Graphical Abstract

   [41]ga1
   [42]Open in a new tab

1. Introduction

   Non-alcoholic fatty liver disease (NAFLD) is the most common chronic
   liver disease globally, affecting 25–33% of the world’s population
   [43][1]. NAFLD encompasses a wide histological spectrum ranging from
   simple steatosis (SS), which often has a benign course, to
   non-alcoholic steatohepatitis (NASH), the advanced form with a much
   greater risk of progression into fatal hepatic complications such as
   cirrhosis and hepatocellular carcinoma [44][2]. Furthermore, NASH is a
   leading cause of liver transplantation and is closely associated with a
   high incidence of cardiovascular mortality independent of traditional
   risk factors [45][3]. Despite the rapidly increasing prevalence of
   NAFLD and NASH, there is currently no approved pharmacotherapy for this
   disease. Early detection of high-risk individuals, followed by
   lifestyle interventions, remains the most effective strategy.

   Various imaging techniques, including ultrasonography, computed
   tomography (CT), and magnetic resonance imaging (MRI), are currently
   used to measure hepatic fat content. However, these imaging methods
   cannot differentiate SS from NASH, which is typically characterized by
   not only steatosis but also hepatocyte ballooning and lobular
   inflammation [46][2]. Currently, both the diagnosis and staging of NASH
   rely heavily on the histological evaluation of liver biopsies, a
   protocol that has many pitfalls, including sampling error, high cost,
   invasiveness, risk of complications, and inter-observer variation among
   pathologists. [47][4] Although several circulating biomarkers have been
   identified as potential tools for the non-invasive diagnosis and/or
   prognosis of NASH, none of them have been clinically implemented at
   present.

   Metabolic disturbance is a primary driver for the onset and progression
   of NAFLD. NAFLD often co-occurs with obesity and other cardiometabolic
   disorders such as insulin resistance, type 2 diabetes, dyslipidemia,
   hypertension, and atherosclerosis [48][5]. Recognizing its close
   association with metabolic disorders, international experts have
   suggested renaming it to metabolic associated fatty liver disease
   (MAFLD) [49][6], and more recently, to metabolic dysfunction associated
   steatotic liver disease (MASLD) [50][7]. A large number of both
   targeted and untargeted metabolomic studies have been performed to
   identify metabolic biomarkers associated with the pathophysiology of
   NAFLD and NASH [51][8], [52][9], [53][10], [54][11]. Distinct
   alterations in amino acids, glutathione metabolism, and various lipid
   species in NAFLD and NASH have been reported in different clinical
   cohorts [55][12]. Although these studies provided insightful evidence
   of dysregulated metabolic pathways in NAFLD and NASH, there are large
   variations and even inconsistencies due to limited sample size,
   differences in the study population, and varying diagnostic criteria.
   Few studies have been performed in the Asian population with liver
   biopsy-confirmed NAFLD and NASH.

   In this study, we aimed to leverage our unique clinical cohort of obese
   Chinese, which included the entire disease spectrum from normal liver
   to SS to borderline NASH to NASH, to systematically search for
   metabolites and lipid species associated with different stages of the
   disease and to explore the possibility of using metabolites and lipid
   biomarkers for predicting the risk of NAFLD and NASH in the obese
   population.

2. Experimental procedures

2.1. Study population

   Five hundred and nine patients with severe obesity were recruited at
   Huaqiao Hospital, affiliated with Jinan University, Guangzhou, China,
   during the period from January 6, 2017, to October 23, 2020. They were
   then further screened for eligibility (see [56]Supplementary methods).
   A total of 250 subjects spanning the full spectrum of NAFLD were
   finally enrolled and subjected to untargeted metabolomic and lipidomic
   profiling. Demographic, clinical, and laboratory data (after an
   overnight fast of at least 10 h) for each patient were collected within
   one week prior to surgery, and liver biopsies were obtained from the
   middle of the right lobes during the bariatric surgery.

2.2. Histological assessment of liver biopsies

   Thin sections of liver biopsy specimens were stained with H&E or Masson
   trichrome and were evaluated by three independent liver pathologists as
   described previously (Qiu H, CMGH, 2022). The scores of steatosis,
   hepatocellular ballooning, lobular inflammation, and fibrosis were
   evaluated according to the NASH Clinical Research Network scoring
   system [57][13]. The definitions of histologically normal liver, SS,
   borderline NASH, and NASH were based on the Fatty Liver Inhibition of
   Progression (FLIP) algorithm [58][14].

   Further experimental procedures (e.g., metabolomics, lipidomics, and
   statistical analysis) are provided in the [59]supplementary methods.

