An Interpretable Deep Learning Framework Reveals Frontoparietal Control Network Hyperactivation Underlying Autism Diagnosis and Symptom Severity

BACKGROUNDAutism spectrum disorder (ASD) is marked by profound neurobiological heterogeneity, which drives inconsistent neuroimaging findings and impede the discovery of reliable biomarkers for precise diagnosis and phenotypic prediction. Although deep learning has shown promising predictive power, its black-box nature obscures the mechanistic interpretability underlying high-dimensional learned representations, limiting their translation into actionable neurobiological insights. METHODSWe present IBSS-GAT, a novel interpretable deep learning framework that explicitly models the spatiotemporal landscape of individual-specific internal brain states and integrates a two-stage mechanistic interpretability pipeline to bridge model-derived features to well-characterized neurodynamic processes and clinical phenotypes. RESULTSAcross three independent large-scale neuroimaging cohorts, IBSS-GAT achieved state-of-the-art classification performance in both cognitive decoding (99.30% accuracy in the HCP-task cohort) and ASD identification (77.26% accuracy in the ABIDE-I, and 77.49% accuracy in the ABIDE-II). Interpretability analyses revealed the frontoparietal control network (FPCN) as a convergent hallmark of ASD, mechanistically anchored in the pathological hyperexpression of an FPCN-dominated metastate. Moreover, both the increased metastate occupancy and model-derived feature strength of FPCN emerged as robust predictors of clinical symptom severity in ASD across ABIDE-I and ABIDE-II. CONCLUSIONSOur work establishes a robust, mechanistically interpretable link between individual high-dimensional brain dynamics and heterogeneous ASD phenotypes, revealing generalizable, neurobiologically grounded brain markers with the potential to inform precision medicine in ASD.