From heterogeneity to translation: data‑driven subtyping of autism in a multilevel framework

Autism spectrum disorder (ASD) exhibits pronounced heterogeneity across genetic, neurobiological, and clinical phenotypic levels, posing substantial challenges for mechanistic elucidation and clinical translation. This review synthesizes advances in data-driven approaches to parsing ASD heterogeneity and centers the discussion on three complementary strata: neural, behavioral, and transdiagnostic subtypes. At the neuroimaging level, studies leveraging features such as functional connectivity and brain structure have consistently identified two core neurosubtypes characterized by increased and decreased neural activity, respectively. These neurosubtypes differ in time-varying dynamics, spatial architecture, and network hierarchy, and they are closely associated with specific symptom dimensions and cognitive functions. At the behavioral level, data-driven methods delineate phenotypes along axes of severity and functional impairment, and further reveal their links to neural circuits. Transdiagnostic investigations indicate that ASD and frequently co-occurring disorders share neurobiological substrates and cognitive endophenotypes. Collectively, these findings argue against a simple one-to-one correspondence between behavioral and neural subtypes; instead, the evidence is more consistent with multi-to-one, one-to-many, or many-to-many mappings that converge on the overall functional impairment. Notwithstanding this progress, major challenges remain, including sample heterogeneity, methodological inconsistency, and the integration of categorical and dimensional models. Future research should prioritize large samples, multi-site collaboration, longitudinal designs, and transdiagnostic frameworks, coupled with reverse validation via intervention response, to build robust evidence for mechanism-informed individualized assessment and intervention in ASD.