When biomarkers for major depressive disorder remain elusive

We read Winter et al.’s recent article published in JAMA Psychiatry with strong interest [1]. Despite that their systematical evaluations on machine learning-based approach were incapable of identifying multivariate neuroimaging biomarker for major depressive disorder (MDD), we argue that achieving no biomarker does not necessarily equate to their non-existence. Rather than dwelling on how Winter et al.’s study should be optimized, however, the focal point in the current psychiatric context might primarily revolve around the utilization of neuroimaging in the absence of diagnostic biomarkers, the justification for such warrants, and how current findings could be effectively translated into clinical applications for MDD.

In brief, we underscore the priority of identifying subtyping rather than diagnostic neurobiomarkers for MDD, supported by the following rationale. Instead of being one single disease, MDD is a highly complex clinical syndrome with up to 1,500 Diagnostic and Statistical Manual of Mental Disorders-IV (DSM-IV) symptomatic combinations [2]. Patients could manifest with non-overlapping or even reverse symptoms while being equally diagnosed with MDD. As achieving an exhaustive combination of symptoms and behaviors would be practically unattainable, it is crucial to firstly categorize heterogeneous MDD patients into homogeneous subtypes to better understand this complex population. However, identical clinical characteristics may stem from distinct neurobiologies; neuroimaging-based subtyping, on the other hand, could increase the neuroscientific validity of MDD subtyping in an objective way [3].

Subsequently, as the ultimate goal of biomarker identification rests with achieving more precise diagnosis and treatment in MDD, drawing on our prior work, we would like to address how the subtyping biomarker could be further translated into clinical practice. By applying deep learning to a large-scale sample, we have previously identified a subtyping neurobiomarker with underlying genetic and environmental bases, which further revealed outperforming reproducibility in an independent MDD population [4], 5]. Using machine learning methods, we then leveraged neural deficits recognized via the biomarker into a more precise neuromodulation strategy for MDD [5]. Notably, the subtyping biomarker in both MDD subtypes was significantly normalized post-treatment, indicating a MDD identification, treatment, and evaluation closed-loop framework as guided by neuroimaging.

Acknowledging potential limitations, the subtyping neurobiomarker has provided another train of thought beyond traditional methodological measures on how neuroimaging could perform, minding the huge gap in translational neuroimaging for better clinical guidance on MDD. Moreover, the revolutions of large language models in modern medicine hold strong promise for psychiatry, signaling a fundamental shift toward more precise diagnosis and treatment in psychiatric fields.