3. Results

3.1. Distinct clinical characteristics in subjects with normal liver, SS, and
NASH

   The clinical characteristics of the study cohort (n = 250) are shown in
   [60]Table 1, including 44 obese individuals with normal livers and 196
   NAFLD patients (55 SS, 75 NASH, and 76 borderline NASH). There were no
   significant differences in the prevalence of different NAFLD stages
   between males and females (p > 0.05, Chi-square test). We identified 17
   out of 20 metabolic-related parameters that were different between
   normal liver and NAFLD (p < 0.05, Student t-test). As expected, the
   levels of body mass index (BMI), alanine transaminase (ALT), aspartate
   aminotransferase (AST), insulin, and Homeostatic Model Assessment for
   Insulin Resistance (HOMA-IR) increased progressively from normal liver
   to SS and to NASH. Another seven parameters, including Apolipoprotein A
   (APOA), Apolipoprotein B (APOB), systolic blood pressure (SBP),
   C-Reactive Protein (CRP), uric acid, amylase, and haemoglobin A1C
   (HbA1c), were higher in SS compared to normal liver, whereas the
   differences between the SS and NASH groups were not significant.
   Furthermore, high-density lipoprotein (HDL) was higher in SS compared
   to NASH, and the level of total protein was highest in SS compared to
   both normal liver and NASH.

Table 1.

   Clinical characteristics in different stages.
   Characteristics Total Population Normal NAFLD
     __________________________________________________________________

   P value
     __________________________________________________________________

   Total SS NASH Normal vs NAFLD Normal vs SS SS vs NASH
   Male (%) 107 (41.3%) 12 (27.2%) 95 (44.2%) 27 (48.2%) 31 (39.2%) 0.056
   0.054 0.389
   APOB (g/L) 1.04 ± 0.23 0.92 ± 0.2 1.06 ± 0.22 1.01 ± 0.21 1.08 ± 0.23
   < 0.001 0.034 0.1
   Age (years) 30 (25, 36) 30 (25, 39) 30 (24.5, 36) 30.5 (25.25, 35.25)
   30 (23, 35) 0.418 0.634 0.689
   BMI (kg/m2) 38.33 (33.58, 43.62) 34.15 (31.8, 38.5) 39.27 (34.71,
   44.86) 37.95 (33.1, 41.82) 41.2 (36.25, 46.96) < 0.001 0.013 0.008
   Systolic BP (MmHg) 126 (118.5, 137) 120 (113.75, 128.25) 127 (120, 138)
   128 (119, 139) 125 (119.5, 134) < 0.001 0.004 0.323
   Diastolic BP (MmHg) 80 (72, 89) 75.18 (71.75, 84.25) 81 (74, 90) 80
   (72, 87) 82 (74, 89) 0.012 0.22 0.257
   CRP (mg/L) 5.68 (3.01, 9.01) 3.72 (1.35, 6.89) 5.95 (3.48, 10.8) 5.28
   (3.03, 8.61) 7.25 (3.86, 11.95) < 0.001 0.025 0.069
   Uric Acid (μmol/L) 453.2 (376.05, 541.2) 402.02 (357.48, 473.2) 462
   (382.3, 554.35) 460.35 (385, 539.07) 465 (379.65, 567) 0.003 0.041
   0.527
   Amylase (U/L) 41 (32.7, 51.2) 49.5 (39, 61.23) 39.7 (31.85, 48.6) 40
   (36.22, 48.75) 40 (32.15, 46.5) < 0.001 0.011 0.385
   ALT (U/L) 43 (24, 68) 22.7 (17, 30.25) 50 (29.5, 71) 39 (23, 53.75) 55
   (35, 74) < 0.001 < 0.001 0.004
   AST (U/L) 25 (19, 37) 20 (16, 22.25) 28 (20, 39.5) 23 (18.75, 30.25) 32
   (22, 46) < 0.001 0.004 0.002
   Total_Protein (g/L) 72.36 (67.8, 75.45) 71.2 (66.8, 73.82) 72.6 (68.1,
   76.05) 73.65 (70.06, 76.65) 71.1 (67.65, 75.45) 0.035 0.003 0.043
   Total Cholesterol (mmol/L) 4.93 (4.5, 5.5) 4.89 (4.32, 5.36) 4.95
   (4.51, 5.52) 4.9 (4.4, 5.48) 5.03 (4.58, 5.46) 0.242 0.567 0.453
   Total Triglycerides (mmol/L) 1.78 (1.32, 2.47) 1.34 (1.08, 1.91) 1.82
   (1.39, 2.65) 1.69 (1.26, 2.18) 1.81 (1.41, 2.54) 0.002 0.114 0.235
   HDL (mmol/L) 0.98 (0.86, 1.15) 1.04 (0.95, 1.24) 0.97 (0.84, 1.1) 1
   (0.92, 1.18) 0.91 (0.82, 1.08) 0.001 0.134 0.021
   LDL (mmol/L) 3.01 (2.66, 3.46) 2.77 (2.48, 3.3) 3.04 (2.7, 3.47) 2.97
   (2.65, 3.43) 3.1 (2.7, 3.52) 0.038 0.29 0.391
   APOA (g/L) 1.25 (1.12, 1.37) 1.32 (1.21, 1.5) 1.23 (1.1, 1.35) 1.22
   (1.03, 1.35) 1.23 (1.12, 1.33) 0.013 0.031 0.815
   Insulin (mIU/L) 19.89 (13.3, 29.15) 12.5 (9.02, 18.08) 21.41 (15.02,
   30.11) 18.59 (13.77, 24.17) 25.66 (18.53, 33.68) < 0.001 0.005 0.001
   HbAIc (%) 5.8 (5.4, 6.45) 5.4 (5.2, 5.9) 5.8 (5.5, 6.67) 5.8 (5.47,
   6.23) 5.87 (5.6, 6.67) < 0.001 0.006 0.241
   Fasting Glucose (mmol/L) 5.47 (4.94, 6.56) 5.15 (4.88, 5.62) 5.52
   (4.96, 6.72) 5.42 (4.96, 6.32) 5.66 (5.22, 6.66) 0.007 0.114 0.089
   HOMA-IR 5.25 (3.29, 8.25) 3.09 (1.93, 5.31) 5.65 (3.82, 8.58) 4.48
   (3.05, 6.57) 6.65 (4.68, 10.37) < 0.001 0.008 < 0.001
   [61]Open in a new tab

