Plasma Sphingomyelin and Ceramide Profiles as Potential Biomarkers in Young Adults with Bipolar Depression: A Lipidomics Study

Abstract

Background

Bipolar depression (BPD) lacks reliable peripheral biomarkers for early diagnosis. Sphingomyelin (SM) and ceramide (Cer) are bioactive lipids implicated in neuronal membrane signaling, yet their diagnostic utility in young adults with BPD remains unclear.

Patients and Methods

We conducted a single-center exploratory untargeted lipidomics study focusing on SM and Cer species in plasma from 26 young adults (18–26 years) with BPD at baseline and after three weeks of quetiapine and magnesium valproate treatment (BPD3w), alongside 27 sex- and age-matched healthy controls (HC). Group comparisons of total SM, total Cer, and SM/Cer ratio were performed; ROC curves assessed discriminative ability. A Venn diagram approach, followed by LASSO regression with bootstrap averaging, identified a refined candidate biomarker panel from 140 SM/Cer species.

Results

Both BPD and BPD3w groups exhibited elevated total SM and Cer relative to HC, whereas the SM/Cer ratio was significantly reduced. Paired analysis revealed that total SM and the SM/Cer ratio decreased after treatment (both P < 0.05), while total Cer remained unchanged. Intersection analysis yielded 14 candidate lipids, which were further refined via LASSO regression (λ = 100) with bootstrap averaging to a 4-species panel: Cer(t20:1/24:3), Cer(d19:1/25:3), SM(t18:0/22:0), and Cer(t17:0/22:0), with an AUC of 0.975 (95% CI: 0.917–1.000), representing an unvalidated internally derived estimate pending further validation. Correlation analysis revealed that Cer(d19:1/25:3), Cer(t17:0/22:0), and SM(t18:0/22:0) were positively correlated with depressive, anxiety, and manic symptom severity (HAMD, HAMA, and YMRS). Cer(t20:1/24:3) showed a modest negatively correlation with HAMD, but not with HAMA or YMRS. Additionally, Cer(t20:1/24:3) and SM(t18:0/22:0) exhibited sex-dependent differences.

Conclusion

Plasma SM and Cer profiles are altered in young adults with BPD and show a partial trend toward normalization following short-term treatment. A refined 4-species panel showed promising but internally derived discriminative performance in this exploratory study. While peripheral sphingolipid signatures may serve as potential biomarkers for BPD, these findings require validation in independent cohorts before any clinical application.

Introduction

Bipolar disorder (BD) is a severe, chronic, and recurrent psychiatric illness affecting approximately 1–2% of the global population.¹ Characterized by alternating manic/hypomanic and depressive episodes, BD severely impairs daily functioning and represents a leading cause of disability and premature mortality worldwide.^2,3 The disorder typically manifests during adolescence or young adulthood, a critical neurodevelopmental window when long-term psychosocial trajectories are established,^4,5 and exhibits markedly heterogeneous clinical presentations across individuals and lifespans.⁶ Depressive episodes dominate the course of BD. More than 50% of patients experience depression as their initial presenting symptom.^7,8 In addition, among the bipolar spectrum, bipolar depression (BPD) is the leading contributor to disease-related morbidity, and patients with bipolar depression demonstrate significantly higher rates of functional impairment, psychosocial disability, and lost work productivity compared to those with bipolar mania.⁹ Longitudinal studies indicate that patients spend substantially more time in depressive states (approximately 34% of the illness course) than in manic or mixed states (approximately 12%), with depressive symptoms accounting for the majority of subthreshold symptomatic periods.^10,11 Furthermore, approximately 25–30% of individuals with BPD develop treatment-resistant forms, characterized by severe, persistent depressive symptoms that do not respond to standard pharmacological interventions, thereby exacerbating the cumulative disease burden.^12,13 Critically, current diagnosis still relies primarily on clinical interviews and phenomenological criteria, with a lack of objective laboratory-based biomarkers to support decision-making—highlighting the urgent need to develop reliable peripheral biomarkers to facilitate diagnostic assessment and precision intervention in young patients with bipolar depression.

Sphingomyelin (SM) is the most abundant eukaryotic sphingolipid and a major component of the plasma membrane.¹⁴ Naturally occurring SM species vary in fatty acid chain length (typically 16–24 carbons) and degree of saturation/hydrogenation, and may contain saturated (SFAs), monounsaturated (MUFAs), or polyunsaturated (PUFAs) fatty acids.¹⁵ SM is essential for maintaining membrane fluidity and cholesterol binding. Hydrolysis of SM by sphingomyelinases (SMases) increases the concentration of ceramide (Cer), a bioactive molecule involved in cellular proliferation, growth, apoptosis, and neuronal differentiation.^16,17 Accumulating evidence indicates that dysregulation of Cer or SM metabolism is associated with several neurological and psychiatric disorders, including Alzheimer’s disease, depression, schizophrenia, and BD,^18,19 and the Cer/SM pathway has been proposed as a therapeutic target for these conditions.^20,21 Moreover, plasma Cer levels have been shown to correlate with depressive severity,²² and administration of imipramine—a tricyclic antidepressant known to inhibit acid sphingomyelinase (ASMase)—significantly reduces hypoglycemia-induced neuronal death and improves cognitive function.²³ These findings collectively suggest that the peripheral Cer/SM may reflect pathophysiological processes relevant to mood regulation and could serve as a quantifiable biomarker for depressive states. In recent years, lipidomics has been widely applied to biomarker discovery in psychiatric disorders. In the field of BD, however, most previous studies have focused on manic or euthymic states without independent analyses of depressive episodes. Consequently, lipidomic investigations specifically targeting patients with BPD remain scarce. In particular, the characteristic alterations in peripheral SM/Cer profiles and their dynamic changes following pharmacotherapy in this population are currently unclear.

