Introduction

Adolescent Major Depressive Disorder (MDD) is a common and disabling psychiatric condition characterized by persistently low mood, hopelessness, and loss of interest to life to the degree that impairs daily functioning [1]. Epidemiological estimates prior to the COVID-19 pandemic placed the prevalence of having at least one major depressive episode during adolescence at roughly 13–15%, with consistently higher rates in adolescent females than males [2, 3]. Onset of MDD during adolescence is associated with poorer academic performance, elevated risk of substance use and suicidal behavior, both contributing to substantial short- and long-term impairment in life perspectives [4]. Together, these observations highlight the need for early identification and timely intervention in youth with depressive disorders [1].

Reliable, objective biomarkers could improve diagnostic accuracy, stratification, and treatment selection for adolescent MDD, but none have yet achieved broad clinical utility. Candidate markers have emerged from diverse modalities — electrophysiology (e.g., resting qEEG coherence), neuroimaging, molecular measures (transcriptomic and proteomic alterations), and metabolomic/lipidomic profiling — yet results of these studies remain inconsistent and limited by small sample sizes or platform-specific effects [5,6,7]. Heterogeneity in adolescent biology, including sequential neurophysiological changes and a variety of early-life exposures further complicates biomarker discovery and their translation into clinic [8, 9].

Lipidomics is a promising avenue for biomarker discovery in MDD. Serving both as signaling molecules and as inflammatory/ anti-inflammatory effectors, lipids are central to brain metabolism. Early lipidomic studies in mood disorders report increased levels of oxidized fatty acids and decreased levels of acyl-carnitines. Several lipid species have shown diagnostic potential in adolescent cohorts, with differences in multiple lipid species shown, including that for LPC (18:0) [10, 11]. However, findings reported to date remain preliminary and often platform- or species-specific, thus, motivating one for a search of reproducible, translational lipid biomarkers [12, 13].

In this study, we apply comprehensive lipidomic profiling across multiple analytical platforms and species to identify and validate lipid biomarkers associated with adolescent MDD. We hypothesize that specific, reproducible lipid signatures could differentiate adolescents with MDD from age-matched controls and illuminate metabolic pathways relevant to disease biology, thereby advancing translational diagnostics and informing personalized therapeutic strategies [11, 14, 15].

Materials and methods

This study employed a multi-stage, cross-platform, and cross-species strategy to identify and validate lipid biomarkers associated with adolescent MDD. Lipidomic profiling was performed in training and validation cohorts, followed by cross-species evaluation in a chronic unpredictable mild stress (CUMS) rat model. Candidate lipids were selected using combined multivariate and univariate statistical approaches, and their diagnostic performance was tested via cross-validation and independent replication. A detailed description of study design, recruitment, sample processing, analytical platforms, and statistical pipelines is provided in the Supplementary Materials and Methods.

Lipid marker section using human data

Study population and sample collection

The training cohort included 95 adolescents with first-episode, antidepressant-naïve MDD and 40 age-matched healthy controls (HC), 56 adolescents MDD and 37 healthy controls in the validation cohort, all recruited from the outpatient and inpatient departments of the Department of Psychiatry at the First Affiliated Hospital of Shanxi Medical University between 2022 - 2023. The study protocols were approved by the Ethics Committee of the First Affiliated Hospital of Shanxi Medical University (Approval No. K062) and were conducted in accordance with the Declaration of Helsinki. Written informed consents were obtained from all participants and their legal guardians. Serum samples were obtained and processed for lipidomic analysis using complementary UPLC–MS/MS (training) and UPLC–HRMS (validation) platforms. Ethical approval and written informed consent were obtained. Detailed inclusion/exclusion criteria, diagnostic procedures, and sample handling are provided in the Supplementary Materials and Methods.

Lipidomic data generation

Serum lipids were extracted using a standardized methanol–MTBE–water protocol and analyzed on complementary platforms: targeted UPLC–MS/MS for the training dataset and untargeted UPLC–HRMS for the validation dataset. Rigorous quality control procedures, including pooled QC samples and internal standards, confirmed reproducibility across runs. Lipid features were annotated against standard databases (LipidMaps, HMDB) and retained for analysis after QC-based filtering. Full technical details of sample processing, chromatographic conditions, QC thresholds, and bioinformatics workflows are provided in the Supplementary Materials and Methods. It is important to note that the human and rat lipidomic datasets were generated independently using different analytical objectives. The adolescent human cohort was analyzed using a targeted UPLC–MS/MS platform optimized for high-sensitivity quantification of predefined lipid classes, whereas the rat CUMS dataset was produced earlier using a non-targeted UPLC–HRMS workflow designed for broad metabolic screening. Because these platforms were not harmonized prospectively, direct metabolite-level matching is not feasible. Accordingly, our cross-species assessment focuses on pathway-level concordance and lipid-class trends rather than one-to-one metabolite replication, consistent with current best practices in cross-platform lipidomics.