   Data are presented as mean ± SD or median (interquartile range)

   APOB, Apolipoprotein B; BMI, Body mass index; BP, Blood pressure; CRP,
   C-reactive protein; ALT, Alanine aminotransferase; AST, Aspartate
   aminotransferase; HDL, High-density lipoprotein; LDL, Low-density
   lipoprotein; APOA, Apolipoprotein A; HbA1c, Haemoglobin A1C; HOMA-IR,
   Homeostatic Model Assessment for Insulin Resistance.

3.2. Differentially changed metabolites and lipid species across different
histological stages of NAFLD

   In total, 263 metabolites and 550 lipid species from 12 lipid classes
   were detected in serum samples. Comparison between normal liver and SS
   showed 12 metabolites and 63 lipid species enriched in normal liver,
   while 9 metabolites and 83 lipid species were elevated in SS (p < 0.05,
   MaAsLin2, [62]Fig. 1A & [63]Table S1). Meanwhile, a comparison between
   SS and NASH identified 17 metabolites and 78 lipid species enriched in
   SS and 13 metabolites and 5 lipid species increased in NASH patients,
   respectively (p < 0.05, MaAsLin2, [64]Fig. 1B & Suppl. [65]Table 1).
   Further analysis across the entire NAFLD spectrum demonstrated several
   lipid species and metabolites were progressively altered across disease
   severity (p < 0.05, ordinal regression, Suppl. [66]Table 2).
   Phosphatidic acid (PA), 12 phosphatidylcholine (PC), 8
   phosphatidylethanolamine (PE), 5 phosphatidylglycerol (PG), 3
   phosphatidylinositol (PI), 6 sphingomyelin (SM), and tridecylic acid
   displayed stepwise decreases from normal liver to SS to NASH ([67]Fig.
   2A). In contrast, 10 triglyceride (TG) species, 1 PE, xanthine,
   L-valine, creatine, and N-palmitoyltaurine exhibited progressive
   elevations from normal liver to SS to NASH ([68]Fig. 2B).
   Interestingly, L-cystine, a precursor for the synthesis of glutathione,
   an important antioxidant, showed a significant decrease in SS compared
   to both normal liver and NASH ([69]Fig. 2C), while another 2 PC and 1
   PG were upregulated in SS ([70]Fig. 2D).

Fig. 1.

   [71]Fig. 1
   [72]Open in a new tab

   Significantly changed metabolites and lipid species from Normal to SS
   and from SS to NASH. The volcano plot shows the significant metabolites
   and lipid species in the comparisons between normal liver and simple
   steatosis (A) and between simple steatosis and NASH (B). Variables with
   adjusted q value higher than 0.25 are labelled in text. The two
   horizontal dotted lines represent the two cut-offs for significance
   (p = 0.05, and q = 0.25). Variables coloured in dark green (q < 0.25)
   and light green (p < 0.05) represent the enrichment in normal liver (A)
   and simple steatosis (B), respectively. Variables coloured in red (q <
   0.25) and light yellow (p < 0.05) represent the enrichment in simple
   steatosis (A) and NASH (B), respectively.

Table 2.

   Detailed information of the enriched pathways at each stage. The
   enriched pathways and the corresponding.
   encoding metabolites are presented.
   Pathway Encoded_metabolites P_value Enrichment Type Comparisons
   Arginine biosynthesis Citrulline 0.044 Normal Metabolites Normal_vs_SS
   Sphingolipid signaling pathway Cer, SM 0.001 SS Lipids Normal_vs_SS
   Necroptosis Cer, SM 0.004 SS Lipids Normal_vs_SS
   Sphingolipid metabolism Cer, SM 0.014 SS Lipids Normal_vs_SS
   Valine, leucine and isolueucine biosynthesis L-Valine 0.021 SS
   Metabolites Normal_vs_SS
   Pantothenate and CoA biosynthesis L-Valine 0.048 SS Metabolites
   Normal_vs_SS
   Pantothenate and CoA biosynthesis Pantothenic acid, L-Valine 0.002 NASH
   Metabolites SS_vs_NASH
   Purine metabolism Xanthine, Allantoin 0.024 NASH Metabolites SS_vs_NASH
   Valine, leucine and isoleucine biosynthesis L-Valine 0.031 NASH
   Metabolites SS_vs_NASH
   GnRH signaling pathway PA 0.033 NASH Lipids SS_vs_NASH
   D-Arginine and D-ornithine metabolism D-Ornithine 0.008 SS Metabolites
   SS_vs_NASH
   Retrograde endocannabinoid signaling PC, PE 0.014 SS Lipids SS_vs_NASH
   Phosphatidylinositol signaling system PA, PI 0.027 SS Lipids SS_vs_NASH
   Biosynthesis of unsaturated fatty acids Eicosapentaenoic acid 0.049 SS
   Metabolites SS_vs_NASH
   [73]Open in a new tab

