The World Health Organization (WHO) projects that by 2030, major depressive disorder (MDD) will be the leading cause of disease burden on a global scale (Bains & Abdijadid, 2026). So why do we still understand so little about how it works biologically?
Researchers have long attempted to identify brain-based markers of MDD using neuroimaging, with some evidence linking depression to structural changes in regions such as the hippocampus; an area important for memory and emotional processing (Campbell & MacQueen, 2004; Roddy et al., 2019).
One of the largest neuroimaging studies to date, the ENIGMA MDD consortium, analysed thousands of people with depression across 45 cohorts in 14 countries (Schmaal et al., 2020). Although this work helped recognise structural changes in the brain related to MDD, findings from broad brain regions have generally shown limited ability to explain depressive symptoms or predict clinical outcomes. Essentially, we are back to square one. It appears that developing brain predictors for MDD is a hopeless case… or is it?
Turns out, us researchers are not willing to give up just yet. Jiang et al. (2026) recognised that these limitations may partly reflect the low spatial resolution of previous studies. Using machine learning and deep-learning methods, the authors aimed to identify more subtle and localised brain patterns that could improve prediction of MDD.
Large neuroimaging studies have struggled to identify reliable brain markers of depression, but newer artificial intelligence approaches may detect more subtle and clinically useful brain patterns.
Methods
The researchers applied their machine learning and deep learning approaches to two separate brain imaging datasets. Machine learning is a type of artificial intelligence which can learn patterns in data to make predictions. Deep learning is a subset of machine learning which can automatically extract learned features without any manual input, therefore offering value to larger, more unstructured datasets.
The first dataset was the UK Biobank and included 1,496 MDD cases and 27,741 controls. The data was split into training and testing samples, with four controls matched to every one MDD case. Grey matter (i.e., the outer surface of the brain focused on information processing) images were divided into 3D sections known as voxels. The authors then trained a machine learning model, called the Best Linear Unbiased Prediction (BLUP), to predict MDD status from voxel-level brain measures.
For greater detail, region-of-interest (ROI) analyses were used to identify specific brain areas linked to MDD risk, and both models were compared using polygenic scores (i.e., a number that summarises the level of predisposition in a person’s specific genes for MDD). Findings were replicated in a smaller independent dataset (DEP-ARREST CLIN), consisting of 64 hospital patients and 32 controls.
Did I lose any of you? In short, the study used machine learning and deep learning on brain imaging data to test whether MDD could be predicted from brain patterns and genetic risk.
Results
If there’s one takeaway you need from this study, it’s this:
The machine-learning (BLUP) model was strongly associated with MDD risk, explaining around 6.1% of variation in case status across more than 415,000 voxel measures.
This finding was consistent across both males and females and applied to depressive episodes occurring up to 5 years before imaging.
Unfortunately, the same success cannot be said for the deep-learning model, which has an AUC of 0.53. AUC refers to Area Under the Curve and tells you how good a model is at distinguishing two results. An AUC of 0.5 means the model essentially distinguishes them completely by chance. In this instance, the results had a p-value of less than 0.05 (which is typically used to indicate statistical significance). However, when we are dealing with these large datasets, the risk of false positives increases. Therefore, the researchers applied multiple testing corrections, lowering the p-value threshold for significance, of which the deep-learning results did not pass (unlike BLUP).
Remember those regions of interest (ROIs) I spoke about in the methods? Well, a total of 17 ROIs were identified that associated with MDD risk prediction within the cerebellum, cortex, and subcortical structures. Although these associations did not remain statistically significant after multiple testing correction, the ROIs aligned well with previous findings, such as the reduced hippocampus volume in the ENIGMA study. Even better, the researchers actually found additional associations that have not been previously recognised, such as an additional genetic component associated with MDD risk.
Speaking of genetics, this may be one of the most interesting elements. It is widely acknowledged that genetics play a considerable role in MDD risk (Alshaya, 2022). Both the BLUP predictor and deep-learning predictor were significantly correlated with the polygenic scores. The significance of this did, however, vary across demographics, with the most success occurring in the mixed-sex and female analyses. When these polygenic scores were added into the BLUP model, it actually improved predictive accuracy.
