We collected clinical data from 2013 subjects, divided into 4 groups based on BMI and disease status, and randomly selected proportionally balanced samples from each group for serum metabolomics analysis. Figure 1 shows the overall study design. Tables 1–6 present the demographic, clinical, and biochemical characteristics of the subjects. Our results revealed a significant increase in the incidence of MASLD among middle-aged populations, with a higher prevalence in this age group (Table 1). Among overweight patients with MASLD, 50.3% were male (546 individuals) and 49.7% were female (539 individuals), whereas in non-overweight patients with MASLD, the proportion of females was as high as 73.1% (Table 1). Compared to their respective healthy control groups, both non-overweight and overweight patients with MASLD exhibited a slightly increased BMI. Although the average levels of transaminases in non-overweight patients with MASLD were within the normal range, both non-overweight and overweight patients with MASLD demonstrated elevated levels of alanine aminotransferase (ALT), aspartate aminotransferase (AST), alkaline phosphatase (AKP), γ-glutamyltransferase (γ-GT), and cholinesterase (CHE) compared with healthy controls (Table 3).

Complete blood count analysis showed that both the lean and overweight MASLD groups had significantly elevated levels of hemoglobin (HGB), hematocrit (HCT), red blood cell count (RBC), and white blood cell count (WBC) (Table 2). Overweight patients with MASLD showed the most significant increases in these indicators. Additionally, individuals with MASLD exhibited markedly elevated fasting blood glucose concentrations compared with healthy controls, suggesting a poorer metabolic status. Although platelet counts were markedly elevated in overweight MASLD patients compared with healthy controls, the differences between the lean and overweight MASLD groups and lean healthy controls were not statistically significant.

Lipid profile analysis (Table 4) revealed significant differences in high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), triglyceride (TG), and apolipoprotein A1 (APOA1) levels among the four groups. This study excluded patients with hyperlipidemia and hypertriglyceridemia; however, both the overweight and lean disease groups exhibited substantially elevated LDL-C and TG concentrations compared with the healthy control group, while HDL-C and APOA1 levels were relatively lower. Furthermore, although total cholesterol (TC) and apolipoprotein B (APOB) levels in both the lean and overweight MASLD groups were elevated compared with their respective healthy control counterparts, these differences were not statistically significant.

A substantial number of observational studies have shown that the incidence of chronic kidney disease (CKD) is markedly higher in patients with MASLD than in those without the condition (34, 35). In contrast to matched healthy controls, individuals with lean MASLD demonstrated substantially heightened concentrations of uric acid (UA) and creatinine (Cr), but only UA concentrations were notably increased in overweight patients with MASLD (Table 3). Several studies have reported that creatinine is a contributing factor to MASLD (36, 37). Our study also observed statistically meaningful variations in creatinine levels between overweight and lean cohorts of patients with MASLD.

A total of 110 lipoprotein subfraction concentrations and 31 metabolite concentrations were quantified using NMR and data filtering. Lipoprotein subclasses and metabolites with a fold change (FC) > 1.2 or FC < 0.85 (p < 0.05) between any two groups were considered differential markers (Tables 7–13). Differentially abundant substances were identified across the BMI groups between the MASLD and control groups. Nine unique lipoprotein subclasses were identified in the lean MASLD group, 32 in the overweight MASLD group, and 39 shared differential lipoproteins between the two groups. (Figure 2D). Specific details are provided in Tables 7–13. Specifically, Tables 7–9 summarize the lipoprotein subclasses unique to the lean and overweight groups, respectively. Tables 10–13 list the lipoprotein subclasses common to both groups, presenting their fold-change (FC) values and direction of change in the disease groups.

