The decades-long failure to identify reliable brain biomarkers for depression isn't simply a data-volume problem — it may be fundamentally a representation problem. That reframing has significant implications for how clinicians and researchers think about psychiatric diagnosis, treatment stratification, and the design of future neuroimaging studies.
This comprehensive review in Biological Psychiatry traces a methodological evolution in psychiatric neuroimaging from univariate brain-symptom correlations — which have produced only small, inconsistent effect sizes despite large samples — toward multivariate and multimodal latent-representation frameworks. The authors document how linear multiview embedding approaches have successfully identified reproducible biological subtypes of depression, yet these methods falter in smaller samples or cohorts with predominantly mild symptoms. The review then examines deep learning and graph-based neural architectures capable of extracting nonlinear relational patterns across brain regions, and crucially, multimodal extensions that fuse imaging with genomic, behavioral, or clinical data to reveal latent structure invisible to single-modality analyses. A particularly forward-looking section addresses foundation models — large-scale pretrained architectures that can transfer representational knowledge to small, privacy-constrained clinical datasets, potentially circumventing the perpetual sample-size bottleneck in psychiatric research.
From a broader scientific perspective, this review arrives at an inflection point. The field has struggled with what some call the 'biomarker crisis' in psychiatry — a pattern where promising imaging findings rarely replicate across sites or populations. The shift to multimodal latent representations doesn't eliminate this problem but may address one of its root causes: models that are too simple to capture the biological heterogeneity of psychiatric conditions. The emphasis on foundation models is especially noteworthy, as transfer learning has already transformed fields like radiology and genomics. Key limitations acknowledged in the review include the continued scarcity of longitudinal data, interpretability challenges in deep architectures, and the absence of head-to-head clinical validation. This is an ambitious, potentially paradigm-orienting synthesis, though translating these computational advances into actionable clinical biomarkers will require substantial prospective validation work.