Considering the above, we performed an exploratory untargeted lipidomic analysis focusing on plasma SM and Cer species using liquid chromatography-mass spectrometry (LC-MS/MS). Plasma samples were collected from young adults with bipolar depression (BPD, n = 26) at baseline (BD) and after three weeks of full-dose pharmacological treatment (BD3w), as well as from sex- and age-matched healthy controls (HC, n = 27). We first compared the relative levels of total SM, total Cer, and the SM/Cer ratio among the three groups (BPD, BPD3w, and HC) and evaluated their discriminative ability using receiver operating characteristic (ROC) curves. Volcano plot analysis was then employed to screen approximately 140 individual SM and Cer species for differential abundance patterns. Candidate species were identified through Venn diagram intersection of differential lipids from two comparisons (BPD vs HC and BPD vs BPD3w). Furthermore, we applied LASSO regression with 100-fold bootstrap averaging to refine the candidate panel and examined the correlations between identified lipid species and clinical symptom severity. The clinical application of this study is positioned as: (1) exploring SM/Cer as an auxiliary screening marker for distinguishing BPD patients from healthy individuals; and (2) evaluating the dynamic changes of these lipids during pharmacotherapy as potential monitoring indicators of treatment response. Differentiation between BPD and MDD is beyond the scope of this study and requires an independent cohort with MDD controls. The decrease in AUC from 0.997 (14-species panel) to 0.975 (4-species panel) reflects the effect of bootstrap averaging, which reduces overfitting and provides a more conservative estimate of model performance.

Materials and Methods

Sample Size Calculation

Sample size was estimated using G*Power software (version 3.1.9.7).²⁴ Given that the primary comparison is between the bipolar depression (BD) group and healthy controls (HC) using an independent-samples t-test, we adopted two complementary strategies to justify the sample size. First, based on a previous lipidomics study in BD that reported a large effect size for lipid species in discriminating patients from controls,²⁵ we considered a very large effect size scenario (Cohen’s d > 3.0). With such a very large effect size, a minimal sample size would be sufficient to achieve 80% power. A Cohen’s d of 0.80 was chosen as a conservative and commonly adopted effect size in psychiatric lipidomic studies, independent of our own results. Assuming a two-sided test, α = 0.05, power (1-β) = 0.80, and an allocation ratio of 1:1, the required sample size was calculated to be 26 participants per group. The final sample comprised 26 BPD patients (with baseline and week-3 follow-up assessments) and 27 HC, satisfying the estimated statistical requirement.

Subjects and Plasma Sampling

Patients who met the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5) criteria for bipolar depression, aged 18–26 years, were recruited from the General Hospital of Ningxia Medical University, Yinchuan, China. Twenty-nine healthy individuals, aged 19–27 years and matched for sex and age distribution, were enrolled as controls. The enrollment age for this study was set at 18–29 years, based on the following considerations: this age range represents the peak period of onset for bipolar disorder, and enrolled patients within this range typically have a relatively short illness duration and limited prior medication exposure. This approach minimizes potential confounding effects of chronic disease course and long-term polypharmacy on plasma lipid metabolism. During the actual recruitment process, the age range of enrolled BPD patients was 18–26 years, and that of HC participants was 18–27 years. Both groups were predominantly composed of young adults, and the minor asymmetry in age ranges reflects the natural distribution of the recruited sample rather than a systematic selection bias. All patients were drug-naive at baseline, defined as no prior exposure to antipsychotics, mood stabilizers, or antidepressants before enrollment. Patients in the BPD group were treated with quetiapine and magnesium valproate for 8 weeks. The 3-week follow-up time point was selected based on the known onset of action of quetiapine and magnesium valproate, which typically requires 2–4 weeks to achieve therapeutic effects. Specifically, quetiapine was initiated at 50 mg/d and titrated to a maintenance dose of 200–300 mg/d (administered orally once daily) within 1–2 weeks based on clinical response and tolerability. Magnesium valproate was initiated at 250 mg/d and titrated to a maintenance dose of 500 mg/d (administered orally in two divided doses). No other psychotropic medications were permitted during the study period. Detailed dosing information is summarized in Supplementary Table 1. Adherence was monitored through self-report and clinical assessment at each visit. No treatment changes were made during the 3-week follow-up period. Screening for pre-existing mental disorders and the diagnosis of bipolar depression were conducted by two senior psychiatrists using the Mini International Neuropsychiatric Interview. Clinicians specialized in psychiatry, who were masked to lipidomic profiling data, assessed clinical symptoms via the 17-item Hamilton Depression Rating Scale (HAM-D), Young Mania Rating Scale (YMRS) and Hamilton Anxiety Scale (HAM-A). The exclusion criteria included: (i) presence of other mental disorders according to DSM-5 criteria; (ii) pregnancy, lactation, or menstrual period; (iii) hypertension (systolic blood pressure ≥ 140 mmHg and/or diastolic blood pressure ≥ 90 mmHg); (iv) other organic diseases affecting lipid metabolism (eg, diabetes mellitus, thyroid dysfunction); (v) obesity (body mass index ≥ 28.0 kg/m²); and (vi) high-fat diet or long-term vegetarianism. This study was conducted in accordance with the Declaration of Helsinki and was registered with the Chinese Clinical Trial Registry (ChiCTR2400089218; registered 4 September 2024) and approved by the Ethics Committee of the General Hospital of Ningxia Medical University (approval number: KYLL-2024–0815). Written informed consent was obtained from all study participants before recruitment. Venous blood specimens were harvested after an overnight fast between 08:00 and 10:00 hours each morning to minimize circadian variation in lipid metabolism. Samples were immediately centrifuged at 1600 rpm for 15 minutes at 4°C, and plasma was aliquoted and stored in liquid nitrogen until lipidomic analysis. A total of 29 healthy controls were enrolled, of whom two were excluded from analysis due to unqualified blood samples, yielding a final HC group of 27 participants.

