Abstract

   Altered gut metabolites are important for the inflammatory progression
   in children with NAFLD. Fecal and plasma samples were collected from
   145 subjects including 53 non-alcoholic fatty liver (NAFL), 39
   nonalcoholic steatohepatitis (NASH) and 53 obese controls. We performed
   G350 targeted integrative metabolomics using high performance liquid
   chromatography mass spectrometry for fecal and plasma analysis of NAFL,
   NASH, and obese children. We found 9 metabolites involved in metabolic
   reprogramming of inflammation in NAFLD, such as lipid, carbohydrate,
   amino acid metabolism, and TCA cycle pathway. Moreover, 7
   inflammation-related metabolites could discriminate NAFLD severity by
   machine learning model. This study identified three novel elevated
   inflammatory pathogenic metabolites and the relationship between
   increased inflammation, may be involved in TLR5/MYD88/NFκB pathway.
   These findings reveal that specific inflammatory metabolites entering
   the blood circulation from the gut are associated with disease severity
   and inflammatory pathogenesis in children with NAFLD.

   Subject terms: Microbiology, Health care

Introduction

   Non-alcoholic fatty liver disease (NAFLD) is the leading cause of
   chronic liver disease worldwide in childhood^[42]1,[43]2. Children with
   non-alcoholic fatty liver simple steatosis (NAFL) may develop
   nonalcoholic steatohepatitis (NASH), which is characterized by hepatic
   inflammation and cell injury, and increases the risk of hepatocellular
   carcinoma in adulthood^[44]3. The pathogenesis of NAFLD is poorly
   characterized, emerging evidence suggests that the gut microbiome
   influences the severity and inflammatory progression of
   NAFLD^[45]4,[46]5.

   Although some studies have demonstrated that the microbiota influence
   on inflammatory progression by gut metabolites, and identified some
   inflammatory gut metabolites for NAFLD, such as ursodeoxycholic acid,
   arachidonic acid, phenylacetic acid, branched-chain amino acids, and
   indolic^[47]6–[48]11. However, in view of the diversity and magnitude
   of the gut metabolites network, there are many unknown metabolites that
   have not been explored. Identifying inflammatory gut metabolites using
   low through-put targeted analyses is limited to a certain class of
   metabolites, and cannot systematically explore the phenotypes of gut
   metabolite profiles. The identification and validation of inflammatory
   gut metabolites in gut-liver axis requires “high-throughput omics”
   approach, such as fatty acids, amino acids, bile acids, benzenoids,
   carbohydrates, indoles, and purine nucleotides. It is also noteworthy
   that the gut metabolites species from gut into blood circulation in
   children with NAFLD are still not fully understood. Furthermore,
   specific significant gut metabolite pathways that may be involved in
   the inflammatory progression of NAFLD remain unexplored.

   The aims of this study were as follows: (i) to perform a “G350
   high-throughput omics” approach to reveal potential metabolic
   alterations in feces and plasma of children with NAFLD; (ii) to reveal
   potential metabolic alterations from the gut into blood circulation,
   then evaluate the potential inflammatory pathogenesis pathways and
   inflammatory changes related to these metabolites; and (iii) to apply
   multiple validations of the inflammatory effect for these metabolites,
   such as machine learning and in vitro cell experiment. The findings of
   this study would provide insights into the gut-liver axis, evaluation
   of potential inflammatory pathogenesis, and the search for
   immunotherapeutic biomarkers for children with NAFLD.

Results

Participant characteristics

   The baseline demographics and laboratory parameters are shown in
   Supplementary Table [49]1 and Supplementary Table [50]2. According to
   the distribution of characteristics data, there was good comparability
   between the case and control groups.

