SUMMARY: The discovery and refinement of brain-based signatures of autism spectrum disorder (ASD) has for many years been a highly desired, but as yet elusive, goal. One key challenge identified has been that numerous levels and sources of variability — between sites, between participants, and within participants — obscure the search for these reproducible neural signatures, complicating the search for biomarkers and undermining the elucidation of cognitive and neural mechanisms. In this renewal application we propose a sequence of studies to dissociate and quantify these sources of variability. We will acquire a new fMRI dataset that is partly continuous with data accrued during the prior funding period and has 3 key features. First, we will scan participants with ASD and matched controls while they watch complex videos with rich narrative content. Evoked responses to videos constrain variability within and across participants and data collection sites, as borne out by strong pilot data, and naturalistic videos better approximate the demands of processing complex real-world social situations. Second, we will use densely-sampled, longitudinally-acquired, high- quality neuroimaging data that will permit precise, stable, and reliable measurements of an individual's brain function. Third, we will collect primary data at two sites (Indiana University and Caltech) in order to ensure broader generalizability. Using machine learning techniques, Aim 1 will learn where in the brain and when, in response to the video, individuals with ASD diverge most from control participants. Extending beyond group- level averages, we will also take a dimensional approach to link brain differences to phenotypic variation, and a clustering approach to identify variation consistent with the presence of ASD subgroups. In Aim 2, we will leverage these results together with state-of-the-art computer vision and speech algorithms to quantify the stimulus features of the videos that evoke these neural differences, both at the level of the group and of the individual, and examine their relationship to phenotypic differences. Our comprehensive feature decomposition of the videos will query high-level semantic features, object-level features like faces, and low- level perceptual features. Finally Aim 3 will share the all the products of this work — e.g., raw and processed data, acquisition tools, annotations, analysis scripts — on OpenNeuro, NDA, and Github in modern formats (e.g., using BIDS format, and with fMRIprep and MRIQC as processing and quality assurance tools) to make them maximally accessible to others. Uniquely, this data release will be validated and refined at yet a third site, the University of Iowa, on a smaller sample of control participants. This renewal application will thus build upon tools, data, and progress from the current funding period, and capitalize on state-of-the-art computer vision and neuroimaging analysis methods to identify audiovisual stimulus features ...