Lipidomic Analyses

Lipidomic analyses were performed by Majorbio Bio-Pharm Technology Co., Ltd. (Shanghai, China) using an established protocol.^26,27 For lipid extraction, 200 μL of plasma was precisely pipetted into an Eppendorf tube, followed by sequential addition of 80 μL methanol and 400 μL methyl tert-butyl ether (MTBE). The mixture was vortexed for 30 seconds and subjected to ultrasonic extraction at 5°C (40 kHz) for 30 minutes. Samples were then incubated at −20°C for 30 minutes and centrifuged at 13,000 × g for 15 minutes at 4°C. The supernatant (350 μL) was transferred and dried in a vacuum concentrator. The dried lipid extract was reconstituted in 100 μL of isopropanol:acetonitrile (1:1, v/v), vortexed for 30 seconds, and sonicated in an ice-water bath (40 kHz) for 5 minutes. Following centrifugation at 13,000 × g for 15 minutes at 4°C, the cleared supernatant was transferred to sample vials for LC-MS/MS analysis. Pooled quality control samples (QC, n = 6) were prepared and tested in the same manner as the analytic samples, which were injected at regular intervals (every 10 samples) to monitor the stability of the analysis. The injection order was arranged to alternate between the BPD and HC groups. Lipid separation was performed on a Thermo Scientific Vanquish Horizon UHPLC system equipped with an Accucore C30 column (100 mm × 2.1 mm i.d., 2.6 μm; Thermo Fisher Scientific, Waltham, MA, USA). Mass spectrometric detection was conducted using a Q Exactive HF-X hybrid quadrupole-Orbitrap mass spectrometer equipped with a heated electrospray ionization (HESI) source operating in both positive and negative ionization modes. Data-dependent acquisition mode was employed to acquire MS/MS spectra for confident lipid identification. The mass detection range was set at m/z 200–2000. Raw data were imported into LipidSearch software (version 4.2; Thermo Fisher Scientific) for peak alignment, feature detection, and lipid identification based on MS/MS fragmentation patterns.^28,29 Lipidomic features detected in at least 80% of samples within any experimental group were retained. Missing values were imputed using the minimum observed value for the respective metabolite. Each lipid feature was normalized to the total ion sum to account for technical variation. No internal standards were used in this study, and the lipidomic data are presented as relative quantification rather than absolute concentrations. Data preprocessing and statistical analyses were conducted using the Majorbio Cloud Platform (cloud.majorbio.com). Lipid species are annotated according to the LIPID MAPS shorthand notation. The prefix “d” denotes a 1,3-dihydroxy long-chain base, and the prefix “t” denotes a 1,3,4-trihydroxy long-chain base (phytosphingosine).

Statistical Analyses

GraphPad Prism 9.0 was used for plotting and analyzing data. Categorical variables (sex) were analyzed using the chi-squared test. Continuous variables (age) were compared using the Mann–Whitney U-test. Normality of distribution was assessed using the Shapiro–Wilk test. For normally distributed continuous variables, one-way analysis of variance (ANOVA) was employed for omnibus comparisons across the three groups (HC, BPD, BPD3w). When significant main effects were detected, post-hoc pairwise comparisons were conducted with Bonferroni correction. For non-parametric data, the Kruskal–Wallis H-test was used, followed by Dunn-Bonferroni pairwise comparisons. Paired comparisons between BPD and BPD3w were analyzed using two-tailed paired t-tests for normally distributed variables or Wilcoxon signed-rank tests for non-parametric variables. Differentially abundant lipids were identified primarily based on a Benjamini-Hochberg FDR-adjusted q < 0.05 and VIP > 1, with fold change reported as a descriptive measure rather than a strict screening criterion. VIP scores were calculated using the orthogonal partial least squares discriminant analysis (OPLS-DA) model implemented in the Majorbio Cloud Platform (https://cloud.majorbio.com). A Venn diagram was employed to identify common and unique lipids across different comparisons and to establish a diagnostic lipid set. Subsequently, a multivariate classifier model based on a logistic regression model with LASSO regularization was applied to establish an optimal diagnostic lipid panel. The optimal regularization parameter λ was selected via 100-fold bootstrap resampling. The LASSO model was implemented using the Majorbio Cloud Platform. The LASSO model was implemented using the Majorbio Cloud Platform, with the regularization parameter λ selected via 100-fold bootstrap resampling. The 14-lipid panel score was computed as the sum of the normalized intensities of the 14 selected lipid species, weighted by their LASSO regression coefficients. The relationships of lipid concentrations to clinical parameters were explored via Spearman rank correlation test. The Area Under Curve (AUC) value of the combination of lipids was evaluated by receiver operating characteristic (ROC) analysis. All lipidomic statistical analyses, including differential lipid screening, volcano plot generation, Venn diagram intersection, and Spearman correlation tests, were performed using the Majorbio Cloud Platform (https://cloud.majorbio.com).

Results

Comparison of Clinical Characteristics of the Recruited Participants in Each Group

A total of 55 participants were screened (26 BPD patients and 29 healthy controls). Two healthy controls were excluded due to unqualified blood samples. The final analytical sample consisted of 53 participants, including 26 BPD patients and 27 healthy controls. No significant difference was found between HC and BPD groups in age (U = 250.5, P = 0.072), sex distribution (χ² = 0.058, P = 0.809) or body mass index (BMI; t = 1.291, P = 0.203). Meanwhile, the scale score of HAMD, HAMA, and YMRS in the BPD group was significantly higher than that of the HC group. Paired analysis further revealed that all three scale scores were significantly decreased after treatment (BPD vs BPD3w, P < 0.01) (Table 1), indicating that three weeks of pharmacological intervention can alleviate the clinical symptoms of BPD patients. The detailed clinical and demographic data are displayed in Supplementary Table 2. After 3 weeks of treatment, the response rate (≥50% reduction in HAMD) was 19.2% (5/26), and the remission rate (HAMD ≤ 7) was 0% (0/26).