Fig. 2.

   [74]Fig. 2
   [75]Open in a new tab

   Comparisons of the significant metabolites and lipid species among
   different NAFLD stages. The bar chart shows the significant metabolites
   and lipids that are progressively decreased from normal liver to simple
   steatosis to NASH (A), that are progressively increased from normal
   liver to simple steatosis to NASH (B), that are decreased in simple
   steatosis compared to normal liver and NASH (C), and that are enriched
   in simple steatosis compared to normal liver and NASH (D).

3.3. Metabolomic/lipidomic pathway enrichment analysis for different stages
of NAFLD

   To delineate the potential metabolic mechanisms underlying the onset
   and/or progression of NAFLD, we performed pathway enrichment analysis
   using the significantly altered metabolites and lipid species in
   different disease stages. This analysis found that the pathway related
   to arginine biosynthesis was enriched in normal liver and SS, while
   pathways for D-arginine and D-ornithine metabolism were enriched in SS
   compared to NASH ([76]Fig. 3 & [77]Table 2). In contrast, the
   sphingolipid signaling pathway, sphingolipid metabolism, and
   necroptosis were enriched in SS compared to normal liver, suggesting
   the potential involvement of these pathways in triggering the
   initiation of NAFLD. Notably, the metabolism of purine encoded by
   xanthine and allantoin and the GnRH signaling pathway encoded by PA
   were uniquely enriched in NASH compared to SS. Retrograde
   endocannabinoid signaling, the phosphatidylinositol signaling system,
   and biosynthesis of unsaturated fatty acids were enriched in SS
   compared to NASH. Importantly, valine, leucine, and isoleucine
   biosynthetic pathways and pantothenate and CoA biosynthesis pathways
   were progressively enriched from normal liver to SS to NASH, indicating
   that increased branched-chain amino acids and pantothenic acid are key
   contributors to the pathogenesis of NAFLD.

Fig. 3.

   [78]Fig. 3
   [79]Open in a new tab

   Pathway enrichment analysis for different histological stages of NAFLD
   using the significant metabolites and lipid species. The bar charts
   show the enriched pathways in each NAFLD stage with the significant
   metabolites and lipid species in the comparisons between normal and
   simple steatosis (A) and between simple steatosis and NASH (B). The
   length of the bar represents the p value of each pathway. The colour of
   the bar represents the type of pathway. The stages of enrichment are
   labelled in text.

3.4. Co-expression network analysis of lipid species in relation to NAFLD and
its associated clinical parameters

   In light of the complexity of dynamic changes in lipid species and
   metabolites during NAFLD progression, we created co-expression networks
   for metabolites and lipid species, respectively. Eight co-occurrence
   modules represented in different colors were constructed successfully
   among 481 lipid species ([80]Fig. 4A). The remaining 78 lipid species
   failed to cluster and were marked in grey; they were not included for
   further analysis. The 8 modules were named according to the assigned
   color from module blue to module yellow. The number of lipid species
   among the 12 lipid categories in each module was included in [81]Fig.
   4B.

Fig. 4.

   [82]Fig. 4
   [83]Open in a new tab

   The co-expression clusters of lipid species and their correlation with
   NAFLD-related parameters. (A) the cluster dendrogram shows the
   constructed modules in different colours. The module coloured in grey
   represents the lipid species that failed to be grouped. (B) The heatmap
   on the left shows the correlation results between the eigenvalues of
   each module and the significant clinical variables. The rho value and
   the p value are labelled in each cell. The colour of each cell
   indicates the rho value. The significance of each parameter in the two
   comparisons is labelled with a star in the upper cell. The heatmap on
   the right shows the number of lipid species in each category for the
   corresponding modules.

   The correlation between modules and clinical parameters showed that
   module blue, mainly consisting of PC and SM species, and module pink,
   with PG species, were negatively correlated with the occurrence of
   NAFLD. These two modules also exhibited negative correlations with
   different metabolic risk factors for NAFLD, including APOB, ALT, AST,
   TG, Insulin, HbA1c, fasting glucose, and HOMA-IR, and positive
   correlations with amylase, suggesting that these two lipid clusters
   might be protective against NAFLD. Meanwhile, module red (mainly TG
   species) and module turquoise (mainly PC, PE, and PG species) were
   positively associated with the occurrence of NAFLD. Module red
   exhibited positive correlations with DBP, TG, insulin, HbA1, fasting
   glucose, and HOMA-IR, while module turquoise displayed positive
   associations with SBP, DBP, insulin, and HOMA-IR, suggesting that these
   two lipid clusters may contribute to NAFLD development.

