Machine Learning-Derived Subtypes of Autism Reveal Distinct Genetic Contributions from Common and Rare Variants
Abstract
BackgroundAutism spectrum disorder (ASD) is characterized by substantial heterogeneity in core symptom severity and co-occurring conditions, complicating efforts to understand its genetic basis. Most genetic studies rely on binary case-control ASD definitions, which may obscure clinically meaningful variation and distinct genetic contributions across the autism spectrum.
MethodsWe constructed machine learning (ML)-based categorical ASD subtypes and corresponding continuous ASD subtype scores using 14 key clinical symptom measures with adjustment of sex and age among 6,355 autistic children in the Simons Foundation Powering Autism Research for Knowledge (SPARK) cohort by Discriminative Dimensionality Reduction via learning a Tree (DDRTree). We then replicated and validated these subtypes and scores among 2,717 autistic children in the Simons Simplex Collection (SSC) using a modified k-nearest-neighbor (k-NN) approach. In SPARK and SSC, we further clinically validated the framework by examining associations of subtype classifications and scores with support needs, service utilization and medication use by regression models with adjustment for sex, age, genetic ancestry and socioeconomic factors. We further assessed SNP-based heritability, cell-type-specific heritability enrichment, polygenic scores (PGS) for psychiatric and cognitive traits, as well as rare copy-number variant (CNV) and de novo variant (DNV) burdens across the ML-derived ASD subtype scores adjusting for sex, age and genetic ancestry.
ResultsWe identified three clinically distinct ASD subtypes in SPARK varying in the types of supports utilized: a high-support-needs (HSN) subtype characterized by broad developmental and functional impairments and very substantial support needs (OR = 1.63, P = 3.71 x 10-98); an emotional-behavioral-support-needs (EBSN) subtype marked by severe repetitive behaviors, elevated psychiatric comorbidity; and a low-support-needs (LSN) subtype with milder symptoms and relatively lower overall support needs. The three groups were associated with differential patterns in the specific types of treatments that were most common for them, from higher intensity to lower intensity. Projected ASD subtypes and scores in SSC recapitulated the phenotypic and medication-use patterns observed in SPARK. SNP-based heritability was highest for the EBSN score (h{superscript 2}g = 0.224, SE = 0.051) followed by the HSN (h{superscript 2}g = 0.167, SE = 0.049) and LSN score (h{superscript 2}g = 0.064, SE = 0.044), with suggestive cell-type-specific heritability enrichments in fetal endodermal cells and adult motor neurons for EBSN score and fetal excitatory neurons and astrocytes for HSN score (both P<0.05). PGS of ADHD and major depressive disorder were consistently associated with higher EBSN scores in both SPARK and SSC. Higher DNV and rare CNV burden were associated with higher HSN score in SPARK, and higher rare CNV burden was associated with higher EBSN score in SSC (P<0.05).
ConclusionsOur ML-based transferable ASD subtype framework captured clinically and genetically meaningful dimensions of ASD heterogeneity, warranting further validation in larger cohorts.
Related
- Machine learning clustering of multidimensional clinical features identifies four prognostically distinct sepsis subtypes
- Radiomics-based machine learning models decipher heterogeneity of anoikis-related molecular subtypes for prognostic stratification in head and neck squamous cell carcinoma
- Revisiting brain asymmetry in schizophrenia: insights from machine learning subtypes from 1,761 patients
- A machine learning classifier reveals hidden rare biallelic heterozygous variants in the coding regions of Meniere disease genome
- Integrating Machine Learning to Identify Prognostic Subtypes and the Cisplatin Resistance Gene RRM2 in Lung Adenocarcinoma.