Table 1.

Comparison of Clinical Characteristics Data and Symptom Scale Assessment Between the HC, BPD and BPD3w Groups

Parameter	HC (n = 27)	BPD (n = 26)	BPD3w (n = 26)	Statistical value	P-value
Age [years, M (P25, P75)]a	22 (20, 24.5)	20 (19, 24)	_	U = 250.5	0.072
Sex (male/female)b	7/20	6/20	_	χ2 = 0.058	0.809
BMIc	19.46 ± 4.440	21.29 ± 3.548	_	t = 1.291	0.203
HAMD scored	5.963 ± 1.990	44.81 ± 8.690**	25.12 ± 6.212 **,##	F = 257.0	< 0.001
HAMA scored	5.148 ± 2.230	36.08 ± 5.433**	20.46 ± 5.217**,##	F = 348.5	< 0.001
YMRS scored	2.185 ± 1.415	14.58 ± 4.588**	9.615 ± 3.699**,##	F = 85.27	< 0.001

Notes: Scale scores are shown as mean ± SD or M (P25, P75); ^aMann–Whitney U-test (HC vs BPD); ^bChi-square test (HC vs BPD); ^cUnpaired-t-test (HC vs BPD); ^dOne-way ANOVA for three-group comparison (HC, BPD, and BPD3w). Note that BPD3w represents the same patients as BPD at 3-week follow-up; the primary BPD vs BPD3w comparisons were analyzed using paired tests (paired t-test or Wilcoxon signed-rank test). For age, sex, and BMI, only HC and BPD were compared because BPD3w are the same patients as BPD at 3-week follow-up. **P < 0.01 vs HC; ^##P < 0.01 vs BPD.

Abbreviation: SD, standard deviation.

Comparison of SM, Cer and SM/Cer Among Groups

Significant differences in relative SM levels were observed among the HC, BPD, and BPD3w groups (F₂, ₇₆ = 9.682, P < 0.001). Post hoc tests revealed that both the BPD and BPD3w groups had significantly higher SM levels than the HC group. Furthermore, paired Wilcoxon analysis showed that SM levels were significantly lower in the BPD3w group than in the BPD group (P = 0.027) (Figure 1A). ROC analysis demonstrated that SM levels effectively distinguished BPD from HC (AUC = 0.837, 95% CI: 0.727–0.948), but did not perform well in distinguishing BPD3w from HC (AUC = 0.709, 95% CI: 0.566–0.852) or BPD3w from BPD (AUC = 0.633, 95% CI: 0.482–0.785) (Figure 1B). Significant differences in relative Cer levels were also observed among the three groups (H = 28.78, P < 0.001). Multiple comparisons indicated that both the BPD and BPD3w groups had significantly higher Cer levels than the HC group, whereas no significant difference was found between the BPD3w and BPD groups (P = 0.980) (Figure 1C). ROC analysis showed that Cer levels effectively distinguished BPD from HC (AUC = 0.880, 95% CI: 0.789–0.972) and BPD3w from HC (AUC = 0.859, 95% CI: 0.761–0.957), but failed to distinguish BPD3w from BPD (AUC = 0.516, 95% CI: 0.355–0.677) (Figure 1D). Furthermore, significant differences in the SM/Cer ratio were observed among the three groups (H = 21.74, P < 0.001). Multiple comparisons revealed that both the BPD and BPD3w groups had significantly lower SM/Cer ratios than the HC group, and a significant difference was also observed between the BPD3w and BPD groups (P = 0.018) (Figure 1E). ROC analysis indicated that the SM/Cer ratio effectively distinguished BPD3w from HC (AUC = 0.849, 95% CI: 0.746–0.952), but did not effectively distinguish BPD from HC (AUC = 0.724, 95% CI: 0.585–0.863) or BPD from BPD3w (AUC = 0.694, 95% CI: 0.546–0.842) (Figure 1F). Together, these results suggest that SM and Cer are significantly up-regulated in BPD, while SM levels and the SM/Cer ratio respond to drug treatment. ROC analyses further support their potential utility as diagnostic biomarkers for BPD and as indicators of treatment response.

Figure 1.

Scatter plots and receiver operating characteristic (ROC) curves.

Notes: (A, C and E) are scatter plots illustrating the differences in SM, Cer and SM/Cer levels among the HC (n = 27), BPD (n = 26) and BPD3w (n = 26) groups, with error bars representing the standard error of the mean (SEM) for each group. Black, blue and red circles each denote an individual value from the HC, BPD and BPD3w groups, respectively. (B, D and F) present receiver operating characteristic (ROC) curves for evaluating the diagnostic performance of SM, Cer and SM/Cer. All ROC curves represent apparent performance without cross-validation or external validation. The red, blue and black lines correspond to comparisons of BPD vs HC, BPD3w vs HC, and BPD vs BPD3w, separately. **P < 0.01, *P < 0.05. BPD refers to bipolar depression patients at baseline, and BPD3w represents patients with bipolar depression after 3 weeks of treatment.