Statistical analysis

Lipidomic data were log-transformed, normalized, and analyzed using a combination of unsupervised (Principal component analysis, PCA)and supervised (Orthogonal partial least squares discriminant analysis, OPLS-DA) approaches. Candidate biomarkers were prioritized using univariate testing with FDR correction and feature-selection methods (Boruta, LASSO, RFE). Predictive performance was assessed by ROC analysis, leave-one-out cross-validation (LOOCV), and validation in an independent cohort. Full descriptions of data preprocessing, statistical tests, model validation procedures, and software packages are provided in the Supplementary Materials and Methods. For univariate comparisons of lipid features between groups, both raw p-values and false discovery rate (FDR)–adjusted p-values (Benjamini–Hochberg method) were calculated, with FDR < 0.05 considered statistically significant. For multivariate modeling steps—including logistic regression, cross-validation performance, and ROC analyses—only raw p-values are reported, as these analyses evaluate model-level statistics rather than feature-level multiple comparisons. Pathway enrichment analyses used FDR-adjusted p-values as generated by the enrichment algorithm. This distinction ensures that appropriate multiple-comparison correction is applied where necessary while preserving interpretability of model-level metrics.

Animal study

Animals and experimental design

Twelve specific pathogen-free (SPF) male Sprague-Dawley (SD) rats (3 weeks old, weighing 80–120 g) were purchased from Beijing Weitonglihua Experimental Animal Technology Co., Ltd. (License No.: SCXK (Jing) 2021-0006) and randomly divided into the control group and the chronic unpredictable mild stress (CUMS) group.The CUMS paradigm was adapted from established protocols [16, 17] to induce depression-like phenotypes, and investigators were blinded during assessments. All animal procedures were conducted in accordance with the ARRIVE 2.0 guidelines for experimental design and reporting, and were approved by the Ethics Committee of Shanxi Medical University (Approval No.: DWLL-2024-035). Detailed housing conditions, stressor schedules, randomization procedures, and monitoring protocols are provided in the Supplementary Materials and Methods.

Behavioral assessments

After 21 days of CUMS, rats underwent standardized behavioral testing to evaluate anhedonia, locomotor activity, despair-like behavior, and spatial memory. Tests included sucrose preference, open field, forced swim, and Y-maze paradigms, performed in a fixed order with appropriate acclimation intervals. Outcomes were quantified using automated video-tracking where applicable. Detailed protocols, scoring criteria, and data analysis procedures are provided in the Supplementary Materials and Methods.

Sample collection and tissue processing

Following behavioral assessments, rats were anesthetized, and serum and brain tissues (prefrontal cortex, hippocampus, striatum) were collected for lipidomic profiling. Samples were processed using standardized extraction protocols consistent with those applied in human datasets. Detailed procedures for blood collection, tissue dissection, storage conditions, and lipid extraction are provided in the Supplementary Materials and Methods.

Quality control (QC) for animal dataset

QC samples were prepared by pooling aliquots of all rat serum and brain tissues and injected periodically during the run. Features with >30% relative standard deviation across QC replicates were excluded. Internal standards confirmed stable retention times and intensities, and PCA showed tight QC clustering. LOESS correction was applied when necessary to minimize batch effects. Detailed QC procedures are provided in the Supplementary Materials and Methods.

Statistical analysis for animal data

Behavioral and lipid data were tested for normality and compared between CUMS and control groups using independent t-tests or Mann–Whitney U tests, with categorical variables assessed by chi-square. Multiple behavioral endpoints were corrected using FDR (adjusted P < 0.05). Correlations between lipid levels and behavioral outcomes were analyzed by Spearman’s rank correlation with FDR correction. Detailed statistical procedures are provided in the Supplementary Materials and Methods.

Metabolomics data analysis

Data preprocessing and multivariate analysis

Raw MS data from human and rat lipidomics were processed for peak detection, alignment, normalization, and compound annotation against LipidMaps and HMDB ( ± 5 ppm mass tolerance, ±0.2 min RT tolerance). QC-confirmed stability was achieved as described in Sections 2.1.2 and 2.2.4. Data were log-transformed and Pareto-scaled prior to multivariate analysis. PCA was used for unsupervised visualization, and OPLS-DA for supervised discrimination, with VIP > 1 considered indicative of relevance. Separate OPLS-DA models were constructed for each dataset (human training, human testing, rat serum, rat brain regions). Detailed preprocessing workflows and parameter settings are provided in the Supplementary Materials and Methods.

