**4. Discussion**

In this pilot study, longitudinal analysis with a small sample size (N = 10) may estimate biased variation and cause inflation of type I error. Statistical techniques using a small sample number have been developed in psychiatry for circumstances wherein sample collection is not easy [50]. To discuss this part, we built a GEE model in addition to the LMM model. Compared to LMM, GEE showed greater statistical validity with a devised variance estimate even in a small number of samples [18]. We applied one of them, the Wang and Long method [51], modified bias-correction and efficiency improvement. Consequently, we found seven significant proteins for the time/response interaction term (adjusted *p*-value < 0.05; Wald test and corrected by SGoF; Supplementary Table S6), overlapping with five (AHSG, IL6ST, APOD, PHOX2B, and SHBG) of 37 significant proteins found in LMM. Unlike a GEE, which is a population level-based model and relatively easy to compute, LMM can consider random effects with technical or biological variation obtained from the whole data, and thus, we considered the LMM technique more suitable than GEEs [15,50,52].

Regarding protein biomarker candidates, we identified 37 plasma proteins significantly associated with MDD by the LMM analysis, and these protein biomarkers could be biologically or physiologically divided into four functional categories through the literature search on PubTabor central [40]. First, six plasma markers, GC, LCN2, ITIH3, VWF, PHOX2B and YWHAE, were previously reported to be associated with SSRI efficacy, with the abundances of GC, LCN2, and ITIH3 in plasma samples associated with response to SSRIs [21,53–55]. Among them, PHOX2B and YWHAE were validated by LC-MRM/MS analysis in this study. SSRI stress in human brain cells was reported to be associated with a PHOX2B transcription factor [56]. The YWHAE genes have been reported to play a significant role in MDD in the Han Chinese population, with alterations in their protein–protein interactions [57]. The second category is that sex hormones, neurotransmitters, and related proteins have been strongly associated with depression [58]. In this study, plasma SHBG, which has been previously linked to depression [59–61], showed the sharpest difference over time between responders and non-responders (adjusted *p*-value = 2.70 <sup>×</sup> 10-3), decreasing gradually with time in responders and increasing gradually with time in non-responders. We also found that plasma dopamine beta-hydroxylase (DBH) concentrations increased in responders from 4 to 10 weeks, consistent with low plasma DBH levels associated with low activity of the noradrenaline system in patients with depression [62–66]. Third, we found that the plasma proteins GSN and C4B were biomarkers of depression, similar to findings in previous studies using the same LC-MS platform [67–69]. The level of C4B was significantly higher in responders than in non-responders and showed a significant change over time. Subsequently, APOD, PON1, BCHE, and IL6ST were reported to be related to depression, a finding consistent with our results [21,70–74]. Finally, APOE, CSTD and MMP2 that varied genetically and in mRNA level of abundance were reported to be associated with depression. We found that the levels of abundance of two of these proteins, APOE and CSTD, with single nucleotide polymorphisms (SNPs) differed significantly in responders and non-responders [75–78]. The expression of the MMP2 gene in the brain was associated with recurrence of depression [55].

Moreover, 11 plasma proteins that strongly correlated with two or more psychiatric indexes were related with neurological mechanisms and SSRI response. EXT1 was involved in the biosynthesis of heparan sulfate, which played an important role in the development of the nervous systems in the brain, and its deletion caused autism-like behavior in mice [79,80]. NUCB1 is known as a Golgi-resident marker of neurons [81] and interrupts amyloid fibrillation in the brain [82]. PROS1 turned out to be a novel Aβ-responsive protein based on proteome profiling of the hippocampus in the 5XFAD mouse model [83]. LYVE1 was the upregulated gene expressed in SSRI responders to non-responders [84], and differential LYVE1 and MHC II expression was used to identify CNS border-associated macrophages in single cell experiments [85]. ATRN, a neuroprotectant [86], was high in the SSRI responder in blood proteins. CD109 was higher in the disease group in the plasma proteome comparison between the psychotic disorder and the normal group [87].

This preliminary retrospective study had several limitations, including its small sample size, the lack of racial diversity among the study subjects, and the collection of plasma samples at a single center. Proteomics studies using small specimens are frequent. Typically, 10–50 samples are used during the preclinical discovery and validation phase, given the analysis of large data sets and limited timelines [88]. Thus, our results should be considered preliminary findings. All participants were Korean population, and our results may not be generalizable to other ethnic groups. In addition, plasma protein abundance may be affected by the plasma preparation method [89,90], but plasma was rapidly prepared from blood and stored frozen at −80 ◦C to avoid any pre-analytical effect [91–93]. Moreover, alterations in plasma protein abundance may be dependent on the SSRI type and dosage. In this study, we used samples treated with the same antidepressant (escitalopram) in a relatively certain range of doses, and this may be a limitation in using the results of this study to predict treatment responses with other antidepressants. However, this can be considered the strength of this study. As it is clinically difficult to collect plasma samples using the same type and dose of antidepressants for patients with major depressive disorders in a prospective design, so far, most studies on protein biomarkers for antidepressant treatment response have not been able to control the types or doses of antidepressants [69]. In the view of personalized treatment, predicting whether an individual with depression will benefit from a particular antidepressant is critical in choosing the right antidepressant; furthermore, how different types of antidepressants affect plasma proteins should be considered. In this study, the use of samples treated with the same antidepressant (escitalopram) in a relatively certain range of doses is considered a strength of this study. A controlled prospective study with a large

sample size is necessary to establish a clear differential influence of several types of antidepressants on plasma proteins. Therefore, prospective studies in larger patient cohorts are needed to validate our findings.