We first compared the lipoproteins and metabolites between the lean MASLD and lean control groups, identifying 48 lipoproteins (33 upregulated and 15 downregulated in the lean MASLD group) (Figure 2B) and eight metabolites with significant differences (Supplementary Figure 2B). The lipoproteins that were elevated in the lean MASLD group primarily consisted of the VLDL and IDL subclasses, whereas the downregulated lipoprotein subclasses were mainly HDL and LDL-2. Changes in lipoprotein profiles between the MASLD and healthy groups are shown in a heat map (Supplementary Figure 1A).

When comparing lipoprotein subclasses and metabolites between the overweight MASLD and overweight control groups, 71 differential lipoproteins (Figure 2A) and 18 differential metabolites (Supplementary Figure 2A) were identified. Among these, 51 upregulated lipoproteins were predominantly associated with LDL, VLDL, and IDL subclasses, whereas 20 downregulated lipoproteins were primarily linked to HDL and smaller, denser LDL particles. Changes in lipoprotein profiles are illustrated in Supplementary Figure 1B.

To explore the biological differences between lean and overweight patients with MASLD, statistical analyses were conducted for the two groups. According to the volcano plot results, 53 lipoproteins (35 downregulated and 18 upregulated in lean MASLD) (Figure 2C) and six differential metabolites (Supplementary Figure 2C) were identified, indicating that lean MASLD represents a distinct pathological condition compared to overweight MASLD. Additionally, cluster heatmap analysis based on FC > 1.2 or FC < 0.85, effectively differentiated the two groups (Supplementary Figure 1C).

To assess the contribution of significantly different lipoproteins to the subgroups, we performed a random forest analysis of the lipoprotein subclasses across the three groups. The contributions of lipoprotein subclasses are illustrated in an importance plot (Figure 3). The model scores were based on the mean decrease in the Gini impurity, with a higher mean value indicating a greater contribution of each feature to the model’s classification accuracy. To evaluate the robustness of the model during the random forest feature selection process and to mitigate the risk of overfitting, we utilized the algorithm’s built-in out-of-bag (OOB) error estimation as an internal validation method. The final model achieved OOB-estimated accuracies of 76.92% (error rate 23.08%) for the lean MASLD group versus the lean control group and 79.01% (error rate 20.99%) for the overweight MASLD group versus the overweight control group (Supplementary Table 1; Supplementary Figure 9). These results demonstrate that, even without introducing an external independent cohort, the model exhibited acceptable generalizability through its intrinsic validation mechanism during the training process.

In the analysis distinguishing the lean MASLD group from the control group, 39 lipoprotein parameters had a contribution score > 1. Among these, nine differential lipoproteins were unique to the lean group, primarily H1A1, H1PL, H1FC, H1CH, H2FC, and H2A2 from the HDL subclasses, V5TG and V5PL from the VLDL subclasses, and L6TG from the LDL subclasses.

In the analysis comparing the overweight MASLD group to the overweight control group, 45 lipoprotein parameters had a contribution score > 1, encompassing the LDL, VLDL, HDL, and IDL subclasses. Among these, 22 differential biomarkers were unique to the overweight group and primarily consisted of LDL subclasses, including LDTG, L1PN, L1CH, L1FC, L1PL, L1AB, L2TG, L3AB, L5PL, L5TG, L5FC, L5CH, L5AB, L5PN, L6FC, L6AB, L6PL, L6PN, and L6CH. HDL subclasses included H4A2, H4PL, and H2A1.

We conducted receiver operating characteristic (ROC) curve analysis of the clinical lipid testing parameters and found a low diagnostic rate (Supplementary Figure 6). To identify clinically relevant biomarkers, we performed ROC analysis on differential lipoprotein subclasses using an AUC > 0.7, to evaluate the diagnostic capability of candidate biomarkers, ultimately identifying 29 lipoproteins that met the criteria. Figure 4 depicts lipoprotein subclasses with an AUC > 0.75. The remaining subclasses are shown in Supplementary Figure 8. Violin plots of lipoprotein subclasses are shown in Supplementary Figure 3. In order to verify the stability of these 29 lipoprotein subclasses, we conducted Bootstrap analysis, and the results showed that they had high stability (Supplementary Figure 10). In the context of overweight MASLD, lipoprotein V5TG, which is specific to lean MASLD, was elevated in the lean MASLD group. Additionally, the concentrations of HDL subclasses H1FC, H1PL, H1A1, H2FC, and H2A2 decreased in the disease group. Interestingly, although most HDL subclasses exhibited reduced concentrations in the disease group, the H4TG concentration increased. Furthermore, the differential lipoproteins displayed in both the lean and overweight groups exhibited a heightened diagnostic value only in the lean group.

