanisms, with secondary effects propagating through interconnected pathways. The presence of cross-cluster pathways sharing reflects that these core mechanisms do not operate in isolation but extend functionally through their interacting partners, consistent with the understanding that biological pathways are extensively interconnected [24].

By stratifying the SPARK cohort into genetically defined subgroups based on molecular subtypes, we established a link between molecular subtypes and clinical outcome. Our four subgroups showed partial convergence with the four ASD classes identified by Litman et al. [37] using phenotype-driven clustering. Subgroup 1 resembled Litman’s Mixed-DD group, exhibiting the most pervasive functional impairment, lowest Vineland-3 scores and poor socialization. Subgroup 2 aligned with their Social/behavioral group, marked by severe core symptoms and prominent ADHD comorbidities despite relatively preserved development condition. Subgroup 3 overlapped with Broadly affected class, sharing psychiatric burden across ODD, CD, and AD domains with both elevated polygenic risk and PTV, but differed in that our subgroup preserved adaptive function and showed specific enrichment in histone modification. Subgroup 4 most closely resembled the Moderate challenges group, both showing preserved adaptive skills, but our Subgroup 4 showed levels of RBS-R comparable to Subgroup 1 and prominent internalizing profiles.

Beyond these correspondences, our study and that of Litman et al. represent fundamentally different approaches to parsing ASD heterogeneity, with distinct discovery orientations. Litman employed phenotype-anchored strategy that they clustered individuals based on clinical presentation, then examined genetic differences between the resulting classes. It is well-suited to capture how ASD presents in real-world settings and has direct clinical interpretability. Our study adopts a biology-driven strategy that we define molecular subtypes from transcriptomic data, then comparing phenotypes across genetically defined subgroups. This reveals that ASD risk genes are organized into molecularly connected gene clusters that modulate specific phenotypic constellations. Moreover, the resilience of our subgroups to the adjustment for variant type and the exclusion of top genetic drivers reinforces the validity of the subgroup patterns. Thus, we were able to point to potential intervention targets by anchoring clinical phenotypes in molecularly distinct subtypes. For instance, the primary biological feature in histone modification of Cluster 3 associated with Subgroup 3 provides a possible biological target for leveraging epigenetic modulators to restore proper chromatin regulation. While translational application remains distant, our framework provides a molecularly anchored starting point for mechanistic understanding and the precision intervention tailored to clinically distinct subgroups.

Previous study showed both rare and common variants contribute additively to ASD risk [64, 65], that is biologically informed threshold model [66]. Here we examined whether this additive burden differs across subgroups with mutation of distinct gene clusters. Consistent with the additive model, we observed that Subgroup 1 exhibited a markedly higher fraction of damaging de novo missense variants but low polygenic risk, whereas Subgroup 3 showed lower-impact rare variants coupled with significantly higher polygenic risk scores. Notably, while the global comparison of adjusted PRS as a continuous measure showed a consistent directional trend, the divergence was most robustly captured at the extremes of the distribution. This pattern is consistent with the established analytical framework in ASD genomics, where decile-based stratification is utilized to enhance the detection of polygenic signals that may be attenuated across the full continuous spectrum [67, 68]. By focusing on these risk strata, it provides empirical evidence for biologically informed threshold model. The pervasive developmental delay and cognitive impairment observed in Subgroup 1 align with a model where high-impact rare variants drive severe impairment even with low polygenic background. Conversely, the lower predicted functional impact of Subgroup 3 is consistent with common variants acting synergistically with lower-impact rare variants to push individuals across the diagnostic threshold [66, 69, 70]. Subgroup 4 presents well-preserved functioning and fewer symptoms, a profile consistent with relatively lower overall genetic burden across both rare and common variations.

Limitations

Despite these advances, our study has several limitations. First, the cohort utilized for clinical stratification was primarily restricted to individuals of European descent, which may limit the generalizability of our findings across diverse ancestral backgrounds. Furthermore, while our transcriptomics-guided decomposition provided robust gene clusters, these results are based on post-mortem transcriptomic data, consequently, the in vivo functional dynamics of these molecular subtypes may not be fully captured within this framework. At present, only about 14% of ASD cases are explained by clearly pathogenic variants, however, this proportion is likely to increase as additional ASD-associated genes are discovered, underscoring the need for ongoing refinement of subgroup definitions. Lastly, while our analyses focused on brain transcriptome data, whole blood constitutes a more accessible tissue source, offering greater feasibility for large-scale studies and potential clinical applications. Nevertheless, gene expression in whole blood may not fully reflect the molecular alterations occurring in the brain, which is the primary site of ASD pathology. Therefore, we emphasize the importance of future investigations that independently evaluate molecular dimensions derived from brain and whole blood transcriptomic profiles to delineate tissue-specific contributions and evaluate the translational relevance of peripheral biomarkers.

Conclusion

Our transcriptomics-guided decomposition of highly reliable ASD risk genes revealed three molecular subtypes that stratify patients into clinically distinct subgroups. These findings support a liability threshold model, linking de novo variants and polygenic background to diverse outcomes. By bridging molecular subtypes with phenotypic constellation, this framework offers translational value for precision stratification and guides individualized interventions in ASD.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Data availability

All data utilized in this study are derived from publicly available resources and are described in detail within the Materials and Methods section. Qualified researchers may request the SPARK population dataset utilized in this study by applying through https://base.sfari.org.

Declarations

Ethical approval

Our study was conducted in accordance with the Declaration of Helsinki and relevant institutional guidelines and regulations. The genotyping array data and clinical data were applied from SPARK and our application was approved by the Institutional Review Board of Peking University Sixth Hospital.

Consent to participate

Not applicable.

Competing interests

The authors declare that they have no competing interests.