BibTeX
@misc{tan2026machine,
title = {{Machine Learning-Derived Subtypes of Autism Reveal Distinct Genetic Contributions from Common and Rare Variants}},
author = {Qing Tan and Jingyu Xie and LeeAnne Green Snyder and Runpu Chen and Le Yang and Fei Chen and Brett Trost and Bridget Fernandez and Charleston Wen-Kai Chiang and Nicholas Mancuso and Chang April Shu},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.19.26365670},
url = {https://www.medrxiv.org/content/10.64898/2026.10.19.26365670v1},
note = {CC BY-NC},
}
RIS
TY - UNPB
TI - Machine Learning-Derived Subtypes of Autism Reveal Distinct Genetic Contributions from Common and Rare Variants
AU - Qing Tan
AU - Jingyu Xie
AU - LeeAnne Green Snyder
AU - Runpu Chen
AU - Le Yang
AU - Fei Chen
AU - Brett Trost
AU - Bridget Fernandez
AU - Charleston Wen-Kai Chiang
AU - Nicholas Mancuso
AU - Chang April Shu
PY - 2026
DA - 2026/10/19
PB - medRxiv
DO - 10.64898/2026.10.19.26365670
UR - https://www.medrxiv.org/content/10.64898/2026.10.19.26365670v1
AB - BackgroundAutism spectrum disorder (ASD) is characterized by substantial heterogeneity in core symptom severity and co-occurring conditions, complicating efforts to understand its genetic basis. Most genetic studies rely on binary case-control ASD definitions, which may obscure clinically meaningful variation and distinct genetic contributions across the autism spectrum. MethodsWe constructed machine learning (ML)-based categorical ASD subtypes and corresponding continuous ASD subtype scores using 14 key clinical symptom measures with adjustment of sex and age among 6,355 autistic children in the Simons Foundation Powering Autism Research for Knowledge (SPARK) cohort by Discriminative Dimensionality Reduction via learning a Tree (DDRTree). We then replicated and validated these subtypes and scores among 2,717 autistic children in the Simons Simplex Collection (SSC) using a modified k-nearest-neighbor (k-NN) approach. In SPARK and SSC, we further clinically validated the framework by examining associations of subtype classifications and scores with support needs, service utilization and medication use by regression models with adjustment for sex, age, genetic ancestry and socioeconomic factors. We further assessed SNP-based heritability, cell-type-specific heritability enrichment, polygenic scores (PGS) for psychiatric and cognitive traits, as well as rare copy-number variant (CNV) and de novo variant (DNV) burdens across the ML-derived ASD subtype scores adjusting for sex, age and genetic ancestry. ResultsWe identified three clinically distinct ASD subtypes in SPARK varying in the types of supports utilized: a high-support-needs (HSN) subtype characterized by broad developmental and functional impairments and very substantial support needs (OR = 1.63, P = 3.71 x 10-98); an emotional-behavioral-support-needs (EBSN) subtype marked by severe repetitive behaviors, elevated psychiatric comorbidity; and a low-support-needs (LSN) subtype with milder symptoms and relatively lower overall support needs. The three groups were associated with differential patterns in the specific types of treatments that were most common for them, from higher intensity to lower intensity. Projected ASD subtypes and scores in SSC recapitulated the phenotypic and medication-use patterns observed in SPARK. SNP-based heritability was highest for the EBSN score (h{superscript 2}g = 0.224, SE = 0.051) followed by the HSN (h{superscript 2}g = 0.167, SE = 0.049) and LSN score (h{superscript 2}g = 0.064, SE = 0.044), with suggestive cell-type-specific heritability enrichments in fetal endodermal cells and adult motor neurons for EBSN score and fetal excitatory neurons and astrocytes for HSN score (both P<0.05). PGS of ADHD and major depressive disorder were consistently associated with higher EBSN scores in both SPARK and SSC. Higher DNV and rare CNV burden were associated with higher HSN score in SPARK, and higher rare CNV burden was associated with higher EBSN score in SSC (P<0.05). ConclusionsOur ML-based transferable ASD subtype framework captured clinically and genetically meaningful dimensions of ASD heterogeneity, warranting further validation in larger cohorts.
ER -