Characterizing Autism Heterogeneity by Machine Learning Methods and Its Genetic and Environmental Drivers

About This Grant

PROJECT SUMMARY Autism spectrum disorder (ASD) affects 1 in 36 children in the U.S. and is characterized by substantial heterogeneity in core symptom severity and co-occurring conditions. These variations complicate diagnosis, intervention, and treatment planning, highlighting the need for characterizing ASD clinical traits that better reflect real-world support needs While ASD has a strong genetic basis with heritability of 0.83, specific prenatal environmental exposures also contribute to risk. However, most existing ASD genetics studies rely on binary ASD case definitions and overlook the role of prenatal risk factors. We propose to use a novel unsupervised machine learning method, discriminative dimensionality reduction tree (DDRTree), to construct ASD subtypes categories, continuous ASD subtype scores and severity based on ASD clinical traits. This approach offers a clear visualization of heterogeneous ASD clinical traits. It also helps identify ASD subtypes and severity based on the natural order of the symptom scores represented in the tree structure. By applying DDRTree model in diabetes, researchers identified patterns of differential drug responses among heterogeneous diabetes patients. In a preliminary analysis of 6,356 autistic individuals from the Simons Foundation Powering Autism Research for Knowledge (SPARK), we identified three clinically relevant subtypes with varying support needs in language, social functioning, and medication use. The proposed study will expand this work using multiple ASD cohorts (N=~55,000) and link subtype variation to genetic and prenatal risk factors. In Aim 1, we will construct ASD subtypes and severity scores using ML models based on symptom and comorbidity data from ~45,000 autistic individuals in SPARK, and validate them in external cohorts (N=~10,000). We will also examine differential use of educational and therapeutic services across subtypes. In Aim 2, we will identify common and rare genetic variants associated with the subtypes and severity using GWAS and whole exome/genome sequencing data. We will map common variant effects to neuronal cell types using public single- cell brain atlases, and assess the contribution of rare inherited and de novo variants using family-based analyses. In Aim 3, we will assess prenatal exposures (e.g., maternal complications and exposures) associated with ASD clinical profiles and build risk prediction models that integrate genetic and environmental data. This project will be led by a multidisciplinary team of genetic epidemiologists, ASD clinicians and statisticians. We aim to improve our understanding of ASD heterogeneity by linking symptom profiles to underlying genetics and early-life exposures. The proposed work will support development of individualized risk prediction tools and inform early screening and intervention. By identifying distinct ASD subtypes with unique genetic and prenatal risk factors, we aim to advance personalized medicine approaches in ASD research and care.