Feature selection and model validation

Feature selection in the human datasets used three complementary methods: Boruta, recursive feature elimination (RFE), and Least Absolute Shrinkage and Selection Operator (LASSO) regression. Candidate lipids with VIP > 1, adjusted P < 0.05, and AUC ≥ 0.70 were retained, and those identified by at least two methods were considered promising biomarkers. Model performance was assessed by LOOCV. Cross-species validation was performed by matching human biomarkers to rat lipidomic datasets (serum and brain regions) using exact mass, retention time, and class annotation, with analog matches flagged and excluded from concordance testing. Detailed feature selection algorithms, validation procedures, and matching criteria are provided in the Supplementary Materials and Methods.

Pathway enrichment analysis

Pathway enrichment and topology analyses were conducted in MetaboAnalyst 6.0 using KEGG databases for Homo sapiens and Rattus norvegicus. Lipid features were mapped to KEGG Compound IDs; when exact matches were unavailable, generic class-level entries were used. Statistical significance was defined as raw P < 0.05, with FDR-adjusted P < 0.05 considered significant. Separate analyses were performed for human and rat datasets, and cross-species comparisons were based on shared enriched pathways. Detailed enrichment workflows, parameter settings, and match handling are provided in the Supplementary Materials and Methods.

Regression and association analyses

Logistic regression was used to identify predictors of MDD status, and multiple linear regression to examine associations between lipid levels and demographic/clinical variables. Covariates included sex, age, guardian education, suicide attempt history, and HAMA/HAMD scores. Multicollinearity and model assumptions were checked, and FDR correction was applied for multiple comparisons (adjusted P < 0.05). Analyses were performed separately for training and validation cohorts. Detailed coding of variables, model specifications, and diagnostic procedures are provided in the Supplementary Materials and Methods.

Cross-Species lipid mapping

Human-derived lipid biomarkers (VIP > 1, adjusted P < 0.05, AUC ≥ 0.70) were searched in rat serum and brain datasets using exact mass, retention time, and class annotation. Analog matches were flagged and excluded from concordance testing. For matched features, univariate tests with FDR correction and OPLS-DA were applied to assess group differences and directional concordance between species. Detailed matching criteria and statistical workflows are provided in the Supplementary Materials and Methods.

Results

Training dataset for feature selection

Demographic characteristics

The demographic and clinical characteristics of participants in the training cohort are summarized in Table 1. The relevant information of the validation cohort in Table S1. In either cohort, there were no significant differences between MDD and control groups with respect to sex, age, or guardian education level (all P > 0.05). The proportions of individuals with a history of suicide attempt and high HAMA scores were significantly larger in the MDD group compared with that in controls in both cohorts (all P < 0.05).

Table 1 Comparison of demographic characteristics between MDD patients and HCs.
Full size table

Continuous variables were assessed for normality using the Kolmogorov–Smirnov test prior to comparison; all were normally distributed and analyzed using independent-sample t-tests. Categorical variables were analyzed using chi-square tests. Variable coding and definitions followed those described in Section 2.5. P-values are reported without multiple-comparison correction, as these analyses were limited to baseline descriptive comparisons.

Targeted lipidomics analysis

A total of 1025 lipid species were reliably detected. By species count, the most represented subclasses were triglycerides (n = 271), phosphatidylcholines (n = 104), and phosphatidylethanolamines (n = 70) (Fig. 1A). Principal component analysis (PCA) demonstrated good within-group reproducibility in both MDD and HC samples (Fig. 1B), with hierarchical clustering being supportive of group-level differences (Fig. 1C).

Fig. 1: Targeted lipidomics analysis.
Fig. 1: Targeted lipidomics analysis.The alternative text for this image may have been generated using AI.
Full size image

A Systematic statistics of all detected lipid species; (B): PCA model; (C): Cluster heatmap; (D): OPLS-DA model; (E): Validation of OPLS-DA model; (F): Volcano plot of serum differential metabolites; (G): Pathway analysis of differential metabolites.

OPLS-DA revealed clear separation between groups (Fig. 1D) with model parameters Q² = 0.342, R²X = 0.29, and R²Y = 0.652. Model validity was supported by permutation testing (Fig. 2E), indicating good predictive ability without evidence of overfitting.

