Machine learning of functional network and molecular mechanisms in autism spectrum disorder subtypes

Autism Spectrum Disorder (ASD) describes a diverse group of neurodevelopmental disorders encompassing a wide range of clinical impairments. The two core symptoms that define ASD are social communication impairments and restricted and repetitive behaviors, and there is a wide range of cognitive and language abilities. How distinct neurobiological substrates give rise to differing clinical symptoms in subsets of ASD patients is unknown. Using a large, publicly available neuroimaging dataset comprising resting state functional magnetic resonance imaging (rsfMRI) scans from N=299 subjects with ASD and N=907 neurotypical controls, we identified three latent dimensions of functional brain network connectivity that predict individual differences in ASD symptoms and behaviors. We show that patients with ASD can be grouped into distinct neurophysiological subgroups based on patterns of dysfunctional connectivity and clinical behaviors. In this cohort, functional connectivity features were extracted from rsfMRI data, regularized canonical correlation analysis was used to identify associations between connectivity features and behavioral data, and ASD subjects were clustered along these dimensions. Cross-validation analyses showed high stability in the brain-behavior dimensions, with replicable clusters in held-out data and a second out-of-sample dataset (National Database for Autism Research; NDAR). Next, we integrated neuroimaging data with gene expression data from the Allen Human Brain Atlas, and found that within each subgroup, ASD-related functional connectivity was explained by regional differences in the expression of distinct gene sets. We replicated these findings using the BrainSpan developmental atlas. These were enriched for transcriptionally-regulated and ASD-associated genes along with immune and synaptic signaling pathways. In sum, our results identify discrete ASD subgroups associated with specific ASD behaviors and neurophysiological signatures, and these different forms of ASD implicate distinct genetic mechanisms. The results of this study suggest a promising new approach for understanding the neurobiological substrates of ASD.