Topological data analysis communities reveal gene-environment-brain subtypes of major depression in UK Biobank and multi-site cohorts

Background: Major depressive disorder (MDD) exhibits substantial clinical heterogeneity complicating prognosis and treatment. Characterizing MDD subtypes could enhance personalized approaches. We developed a topological data analysis (TDA) framework with graph-based community detection to identify patient subgroups using multimodal data. Methods: We implemented a TDA pipeline in MDD UK Biobank participants with gene-environment (G-E, N=20,715) and gene-environment-neuroimaging (G-E-I, N=3,044) data. We systematically compared genetic, environmental, and neuroimaging features, alone and combined, to stratify MDD individuals across 18 health-related outcomes. For each outcome's best-performing feature set, a novel feature ranking approach identified features driving graph construction and community-based outcome differentiation. Cross-cohort validation through selective, heterogeneous replication utilized two independent datasets: GSRD (G-E data, N=1,017) and HSR (G-E and imaging data, N=71-87). Results: G-E combination demonstrated superior stratification performance for 13 outcomes, including treatment-resistant depression (TRD), symptom subtypes, and suicidal phenotypes. Community profiling revealed distinct patterns: trauma-stress exposures linked to TRD and episode severity, while substance-behavioral profiles associated with anxious symptoms. Environmental factors primarily determined most health-related outcomes, whereas neuroimaging features best discriminate medical comorbidities. Partial replication was observed for gene-environment sets in GSRD (self-harm behavior, anxious features) and preliminary imaging-based replication in HSR (vascular problems), with limited statistical power for imaging analyses. Environmental stress-related top-ranked features were consistent across cohorts. Conclusions: TDA successfully identified relevant MDD subgroups with domain-specific multimodal contributions. These findings underscore the value of multimodal integration for comprehensive health-related outcome stratification, with modalities contributing selectively to specific outcome domains. TDA-based community detection is a promising framework for MDD stratification and precision medicine.