In the overweight group, we identified 37 lipoproteins with an AUC > 0.7 (Figure 5; Supplementary Figure 9), of which 11 demonstrated high diagnostic capability (AUC > 0.8) (Figure 5). Changes in the concentrations of lipoprotein subclasses in the disease group compared with those in the healthy group are shown in Supplementary Figure 4. Bootstrap analysis of 37 lipoproteins also showed that these lipoproteins also had high stability (Supplementary Figure 11). Upregulated lipoproteins unique to the overweight group included ABA1, LDTG, IDPL, L1PN, L1CH, L1FC, L1PL, L1AB, L2TG, and L3TG, whereas downregulated lipoproteins included L5CH, L5FC, L5PL, L6CH, L6FC, L6PL, and H3PL.

ROC curve analysis was similarly conducted for the overweight and lean MASLD groups, resulting in the identification of eight lipoprotein subclasses with good diagnostic performance (AUC > 0.7) (Supplementary Figure 5).

Significant differences in lipoprotein levels with respect to disease severity and obesity were supported by their correlations with complete blood counts as well as liver and kidney function test parameters (Figure 6; Supplementary Table 2). We selected lipoprotein subclasses with AUC > 0.7 and parameters exhibiting statistical differences in blood count and liver and kidney function tests for correlation analysis. The results indicated that in the lean group, HDL subclasses were negatively correlated with UA and fasting blood glucose (FBG), whereas VLDL subclasses were positively correlated with UA, blood urea nitrogen (BUN), Cr, FBG, eosinophils, and ALT, and negatively correlated with total protein (TP). In the overweight group, only the VLDL subclasses were positively correlated with ALT levels.

We summarized the lipoproteins with a contribution score > 1 from the random forest analysis, as well as the lipoprotein subclasses with an AUC > 0.7 from the ROC analysis, and identified several specific lipoprotein subclasses that differentiated the lean group from the overweight group (Figures 7,8). Based on the uniqueness of these lipoproteins, we propose that lean and overweight patients with MASLD have different pathogenic mechanisms. The results revealed that lipoproteins in the lean MASLD group were primarily concentrated in the HDL and VLDL subclasses, whereas those in the overweight group were primarily comprised of LDL and VLDL. We hypothesized that the unique pathogenic mechanism of lean MASLD may be related to its difficulty in forming HDL-associated subclasses, whereas overweight MASLD may progress owing to the excessive generation of LDL subclasses. Moreover, the concentrations of LDL-5 and LDL-6 were reduced in the disease group, and their densities were similar to those of HDL, leading us to speculate that their functions may be similar to those of HDL subclasses. Our findings suggest significant differences in the lipoprotein composition between these groups, which may have important implications for diagnosis and therapy.

Participants were enrolled from Putuo Hospital, which is affiliated with Shanghai University of Traditional Chinese Medicine in China, from March 2019 to June 2024. In total, 2013 eligible subjects were included in this study. MASLD was diagnosed based on the presence of significant hepatic steatosis detected by abdominal ultrasound in the absence of other causes of fatty liver disease according to the “Guidelines for Prevention and Treatment of Metabolism related (Non alcoholic) Fatty Liver Disease (2024 Edition)” established by the National Seminar on Fatty Liver and Alcoholic Liver Disease of the Chinese Medical Association and the Expert Committee of Fatty Liver of the Chinese Medical Doctor Association (2024, China). To minimize subjective bias, each examination in this study was performed by two ultrasound physicians. The study protocol was approved by the Medical Ethics Committee, and all participants provided written informed consent.

