Decomposition of phenotypic heterogeneity in autism reveals distinct and coherent genetic programs
Aviya LitmaniD, Natalie SauerwaldiD, LeeAnne Green SnyderiD, Jennifer H. Foss‐FeigiD, Christopher Y. ParkiD, Yun HaoiD, Ilan DinsteiniD, Chandra L. TheesfeldiD, Olga G. TroyanskayaiD
Abstract
Unraveling the phenotypic and genetic complexity of autism is extremely challenging yet critical for understanding the biology, inheritance, trajectory, and clinical manifestations of the many forms of the condition. Here, we leveraged broad phenotypic data from a large cohort with matched genetics to characterize classes of autism and their patterns of core, associated, and co-occurring traits, ultimately demonstrating that phenotypic patterns are associated with distinct genetic and molecular programs. We used a generative mixture modeling approach to identify robust, clinically-relevant classes of autism which we validate and replicate in a large independent cohort. We link the phenotypic findings to distinct patterns of de novo and inherited variation which emerge from the deconvolution of these genetic signals, and demonstrate that class-specific common variant scores strongly align with clinical outcomes. We further provide insights into the distinct biological pathways and processes disrupted by the sets of mutations in each class. Remarkably, we discover class-specific differences in the developmental timing of genes that are dysregulated, and these temporal patterns correspond to clinical milestone and outcome differences between the classes. These analyses embrace the phenotypic complexity of children with autism, unraveling genetic and molecular programs underlying their heterogeneity and suggesting specific biological dysregulation patterns and mechanistic hypotheses.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 106
- FWCI
- 60.03
- Field percentile
- top 1%
- Organism
- Assay
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (16 Aug 2024) · CC BY 4.0
Mixture modeling of 239 phenotype features in 5,392 SPARK autistic children defines four replicable classes carrying distinct common and rare variant signatures and developmental gene-timing patterns.
Autism varies widely in both phenotype and genetics, and earlier genotype-phenotype studies used trait-centric analyses that ignore how traits co-occur within each person. This work applies a person-centered decomposition to a large cohort with broad phenotyping and sequencing, linking holistic clinical presentations to class-specific genetic and developmental signals.
Key results
ShownThe mixture model finds four phenotypic classes (Social/Behavioral, Mixed ASD with DD, Moderate Challenges, Broadly Impacted). Their profiles match medical diagnoses held out from training, and the classes replicate in the independent Simons Simplex Collection.
Four classes were chosen from 2-10 class models using BIC, validation log likelihood, LMR-LRT and clinician review; class sizes n = 1976/1860/1002/554. A SPARK model on 108 shared features, applied to SSC (n = 861), gave an overall correlation of 0.927 that no run among 10,000 label permutations exceeded (p-value < 1e-4); per-category Pearson r was 0.89 to 0.98. · Identifying the structures of autism phenotypes; Phenotype classes display distinct clinical…; Exploratory class enumeration; Phenotypic Replication on the Simons Simplex…
ShownPolygenic scores (PGS) for common variants differ by class in ways that track phenotype. Moderate Challenges has the highest autism PGS and Social/Behavioral the highest depression PGS. The two developmentally delayed classes have lower educational attainment PGS than siblings.
European-ancestry subset (n = 2,293 probands, n = 3,179 siblings). Autism PGS highest in Moderate Challenges (FDR = 0.087); ADHD PGS high in Broadly Impacted (FDR = 2.44e-05) and Social/Behavioral (FDR = 4.12e-06); MDD PGS in Social/Behavioral FDR = 7.14e-4; Educational Attainment lower in Broadly Impacted (FDR = 9.78e-08) and Mixed ASD with DD (FDR = 1.69e-4). · Distinct genetic signals underlie phenotypic…; Polygenic scores
ShownRare variation separates the two delayed classes: Broadly Impacted carries the most high-impact de novo variants, and only Mixed ASD with DD shows significant rare inherited burden. Moderate Challenges shows an excess of de novo loss-of-function (LoF) variants in genes of intermediate constraint.
Exome trios (n = 1,680 probands, n = 801 siblings). High-impact de novo burden exceeds siblings in all classes (FDR = 0.01/0.01/0.07/0.02) and is largest in Broadly Impacted. Rare inherited LoF/missense burden is significant only in Mixed ASD with DD (FDR = 0.034). Moderate Challenges dnLoF in genes with 0.5 <= pLI < 0.995: FDR = 0.048, OR = 2.8. · Distinct genetic signals underlie phenotypic…; De novo variant calling and filtering; Count-based burden analyses
ShownGene-set and Gene Ontology (GO) analyses point to different biology in each class. De novo hits in FMRP targets concentrate in Broadly Impacted. Chromatin, DNA repair and synaptic terms mark Social/Behavioral, while neuronal excitability and sodium-channel terms mark Mixed ASD with DD.
