Depression May Be One Continuous Spectrum, New Protein Study Suggests

Major depressive disorder has long resisted the tidy categories that clinicians and researchers would like to impose on it. Two patients with the same diagnosis can carry strikingly different burdens of inflammation, metabolic disruption, and hematological change, and conventional case-control studies, which compare a single depressed group against healthy controls, tend to average those differences into noise. A new study published in BMC Medicine argues that this averaging may be hiding the real structure of the disease. Using a machine learning method called BioNet-SPACC, a team at Macao Polytechnic University found that the plasma proteins of people with major depression are organized not into distinct subtypes but along a single, predominantly continuous molecular axis, one that a locked panel of just 15 proteins can reproduce with near-perfect fidelity in held-out data.

The method itself is as much a statistical argument as a biological one. BioNet-SPACC, short for biologically informed network-regularized sparse consensus clustering, was built to handle the twin problems that plague high-dimensional proteomics: thousands of correlated measurements and no guarantee that a clustering algorithm will pick features that mean anything biologically. The researchers integrated protein-protein interaction priors, essentially a map of which proteins physically and functionally associate with one another, directly into the regularization of the clustering procedure. Sparse feature weighting then forces the algorithm to concentrate on a small subset of proteins rather than spreading its attention across the entire plasma proteome. The result is a clustering framework that is simultaneously parsimonious and biologically coherent, designed to find molecular structure that survives scrutiny rather than artifacts that evaporate on retesting.

The discovery cohort was drawn from the UK Biobank Pharma Proteomics Project and comprised 4,123 participants with major depressive disorder, divided into discovery and held-out test samples in a leakage-controlled design. That phrase, leakage-controlled, matters enormously here. Many molecular psychiatry studies have been undermined by subtle information leaks between training and test data, which inflate apparent performance. In this design, the clustering was performed only on the discovery sample, and the held-out sample was used to ask whether the same structure emerged independently. Healthy controls, matched to the depressed participants using propensity scores on demographic and metabolic variables, were introduced only after the MDD-only discovery phase was complete, so that the depressed-group structure could be defined without reference to the control distribution.

What the algorithm found was a stable two-way partition, with a median adjusted Rand index of 0.986, an unusually high value indicating that repeated runs of the clustering converge on essentially the same grouping. The partition was defined by a Core 85 protein signature, a set of 85 proteins whose coordinated variation carries the signal. But the more striking finding came when the researchers examined the geometry of that signature. The first principal component of the Core 85 representation explained 91.3 percent of its variance, an extraordinary degree of compression that indicates the 85 proteins move almost as one. Complementary distributional analyses found no evidence against unimodality, meaning the data did not show two separate peaks that would justify calling these genuine disease subtypes. Instead, the two operational groups, labeled low-axis and high-axis, occupy opposing ends of a single continuous dimension.

The held-out test sample provided the crucial replication. When de novo clustering was applied to participants who had played no role in discovery, the resulting assignments agreed with the discovery-projected labels with an adjusted Rand index of 0.789, normalized mutual information of 0.742, and Cohen’s kappa of 0.888. Notably, the proteins selected in the held-out run overlapped only partially with the discovery signature, with a Jaccard index of 0.417, yet the axis they defined was the same. This is a subtle but important point: the individual proteins that a sparse algorithm selects can vary between samples, especially when thousands of correlated features compete for selection, but the dominant dimension of variation is robust. The structure, not the feature list, is what replicates.

The relationship between the two operational groups and the underlying continuum was quantified using consensus-based conditioning, a procedure that averages assignments across many clustering runs to stabilize the labels. Consensus-based conditioning strongly preserved the continuous molecular-axis scores within the depressed group, with a Pearson correlation of 0.940 between conditioned and unconditioned scores, while the binary group assignments showed only moderate agreement with the primary solution. In other words, the labels are convenient handles for a continuum, not hard biological categories. When the researchers anchored the axis to the healthy control distribution, the overall mean of the depressed group was almost identical to that of the controls, with a Cohen’s d of just 0.01, but the low-axis and high-axis depressed groups sat on opposite sides of that control mean. Depression, at the level of circulating proteins, is not a uniform shift away from health; it is a spread in both directions along a molecular dimension.

That dimension has peripheral correlates that reach beyond the proteome. Analyses of complete blood count and nuclear magnetic resonance metabolomics data identified associations spanning platelet-related hematological traits and metabolic measures, suggesting that the proteomic axis intersects with established biological systems long implicated in depression, including inflammatory and cardiovascular pathways. The researchers also checked the axis against external proteomic datasets from the HERITAGE Family Study and the Molecular Transducers of Physical Activity Consortium, which provided cardiorespiratory fitness proteomics data, to ensure the dimension was not simply an artifact of fitness or body composition. The axis survived as a depression-relevant structure.

Perhaps the most clinically provocative result is the 15-protein panel. After the full discovery analysis, the team locked a minimal signature of 15 proteins and assessed it in a non-overlapping UK Biobank Wave 3 MDD sample of 42 participants. The panel recovered the discovery-derived assignments in held-out data with an area under the curve of 0.997, and showed preliminary within-UKB, same-platform cross-wave concordance. A separate exploratory analysis used an independent plasma LC-MS/MS dataset from a Korean psychiatric cohort, comprising 38 participants with major depression and 40 healthy controls, to compare the axis at the pathway level across platforms. Cross-platform agreement was preliminary and assessed at the level of biological pathways rather than individual proteins, an honest acknowledgment that proteomic platforms measure overlapping but not identical sets of molecules.

The authors are careful about what these numbers do and do not mean. The cross-wave and cross-platform analyses are explicitly exploratory, the Wave 3 sample is small, and the LC-MS/MS comparison is pathway-level. The two operational groups are described as positions along a dominant continuous dimension rather than disease subtypes, and the near-perfect AUC of the 15-protein panel reflects within-cohort, same-platform recovery of assignments rather than a validated diagnostic test. Still, the conceptual shift is substantial. If the molecular biology of major depression is organized along a continuum, then research designs that search for discrete depressed subtypes may be chasing the wrong target, and treatment stratification may eventually depend on where a patient sits along an axis rather than which box they occupy.

For a field that has struggled to translate molecular findings into clinical tools, the study offers a methodological template as much as a biological claim. BioNet-SPACC demonstrates that protein-protein interaction priors can be woven into consensus clustering to produce molecular structure that is sparse, stable, and replicable in held-out data, providing what the authors describe as a biologically informed framework for reproducible molecular structure discovery in high-dimensional proteomic data. Whether the plasma proteomic axis identified here will prove useful for predicting treatment response, tracking disease course, or bridging peripheral biology and brain function remains an open question. But the demonstration that depression’s peripheral molecular signature is continuous, replicable, and reducible to a small protein panel gives researchers a concrete target, and a statistical toolkit, for the next round of investigation.

Subject of Research: Genetics

Article Title: Biologically informed network-regularized sparse consensus clustering (BioNet-SPACC) identifies a continuous plasma proteomic axis in major depressive disorder

Article References: Pan, H., Luo, G., Zhai, X., Xu, W., & Li, K. (2026). Biologically informed network-regularized sparse consensus clustering (BioNet-SPACC) identifies a continuous plasma proteomic axis in major depressive disorder. BMC Medicine. https://doi.org/10.1186/s12916-026-05344-3

Image Credits: AI Generated

DOI: 10.1186/s12916-026-05344-3

Keywords: Biologically, informed, network-regularized, sparse, consensus, clustering, BioNet-SPACC, identifies, continuous, plasma, proteomic, axis