Multi-axis metabolic 3D modelling identifies a distinct phenotype in double negative demyelinating disease

Abstract Background and objectives Accurate discrimination of inflammatory demyelinating diseases remains clinically challenging, particularly in seronegative cases where conventional diagnostic biomarkers are absent. Existing metabolomic studies have largely relied on pairwise classification approaches, which do not capture shared and disease-specific metabolic components across related disorders. Here, we investigated whether integrated metabolomics and lipidomics combined with a multi-axis modelling framework could resolve disease-specific and overlapping metabolic structure across inflammatory demyelinating diseases, including double-negative cases (DN). Methods Serum samples from relapsing-remitting multiple sclerosis (RRMS, n = 29), secondary progressive MS (SPMS, n = 20), primary progressive MS (PPMS, n = 4), aquaporin-4 antibody-positive neuromyelitis optica spectrum disorder (AQP4-NMOSD; n = 49), myelin oligodendrocyte glycoprotein antibody-associated disease (MOGAD; n = 18), and DN cases ( n = 41) were analysed using nuclear magnetic resonance (NMR) metabolomics and liquid chromatography-mass spectrometry (LC-MS) lipidomics. Independent supervised models were trained, and outputs were transformed to logit probabilities (P_MOGAD, P_AQP4-NMOSD, P_SPMS) to construct a three-dimensional metabolic disease space, into which all individuals were projected. Results The combined model showed strong discrimination across disease groups (MS vs. antibody-mediated disease AUC = 0.938; RRMS vs. NMOSD AUC = 0.962; RRMS vs. MOGAD AUC = 0.920), with lower separation for RRMS vs. SPMS (AUC = 0.887). In the two-axis space, RRMS, NMOSD, and MOGAD formed clearly distinct regions, while DN and SPMS cases associated with the MOGAD axis. Adding a third SPMS axis separated SPMS from DN and captured an additional progression-related component. The position in this 3D space reflects how closely individuals aligned with different disease components rather than forcing discrete classification. Key discriminatory features included myo-inositol, glucose, histidine, and lipid-related measures (phosphatidylcholines, high-density lipoproteins), with consistent differences between MS and antibody-mediated disease. Discussion Multi-axis probabilistic modelling defines a continuous metabolic disease space that captures both shared and disease-specific components across inflammatory demyelinating disorders. Rather than enforcing discrete classification, this framework resolves overlapping and clinically ambiguous phenotypes, improves separation of progressive and seronegative cases, and reveals DN as a heterogeneous entity that is metabolically distinct from MS while demonstrating closer proximity to MOGAD. The compatibility of this analytical approach with clinically deployable metabolomics platforms highlights its potential utility for improving disease classification and patient stratification in inflammatory demyelinating disorders.