Diagnosis-informed neuro-subtyping reveals subgroups of autism spectrum disorder with reliable and distinct functional connectivity profiles

• Published in: Progress in neuro-psychopharmacology & biological psychiatry Vol. 141; p. 111452
• Main Authors: Wang, Yaping, Chen, Zehua, Song, Peilun, Lam, Gary Yu-Hin, Kang, Xin, Wong, Patrick C.M., Geng, Xiujuan
• Format: Journal Article
• Language: English
• Published: England Elsevier Inc 30.08.2025
• Subjects: Adolescent; Autism spectrum disorder; Autism Spectrum Disorder - classification; Autism Spectrum Disorder - diagnosis; Autism Spectrum Disorder - diagnostic imaging; Autism Spectrum Disorder - physiopathology; Brain - diagnostic imaging; Brain - physiopathology; Child; Cluster Analysis; Connectome - methods; Female; Functional connectivity; Functional MRI; Humans; Magnetic Resonance Imaging; Male; Nerve Net - diagnostic imaging; Nerve Net - physiopathology; Neural Pathways - physiopathology; Neuro-subtyping; Semi-supervised clustering; Functional connectivity; Neuro-subtyping; Functional MRI; Semi-supervised clustering; Autism spectrum disorder
• ISSN: 0278-5846; 1878-4216; 1878-4216
• DOI: 10.1016/j.pnpbp.2025.111452

Autism Spectrum Disorder (ASD) is a complex neurodevelopmental disorder characterized by heterogeneous symptoms and neurobiological features, which hinders the identification of reliable biomarkers. Until recently, ASD neuro-subtyping has emerged to detect neural features in each subgroup. We implemented neuro-subtyping of ASD using a semi-supervised clustering method, HeterogeneitY through DiscRiminative Analysis (HYDRA), guided by the labeling information of ASD/controls, together with a multi-scale dimension reduction method of high-dimensional input features. Functional connectivity was estimated as neural features for subtyping subjects from a large dataset with ∼2000 subjects. Systematic evaluation of clustering performance was conducted and the semi-supervised approach was compared with unsupervised K-means, commonly used for neuro-subtyping, combined with different types of feature reduction methods. We successfully detected two clusters, the hyper-connectivity subtype and hypo-connectivity subtype, each exhibiting distinct connectivity patterns between and within large networks, with high reliability. The semi-supervised clustering approach demonstrated superior performance compared to the unsupervised approach. We observed cluster effect on functional connectivities, for instance, the hyper-connectivity cluster shows hyper-connectivity within major large networks and hyper/hypo-connectivities between networks, such as hyper-connectivity between default mode and attention networks, and hypo-connectivity between default mode and visual/auditory networks. In contrast, the hypo-connectivity cluster displayed the opposite connectivity patterns. Furthermore, we found varying correlations between connectivities and main symptoms of ASD across subtypes. Our findings indicate that the semi-supervised approach has the potential to subtype ASD into distinct and reliable clusters. The clusters effectively differentiate heterogeneous neural markers based on functional connectivity patterns, meanwhile establish distinct neurobehavioral relationships across each subtype, which is a critical step towards developing individualized diagnosis and treatment strategies in the future. •Implemented semi-supervised clustering method with diagnosis labels with high-dimension reduction method for ASD subtyping.•Discovered two distinct and robust subtypes of ASD based on functional connectivity maps from ∼1800 individuals.•Each subtypes show distinct neuro-behavioral correlations critical to develop individualized treatment strategies.