Uncovering Altered Developmental Patterns in Structural Connectivity for Autism Spectrum Disorder through Deep Generative Modeling of Normative Development
Penn collection
Degree type
Discipline
Computer Sciences
Neuroscience and Neurobiology
Subject
Normative Model
Diffusion Tensor Imaging
Generative Network Modeling
Deep Learning
Funder
Grant number
Copyright date
Distributor
Author
Contributor
Abstract
The evolution of structural connectivity during development involves intricate biological processes, disruptions in which are the hallmark of several disorders, including autism spectrum disorder (ASD). To characterize the normative developmental patterns and detect deviations therein, we propose a novel deep generative model which employs a variational auto-encoder with biologically meaningful wiring constraints to infer the latent embedding of structural connectivity at different ages, as shown in Fig.1. This approach allows us to 1) obtain latent features and map normative variation for typical developing controls (TDCs), 2) estimate brain age for each subject, 3) generate brain networks that mimic observed ones, and 4) detect individual deviation from the normative model to derive an overall deviation score. We assessed the model’s performance in two datasets: PNC (TDC=968, age 8-12) and CAR (TDC=196, ASD=229, age 6-19). We compared our model to 13 classic generative models and 5 machine/deep learning models. Our model consistently achieved the best performance with smallest discrepancy between generated and observed data, and the lowest MAE & RMSE between estimated brain age and chronological age. The gap between brain and chronological age is significantly correlated with autistic traits (Spearman R = −0.29, p < 0.01) and IQ (Pearson R = 0.20, p < 0.01). We observe substantial variability in regional differences among autistics, indicating notable heterogeneity in our sample. Generally, youth with more autistic traits exhibit larger brain deviations across more brain regions. The visual (right medial visual cortex), default mode (bilateral prefrontal cortex), dorsal attention (bilateral frontal eye fields), somatomotor (right pericentral cortex), and executive control networks (bilateral lateral prefrontal cortex) are most frequently associated with autism. Overall, our model is effective in characterizing normative and altered developmental brain trajectories, and brain deviation is a promising subject-specific marker of autism.