Identification of Differential SM and Cer Species Among Groups

A total of 140 SM (57) and Cer (83) species were identified in the samples of each group. There were 66 differential species between BPD and HC group (63 up-regulated and 3 down-regulated in BPD, Supplementary Table 3), 36 differential species between BPD and BPD3w group (22 up-regulated and 14 down-regulated in BPD, Supplementary Table 4), and 63 differential species between HC and BPD3w group (58 up-regulated and 5 down-regulated in BPD3w, Supplementary Table 5). To further refine the lipid signature specific to BPD, we performed a Venn diagram analysis intersecting the differential species from two comparisons (BPD vs HC and BPD vs BPD3w) which we defined as BPD specific set. As shown in Figure 2A, 14 candidate species showed differential changes in both comparisons, with an internally derived AUC of 0.997 (95% CI: 0.995–0.998); however, this estimate is unvalidated and likely overfitted given the small sample size (Figure 2B).

Figure 2.

Venn diagrams and receiver operating characteristic (ROC) curves.

Notes: (A) shows overlapping and unique lipids identified in comparisons of BPD vs HC and BPD vs BPD3w. (B) presents the ROC curve for the 14-species panel, with an apparent AUC of 0.997 (95% CI: 0.995–0.998). This curve represents apparent performance without cross-validation or external validation. This estimate is unvalidated and likely overfitted given the small sample size. The subsequent LASSO regression with bootstrap averaging yielded a more conservative AUC of 0.975 (95% CI: 0.917–1.000) for the 4-species panel.

Abbreviations: HC, healthy controls; BPD, patients with bipolar depression at baseline; BPD3w, patients with bipolar depression after 3 weeks of intervention; AUC, area under the curve; LASSO, least absolute shrinkage and selection operator.

There were significant differences in the relative contents of Cer (d16:0/18:2) (H = 10.24, P = 0.006; Figure 3A), Cer (t18:1/22:0) (H = 30.99, P < 0.001; Figure 3B), Cer (t20:1/24:3) (H = 26.75, P < 0.001; Figure 3C), Cer (t17:0/22:0) (H = 28.70, P < 0.001; Figure 3D), Cer (d18:0/23:0) (H = 14.27, P < 0.001; Figure 3E), Cer (d18:0/24:0) (H = 13.48, P = 0.0012; Figure 3F), Cer (d19:1/25:3) (H = 51.92, P < 0.001; Figure 3G), Cer (d18:2/25:0) (H = 13.20, P = 0.0014; Figure 3H), Cer (d18:1/25:0) (H = 11.87, P = 0.0026; Figure 3I), Cer (t17:0/23:0+O) (H = 17.75, P < 0.001; Figure 3J), SM (t18:1/24:0) (H = 28.21, P < 0.001; Figure 3K), SM (t18:0/22:0) (H = 36.50, P < 0.001; Figure 3L), SM (d18:2/22:1) (H = 34.14, P < 0.001; Figure 3M), and SM (d20:0/24:2) (H = 26.48, P < 0.001; Figure 3N) among the three groups. Among these lipids, the relative contents of Cer(d16:0/18:2) were decreased in BPD (BPD vs HC), and it increased after 3 weeks of treatment (BPD vs BPD3w). Cer(t17:0/22:0) was increased in BPD (BPD vs HC), and it decreased after 3 weeks of treatment (BPD vs BPD3w). Moreover, the relative contents of Cer(t18:1/22:0), Cer(t20:1/24:3), Cer(d19:1/25:3), Cer(d18:1/25:0), Cer(d18:0/24:0), Cer(d18:2/25:0), Cer(d18:0/23:0), Cer(t17:0/23:0+O), SM(d18:2/22:1), SM(t18:1/24:0), SM(d20:0/24:2), and SM(t18:0/22:0) were increased in both BPD and BPD3w when compared with HC. Notably, the relative contents of SM(d18:2/22:1), SM(d20:0/24:2), and SM(t18:0/22:0) in the BPD group decreased partially toward HC levels but remained higher than HC.

Figure 3.

Scatter plots of individual ceramide and sphingomyelin species.

Notes: (A–N) are scatter plots depicting the relative levels of distinct ceramide (Cer) and sphingomyelin (SM) species across the HC (n = 27), BPD (n = 26) and BPD3w (n = 26) groups. Open circles represent individual sample values, and horizontal lines denote group mean values. (A) Cer(d16:0/18:2); (B) Cer(t18:1/22:0); (C) Cer(t20:1/24:3); (D) Cer(t17:0/22:0); (E) Cer(d18:0/23:0); (F) Cer(d18:0/24:0); (G) Cer(d19:1/25:3); (H) Cer(d18:2/25:0); (I) Cer(d18:1/25:0); (J) Cer(t17:0/23:0+O); (K) SM(t18:1/24:0); (L) SM(t18:0/22:0); (M) SM(d18:2/22:1); (N) SM(d20:0/24:2). *p < 0.05; **p < 0.01.

Abbreviations: HC, healthy controls; BPD, patients with bipolar depression at baseline; BPD3w, patients with bipolar depression after 3 weeks of intervention; Cer, ceramide; SM, sphingomyelin.

Refinement of the Biomarker Panel and Correlation Analysis

To provide a robust estimate of model performance and mitigate overfitting concerns, we applied LASSO regression with 100-fold bootstrap averaging to refine the candidate panel and examined the correlations between the identified lipid species and clinical symptom severity (Figure 4). From the 14 candidate lipids identified by intersection analysis, LASSO regression (λ = 100) with bootstrap averaging selected a parsimonious 4-species panel: Cer(t20:1/24:3), Cer(d19:1/25:3), SM(t18:0/22:0), and Cer(t17:0/22:0), with a bootstrap-averaged AUC of 0.975 (95% CI: 0.917–1.000). Although bootstrap averaging reduces overfitting, this estimate remains internally derived and requires external validation. Correlation analysis revealed that Cer(d19:1/25:3), Cer(t17:0/22:0), and SM(t18:0/22:0) were positively correlated with depressive, anxiety, and manic symptom severity (HAMD, HAMA, and YMRS). In contrast, Cer(t20:1/24:3) showed a modest negative correlation with HAMD, but not with HAMA or YMRS. Additionally, Cer(t20:1/24:3) and SM(t18:0/22:0) were negatively correlated with sex, indicating lower levels in females than in males (male = 1, female = 2).

