Subtyping Autism Spectrum Disorder With a Population Graph-Based Dual Autoencoder: Revealing Two Distinct Biotypes

Aim: Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by significant heterogeneity in clinical symptoms and underlying neurobiology. This study aimed to identify distinct ASD biotypes and uncover their neurobiological underpinnings using a novel graph-based subtyping approach. Methods: Resting-state fMRI and clinical data from 443 males with ASD (17.22 ± 8.63 years) were analyzed. We proposed a population graph-based dual autoencoder for subtyping (PG-DAS), a deep clustering framework that integrates imaging data and nonimaging data to extract deep features for biotype identification. Statistical analyses were conducted to compare clinical scores and functional connectivity patterns between biotypes. Correlation analyses examined the associations between intra- and internetwork connectivity and clinical symptoms. Predictive modeling using support vector regression assessed the ability of network connectivity to predict clinical scores. Results: Two distinct ASD biotypes were identified. ASD1 exhibited significantly lower clinical scores and reduced network integration, characterized by weaker intra- and internetwork connectivity, particularly in core networks such as the cingulo-opercular network, linked to communication symptom scores. In contrast, ASD2 exhibited greater network segregation, with internetwork connectivity in sensorimotor-related networks correlating with total symptom scores. Predictive modeling further revealed biotype-specific brain-behavior associations, with ASD1 and ASD2 showing positive correlations with social and communication scores, respectively. Conclusion: This study underscores the critical role of biotype-specific brain network patterns in understanding ASD heterogeneity. The proposed PG-DAS framework proved effective in ASD subtyping and holds promise for broader applications in exploring other neuroheterogeneous disorders.