Interpretable connectome modelling reveals distributed dysconnectivity and developmental transcriptomic associations in autism

Background: Autism spectrum disorder (ASD) is associated with heterogeneous alterations in large-scale brain connectivity, but the biological relevance of machine-learning-derived connectivity patterns remains uncertain. We developed an interpretable resting-state functional connectivity framework to identify distributed connectivity patterns associated with ASD and to examine their spatial correspondence with autism-risk gene expression and developmental neuronal perturbation signatures. Methods: Resting-state functional connectivity was analysed in 778 participants from the ABIDE I cohort after motion-based quality control, including 350 autistic individuals and 428 typically developing controls. Pairwise connectivity among 116 AAL regions yielded 6,670 Fisher z-transformed features. Elastic Net feature selection was performed using the training data, and four candidate pipelines combining the top 68 or top 100 selected connections with support vector machine or LightGBM classifiers were compared in a held-out evaluation set. The top100 + LightGBM pipeline was retained as the reference model for TreeSHAP interpretation, covariate-adjusted directionality analysis and network-level summarization. Model-derived dysconnectivity axes were subsequently examined using Allen Human Brain Atlas regional expression data, SFARI autism-risk gene sets, Geneformer virtual deletion and targeted scTenifoldKnk virtual knockout analysis. Results: Among the four candidate pipelines, the Elastic Net top100 + LightGBM model showed the highest observed discrimination in the held-out evaluation set, with an AUC of 0.791, accuracy of 0.744, balanced accuracy of 0.744, sensitivity of 0.746 and specificity of 0.741. Repeated five-fold cross-validation supported moderate discrimination, whereas leave-one-site-out validation yielded more conservative estimates of cross-site generalizability. SHAP interpretation identified a distributed between-network signature involving limbic, cerebellar, frontoparietal/executive, visual, sensorimotor and parietal/attention systems, with ASD-related hypoconnectivity predominating. The frontoparietal/executive–visual axis showed the strongest exploratory enrichment for SFARI genes. In silico perturbation analyses prioritized TSHZ3 in neuroblasts, RORB in neuronal intermediate progenitors and CUX2 in neuroblasts. Conclusions: Interpretable functional connectivity modelling identified a distributed, predominantly hypoconnected between-network pattern associated with ASD. The spatial correspondence of the frontoparietal/executive–visual axis with autism-risk gene expression, together with complementary in silico perturbation findings, provides biologically informed candidates for future investigation.