Distinct distribution of gut metabolites in children with NAFLD and control
in fecal samples

   The principal components analysis of fecal (Supplementary Fig. [51]1a)
   and plasma (Supplementary Fig. [52]1b) results shown that QC samples
   are closely gathered, indicating that the data quality was stable and
   accurate. The TIC plots showing the results of the differential fecal
   and plasma metabolites on Amide column and C18 columns (Supplementary
   Fig. [53]2). A total of 266 annotated metabolites were detected in
   fecal samples (Fig. [54]1A). Distinct clusters of gut metabolites were
   demonstrated in fecal samples (R2Y = 0.802 and Q2Y = 0.841) in children
   with NAFL compared with control children by OPLS-DA score plots (Fig.
   [55]1B). Significant differences were also found for the following
   groups: NASH children compared with control children (R2Y = 0.742 and
   Q2Y = 0.834); and NASH children compared with NAFL children
   (R2Y = 0.659 and Q2Y = 0.698). Moreover, the intercepts for
   goodness-of-prediction (Q2) and goodness-of-fit (R2) indicated that the
   OPLS-DA model containing metabolomics information was reliable and not
   overfitting (Fig. [56]1C).

Fig. 1. Gut metabolic alterations in NAFLD.

   [57]Fig. 1
   [58]Open in a new tab

   Chemical class metabolite sets plots (A) shown the chemical class
   metabolite of metabolites. B OPLS-DA score plots for differential
   metabolites in fecal samples. X axis and Y axis represent contributions
   of persons to the first two principal components (PC1 and PC2). C
   Cross-validation plot with a permutation test repeated 200 times. The
   intercepts of R2 and Q2 suggest that the OPLS-DA model is not
   overfitting. D Volcano plots shown the results of pairwise comparisons
   of metabolites in each case sample’s group relative to controls. The
   abscissa represents the variation of variation multiple of metabolites
   in different groups (log2 Fold Change), and the ordinate represents the
   significance level (−log10 p-value). The vertical dashed lines indicate
   the threshold for the abundance difference. The horizontal dashed line
   indicates the p = 0.05 threshold. Each point in the graph represents
   metabolites, the up-regulated metabolite red dot, the down regulated
   metabolite green dot. Between-group comparisons were performed using
   empirical Bayes hierarchical model and Kruskal-Wallis rank sum test.
   OPLS-DA Orthogonal Partial Least Squares Discrimination Analysis, NAFL
   nonalcoholic fatty liver, NASH non-alcoholic steatohepatitis.

   The volcano plots showed that 16 metabolites were significantly
   increased and 12 metabolites were significantly decreased in the NAFL
   group compared with the control children, whereas 62 metabolites were
   significantly increased and 17 metabolites were significantly decreased
   in the NASH children compared with the control children. Compared with
   the NAFL children, 34 metabolites were significantly increased and 12
   metabolites were significantly decreased in the NASH children (Fig.
   [59]1D).

Distinct distribution of gut metabolites in children with NAFLD and control
in plasma samples

   A total of 212 annotated metabolites were detected in plasma samples
   (Fig. [60]2A). Distinct clusters of gut metabolites in NAFL children
   compared with control children were demonstrated in plasma samples
   (R2Y = 0.629 and Q2Y = 0.871) by OPLS-DA score plots (Fig. [61]2B).
   Significant differences were also found for the following groups: NASH
   children compared with control children (R2Y = 0.609 and Q2Y = 0.923);
   and NASH children compared with NAFL children (R2Y = 0.639 and
   Q2Y = 0.886). Furthermore, the intercepts of goodness-of-prediction
   (Q2) and goodness-of-fit (R2) indicated that the OPLS-DA model
   containing metabolomics information was reliable and not overfitting
   (Fig. [62]2C).

Fig. 2. Plasma metabolic alterations in NAFLD.

   [63]Fig. 2
   [64]Open in a new tab

   Chemical class metabolite sets plots (A) shown the chemical class
   metabolite of metabolites. B OPLS-DA score plots for differential
   metabolites in Plasma samples. X axis and Y axis represent
   contributions of persons to the first two principal components (PC1 and
   PC2). C Cross-validation plot with a permutation test repeated 200
   times. The intercepts of R2 and Q2 suggest that the OPLS-DA model is
   not overfitting. D Volcano plots shown the results of pairwise
   comparisons of metabolites in each case sample’s group relative to
   controls. The abscissa represents the variation of variation multiple
   of metabolites in different groups (log2 Fold Change), and the ordinate
   represents the significance level (−log10 p-value). The vertical dashed
   lines indicate the threshold for the abundance difference. The
   horizontal dashed line indicates the p = 0.05 threshold. Each point in
   the graph represents metabolites, the up-regulated metabolite red dot,
   the down regulated metabolite green dot. Between-group comparisons were
   performed using empirical Bayes hierarchical model and Kruskal-Wallis
   rank sum test. OPLS-DA Orthogonal Partial Least Squares Discrimination
   Analysis, NAFL nonalcoholic fatty liver, NASH non-alcoholic
   steatohepatitis.