Fig. 2: LOOCV Classification Performance for Selected Lipid Feature Sets.
Fig. 2: LOOCV Classification Performance for Selected Lipid Feature Sets.The alternative text for this image may have been generated using AI.
Full size image

Four lipid feature sets were evaluated for diagnostic classification performance via LOOCV, with metrics including accuracy, precision for HC, recall for HC, and specificity. The 29‑feature set (Boruta+LASSO+RFE) and 7‑feature set (LASSO+RFE) were validated in the targeted UPLC–MS/MS training cohort; the 8‑matched‑feature set and 2‑core‑feature set (Boruta only) were validated in the untargeted UPLC–HRMS independent validation cohort.

Differential analysis identified a total of 244 significantly altered lipid species based on VIP > 1 and FDR-adjusted P < 0.05 (Fig. 1F), of which 147 were upregulated and 97 downregulated (Table S2). The volcano plot is shown for visualization of fold change versus significance; no additional fold-change threshold was imposed. For presentation, the Top 50 species ranked by absolute fold change were predominantly glycerolipids and glycerophospholipids. KEGG pathway enrichment of these 244 species using exact and, where necessary, generic class-level mappings as specified in Methods (Section 2.4), has highlighted glycerophospholipid, glycerolipid, and sphingolipid metabolism (Fig. 1G).

Feature selection

From the 244 significant metabolites (VIP > 1 and FDR-adjusted P < 0.05), univariate ROC analysis in the training cohort identified 30 candidates with AUC ≥ 0.70 (Table S3). Three complementary feature-selection methods—Boruta, RFE, and LASSO—were then applied within the training cohort using LOOCV. The intersection of features retained by at least two methods yielded a final panel of 29 robust biomarkers spanning multiple lipid classes, including lysophosphatidic acids, free fatty acids, hexosylceramides, and triglycerides (Table 2).

Table 2 The 29 selected metabolites using combined feature selection approaches.
Full size table

For each biomarker, we report AUC with 95% bootstrap confidence intervals, sensitivity, specificity, and accuracy at the Youden-optimal cutoff. Cutoff values are expressed on log-transformed, Pareto-scaled intensities (as in Methods 2.3.1). No data from the testing cohort were used during feature selection.

LOOCV model performance

LOOCV using the full 29-feature panel (Boruta + LASSO + RFE; Table 2) achieved 90.4% accuracy in the training dataset, with precision, recall, and specificity of 91.8, 94.7, 80% for MDD and 86.5, 80.0, 95% for HC, respectively (Fig. 2; Table S4).

A reduced 7-feature panel (Table 3) selected jointly by LASSO and RFE achieved 94.8% accuracy. For MDD, precision, recall, and specificity were 96.8, 95.8, and 92.0%; for HC, they were 90.2, 92.5, and 96.0% (Fig. 2; Table S5).

Table 3 Lipid Features Included in Each LOOCV Model.
Full size table

In the independent validation dataset, which was obtained using untargeted lipidomics on a different platform, an 8-feature subset (Table 3) achieved 71.2% accuracy (Fig. 2; Table S6). A minimal 2-feature subset, including LPA(18:2) and SPH(d16:1) molecules, has achieved 72.1% accuracy, suggesting that minimizing of a biomarker panels may be feasible for easier clinical implementation (Fig. 2; Table S7).

The use of two different analytical platforms for the training and validation measurements was deliberate. This approach improved biomarker robustness, as truly stable biomarkers should maintain diagnostic utility across multiple analytical technologies. However, as a trade-off, only 8 of the 29 training-derived features were covered in the validation dataset, likely contributing to the decrease in external accuracy (71.2 vs. 90.4%).

All LOOCV models, including the reduced and minimal panels, are described in Table 3. In both datasets, smaller feature subsets have achieved higher or comparable predictive accuracy compared with the full feature sets, suggesting that minimized biomarker panels may retain strong discriminatory power while offering practical advantages for clinical application.

Cross-Species validation of top lipid biomarkers

Behavioral performance in rat models

Compared to controls, CUMS rats showed significantly reduced body weight, which was evident after 3 weeks of stress exposure (Fig. 3A). In the open field test, CUMS rats exhibited significantly decreased central movement time and total travel distance (Fig. 3B and 3C). In the sucrose preference test, CUMS rats consumed significantly lesser amounts of sweetened water (Fig. 3D). In the forced swim test, CUMS rats displayed prolonged immobility time (Fig. 3E). In the Y-maze test, CUMS rats demonstrated a lower spontaneous alternation percentage relative to controls (Fig. 3F).