Figure 4.

Lasso regression, ROC analysis, and correlation clustering heatmap.

Notes: (A) shows the LASSO binomial deviance plot for feature selection. (B) shows the LASSO coefficient path plot. (C) shows the ROC curve (bootstrap average) for the 4-species panel, with an AUC of 0.975 (95% CI: 0.917–1.000). This curve represents bootstrap-averaged performance, which reduces overfitting but remains internally derived without external validation. (D) shows a correlation clustering heatmap illustrating correlations between clinical parameters and the levels of the 14 lipid species. The horizontal axis corresponds to clinical parameters (YMRS, HAMA, HAMD, age, and sex), and the vertical axis corresponds to differentially abundant lipid molecules. Blue squares indicate negative correlations, while orange squares indicate positive correlations. The color gradient shifts from blue to red, corresponding to Spearman correlation coefficients; white squares represent zero correlation. *p < 0.05; **p < 0.01; ***p < 0.001.

Abbreviations: HC, healthy controls; BPD, patients with bipolar depression at baseline; BPD3w, patients with bipolar depression after 3 weeks of intervention; ROC, receiver operating characteristic; AUC, area under the curve.

Discussion

The present study establishes that plasma sphingomyelin (SM) and ceramide (Cer) profiles are significantly perturbed in young adults with bipolar depression (BPD) and identifies a 14-species lipid panel with high, but internally derived and unvalidated, discriminative performance (AUC = 0.997). These findings position the SM-Cer metabolic axis as a promising peripheral window into the pathophysiology of BPD, with implications for both mechanistic understanding and clinical translation.

The elevation of total Cer in BPD aligns with emerging evidence implicating ceramide accumulation in stress-related neuropsychiatric disorders.³⁰ Ceramide functions as a second messenger in pro-apoptotic signaling, mediating neuronal atrophy and synaptic pruning through activation of protein phosphatase 2A (PP2A), inhibition of protein kinase B (Akt), and stimulation of the mitochondrial intrinsic apoptosis pathway.^31,32 The observation that total Cer remained elevated post-treatment, despite symptomatic improvement, suggests that ceramide accumulation may reflect a persistent pathophysiological process that is partially refractory to short-term pharmacological modulation. While this observation may be consistent with the neuroprogression hypothesis of bipolar disorder,^33,34 the short 3-week follow-up does not allow us to distinguish between neuroprogressive changes and incomplete normalization within this treatment period. Longer-term longitudinal studies are needed to clarify this.

Conversely, the significant reduction in total SM and the corresponding increase in SM/Cer ratio following three weeks of pharmacotherapy indicate that SM dynamics are more responsive to acute intervention. This dissociation between Cer and SM trajectories underscores the complexity of sphingolipid regulation: while Cer synthesis via de novo pathways (serine palmitoyltransferase, ceramide synthase) and SM hydrolysis (ASMase, neutral sphingomyelinase) may be constitutively upregulated in BPD, SM replenishment through ceramide phosphorylation (sphingomyelin synthase) or de novo synthesis appears to be pharmacologically modifiable.^35,36 The differential turnover kinetics of specific acyl chain species—evidenced by treatment-responsive decreases in Cer(t17:0/22:0) alongside persistent elevations in very-long-chain species—further highlight the necessity of species-resolved lipidomics over aggregate measures.³⁷ The SM/Cer ratio integrates two functionally antagonistic lipid classes and thus serves as a composite index of membrane integrity vs catabolic stress. In healthy neuronal membranes, SM constitutes approximately 10–20% of total phospholipids, forming lipid raft microdomains that scaffold neurotransmitter receptors, ion channels, and cell adhesion molecules. Cer enrichment, by contrast, displaces cholesterol and promotes raft coalescence into larger, less fluid domains, thereby impairing receptor trafficking and signal transduction.³⁸ The reduced SM/Cer ratio observed in BPD, and its partial normalization with treatment, may hypothetically reflect a shift from rigid, Cer-enriched membranes toward more fluid, SM-rich configurations conducive to synaptic plasticity. However, membrane fluidity and lipid raft dynamics were not directly measured in this study, and this interpretation remains speculative.

