777. Advancing biological subtypes of treatment-resistant depression

Abstract

Abstract Background Major depressive disorder (MDD) shows substantial clinical and neurobiological heterogeneity, yet current diagnostic frameworks treat this highly disabling condition as a unitary syndrome. This mismatch has limited efforts to identify reproducible neural markers and to develop biologically informed treatment strategies. Data-driven approaches that integrate symptom measures with functional neuroimaging offer a promising path toward parsing this heterogeneity, but many reported subtypes fail to generalize beyond the original sample. Robust and replicable methods are therefore needed to derive clinically meaningful and biologically grounded dimensions of depression that can inform personalized treatment strategies like non-invasive neuromodulation. Aims & Objectives We aimed to identify stable, generalizable dimensional and categorical representations of depression heterogeneity using multivariate modeling of clinical symptoms and resting-state functional connectivity (RSFC). Specifically, we sought to (1) derive latent brain–behavior dimensions associated with core depressive symptom domains, (2) assess the stability and generalizability of these dimensions using cross-validation, and (3) determine whether dimensional solutions yield clinically meaningful categorical subtypes with distinct neurobiological profiles and differential treatment response. Method We analyzed clinical and RSFC data from 328 individuals with MDD. We applied regularized canonical correlation analysis (rCCA) to identify latent dimensions capturing shared variance between symptom measures and RSFC while minimizing overfitting. We evaluated model performance using cross-validated held-out test sets. We then clustered individuals with MDD based on their dimensional scores using hierarchical clustering to identify categorical subtypes. Finally, we compared subtypes on symptom profiles, functional connectivity patterns, and response to repetitive transcranial magnetic stimulation (rTMS). Results The optimal rCCA model identified three reproducible brain–behavior dimensions that generalized to held-out data. These dimensions primarily reflected variation in (1) depressed mood and somatic symptoms, (2) anhedonia, and (3) insomnia. Each dimension mapped onto distinct patterns of RSFC involving default mode, limbic, and frontoparietal networks. Clustering of dimensional scores revealed four depression subtypes characterized by dissociable clinical profiles and connectivity signatures. Subtypes differed in the relative prominence of affective versus somatic symptoms and showed distinct alterations in default mode network connectivity and limbic–cortical interactions. Importantly, subtype membership predicted differential response to rTMS, indicating that the identified neurobiological patterns carried treatment-relevant information. Subtype assignment was stable post-rTMS, irrespective of response or remission status. Discussion & Conclusions By combining multivariate brain–behavior modeling with clustering, we identified robust dimensional and categorical representations of depression heterogeneity that generalized beyond the training data. The derived dimensions captured clinically salient symptom domains and corresponded to distinct functional connectivity patterns, supporting their neurobiological relevance. The resulting subtypes showed differential treatment response, underscoring their potential utility for stratifying patients and guiding personalized intervention strategies. These findings demonstrate that integrating dimensional and categorical approaches can bridge symptom heterogeneity with circuit-level variation and provide a scalable framework for biologically informed subtyping in MDD. This approach advances efforts toward precision neuropsychopharmacology by identifying reproducible brain–behavior relationships that may inform treatment selection and mechanistic targeting.