   The same correlation analysis was performed using the metabolite
   profile ([84]Fig. S2). However, only 113 out of 263 metabolites
   (∼43.0%) can be grouped into two clusters. Module blue, containing 48
   metabolites, showed negative correlations with ALT, APOA, SBP, AST, and
   insulin and a positive correlation with amylase ([85]Table S5). By
   performing a pathway analysis with these 48 metabolites, the metabolism
   of glycine, serine, and threonine was significantly enriched. Further
   investigation showed that the abundance of these three metabolites was
   negatively correlated with both ALT and AST, suggesting the metabolism
   of glycine, serine, and threonine may inhibit the progression of NAFLD
   by suppressing the level of ALT and AST.

3.5. Integrated machine learning (ML) model for identifying important
metabolites and lipids features in NAFLD and NASH

   To further investigate the roles of circulating metabolites and lipid
   species in the onset and progression of NAFLD, we next attempted to
   construct classification models for differentiating NAFLD from normal
   liver and NASH from SS, respectively, using clinical parameters,
   metabolites, and lipids species. We used a stepwise model to select the
   best clinical variables for these two purposes (Suppl. [86]Table 3).
   For the comparisons between normal liver and NAFLD, ALT, APOA, APOB,
   direct bilirubin, and amylase were selected, while total protein (TP),
   AST, ALT, and insulin were identified for differentiating NASH from SS.
   Furthermore, 22 metabolites and 27 lipid species were selected for
   distinguishing NAFLD from normal liver, while 10 metabolites and 3
   lipid species were identified for differentiating NASH from SS using
   Boruta analysis ([87]Table S4).

Table 3.

   TG and PC species enriched in different modules with different
   contribution to NAFLD progression.
   TG species
     __________________________________________________________________

   PC species
     __________________________________________________________________

   Red Module Black Mdule Blue Module Turqoise Module
   TG(14:0/16:0/16:0) TG(16:0/18:2/22:6) PC(15:0/20:4(5E,8E,11E,14E))
   PC(10:0/26:2(5E,9Z))
   TG(14:0/16:0/18:1) TG(16:0/20:4/22:6) PC(17:0/0:0) PC(14:0/0:0)
   TG(14:0/16:1/18:1) TG(16:0/22:6/22:6) PC(O-16:1(7Z)/0:0) PC(16:0/0:0)
   TG(16:0/16:0/16:0) TG(16:1/18:2/18:3) PC(O-17:0/0:0) PC(16:1(7Z)/0:0)
   TG(16:0/16:0/18:1) TG(16:1/18:2/22:6) PC(O-18:0/0:0) PC(18:0/0:0)
   TG(16:0/16:0/18:2) TG(16:1/20:5/22:6) PC(O-24:0/0:0) PC(18:0/18:1(11E))
   TG(16:0/16:1/18:1) TG(17:0/22:6/22:6) PC(P-16:0/13:0) PC(18:1(11E)/0:0)
   TG(16:0/16:1/18:2) TG(18:0/22:5/22:6) PC(P-16:0/14:0)
   PC(18:1(11Z)/16:1(7Z))
   TG(16:0/18:0/16:0) TG(18:1/18:2/22:6) PC(P-16:0/16:0)
   PC(18:3(6Z,9Z,12Z)/0:0)
   TG(16:0/18:0/18:0) TG(18:1/20:4/22:6) PC(P-16:0/18:2(2E,4E))
   PC(18:3(9Z,12Z,15Z)/18:1(13Z))
   TG(16:0/18:0/18:1) TG(18:1/22:5/22:6) PC(P-16:0/20:3(5Z,8Z,11Z))
   PC(20:0/0:0)
   TG(16:0/18:0/22:0) TG(18:1/22:6/22:6) PC(P-16:0/20:4(5E,8E,11E,14E))
   PC(20:1(11E)/0:0)
   TG(16:0/18:1/20:4) TG(18:2/18:2/18:3)
   PC(P-16:0/22:5(4Z,7Z,10Z,13Z,16Z)) PC(22:0/0:0)
   TG(16:0/18:2/18:3) TG(18:2/18:2/22:6) PC(P-16:0/23:0)
   PC(22:6(4Z,7Z,10Z,13Z,16Z,19Z)/18:2(2E,4E))
   TG(16:0/18:2/20:4) TG(18:2/18:3/22:6) PC(P-16:0/24:0)
   PC(7:0/18:2(9Z,12Z))
   TG(16:1/16:1/18:2) TG(18:2/22:6/22:6) PC(P-16:0/26:1(5Z))
   PC(O-17:1(9Z)/0:0)
   TG(16:1/17:1/18:2) TG(18:3/20:5/22:6) PC(P-18:0/16:0)
   PC(O-18:1(11E)/0:0)
   TG(17:0/18:1/18:2) TG(20:4/20:5/22:6) PC(P-18:0/18:1(11E))
   PC(O-20:0/0:0)
   TG(18:0/18:1/20:4) TG(20:4/22:6/22:6) PC(P-18:0/18:2(2E,4E))
   PC(O-22:0/0:0)
   TG(18:1/18:1/18:3) TG(20:5/22:6/22:6) PC(P-18:0/20:3(5Z,8Z,11Z))
   PC(O-22:1(13Z)/0:0)
   TG(22:5/22:6/22:6) PC(P-18:0/22:2(13Z,16Z)) PC(O-24:1(15Z)/0:0)
   PC(P-18:0/22:5(4Z,7Z,10Z,13Z,16Z)) PC(P-16:0/2:0)
   PC(P-18:0/22:6(4Z,7Z,10Z,13Z,16Z,19Z)) PC(P-18:0/26:2(5E,9Z))
   PC(P-20:0/20:3(5Z,8Z,11Z)) PC(P-20:0/17:2(9Z,12Z))
   PC(P-20:0/20:4(5E,8E,11E,14E))
   PC(P-20:0/22:4(7Z,10Z,13Z,16Z))
   PC(P-20:0/22:5(4Z,7Z,10Z,13Z,16Z))
   PC(P-20:0/22:6(4Z,7Z,10Z,13Z,16Z,19Z))
   PC(P-20:0/24:4(5Z,8Z,11Z,14Z))
   [88]Open in a new tab

