Development of machine-learning-driven signatures for diagnosing and monitoring therapeutic response in major depressive disorder using integrated immune cell profiles and plasma cytokines

<b>Background:</b> Diagnosis and treatment efficacy of major depressive disorder (MDD) currently lack stable and reliable biomarkers. Previous research has suggested a potential association between immune cells, cytokines, and the pathophysiology and treatment of MDD. <b>Objective:</b> This study aims to investigate the relationship between immune cells, cytokines, and the diagnosis of MDD and treatment response, further utilizing machine learning algorithms to develop robust diagnostic and treatment response prediction models. <b>Methods:</b> Using mass cytometry by time-of-flight (CyTOF) technology and high-throughput cytokine detection, we analyzed 63 types of immune cells from 134 pre-treatment MDD patients. Among these patients, plasma data for 440 cytokines were obtained from 84 individuals. Additionally, we conducted the same set of immune cell and cytokine analyses on 50 healthy controls (HC). An 8-week follow-up was conducted to observe post-treatment changes in immune cells and cytokines. <b>Results:</b> By combing eight machine-learning algorithms with CyTOF and cytokine data, we constructed a diagnostic model for MDD patient with 16 indicators, achieving an AUC of 0.973 in the internal validation set. Additionally, a treatment response prediction model based 7 cytokines was developed, resulting in an AUC of 0.944 in the internal validation set. Furthermore, Mfuzz time-series analysis revealed that cytokines such as Basic fibroblast growth factor (bFGF), Interleukin 13 (IL-13), and Interleukin 1 receptor, type I (IL1R1) that revert towards normal levels after 8 weeks of treatment, suggesting their potential as therapeutic targets for MDD. <b>Conclusions:</b> Our diagnostic model derived from CyTOF and cytokines demonstrates high diagnostic value. However, relying solely on immune cells may not provide optimal predictions for antidepressant treatment response. In contrast, leveraging cytokines has proven valuable, leading to the construction of a seven-factor treatment response prediction model. Importantly, we observed that several significantly altered cytokines in MDD can normalize following antidepressant treatment, indicating their potential as therapeutic targets.