Contrastive learning enhances the links between functional signatures and antidepressant treatment

Major depressive disorder (MDD) is highly heterogeneous in terms of responses to treatment, which hinders the improvement in treatment effectiveness and outcomes for MDD. Identifying MDD subtypes associated with treatment responses could inform interventions and facilitate personalized treatment. Here, we sought to identify reproducible MDD subtypes characterized by distinct neurofunctional (i.e., neuroimaging) patterns to delineate heterogeneity in MDD and explored the relationship between subtypes and antidepressant treatment response. We used contrastive variational autoencoders (CVAEs) to identify two distinct MDD subtypes with the REST-meta-MDD II dataset (1660 MDD participants, 1340 HCs). Subtype 1 exhibited increased functional activity in occipital, parietal, temporal, and frontal areas, while subtype 2 showed decreased functional activity in these areas. The number and patterns of MDD subtypes were validated in a further large multi-center dataset (1276 MDD participants, 1104 HCs). Notably, patients with subtype 1 could be considered the "treatment-sensitive" group, with a response rate of over 50% to all antidepressants and a better response to repetitive transcranial magnetic stimulation (rTMS) compared to patients with subtype 2. In contrast, patients with subtype 2 could be characterized as the "treatment-resistant" group, with a response rate of below 50% for most medications. The ensuing MDD-specific features from CVAEs may serve as a neuroimaging biomarker for predicting treatment outcomes for both medication and rTMS treatments. Our study shows that contrastive learning can be used to establish the predictive validity of functional brain signatures — in terms of responses to antidepressant treatment — offering potential new targets for optimizing treatment strategies for treatment-resistant depression, and further lay a path toward higher treatment outcomes.