   By using the selected clinical features, metabolites, and lipid
   species, we constructed different logistic regression models for
   classifying differences between NAFLD and normal liver and between NASH
   and SS, respectively. DeLong tests were implemented to compare the area
   under the curve (AUC) among models using different panels. For the
   classification between normal liver and NAFLD ([89]Fig. 5A), the model
   integrating the five routine clinical parameters yielded the AUC of
   0.882, which increased significantly to 0.940 and 0.954 by addition of
   the selected metabolites and lipid species, respectively (p < 0.05,
   DeLong test). Furthermore, a model integrating the selected
   metabolites, lipid species, and clinical parameters achieved an 11.6%
   improvement in the AUC compared to the clinical model (p < 0.05, DeLong
   test). For stratification between NASH and SS, the model built using
   the four clinical parameters (TP, AST, ALT, and insulin) yielded an AUC
   of 0.785, which was further improved to 0.868 and 0.893 by adding the
   selected metabolites and lipid species, respectively (p < 0.05, DeLong
   test, [90]Fig. 5B). When adding both the selected metabolites and lipid
   species in combination, the AUC of the final model significantly
   improves to 0.915.

Fig. 5.

   [91]Fig. 5
   [92]Open in a new tab

   Establishment of machine learning models for risk prediction of NAFLD
   and NASH in obese individuals. (A) Receiver operating characteristic
   (ROC) curves for classifying normal liver and NAFLD using 5 clinical
   parameters (ALT, APOA, APOB, direct Bilirubin, and amylase), 5 clinical
   parameters + selected metabolites, 5 clinical parameters + selected
   lipid species, and 5 clinical parameters + selected metabolites and
   lipid species, respectively. (B) ROC curves for the differentiation of
   NASH from simple steatosis using 4 clinical parameters (total protein,
   AST, ALT, and insulin), 4 clinical parameters + selected metabolites or
   lipid species, and 4 clinical parameters + selected metabolites and
   lipid species, respectively.

4. Discussion

   This study revealed the lipidomic and metabolomic biomarkers associated
   with the onset and progression of NAFLD in a cohort of 250 obese
   Chinese individuals with the full histological spectrum of NAFLD. The
   main findings include: (i) the identification of previously overlooked
   bioactive lipids such as N-palmitoyltaurine, which increase
   progressively with NAFLD severity, and (ii) a distinct correlation of
   different circulating TG and PC species with disease severity as
   determined by co-expression network analyses: saturated and
   monounsaturated TGs correlate positively with NAFLD, whereas a negative
   association links PC plasmalogens to NAFLD. Furthermore, we have built
   and compared multiple ML models classifying differences between NAFLD
   and normal liver and between NASH and SS to identify important
   metabolic and lipidomic features associated with significant
   improvement in disease prediction compared to conventional clinical
   parameters.