FMRP-target odds ratio 13.76 (FDR = 0.001) in Broadly Impacted vs 6.15 to 6.37 in the other classes. ASD risk gene dnLoF OR reaches 27.14, while de novo synonymous ORs show no excess (FDR = 0.78 each). GO (ShinyGO 0.80): Social/Behavioral nuclear matrix FE = 100.4; Mixed ASD with DD voltage-gated sodium channel activity FE = 32.7 (FDR = 2.8e-03). · Unique gene sets and cell biological processes…; Odds ratios; GO term analysis
SuggestedGenes disrupted by de novo LoF variants differ by class in when they are expressed during human prefrontal cortex development, which parallels milestone timing: early-expressed genes in Mixed ASD with DD versus postnatal interneuron genes in Social/Behavioral.
dnLoF burden vs siblings in Herring et al. PFC devDEG trend sets (6 stages, four cell categories). Mixed ASD with DD is enriched in Down/Trans Down genes across cell types; Social/Behavioral only in Up/Trans Up MGE interneuron genes. Genes hit in Moderate Challenges have lower pLI (median 0.75 vs 0.95 in Mixed ASD with DD; p-value = 0.026). · Developmental gene expression patterns align…; Developmental gene expression analysis
Methods, data and reproducibility
A StepMix general finite mixture model (Gaussian, binomial and multinomial densities; sex and age covariates) clusters children on 239 questionnaire and milestone features. Classes are then compared with siblings for polygenic scores, exome de novo and rare inherited burden, gene-set odds ratios and developmental expression sets, with GO enrichment tested against a protein-coding gene background.
- SPARK autistic children aged 4-18 with complete SCQ, RBS-R, CBCL 6-18 and background history data (239 features) · human · n = 5,392 (4,174 males)
- SPARK non-autistic siblings with whole-exome data (only SCQ and background history phenotypes available) · human · n = 1745 paired siblings
- SPARK complete trios with whole-exome VCFs used for de novo and rare inherited variant calling · human · n = 1,680 probands, 801 siblings
- SPARK genotyped European-ancestry subset (Infinium Global Screen Array) used for polygenic scores · human · n = 2,293 probands, 3,179 siblings
- SPARK individuals with complete data on 108 features shared with SSC, used to train the replication model · human · n = 6,393
- Simons Simplex Collection autistic children with complete data on the 108 shared features (replication cohort) · human · n = 861
- European-ancestry GWAS summary statistics for autism and five related traits (ADHD, MDD, educational attainment, IQ, epilepsy) · human
- Herring et al. single-cell human prefrontal cortex developmental DEG sets across 6 stages (Up, Trans Up, Down, Trans Down) · human
- Methods & tools
- StepMix 1.2.5 (General Finite Mixture Model), SCQ, RBS-R, CBCL 6-18 and background history questionnaires, HAT de novo variant pipeline (DeepVariant + GATK HaplotypeCaller), Ensembl VEP Release 111.0, LOFTEE v.1.0.4, AlphaMissense, PLINK clumping for polygenic scores, ShinyGO 0.80, gnomAD v4.1.0 allele frequencies and gnomAD pLI, GTEx v.7 brain expression, Gencode v29, SFARI Gene, Satterstrom and Werling et al. gene sets, Infinium Global Screen Array genotyping
- Data & accessions
- https://base.sfari.org/
- Reagents & environment
- StepMix 1.2.5; Ensembl VEP Release 111.0 with LOFTEE v.1.0.4 and AlphaMissense plugins; gnomAD v4.1.0; Gencode v29; GTEx v.7; ShinyGO 0.80 (run June 3, 2024).
- Reproducibility gaps
- Analysis code is promised only upon publication. SPARK and SSC genetic and phenotypic data need approved SFARI Base access under participant consent terms.
Limitations (5)
- The authors note that Moderate Challenges may show the top autism polygenic score partly because GWAS training cohorts phenotypically resemble it, while the two developmentally delayed classes are underrepresented among GWAS participants. · stated by the authors
- No single statistic determined the number of classes. Fit indices put the elbow at 4-6 components, and the four-class choice also relied on clinicians' judgment of interpretability. · stated by the authors
- The class-genotype links are framed as hypotheses that need larger cohorts, broader and longitudinal phenotyping, and whole-genome sequencing to capture regulatory variation that exome data miss. · stated by the authors
- Genetic associations were tested only in SPARK, with no genetic replication in SSC. Several reported rare-variant and PGS contrasts sit near FDR 0.1 (e.g., 0.07, 0.079, 0.0887) across many tests. · noted in this summary
- Classes rest on parent-reported questionnaires, and CBCL 6-18 data were included for some children recorded as under 6. Siblings had only SCQ and background history data, which limits sibling baselines for most phenotype domains. · noted in this summary
Who should read:Read this if you study autism heterogeneity, person-centered phenotype clustering, or how de novo, inherited and polygenic signals map onto clinically defined patient subgroups.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- Genetic correlates of phenotypic heterogeneity in autism
Varun Warrier, Xinhe Zhang, … Simon Baron‐Cohen·Nature GeneticsPeer-reviewed·Jun 2, 2022
- Rare variation illuminates the distinct and pleiotropic genetic architecture of autism across neuropsychiatric traits
F. Kyle Satterstrom, Chiara Auwerx, … Mark J. Daly·medRxiv·Aug 26, 2026
- Enrichment of genomic variation in pathways linked to autism
Francisco José Esteban, Peter J. Tonellato, Dennis Paul Wall·bioRxiv·Oct 19, 2020
- Structural Variants Are a Major Component of the Missing Heritability of Autism Spectrum Disorder
David Kainer, Alan R. Templeton, … Michael R. Garvin·medRxiv·Oct 10, 2021
- Disentangling the genetic causes of autism
María M. Abad‐Grau·medRxiv·Jul 30, 2025
- A Novel Phenotype-based Approach for Prioritizing Candidate Genetic Variants for Autism Spectrum Disorder.
Noam Levi, Mor Dekel, … Idan Menashe·medRxiv·Jul 2, 2026