A Modular Framework for Data-Driven Neurogenetics to Predict Complex and Multidimensional Autistic Phenotypes

Project Summary/Abstract Autism Spectrum Disorder (ASD) can be viewed through three complementary lenses: neurologically, it is linked to distributed changes in brain structure, function, and connectivity; biologically, it is associated with genome-wide mutations across multiple pathways; and clinically, it manifests as a diverse spectrum of behavioral and cogni- tive impairments. Despite this richness, treatment options for ASD are based on coarse diagnostics and target a few specific symptoms, such as social awareness, irritability, and depression. As a result, they have varied, and often limited, efficacy across patients. Taken together, next-generation therapeutics for ASD will crucially depend on our ability to bridge its neurological, biological, and clinical viewpoints for personalized intervention. Imaging-genetics is an emerging field that attempts to link neuroimaging features with genetic variants. How- ever, most methods focus on a restricted set of biomarkers, and they do not account for clinical phenotype, both of which provide an incomplete picture of the impacted processes. Our long-term goal is to develop a modu- lar platform that fuses multimodal neuroimaging and multi-omics data to unravel the complex etiology of ASD. The overall objective of this proposal is to develop and validate interpretable deep learning models to combine genome-wide variants with whole-brain structural and functional MRI data. Our innovative strategy is to use biologically-informed neural network architectures to project each of the modalities to a shared latent space that is simultaneously predictive of patient-level clinical phenotype. This unique formulation can easily accommodate missing data modalities, thus maximally utilizing all of the available information. We will devise, implement, vali- date, and disseminate our model via four specific aims. In Aim 1 we will develop a graph neural network, whose connections mimic a well-known gene ontology. Hierarchical pooling operations will capture the information flow through the network, while an attention layer will learn the discriminative biological pathways associated with the phenotype. In Aim 2 we will develop and integrate a Bayesian feature selection procedure to identify ROI-based neuroimaging biomarkers and a matrix autoencoder to extract discriminative functional subnetworks from brain connectivity data. In Aim 3 we will use the fused imaging and genetic architectures to uncover the neural and biological bases underlying the observed clinical heterogeneity of ASD. Finally, in Aim 4 we will package and dis- seminate our model as a user-friendly tool for the broader research community. We anticipate the proposed work will have a transformative impact on ASD research by allowing us to develop refined diagnostic instruments and on the field of imaging-genetics by providing a new approach for multimodal biomarker discovery.