   The volcano plots showed that 33 metabolites were significantly
   increased and 32 metabolites were significantly decreased in the NAFL
   group compared with the control children, whereas 52 metabolites were
   significantly increased and 41 metabolites were significantly decreased
   in the NASH children compared with the control children. Compared to
   the NAFL children, 38 metabolites were significantly increased and 31
   metabolites were significantly decreased in the NASH children (Fig.
   [65]2D).

Identification of signature metabolites for NAFLD from gut into blood
circulation

   We further selected the differential inflammation-related metabolites
   in the correlation analysis as independent variables, and established a
   machine learning model to distinguish different states of NAFLD. The
   metabolite, 2-Hydroxy-3-methylbutyric acid, could distinguish, using
   machine learning, NAFL children from the controls in feces and plasma.
   Also, the 2-Hydroxy-3-methylbutyric acid was significantly correlated
   with markers of glycolipid metabolism, and liver function. Furthermore,
   the three metabolites, namely 2-Hydroxy-3-methylbutyric acid,
   L-Thyronine, and L-Alanine could distinguish NASH children from the
   NAFL children by machine learning in feces and plasma. These
   metabolites were also significantly correlated with markers of
   inflammation, gut microbiota, glycolipid metabolism, and liver
   function. The six metabolites, 2-Hydroxy-3-methylbutyric acid, Indole,
   L-Thyronine, w-TMCA, cis-Aconitic acid and Ribonolactone could also
   distinguish NASH children from the controls by machine learning in
   feces and plasma (Supplementary Table [66]3).

The key metabolic pathways in the inflammatory pathogenesis of NAFLD

   We further utilized the differential metabolites for pathway enrichment
   analysis and functional annotation to find the metabolic pathways
   related to inflammation. Pathway analysis of these different gut
   metabolites between children with NAFL and controls revealed that
   several processes were significantly enriched. Pathway analysis
   identified significant differences in amino acid metabolism,
   carbohydrate metabolism, digestive system, and energy metabolism as the
   most enriched pathways in fecal and plasma samples (Fig. [67]3A, B and
   Supplementary Table [68]4). These different gut metabolites, between
   NASH children and controls, are involved in several processes, such as
   amino acid metabolism, carbohydrate metabolism, digestive system,
   energy metabolism and metabolism of cofactors and vitamins; and were
   the most enriched pathways in fecal and plasma samples (Fig. [69]3A, B
   and Supplementary Table [70]5). We compared the metabolomic profiling
   of NAFL to the NASH to identify metabolic pathways involved in the
   progression of NASH. We identified significant differences in amino
   acid metabolism, digestive system, and metabolism of cofactors, with
   vitamins being the most enriched pathways in fecal and plasma samples
   (Fig. [71]3A, B and Supplementary Table [72]6).

Fig. 3. Overview of the key gut metabolic pathways in the pathogenesis of
NAFLD.

   [73]Fig. 3
   [74]Open in a new tab

   Pathway enrichment plots of fecal (A) and plasma (B) shown the pathway
   enrichment analysis by KEGG for metabolites. The pathway enrichment
   analysis result takes KEGG pathway as the unit, and the hypergeometric
   test is applied to find the pathway that is enriched in the
   differential metabolites against the background of all identified
   metabolites. Pathway enrichment analysis can determine the most
   important metabolic pathways and signal transduction pathways involved
   in the differential metabolites. An overview plot shown the key
   metabolic pathways in the pathogenesis of NAFLD (C). Boxplots show
   levels of these metabolites in the plasma samples at different stages.
   AAs amino acids, PCs Phosphatidylcholines, BCAAs branched-chain amino
   acids, ChoE cholesterol esters, CoA coenzyme A, G3P glyceraldehyde
   3-phosphate, GNG gluconeogenesis, CKs Inflammatory cytokines, ROS
   reactive oxygen species, SFA saturated fatty acid, TCA tricarboxylic
   acid. NAFL nonalcoholic fatty liver, NASH non-alcoholic
   steatohepatitis.

