In this project we examined a variety of different types of stratifiers and other continuous models of heterogeneity in autism. Stratifiers included variables such as biological sex, early language outcomes, social visual engagement, and a number of unsupervised data-driven distinctions that can be made with machine learning algorithms applied to phenotypic behavioral data and neuroimaging biomarkers. We also examined continuous models of heterogeneity such as continuous variation in how individuals developed and responded to early intervention. We uncovered several important findings from this work. First, biological sex is a meaningful stratifier that helps disentangle heterogeneity in functional and structural neuroimaging biomarkers (e.g. neural self-representation response, excitation-inhibition imbalance, cortical thickness). Second, early language outcomes are an early developmental distinction that allows for splitting autism into outcome-relevant subtypes, but also subtypes with different developmental and neurobiological mechanisms underpinning them both at the macro-scale level measured by neuroimaging data, but also by biological mechanisms measured in gene expression. Third, we identified that continuous variation in how fast an individual responds to an early behavioral intervention can be predicted from baseline phenotypic and gene expression features measured before the individuals starts the intervention. Fourth, we discovered that subtypes defined by how an individual visually engages with the social world around them is a meaningful distinction that helps parse apart different ways in which neural circuits for social and visual processing are connected. Fifth, we utilized data-driven models of phenotypic and neuroimaging features to define subtypes, and then showcased how subtyping based on data-driven models can help unpack heterogeneity in autism.
