2026 ECNP Congress

PS03-3109

Differentiation of bipolar and unipolar depression using neuroimaging, immune-inflammatory and transcriptomic signatures: a machine learning approach

Tommaso Cazzella — F. Colombo, B. Bravi, D. Betelli, C. Di Carlo, C. Monopoli, S. Spadini, C. Lorenzi, R. Zanardi, C. Colombo, S. Poletti, F. Benedetti, B. Vai · Università Vita-Salute San Raffaele- Milan- Italy, Department of Psychology, Milan, Italy

2026 ECNP Congress

Introduction

The differential diagnosis between major depressive disorder (MDD) and bipolar disorder (BD) still relies largely on clinical observation, contributing to high misdiagnosis rates (Singh and Rajput, 2006). Evidence suggests that MDD and BD may differ in immune-inflammatory profiles, potentially reflecting altered gene-expression patterns and downstream effects on white-matter integrity (Benedetti et al., 2011; Poletti et al., 2021). In parallel, machine learning (ML) approaches trained on biological features emerged as promising tools for improving differential diagnosis in mood disorders (Colombo et al., 2022). In the present study, we assessed whether ML classifiers could distinguish MDD from BD across white matter fractional anisotropy (FA), a panel of cytokines and chemokines, and gene-expression data.

Methods

342 currently depressed inpatients (131 MDD, 211 BD) underwent 3-Tesla diffusion MRI. FA values were extracted according to JHU atlas using tract-based spatial statistics. At MRI scan timepoint, plasmatic concentrations of cytokines, chemokines and growth factors were quantified using Bio-Plex Pro™ Human Cytokine 48-Plex assay. After RNA extraction, gene expression data were obtained from library preparation and sequencing on Illumina NovaSeq6000_S2_100. For each feature, samples were split into 70% training and 30% test sets, balanced for age, sex, and diagnosis. Within the training set, elastic-net logistic regression models were trained with feature standardization and class balancing. Hyperparameters were optimized by 5-fold nested cross-validation. The confounding effect of age and sex was regressed out in each model. Statistical significance was assessed using 5000-permutation testing. Final model performance was evaluated in the independent test set, through balanced accuracy (BA), sensitivity, specificity, and area under the curve (AUC). Group comparisons between MDD and BD were performed for each feature within each modality and split, with false-discovery-rate (FDR) correction.

Results

S:All three modalities showed significant classification in the training set (all p<0.001). The 48-plex cytokine/chemokine model showed the highest training performance, with BA=70.95% and AUC=0.71 (72.97 sensitivity; 68.92 specificity). Gene-expression pathways showed a BA of 63.1% and an AUC of 0.63 (59.1% sensitivity; 67.11 specificity). Similarly, FA showed a BA of 62.68% and an AUC of 0.63 (66.67% sensitivity; 58.7 specificity). In the independent test sets, gene-expression pathways showed the best generalisation, reaching BA=79.27% and AUC=0.79 (73.68% sensitivity, 84.85% specificity). FA showed moderate test performance (BA=66.71%, AUC=0.67, sensitivity=71.88%, specificity=61.54%), whereas the 48-plex model reached BA=60.94% and AUC=0.61 (sensitivity: 53.12%, specificity: 68.75%). Univariate analyses identified FDR-significant diagnostic differences across modalities, including higher FA in BD (t range: -3.25 to -2.6), higher inflammatory/angiogenic and immune-cell recruitment cytokine/chemokine markers in MDD (t range: 1.98 to 5.22), and higher immune/inflammatory and stress-response gene-expression pathway scores in BD (t range: -4.86 to -2.31).

Conclusions

Elastic-net models differentiated MDD from BD among patients assessed during a depressive episode. Although cytokine/chemokine profiles showed the highest training performance, gene-expression pathways outperformed other modalities in the independent test set, suggesting stronger generalizability to unseen data and potential biomarker relevance. Our findings support the utility of ML for improving differential diagnosis in mood disorders. If replicated in independent cohorts, these biological signatures may also highlight druggable pathways for tailored therapeutic strategies.

References

REFERENCES Benedetti, F., Absinta, M., Rocca, M.A., Radaelli, D., Poletti, S., Bernasconi, A., Dallaspezia, S., Pagani, E., Falini, A., Copetti, M., Colombo, C., Comi, G., Smeraldi, E. and Filippi, M. (2011), Tract-specific white matter structural disruption in patients with bipolar disorder. Bipolar Disorders, 13: 414-424. https://doi.org/10.1111/j.1399-5618.2011.00938.x Colombo, F., Calesella, F., Mazza, M. G., Melloni, E. M. T., Morelli, M. J., Scotti, G. M., Benedetti, F., Bollettini, I., & Vai, B. (2022). Machine learning approaches for prediction of bipolar disorder based on biological, clinical and neuropsychological markers: A systematic review and meta-analysis. Neuroscience & Biobehavioral Reviews, 135, 104552. https://doi.org/10.1016/j.neubiorev.2022.104552 Poletti, S., Vai, B., Mazza, M. G., Zanardi, R., Lorenzi, C., Calesella, F., Cazzetta, S., Branchi, I., Colombo, C., Furlan, R., & Benedetti, F. (2021). A peripheral inflammatory signature discriminates bipolar from unipolar depression: A machine learning approach. Progress in Neuro-Psychopharmacology and Biological Psychiatry, 105, 110136. https://doi.org/10.1016/j.pnpbp.2020.110136 Singh, T., & Rajput, M. (2006). Misdiagnosis of bipolar disorder. Psychiatry (Edgmont), 3(10), 57–63.