Notably, although the patients showed clinical improvement, the SM/Cer ratio after treatment remained effective in discriminating between patients and healthy controls (AUC = 0.849), indicating that the sphingolipid metabolic profile had not yet fully returned to the level observed in healthy controls after three weeks of pharmacological treatment. Interestingly, the SM/Cer ratio discriminated BPD3w from HC (AUC = 0.849) better than BPD from HC (AUC = 0.724). This may reflect reduced inter-individual heterogeneity in SM/Cer ratios after treatment, partial normalization without full recovery, and/or the limited sample size. These possibilities warrant further investigation in larger cohorts. However, it remains unclear whether this persistent difference in lipid metabolism reflects a lag in membrane lipid remodeling relative to symptomatic improvement or simply results from an insufficient treatment duration of three weeks to allow complete normalization of the relevant metabolic processes. Persistent alterations in plasma phospholipid metabolism have been associated with residual cognitive deficits and the risk of disease recurrence.^39,40 It should also be considered that the lipid changes observed between BPD and BPD3w may partly reflect direct effects of quetiapine and magnesium valproate on lipid metabolism, rather than treatment response or disease-related changes. Quetiapine has been associated with alterations in plasma lipid profiles, and valproate may also affect lipid metabolism. Our study design cannot distinguish between treatment response and direct drug effects. Future studies with drug-naïve comparison groups or placebo-controlled designs are needed to disentangle these possibilities. Additional factors that may influence plasma lipid profiles-including diet, metabolic status (eg, insulin resistance), and other unmeasured lifestyle variables-were not systematically assessed in this study and may contribute to the observed differences. These residual confounders should be considered when interpreting the lipid alterations. Therefore, longer-term longitudinal studies are needed to track dynamic changes in the SM/Cer ratio during the acute treatment phase and subsequent follow-up, to further clarify the temporal relationship between sphingolipid normalization and clinical or neurofunctional recovery and to evaluate its potential association with sustained remission. Preclinical models demonstrate that chronic unpredictable mild stress (CUMS) elevates cortical ceramide levels, precipitating dendritic retraction and depressive-like behaviors that are reversible upon pharmacological inhibition of acid sphingomyelinase (ASMase). Recent advances in region-specific and systems-level approaches have substantially refined our understanding of ASM-Cer signaling in mood regulation. Sortilin, a multiligand receptor predominantly expressed in neurons, undergoes intracellular trafficking to lysosomes where it facilitates ASMase maturation and subsequent ceramide generation. Conditional deletion of sortilin in the mouse prefrontal cortex and hippocampus markedly attenuates depressive-like behaviors in chronic stress paradigms, concomitant with reduced ASMase activity and ceramide accumulation.⁴¹ These findings establish a causal link between ASM-Cer hyperactivity and depression phenotypes. While emerging evidence suggests a potential link between gut microbiota and sphingolipid metabolism, our study did not assess microbiota composition. Future studies integrating metagenomics with lipidomics may help clarify whether microbiota-derived ceramides contribute to the observed alterations.

The 14-lipid panel identified through Venn diagram intersection provides a multi-marker approach that may offer advantages over single-biomarker strategies. Its constituent species-enriched in very-long-chain polyunsaturated fatty acid (VLCPUFA)-containing ceramides (eg, Cer(d19:1/25:3), Cer(d18:2/25:0)) and dihydrosphingomyelins (eg, SM(t18:0/22:0))-exhibit distinctive biophysical properties. VLCPUFA-ceramides, characterized by acyl chains exceeding 24 carbons with ≥3 double bonds, are highly enriched in mammalian germ cells and specific neuronal subpopulations, where they confer membrane curvature stress and regulate exocytosis.⁴² Their selective upregulation in BPD may indicate perturbation of peroxisomal β-oxidation and elongation pathways, which are essential for VLCPUFA synthesis and have been implicated in neurodevelopmental disorders.⁴³

The diagnostic performance of this panel (AUC = 0.997) appears comparable to previously reported multi-analyte signatures in psychiatry, including a 17-metabolite panel for major depressive disorder (AUC = 0.88)⁴⁴ and a 6-lipid panel for bipolar disorder (AUC = 0.994). However, direct comparison is limited by differences in sample composition, analytical platforms, and validation status. However, such exceptional accuracy warrants cautious interpretation. The panel was derived and evaluated within the same dataset, introducing substantial risk of overfitting. Moreover, the intersection strategy—selecting species differential in both BPD vs HC and BPD vs BPD3w comparisons—may inadvertently exclude trait markers that do not change with treatment. Prospective validation in independent cohorts, preferably with standardized medication protocols, is imperative to establish generalizability.⁴⁵ It should be noted that the Venn intersection strategy employed in this study—selecting lipids differential in both BPD vs HC and BPD vs BPD3w comparisons—preferentially identifies treatment-responsive or state markers rather than stable diagnostic (trait) markers. This approach may inadvertently exclude trait markers that are differentially expressed between BPD and HC but do not change with treatment. Therefore, the 14-lipid panel should be regarded as a candidate state/treatment-response marker set rather than a definitive diagnostic panel. Future studies should separately investigate stable trait markers using cross-sectional comparisons. Importantly, although we applied LASSO regression with 100-fold bootstrap averaging to reduce overfitting, the resulting 4-species panel still requires external validation in larger, independent cohorts to confirm its true diagnostic performance. The AUC of 0.997 for the 14-lipid panel should therefore be regarded as an exploratory, internally derived estimate rather than definitive evidence of diagnostic accuracy. In addition, sensitivity, specificity, calibration, and decision thresholds for the classifier were not fully reported, which further limits the interpretability of the panel’s diagnostic performance. However, the translational trajectory of sphingolipid biomarkers in bipolar depression faces several challenges. First, while LC-MS/MS remains the gold standard for lipidomic profiling, its technical demands and operational costs hinder routine clinical implementation. Targeted assays with lower analytical complexity, such as flow injection analysis-tandem mass spectrometry (FIA-MS/MS) or antibody-based detection of specific species, may offer a scalable alternative.⁴⁶ Second, the optimal timing of biomarker assessment relative to disease stage and treatment initiation remains to be clarified. Our three-week follow-up captured early pharmacodynamic changes, but the predictive value of baseline lipid signatures for long-term outcomes, including treatment resistance, suicide risk, and functional recovery is unknown. From a therapeutic perspective, the SM-Cer axis may represent a potential target for future investigation, although our findings are hypothesis-generating and require validation in independent cohorts before any therapeutic implications can be drawn.