Fig. 3: Behavioral outcomes in CUMS model rats.
Fig. 3: Behavioral outcomes in CUMS model rats.The alternative text for this image may have been generated using AI.
Full size image

A Body weight over 21 days. B Central movement time in the open field test. C Total movement distance in the open field test. D Sucrose preference rate. E Immobility time in the forced swim test. F Correct alternation percentage in the Y-maze test. n = 6 per group; *: P < 0.05, **: P < 0.01,***: P < 0.001.

Collectively, these results confirm successful establishment of the CUMS model, as evidenced by reduced body weight, increased anhedonia-like behavior, impaired locomotor and exploratory activity, increased behavioral despair, and decreased spatial working memory.

Cross-species validation identifies two shared lipid biomarkers

From the human adolescent MDD training dataset, nine top lipid biomarkers were prioritized based on combined feature-selection approaches (Boruta, LASSO, and RFE) and diagnostic performance metrics. These included seven significantly downregulated species, including HexCer(t14:1/17:0(2OH)), SPH(d16:1), TG(14:0_15:0_16:0), FFA(18:3), PC(10:0_18:0), LPC(0:0/16:0), and LPA(18:2), and two significantly upregulated species, namely, lithocholicacid-3-sulfate and TG(18:0_17:1_18:1) (Table 4). These lipids spanned multiple metabolic classes, including glycerophospholipids, sphingolipids, glycerolipids, fatty acyls, and sterol lipids, reflecting broad disturbances in lipid metabolism observed in adolescent MDD.

Table 4 Summary of top lipid biomarkers identified in human and rat datasets.
Full size table

Targeted lipidomic profiling in the CUMS rat model was performed using collected serum samples and three brain regions (hippocampus, striatum, and prefrontal cortex), as described in Methods 2.6. Among the nine human-derived biomarkers, two showed significant decreases in CUMS rats: LPC(0:0/16:0) in the prefrontal cortex and SPH(d16:1) in serum. Notably, SPH(d16:1) was consistently downregulated in serum from both species, suggesting a conserved mode of alteration in sphingolipid metabolism. The prefrontal cortex–specific decrease in LPC(0:0/16:0) aligns with the previously described role of glycerophospholipids in neuronal membrane structure and synaptic signaling.

Overall, these results demonstrate incomplete but biologically meaningful concordance between human and rat lipidomic signatures of MDD, with overlapping molecular changes observed in both central (brain) and peripheral (serum) compartments. SPH(d16:1) and LPC(0:0/16:0) emerge as promising cross-species biomarkers for MDD, while highlighting the importance of tissue context in interpreting lipid dysregulation.

Pathway analysis of the 9 metabolites

Pathway enrichment analysis of the nine selected metabolites revealed three lipid-related pathways of importance, with varying degrees of significance (Fig. 4).

Fig. 4: Pathway enrichment analysis for the nine selected metabolites.
Fig. 4: Pathway enrichment analysis for the nine selected metabolites.The alternative text for this image may have been generated using AI.
Full size image

A Statistical overview of enriched pathways; (B) Glycerophospholipid metabolism; (C) Alpha-linolenic acid metabolism; (D) Sphingolipid metabolism.

Glycerophospholipid metabolism was the most significantly enriched pathway (P = 3.50 × 10⁻⁴, FDR = 0.028, Impact = 0.12), with three metabolites mapped, including phosphatidylcholine (PC 10:0_18:0) and two lysophosphatidylcholines (LPC 16:0, LPC 18:2). Although the pathway impact score was modest, the presence of multiple glycerophospholipids suggests their contribution to membrane lipid composition and phospholipid turnover in adolescent MDD.

Alpha-linolenic acid metabolism demonstrated moderate enrichment (P = 0.0013, FDR = 0.051, Impact = 0.33), with two metabolites mapped, namely, phosphatidylcholine (PC 10:0_18:0) and α-linolenic acid (FFA 18:3). The relatively higher pathway impact indicates that changes in (n-6 PUFA) to (n-3 PUFA) ratio may play a biologically relevant role in the disorder.

Sphingolipid metabolism showed nominal significance (P = 0.0077, FDR = 0.206, Impact = 0.28), with two metabolites mapped, sphingosine (SPH d16:1) and N-acylsphingosine (HexCer t14:1/17:0(2OH)). Despite not surviving multiple testing correction, this finding indicates a potential role for alterations in sphingolipid signaling and ceramide metabolism in adolescent MDD.