So, where are we so far? Although the deep learning prediction was almost completely down to chance, BLUP prediction performed with an AUC of 0.57. Even still, this score is only moderately above 50%, limited by that variance of 6.1%. Combining genetic predictors with the BLUP model produced an AUC of 0.66, compared to 0.65 with polygenic scores alone. You’re probably thinking, “that’s only a difference of 0.1”, and you’d be right. Despite this small difference, it does suggest that there may be some kind of environmental element to genetic predictors of MDD (e.g., being bullied as a child).
Machine learning and deep-learning models applied to large brain imaging datasets found modest but significant brain-based signals of MDD, with limited predictive accuracy and small improvements when combined with genetic data.
Conclusions
In conclusion, this study outlines the modest ability of a BLUP machine learning predictor to distinguish MDD cases from controls. Moreover, combining BLUP with genetic factors could improve upon this predictive accuracy. This additional finding is also an exciting piece of evidence supporting the argument that both genetics and environment contribute to the risk of a diagnosis of MDD, addressing the longstanding “nature vs nurture” debate.
Overall, although the authors acknowledge that brain markers will likely never be used clinically due to the limited level of variance they explain for MDD, their research is invaluable in supporting the enrichment of “current knowledge on the function and pathophysiological links of specific brain regions in MDD.” To put it simply, we can learn more about how our bodies are impacted by MDD on a biological level.
Predictive ability of genetic factors combined with structural brain markers support future research on the pathophysiology of depression.
Strengths and limitations
Overall, this is a robust study with well-thought-out, comprehensive methodology supporting reliable outcomes that have potential to lead future research in expanding our understanding of the causes of MDD. Despite the fairly moderate outcomes, the overarching structural and genetic factors associated with MDD not only support existing evidence, but go beyond that. The study applies multiple testing corrections to reduce the sway of false positives on predictive value, as well as adjusting for covariates with logistic regression. Nevertheless, there are a few limitations that should be acknowledged when assessing their proposed findings.
Firstly, the researchers assign controls to each case according to a range of demographic factors, such as sex, ancestry, and age. Although this is useful to control for any confounders, it also potentially introduces selection bias whereby the population becomes less representative. Even further, the testing group essentially consists of the ‘leftover’ cases and forces remaining controls to be matched, potentially reducing population representation even further.
Additionally, the researchers acknowledge that the sample largely consists of females, with limited male representation. Although they evaluate both sexes separately to account for this, the much smaller male sample may limit appropriate representation of the overall population. This may explain why only the mixed-sex and female groups were significant for MDD risk in the integrative model (BLUP + polygenic scores). Speaking of polygenic scores, these were only calculated for European-ancestry participants, excluding other, potentially meaningful, genetic influences.
Finally, if we focus on the second cohort, DEP-ARREST CLIN, we find that these participants are included if they have experienced a major depressive episode, but do not necessarily have MDD. This makes direct comparison with the UK Biobank dataset challenging. On top of this, the controls used within this cohort are not specified, and we do not know whether these are other hospital patients or how they were recruited. This may account for the missed significance found for this sample.
After assessing these limitations, it is also important to see where they might take their study one step further. For example, they exclude any participants with mental health problems outside of MDD, however, MDD is highly comorbid, and its interaction with other mental health concerns may lead to some interesting findings (Thaipisuttikul et al., 2014). Furthermore, the researchers hint in their methodology that they are keen to explore how antidepressant use may contribute to brain structural changes, however, in their results they merely adjust for antidepressant use as a confounding factor. Similarly, the researchers could have split participants based on the severity of their MDD symptoms, potentially identifying additional correlations and brain structural changes there.
Robust methods and large datasets support modest but meaningful findings, though selection bias, limited representativeness and replication differences constrain interpretation and generalisability.
Implications for practice
Okay, let’s regroup. We have an intriguing study that has not only confirmed previous associations between brain structures and MDD risk, but also identified more localised, specific regions and even an additional genetic element. Yet the question still stands: where do we go from here?
