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

Major depressive disorder (MDD) is clinically heterogeneous, and conventional case-control analyses may obscure continuous molecular variation. We developed biologically informed network-regularized sparse consensus clustering (BioNet-SPACC), integrating protein-protein interaction priors with sparse feature weighting in a leakage-controlled discovery and held-out test design. BioNet-SPACC was applied to 4,123 UK Biobank participants with MDD, divided into discovery and held-out test samples. Propensity-matched healthy controls (HCs) were introduced after MDD-only discovery for anchoring. Analyses assessed clustering stability, held-out concordance, and dimensional structure. A locked 15-protein panel was assessed in a non-overlapping UK Biobank Wave 3 MDD sample ( n = 42) for exploratory cross-wave concordance. An independent plasma LC-MS/MS dataset comprising 38 participants with MDD and 40 HCs was used for exploratory cross-platform comparison at the pathway level. BioNet-SPACC identified an internally stable K = 2 partition (median ARI = 0.986) defined by a Core 85 protein signature. The partition identified low-axis and high-axis operational groups positioned along a dominant continuous proteomic dimension. PC1 explained 91.3% of the variance in the Core 85 representation, and complementary distributional analyses provided no evidence against unimodality. In the held-out test sample, de novo clustering agreed with discovery-projected assignments (ARI = 0.789; NMI = 0.742; κ = 0.888), despite partial overlap in selected proteins (Jaccard = 0.417). CBC conditioning strongly preserved continuous molecular-axis scores within MDD (Pearson r = 0.940), whereas binary assignments showed moderate agreement with the primary solution. On the HC-standardized centroid-defined molecular-axis score, the overall MDD and HC means were similar (Cohen’s d = 0.01), whereas the low-axis and high-axis MDD groups occupied opposing positions relative to HCs. CBC and NMR analyses identified peripheral correlates spanning platelet-related hematological and metabolic traits. The 15-protein panel recovered discovery-derived assignments in held-out data (AUC = 0.997) and showed preliminary within-UKB, same-platform cross-wave concordance. These findings support a reproducible, predominantly continuous peripheral proteomic axis within MDD. BioNet-SPACC integrates PPI-informed regularization with sparse consensus clustering to provide a biologically informed framework for reproducible molecular structure discovery in high-dimensional proteomic data.