   In the last decade, the lipidomic and metabolomic characterization of
   histological stages of liver biopsy-proven NAFLD was accomplished
   mainly in clinical cohorts of Caucasian ethnicity [93][15]. However,
   extrapolation to non-Caucasian individuals is questionable as an
   ethnic-specific metabolic phenotype is generally expected. In this
   regard, it is well established that ethnicity intrinsically influences
   body fat content and distribution [94][16]. Herein, Asians exhibit a
   greater amount of body fat and visceral adipose tissue (VAT) than
   Caucasians at the same BMI [95][16]. In turn, VAT is an independent
   predictor of liver inflammation and fibrosis, leaving definite traces
   on the serum lipidome [96][17]. The potentially distinct metabolic
   phenotype of the Chinese population, along with the projected increase
   in the prevalence of NASH in China, [97][18] highlights the need for a
   large cohort study assessing the lipidomic and metabolomic profiles of
   well-characterized NAFLD Chinese patients. Thus far, scarce evidence is
   available in this population, with only a few small cohorts in which
   serum metabolites and circulating bile acids have been described
   [98][19].

   In the present study, widespread lipidomic between-group differences
   were detected in NAFLD Chinese patients. Lipidomic changes, which
   included increases in the majority of TG species, were of greater
   magnitude in individuals with normal liver versus patients with SS than
   between the former and NASH patients. This is consistent with recently
   reported changes in liver and plasma lipidomes at the early phase of
   NAFLD, which were relatively stable in the transition from SS to NASH
   [99][12]. Moreover, higher phosphatidylinositol 4,5-bisphosphate (PIP2)
   was observed in SS patients compared to controls. PIP2 is the precursor
   of phosphatidylinositol 3,4,5-trisphosphate (PIP3), the second
   messenger in the phosphatidylinositol-3 kinase (PI3K)/protein kinase B
   (Akt) pathway. The increased level of PIP2 contributed to the lower
   abundance of PIP3, which attenuates PI3K signaling and thus induces
   insulin resistance, an important driving factor for NAFLD development
   [100][20]. Unexpectedly, omega-3 eicosapentaenoic acid (EPA) was
   elevated in SS patients relative to normal liver individuals.
   Interestingly, supplementation with EPA has been associated with
   decreased accumulation of hepatic TG via increasing hepatic β-oxidation
   and reducing de novo lipogenesis [101][21]. Therefore, elevated EPA may
   represent a compensatory response to defend the early progression from
   normal liver to SS.

   Another notable observation is the significantly outnumbered changes in
   non-lipid metabolites between SS and NASH compared to the predominant
   alternations in lipid species between normal liver and SS.
   Consistently, a recent study in the Caucasian population found that
   hepatic lipidomic remodeling occurs predominantly in obese individuals
   with steatosis but not with non-alcoholic steatohepatitis. These
   findings suggest that the dynamics of metabolic alteration starts with
   lipidomic species during the progression from normal liver to SS but
   shifts to other metabolites in the later stages of NAFLD development
   [102][12]. Specifically, our results found significantly higher levels
   of L-cystine and citrate in NASH than in SS patients. In this respect,
   oxidative stress is a well-established pathogenic insult that drives
   the progression of NAFLD to NASH [103][22]. Increased L-cystine may
   reflect a compensatory mechanism to counteract the production of ROS by
   stimulating the synthesis of glutathione. In contrast, citrate
   contributes to the formation of hydrogen peroxide in the presence of
   iron [104][23], suggesting that pro-oxidant and antioxidant mechanisms
   may be activated in the progression to NASH. Futhermore,
   1-stearoyl-sn-glycerol 3-phosphocholine (18:0-LPC) and
   1-stearoyl-sn-glycerol were decreased in NASH compared to SS. In line
   with our findings in patients, a previous study found decreased levels
   of 18:0-LPC in mice with NASH, resulting from hepatic up-regulation of
   lysophosphatidylcholine acyltransferase (LPCAT)1–4 [105][24].
   Meanwhile, we observed a significant alteration in
   1-stearoyl-sn-glycerol, the precursor of 18:0-LPC, suggesting a
   possible involvement of this pathway in NASH development.

   Our pathway enrichment analysis found that branched-chain amino acid
   (BCAA) signaling pathways were progressively upregulated with
   increasing disease severity, mainly due to the increased levels of
   L-valine. Consistently, plasma levels of BCAAs (e.g., isoleucine and
   valine) were shown to be closely associated with the degree of liver
   inflammation and ballooning in NAFLD/NASH patients [106][25]. Increased
   circulating BCAAs in NAFLD have been attributed to excessive BCAA
   intake, impairment in BCAA catabolism, as well as changes in the
   composition and function of the gut microbiota. [107][26] Of particular
   interest is the stepwise increase in N-palmitoyltaurine, a bioactive
   lipid that belongs to the family of N-acyl taurine (NAT). A previous
   study reported that the members of this bioactive lipid family regulate
   insulin secretion in pancreatic β-cells [108][27]. Yet, there is only
   one study that linked N-palmitoyltaurine to asthma severity [109][28].
   Further investigations are justified to address the potential role of
   N-palmitoyltaurine in the pathogenesis of NAFLD in both animals and
   humans. Moreover, we identified tridecylic acid, a medium-chain fatty
   acid, was progressively decreased from normal liver to SS to NASH.
   Tridecylic acid has been shown to activate the immunomodulating
   receptor G protein-coupled 84 (GPR84) to suppress lipotoxicity-induced
   hepatic fibrosis [110][29] Therefore, further studies are warranted to
   explore whether supplementation with tridecylic acid ameliorates NAFLD
   development.