Metabolite-to-KEGG mapping combined exact matches and generic entries when specific molecular species were not available. This ensured comprehensive pathway coverage, particularly for complex lipids and sphingolipids (Table S8).

In summary, glycerophospholipid and alpha-linolenic acid metabolic branches emerge as the most relevant to adolescent MDD, with potential involvement of sphingolipid metabolism providing additional, albeit less robust, evidence of altered lipid homeostasis in adolescent MDD.

Regression and association analyses

Logistic and linear regression analyses identified several significant associations between clinical factors, MDD status, and lipid measures (Fig. 5). Regarding MDD status, sex emerged as a predictor, with males showing higher prevalence of disease than females, highlighting sex-related differences in adolescent depression. Guardian education level and history of suicide attempts were both positively associated with MDD, suggesting pathophysiologically meaningful contribution of socio-environmental and clinical factors.

Fig. 5: Dot plot of P-values from regression and association analyses of clinical and demographic factors on MDD status and lipid measures.
Fig. 5: Dot plot of P-values from regression and association analyses of clinical and demographic factors on MDD status and lipid measures.The alternative text for this image may have been generated using AI.
Full size image

The size of the dots represents the p-value (smaller dots indicate a smaller p-value and a stronger correlation). The color represents the direction of the association (positive/negative correlation). The x-axis shows core lipid markers (e.g., TG (14:0_15:0_16:0), LPA (18:2), LPC (0:0/16:0), etc.), and the y-axis lists clinical and demographic factors (gender, age, HAMA, HAMD, suicide attempt, education level, etc).

Lipid measures showed distinct associations with demographic and clinical variables. Age demonstrated opposite effects on two triglyceride species, being negatively associated with one species of TG, TG(14:0_15:0_16:0), but positively with another one, TG(18:0_17:1_18:1), indicating differential age-related regulation of triglycerides. Higher anxiety scores (HAMA) were linked to lower levels of TG(14:0_15:0_16:0) and LPA(18:2). Females exhibited lower LPC(0:0/16:0) and LPA(18:2) when compared to males, underscoring sex-specific differences in lipid profiles.

Overall, these findings highlight the influence of sex, age, and anxiety on lipid metabolism, complementing their associations with MDD status in adolescents, and suggest potential interactions between clinical and metabolic factors in the pathophysiology of depression.

Discussion

This study provides a comprehensive lipidomic characterization of adolescent MDD, identifying a panel of lipid biomarkers with both diagnostic and mechanistic significance. We detected widespread alterations in sphingolipids, glycerolipids, fatty acids, phospholipids, and bile acids, with 29 promising lipid biomarkers achieving high diagnostic accuracy and a refined 7-lipid subset exceeding 94% accuracy. Cross-species validation in a rat model confirmed conserved alterations in SPH(d16:1) and LPC(0:0/16:0), supporting their translational relevance.

The robustness of these biomarkers was further evaluated using an independent testing dataset analyzed on a different analytical platform, providing a stringent test of their stability. Although only 8 of the 29 training-derived markers were captured in the testing datasets, which partly explained the lower external accuracy (71.2 vs. 90.4%), the persistence of key lipid alterations across datasets reinforced their translational and diagnostic utility. Pathway analysis highlighted glycerophospholipid and alpha-linolenic acid metabolism as key modified branches, offering mechanistic insight into MDD pathophysiology.

Key lipid classes and their roles

Sphingolipids, particularly HexCer(t14:1/17:0(2OH)) and SPH(d16:1), were downregulated, consistent with prior findings linking sphingolipid metabolism to stress resilience, synaptic plasticity, and depressive symptoms [18,19,20,21,22]. Glycerolipids exhibited lipid species-specific opposing trends, with TG(14:0_15:0_16:0) decreased and TG(18:0_17:1_18:1) increased, suggesting disruption in energy storage related processes. The levels of free fatty acids and lysophospholipids, including FFA(18:3), LPA(18:2), and LPC(0:0/16:0), were largely reduced, potentially impacting membrane composition, neuroinflammation, and neurotransmission [23]. Phosphatidylcholines such as PC(10:0_18:0) showed subtle downregulation, whereas the levels of lithocholic acid-3-sulfate was markedly upregulated, indicating altered gut–brain and hepatic signaling. Together, these patterns reflect a coordinated dysregulation of lipid homeostasis relevant to adolescent MDD pathology.