As the researchers of this study acknowledge themselves, the limited AUC score (a result of a capped variance explained of 6.1%) means that clinical value of applying a predictive tool like the BLUP predictor is unlikely. We simply could never reliably support a diagnosis of MDD with the relatively slight associations. However, that is not to say these findings are not valuable. This study is phenomenal in increasing our understanding of the biological impact of MDD. It not only expands our knowledge on structural changes in the brain but also informs us of the interplay between genetic and environmental factors. It may be that these discoveries support the determination of mechanisms and brain function regarding MDD, offering potential avenues for additional treatment opportunities and novel targets in the brain.
More broadly speaking, this research is, in my opinion, a huge milestone for reducing the stigma around mental health. The reliable, validated findings in the study evidence the biological, physical changes linked to MDD. This defies outdated criticisms that mental health is ‘only in your head’ or something you can simply ‘get over’ without support. This study allowed MDD to be treated like any other disease, with just as much value to research on how we can better understand, support, and treat it.
Uncovering brain markers connected to depression supports the treatment of this often-stigmatised mental health condition just like any other disease.
Statement of interests
Emily Gillings has no conflicts of interest to report.
Editor
Edited by Éimear Foley. AI tools assisted with language refinement and formatting during the editorial phase.
Links
Primary paper
Jiayue-Clara Jiang, Camille Brianceau, Elise Delzant, Romain Colle, Hugo Bottemanne, Emmanuelle Corruble, Naomi Wray, Olivier Colliot, Sonia Shah, and Baptiste Couvy-Duchesne. (2026). Applying machine-learning and deep-learning to predict depression from brain MRI and identify depression-related brain biology. Translational Psychiatry, 16(1), 171. https://doi.org/10.1038/s41398-026-03889-8
Other references
Alshaya, D. S. (2022). Genetic and epigenetic factors associated with depression: An updated overview. Saudi Journal of Biological Sciences, 29(8), 103311. https://doi.org/10.1016/j.sjbs.2022.103311
Bains, N., & Abdijadid, S. (2026). Major Depressive Disorder. In StatPearls. StatPearls Publishing. http://www.ncbi.nlm.nih.gov/books/NBK559078/
Campbell, S., & MacQueen, G. (2004). The role of the hippocampus in the pathophysiology of major depression. Journal of Psychiatry and Neuroscience, 29(6), 417–426.
Jiang, J.-C., Brianceau, C., Delzant, E., Colle, R., Bottemanne, H., Corruble, E., Wray, N. R., Colliot, O., Shah, S., & Couvy-Duchesne, B. (2026). Applying machine-learning and deep-learning to predict depression from brain MRI and identify depression-related brain biology. Translational Psychiatry, 16(1), 171. https://doi.org/10.1038/s41398-026-03889-8
Roddy, D. W., Farrell, C., Doolin, K., Roman, E., Tozzi, L., Frodl, T., O’Keane, V., & O’Hanlon, E. (2019). The Hippocampus in Depression: More Than the Sum of Its Parts? Advanced Hippocampal Substructure Segmentation in Depression. Biological Psychiatry, Revisiting the Neural Circuitry of Depression, 85(6), 487–497. https://doi.org/10.1016/j.biopsych.2018.08.021
Schmaal, L., Pozzi, E., C. Ho, T., van Velzen, L. S., Veer, I. M., Opel, N., Van Someren, E. J. W., Han, L. K. M., Aftanas, L., Aleman, A., Baune, B. T., Berger, K., Blanken, T. F., Capitão, L., Couvy-Duchesne, B., R. Cullen, K., Dannlowski, U., Davey, C., Erwin-Grabner, T., … Veltman, D. J. (2020). ENIGMA MDD: Seven years of global neuroimaging studies of major depression through worldwide data sharing. Translational Psychiatry, 10, 172. https://doi.org/10.1038/s41398-020-0842-6
Thaipisuttikul, P., Ittasakul, P., Waleeprakhon, P., Wisajun, P., & Jullagate, S. (2014). Psychiatric comorbidities in patients with major depressive disorder. Neuropsychiatric Disease and Treatment, 10, 2097–2103. https://doi.org/10.2147/NDT.S72026
