ASD-FCGCN: A biologically interpretable graph convolutional network for ASD diagnosis from rs-fMRI

Background and Objectives: Autism spectrum disorder (ASD) is a neurodevelopmental condition traditionally diagnosed through subjective behavioral assessments. Resting-state functional MRI (rs-fMRI) provides a potential pathway for objective neuroimaging biomarker discovery; however, existing deep learning methods often struggle to integrate neurobiological priors with data-driven representations. This study aimed to address these limitations by developing a biologically interpretable graph-learning framework for rs-fMRI-based ASD classification. Methods: We propose ASD-FCGCN, a biologically informed graph convolutional network validated on 871 subjects from the ABIDE dataset. The method systematically integrates neurobiological priors with graph learning. Specifically, prior-weighted functional connectivity matrices are constructed to generate hemisphere-specific representations. A Pairwise Association Encoder (PAE) is employed to estimate inter-subject phenotypic relationships, transforming similarity weights into channel-wise gating signals to dynamically regulate graph convolution. Furthermore, a corpus callosum-inspired interaction mechanism is introduced to enable bidirectional information exchange between hemispheres, thereby synthesizing whole-brain features for classification. Results: Experimental results demonstrated that ASD-FCGCN achieved strong classification performance under subject-level 10-fold cross-validation, with an accuracy of 94.60% ± 3.04%, an AUC of 97.66% ± 1.82%, a sensitivity of 94.66% ± 3.72%, a specificity of 94.52% ± 4.16%, and an F1-score of 94.96% ± 2.81%. Conclusion: The results indicate that integrating neurobiological priors, phenotype-guided convolution, and cross-hemispheric coordination can improve subject-level rs-fMRI-based ASD classification. The proposed model provides a promising computational framework for ASD auxiliary diagnosis and offers biologically informed model-level interpretability by integrating ASD-related prior ROIs, phenotype-guided graph propagation, and hemisphere-specific interaction modeling.