Venous blood samples were obtained following an overnight fast of no less than 12 h. The participants underwent a comprehensive physical examination, which included assessment of vital signs and anthropometric measurements. BMI was calculated by dividing the weight in kilograms by the square of the height in meters (kg/m²). Comprehensive medical history was collected using questionnaires. Laboratory tests included routine blood tests, fasting blood glucose levels, lipid profiles, biochemical measurements of liver and kidney function, and thyroid function tests.

The metabolomic sub-cohort was selected via stratified random sampling using SPSS (version 25). After stratifying the cohort by BMI and disease status, random numbers were generated within each stratum to draw proportionally balanced samples. Serum samples from 113 overweight MASLD patients, 107 overweight healthy individuals, 83 lean MASLD individuals, and 99 lean healthy controls were subsequently analyzed using metabolomic profiling performed in accordance with the Bruker IVDr standard operating procedure (SOP) (53). Serum samples were collected on a 600 MHz Avance NEO NMR spectrometer (Bruker Biospin GmbH, Rheinstetten, Germany) equipped with a BBI probe and a Samplejet autosampler with the temperature set at 4 °C. prior to sample collection, the instrument was first subjected to quality control experiments (B. I. BioBank QC™) in order to ensure consistent spectral quality. For each sample, solvent presaturation 1D 1H experiments were performed at 310 K using the NOESY pulse sequence, 32 scans were performed, totaling 98,304 data points covering a spectral width of 17,857 Hz. Corrections for baseline and phase were performed in Topspin 4.1.1 after data acquisition was completed.

Statistical analyses of clinical data were conducted using a two-tailed significance threshold of p < 0.05. Normality was evaluated using the Shapiro–Wilk test. Data that were normally or approximately normally distributed are expressed as mean ± standard deviation, while severely skewed distributions are reported as median (5th and 75th). For comparisons between multiple groups of non-normally distributed quantitative data, the Kruskal–Wallis H test was used, and the Bonferroni correction was applied to account for multiple comparisons of skewed data.

Before statistical analysis, metabolomic data were preprocessed to remove and impute missing values. Lipoprotein subclasses and metabolites with more than 30% of missing data were excluded. To address potential inter-batch variation during NMR detection and to ensure the effectiveness of our quality control (QC) protocol, we implemented a rigorous procedure. For each analytical batch, a QC experiment was first performed to verify instrument performance. Sample analysis proceeded only after the instrument passed this QC check. Furthermore, a detailed QC report was generated for every individual sample analyzed. Significant lipoprotein subcomponents and metabolites were identified by differential fold analysis, t-tests, and volcano plots. In all cases, statistical significance was set at p < 0.05. Considering the characteristics and nonlinear relationship of high-dimensional data, we use random forest as a tool for feature screening instead of building a prediction model. Random forest analysis was performed on the lipoprotein subcomponents data using “mean decrease Gini” as the metric to assess each variable’s contribution to each group, while the out-of-bag (OBB) error estimate was employed to assess model stability. A higher mean decrease in the Gini coefficient indicated a greater contribution of a feature to the classification accuracy of the model. Pearson’s correlation analysis was conducted to investigate the association between metabolites and clinical parameters. The area under the ROC curve (AUC) was calculated to assess the diagnostic efficacy of all the candidate biomarkers for MASLD. Bootstrap resampling stability analysis was utilized to assess the stability of the candidate biomarkers.

Data analysis and visualization were performed using SPSS v.26, R v.4.3.2, and GraphPad Prism v.9.0.

Our study examined both male and female participants, and sex was not considered a biological variable. These findings are not expected to be differentially relevant between males and females.