The altered lysolipid species identified in this study, including decreased LPC(0:0/16:0) and LPA(18:2), may also indicate perturbations in the Lands cycle, the major phospholipid remodeling pathway responsible for interconversion between lysophospholipids and their parent phospholipids. Dysregulated Lands cycle activity can alter membrane phospholipid composition, receptor mobility, and the generation of lipid mediators involved in neuroinflammatory signaling [24]. Because LPC and LPA serve as bioactive lipids with roles in neuroimmune modulation, changes in their levels may reflect upstream disturbances in phospholipid turnover rather than isolated lipid defects. Although Lands cycle enzymes were not directly measured, the observed lysolipid alterations raise the possibility that phospholipid remodeling pathways are involved in the lipidomic phenotype of adolescent MDD.

It is also noteworthy that the altered TG species identified in this study contain odd-chain and short-chain fatty acids (e.g., C14:0, C15:0, C17:1), which may reflect contributions from gut microbiome–derived metabolic processes. Odd-chain fatty acids (OCFAs) such as C15:0 and C17:1 are produced in part through microbial fermentation and can enter host lipid pools via the portal circulation [25]. Their incorporation into triglycerides may therefore indicate shifts in host–microbiome interactions rather than changes in energy storage alone. This interpretation is consistent with emerging evidence linking the gut microbiota, SCFA production, and lipid remodeling to depressive phenotypes. While microbiome profiles were not available in the present study, the observed TG alterations may point toward a potential microbiome–lipid axis relevant to adolescent MDD.

Metabolic pathways and PFC relevance

Glycerophospholipid metabolism emerged as the most significantly enriched pathway (p = 3.50 × 10⁻⁴, FDR = 0.028), with alpha-linolenic acid metabolism also implicated (p = 0.0013, FDR = 0.051) [26,27,28]. Finding of altered LPC(0:0/16:0) levels in the prefrontal cortex (PFC) aligns with previous reports concerning disruption of synaptic signaling and membrane dynamics observed in MDD [29, 30].

Neurobiological mechanisms linking lipid dysregulation to adolescent MDD

The lipid alterations identified in this study intersect with several neurobiological pathways implicated in MDD. Sphingolipids such as SPH(d16:1) and HexCer species are critical regulators of synaptic plasticity, stress signaling, and neuroinflammatory responses [31]. SPH serves as a precursor for sphingosine-1-phosphate (S1P), a signaling lipid that modulates neurogenesis, microglial activation, and HPA-axis responsiveness [32]. Reduced SPH levels observed in both human and rat serum suggest a shift in sphingolipid homeostasis that may impair stress resilience and synaptic adaptability—processes consistently disrupted in depression.

Glycerophospholipids (e.g., LPC(0:0/16:0) and PC(10:0_18:0)) contribute to membrane fluidity, neurotransmitter-receptor mobility, and intracellular signalling, as changes in phospholipid composition are known to influence the biophysical properties of the bilayer and thus modulate the function of embedded receptors [33, 34]. Alterations in LPC species may also reflect dysregulation of the Lands cycle, which governs phospholipid remodeling and is essential for maintaining neuronal membrane composition. Such disturbances can influence glutamatergic and monoaminergic signaling, both of which are implicated in the pathophysiology of MDD.

The observed reduction in FFA(18:3), an omega-3 polyunsaturated fatty acid, aligns with literature showing impaired omega-3 metabolism in depressive disorders [35]. Omega-3 fatty acids regulate neuroimmune interactions, reduce microglial activation, and support synaptic membrane integrity. A shift in omega-6/omega-3 balance may therefore contribute to heightened neuroinflammatory sensitivity—a mechanism strongly linked to depressive symptoms [36].

These lipid disturbances may be particularly consequential during adolescence, a developmental period characterized by synaptic pruning, maturation of the prefrontal cortex, and dynamic remodeling of neuronal membranes. Perturbations in lipid availability during this window may exacerbate vulnerability to stress, impair neurodevelopmental trajectories, and influence the onset or severity of depressive symptoms.

Integration with neuroinflammatory and transcriptomic pathways

The lipid signatures identified in this study intersect with several biological systems implicated in depression, particularly neuroimmune signaling. Sphingolipids such as sphingosine and hexosylceramides regulate microglial activation, cytokine production, and S1P-dependent immune signaling [37], while glycerophospholipids and lysolipids influence membrane composition, receptor function, and inflammatory mediator synthesis [38]. Likewise, omega-3 polyunsaturated fatty acids modulate neuroinflammatory tone by shaping microglial reactivity and suppressing pro-inflammatory cytokine release [36]. These interactions suggest that the lipid alterations observed in adolescent MDD may not only reflect metabolic disruption but also contribute to inflammatory sensitivity and impaired stress adaptation.