   Although the accumulation of TG in hepatocytes is generally regarded as
   the main hallmark of NAFLD, our results from co-expression network
   analyses demonstrated substantial heterogeneity among circulating TG
   species. Notably, TGs in the red module contained saturated and
   monounsaturated fatty acids and were positively associated with NAFLD,
   whereas TGs in the black module included mostly polyunsaturated fatty
   acids and were not associated with NAFLD ([111]Table 3). Saturated and
   monounsaturated TGs in serum may to some extent mirror the hepatic
   triglyceride content, since TGs containing saturated fatty acids are
   elevated in the liver of NAFLD patients mainly due to hepatic de novo
   lipogenesis [112][30]. Similarly, we observed two different clusters of
   PC species: PCs in module blue, mainly PC plasmalogens, were negatively
   associated with NAFLD, whereas PCs without these moieties were
   positively correlated with the progression of NAFLD ([113]Table 3). Of
   note, circulating levels of plasmalogens decreased in a stepwise manner
   in a cohort of biopsy-proven NAFLD patients [114][31]. Furthermore,
   administration of a plasmalogen precursor prevented SS and NASH in mice
   [115][32], suggesting that certain PC species may be protective against
   the development of NAFLD.

   Currently, the diagnosis and staging of NAFLD rely heavily on the
   histological evaluation of liver biopsies, which is invasive,
   subjective, and expensive [116][4]. There is an unmet need to discover
   novel biomarkers to distinguish benign SS from NASH, an advanced form
   of NAFLD with a much higher risk of developing cirrhosis and
   hepatocarcinoma. In this study, we developed and compared multiple ML
   models classifying NAFLD versus normal liver and NASH versus SS by
   integrating common clinical parameters with important metabolomic and
   lipidomic biomarkers. Notably, a saturated TG (16:0/16:0/18:1) was the
   top informative feature to discriminate NAFLD patients from individuals
   with normal liver. A previous study has observed a close association
   between circulating levels of TG (16:0/18:0/18:1) and liver fat content
   in NAFLD patients assessed by proton magnetic resonance spectroscopy
   [117][33]. Interestingly, among the ten TG species retrieved from our
   prediction models, seven of them were also selected by Mayo, et al. for
   the construction of their model to diagnose NASH [118][11].
   Furthermore, the odd-chain saturated fatty acid (SFA) heptadecanoic
   acid (17:0) was the strongest factor in discerning NASH from SS.
   Odd-chain SFAs partly reflect dairy fat intake, and their levels were
   inversely correlated with liver function enzymes as well as T2D
   [119][34]. Further studies are warranted to investigate the role of
   SFAs in NAFLD development and to validate the clinical value of
   metabolite/lipid species-based ML models as a non-invasive diagnostic
   tool for NAFLD and NASH in independent cohorts with larger sample
   sizes.

Statements & Declarations

   none.

Ethics approval and consent to participate

   All participants had given written informed consent. Ethics approval
   was obtained from the Institutional Review Board of the University of
   Hong Kong/Hospital Authority, Hong Kong West Cluster.

CRediT authorship contribution statement

   Jiarui Chen and Ronald Siyi Lu researched the data. Jiarui Chen, Ronald
   Siyi Lu, and Candela Diaz-Canestro wrote the manuscript. Jiarui Chen
   performed statistical analyses. Aimin Xu critically reviewed and edited
   the manuscript. Aimin Xu, Gianni Panagiotou, Erfei Song, Xi Jia, Yan
   Liu, Cunchuan Wang, and Cynthia K.Y. Cheung initiated and supervised
   the study. All authors reviewed the manuscript. Aimin Xu and Gianni
   Panagiotou are the guarantors of this work and, as such, had full
   access to all the data in the study and take responsibility for the
   integrity of the data and the accuracy of the data analysis.

Availability of data and materials

   The datasets generated during and/or analysed during the current study
   are available from the corresponding author upon reasonable request.

Funding

   This work was supported by the Merck Investigator Studies Program (MSD
   MISP 59615), Hong Kong Research Grants Council/Area of Excellence
   (AoE/M/707-18), Germany´s Excellence Strategy – EXC 2051 – Project-ID
   390713860, and the BMBF-funded “PerMiCCion” project (Project ID
   01KD2101D).

Author Statement

   During the preparation of this work, the author(s) used no generative
   AI and AI-assisted technologies.

Declaration of Competing Interest

   None.

Footnotes

   ^Appendix A

   Supplementary data associated with this article can be found in the
   online version at [120]doi:10.1016/j.csbj.2024.01.007.

Contributor Information

   Gianni Panagiotou, Email: Gianni.Panagiotou@hki-jena.de.

   Aimin Xu, Email: amxu@hku.hk.

Appendix A. Supplementary material

   Supplementary material
   [121]mmc1.docx^ (21KB, docx)

   .

   Supplementary material
   [122]mmc2.pdf^ (448.8KB, pdf)

   .

References