The biological effects of ceramide are highly context-dependent, and different species may exert distinct effects: C16 and C18 ceramides are primarily associated with pro-apoptotic effects, whereas very-long-chain ceramides (C22–C24) are more closely related to cell survival and differentiation.⁴⁷ Therefore, overall changes in Cer levels may mask the species-specific biological effects of individual ceramides. In exploring the pathological significance and potential therapeutic implications of sphingolipid alterations in BPD, particular attention should be given to the species-specific effects of different ceramides. In addition, the sex-dependent variation in Cer(t20:1/24:3), SM(t18:1/24:0), and SM(t18:0/22:0), with lower levels in females, which resonates with established sexual dimorphisms in sphingolipid metabolism. The interplay between sex hormones and sphingolipid metabolism is well recognized, albeit the exact regulatory mechanisms remain incompletely defined.⁴⁸ Epidemiologically, bipolar disorder, primarily in bipolar I disorder exhibits sex differences in age at onset, episode polarity, and treatment response, with females experiencing more depressive episodes and rapid cycling.^47,49 Whether the observed lipid sex effects mediate these clinical differences—or merely reflect hormonal confounding—remains to be elucidated. Future studies should stratify analyses by sex and menstrual phase, and consider gonadal steroid measurements to disentangle endocrine from disease effects.⁵⁰

Several limitations constrain the interpretation of our findings. The baseline HAMD mean of 44.81 in our BPD group indicates a severely affected sample; however, all patients were recruited from the psychiatric outpatient clinic of a tertiary general hospital, reflecting a real-world treatment-seeking population. The post-treatment HAMD mean of 25.12 indicates that most patients had not achieved remission after 3 weeks of treatment, which is consistent with the known delayed onset of full antidepressant effects. Longer treatment duration may be required for complete symptomatic remission. Severe depressive symptoms are a core feature of bipolar depression, and our findings may not fully generalize to community-based or less severely affected populations. The modest sample size (n = 26 BPD, n = 27 HC) and single-center design limit statistical power for subgroup analyses and generalizability across ethnicities and healthcare settings. Although our a priori power calculation indicated sufficiency for large effect sizes, biomarker validation typically requires hundreds to thousands of participants to ensure robust calibration.⁵¹ The heterogeneity of pharmacological regimens, encompassing mood stabilizers, atypical antipsychotics, and antidepressants, precludes attribution of lipid changes to specific drug mechanisms. A further statistical limitation is that the three-group comparison of HAMD, HAMA, and YMRS scores in Table 1 used one-way ANOVA, which treated BPD and BPD3w as independent groups and did not fully account for the repeated-measures design. However, the primary comparisons of interest (BPD vs BPD3w) were analyzed using paired tests, which appropriately account for within-subject correlations. A further statistical limitation is that BMI was compared using a parametric unpaired t-test, although the HC group did not pass the Shapiro–Wilk normality test (P = 0.006). BMI was not a primary outcome of this study, and no significant difference in BMI was observed between the HC and BPD groups (P = 0.203), suggesting that this limitation is unlikely to materially affect the main findings. We did not measure ASMase or sphingomyelin synthase activity, nor did we assess central nervous system lipids via cerebrospinal fluid or neuroimaging, leaving the brain-periphery relationship inferential. Finally, it should be acknowledged that the observed SM/Cer differences in this study only reflect the disparities between BPD patients and healthy individuals, and it remains unclear whether these markers possess sufficient specificity to differentiate BPD from MDD. Additionally, multiple comparison correction was not applied to the correlation analyses, which may increase the risk of false-positive findings. Therefore, the correlation results should be interpreted as exploratory and hypothesis-generating rather than definitive. It should also be noted that some lipid species showed statistically significant but small fold changes (0.96–1.10) between BPD and BPD3w. These changes should be interpreted with caution, as they may reflect early or subtle pharmacodynamic effects rather than robust biological shifts. The high proportion of differential lipids identified between BPD and HC (66/140, 47%) may also raise concerns about overfitting given the modest sample size. However, FDR correction (q < 0.05) and VIP > 1 were applied to reduce false-positive findings, and widespread lipid dysregulation is biologically plausible in bipolar disorder. Nevertheless, independent validation in larger cohorts is required to confirm these findings. The 14-lipid panel was derived and evaluated using the same dataset, which may introduce selection bias and data leakage. Nested cross-validation or bootstrap optimism correction was not performed, and sensitivity, specificity, calibration, and decision thresholds were not fully reported. Correlation and sex-stratified findings should be considered exploratory given the small sample size and extensive multiple testing. Independent external validation is required to confirm the diagnostic performance of the panel.

Conclusion

In summary, this study demonstrates that plasma SM and Cer are significantly altered in young adults with bipolar depression. A 14-species lipid panel shows preliminary discriminative ability (internally derived AUC = 0.997) that requires validation in independent cohorts, and several species correlate with symptom severity and sex. These findings suggest that the SM-Cer-related lipid signature may serve as a candidate biomarker for further investigation. However, this signature should currently be considered a discovery-stage candidate panel and does not yet provide sufficient evidence to support its use for the diagnosis of bipolar disorder. In addition, the potential effects of medication could not be fully disentangled from disease-related effects and should therefore be considered when interpreting the observed lipid alterations. Accordingly, further validation in larger, independent cohorts, including individuals with major depressive disorder (MDD) and other psychiatric disorders as clinical controls, is warranted to further evaluate the robustness, specificity, and potential value of this candidate lipid signature for diagnosis.

Funding Statement

This research was supported by the General Program (2025AAC030800) and the Key Research and Development Program of Ningxia Science and Technology Department (2021BEG03057).

Data Sharing Statement

Raw and processed datasets generated and analyzed in this study are available from the corresponding author on reasonable request.

Ethical Approval

This study was registered with the Chinese Clinical Trial Registry (Identifier: ChiCTR2400089218), approved by the Institutional Review Board/Ethics Committee of our hospital, and all study procedures were conducted in accordance with the Declaration of Helsinki.

Disclosure

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Majorbio Bio-Pharm Technology Co., Ltd. provided only analytical services and had no involvement in the research process beyond data generation.