Although our study focused exclusively on lipidomics, future integration with transcriptomic, cytokine, or proteomic datasets will be essential to further elucidate lipid–immune interactions in MDD. Multi-omics approaches could clarify whether lipid abnormalities act as upstream regulators of neuroinflammatory processes, downstream consequences of chronic stress exposure, or intermediate modulators of synaptic remodeling and neuronal resilience. Longitudinal studies incorporating immune markers may also help determine whether specific lipids serve as mechanistic mediators linking metabolic alterations to neuroinflammatory pathways in adolescent depression.

An integrative metabolic–neurobiological framework summarizing these relationships is provided in Supplementary Fig. S1.

Influential clinical and demographic factors

Sex, guardian education, history of suicide attempts, age, and anxiety severity were associated with both a set of specific lipid alterations and the risk of MDD, highlighting multidimensional determinants of depression biology. We show that females exhibit lower LPC(0:0/16:0) and LPA(18:2) levels, whereas age correlates inversely with TG(14:0_15:0_16:0) and positively with TG(18:0_17:1_18:1), emphasizing the need for personalized approaches in MDD diagnosis.

Advantages and limitations

This study leveraged complementary lipidomics platforms, multiple feature selection methods, LOOCV, and cross-species validation to identify promising biomarker candidates for adolescent MDD. However, several limitations should be noted. First, although the training set achieved high diagnostic performance, validation accuracy was lower, in part because the human discovery dataset and the rat validation dataset were generated on different analytical platforms, limiting one-to-one metabolite matching. Larger cohorts measured on harmonized platforms will be required to improve generalizability. Second, the study focused exclusively on lipidomics; we did not assess inflammatory markers, transcriptomic signatures, or other omics layers that could clarify the mechanistic roles of lipid–immune interactions. Third, the specificity of these lipid alterations for MDD is uncertain, as many lipid classes—including lysophospholipids, sphingolipids, and triglycerides—are also perturbed in other psychiatric and metabolic conditions; future work incorporating differential-diagnosis cohorts is needed. Fourth, although the CUMS rat model supported key lipid changes (e.g., SPH(d16:1), LPC(0:0/16:0)), species differences in lipid composition and stress responsivity may limit direct clinical translation, and pathway-level rather than metabolite-level concordance should be emphasized. Finally, several altered triglyceride species contained odd-chain or short-chain fatty acids, suggesting potential involvement of gut microbiome–related metabolism; microbiome profiling was not performed, and future studies integrating lipidomics with microbiome and SCFA analyses are warranted.

Cross-platform validation and generalizability considerations

Although the diagnostic performance of the 29- and 7-lipid panels was high in the training dataset, accuracy decreased to 71–72% in the independent validation cohort. This reduction is expected given the deliberate use of two different analytical platforms (targeted UPLC–MS/MS for discovery and untargeted UPLC–HRMS for validation), which resulted in only 8 out of 29 training-derived lipids being detectable in the validation dataset. Cross-platform variability is well recognized in lipidomics, and performance typically declines when classifiers are transferred across platforms and sample-processing pipelines. Thus, the observed reduction likely reflects platform-related constraints rather than model overfitting. Nonetheless, the ability of a minimal 2-lipid panel to achieve comparable accuracy suggests promising translational potential, while underscoring the need for larger, harmonized cohorts and same-platform replication in future studies.

Sample size considerations and overfitting risk

Both the training and validation cohorts were modest in size, which may increase the risk of overfitting despite the safeguards implemented in this study. To minimize this concern, we employed three independent feature-selection algorithms (Boruta, LASSO, and RFE), rigorous LOOCV within the training dataset, and validation using an entirely separate cohort. These methodological steps reduce, but cannot fully eliminate, overfitting risk in studies with limited sample sizes. Furthermore, adolescent lipidomics exhibits substantial inter-individual variability related to developmental stage, lifestyle factors, and clinical heterogeneity, all of which may require larger sample sizes to adequately capture. Future work involving larger, multi-center cohorts and harmonized lipidomic platforms will be essential to strengthen generalizability, refine the biomarker panels, and assess population-level variability.

Conclusion

In summary, this study presents a comprehensive lipidomic characterization of adolescent MDD, identifying a panel of potential biomarker candidates with diagnostic and mechanistic relevance. Cross-species validation and independent testing confirmed the stability and translational potential of key lipid alterations. These findings highlight dysregulated glycerophospholipid and alpha-linolenic acid metabolism as central to MDD pathophysiology and provide a foundation for future biomarker-guided diagnostic strategies and targeted interventions in adolescent depression.