   These metabolites were significantly changed in the cases, and were
   associated with impaired inflammation, gut microbiota, glycolipid
   metabolism, and liver function through critical metabolic pathways such
   as amino acid metabolism, carbohydrate metabolism, digestive system,
   energy metabolism and metabolism of cofactors and vitamins. Metabolites
   associated with these metabolic pathways have been reported to
   contribute to the pathogenesis of systemic gut inflammation and liver
   inflammation in NAFLD patients (Fig. [75]3C and Table [76]1). After
   adjusting the major confounding factors such as age, gender and BMI, we
   found that metabolites such as 2-Hydroxy-3-methylbutyric acid,
   Indole-3-lactic Acid, 2-Furoic acid, cis-Aconitic acid, and
   L-Isoleucine in plasma may be risk factors for children with NAFLD,
   while L-Thyronine may be a protective factor for children with NAFLD
   (Supplementary Table [77]7).

Table 1.

   Dysregulation of the significantly pathway of metabolites might be
   involved in the inflammatory progression of NAFLD
   Metabolites Carbohydrate metabolism Lipid metabolism Amino acid
   metabolism Inflammatory pathway
   2-Furoic acid
   2-Hydroxy-3-methylbutyric acid * * ✰
   Alpha-ketoisovaleric acid
   Arachidonic acid
   Beta-Alanine * *
   Cis-Aconitic acid * * ✰
   Indole * ✰
   Indole-3-lactic Acid *
   L-Alanine * * ✰
   L-Alloisoleucine *
   L-Homoserine *
   L-Isoleucine * * ✰
   L-Leucine *
   L-Norleucine *
   L-Phenylalanine * *
   L-Thyronine * ✰
   L-Valine *
   N-acetyltryptophan * *
   Omega-muricholic acid * ✰
   Ribonolactone * ✰
   Sarcosine * *
   Tauro-a-muricholic acid
   Tauro-b-muricholic acid *
   Tauro-w-muricholic acid * * ✰
   [78]Open in a new tab

   *significantly metabolism pathway of metabolites; ✰, metabolites
   involved in the inflammatory pathway.

The correlation between selected metabolites with NAFLD indices in plasma
samples

   Considering the 24 differential metabolites, all of them were
   significantly associated with at least one inflammatory factor, with 22
   of them positively correlated with inflammatory factors. Specifically,
   6 metabolites (Beta-Alanine, cis-Aconitic-acid, L-Alanine, L-Valine,
   Sarcosine, Alpha-ketoisovaleric-acid) had a positive correlation with
   the levels of 8 inflammatory factors; 5 metabolites (L-Alloisoleucine,
   L-Isoleucine, L-Leucine, L-Norleucine, and
   2-Hydroxy-3-methylbutyric-acid) had a positive correlation with the
   levels of 7 inflammatory factors; 2 metabolites (L-Homoserine and
   L-Phenylalanine) had a positive correlation with the levels of 6
   inflammatory factors; w-TMCA had a positive correlation with the levels
   of 5 inflammatory factors; 4 metabolites (Indole-3-lactic-Acid,
   N-acetyltryptophan, w-MCA, and 2-Furoic-acid) had a positive
   correlation with the levels of 4 inflammatory factors; ribonolactone
   had a positive correlation with the levels of 3 inflammatory factors; 2
   metabolites (a-TMCA and b-TMCA) had a positive correlation with the
   levels of 2 inflammatory factors; arachidonic acid had a positive
   correlation with the level of 1 inflammatory factor; whereas indole had
   a negative correlation with the levels of 8 inflammatory factors; and
   L-Thyronine had a negative correlation with the levels of 5
   inflammatory factors. Moreover, we compared these selected significant
   metabolites in plasma samples with gut microbiota, glycolipid
   metabolism, and liver function (Supplementary Table [79]8 and Fig.
   [80]4A).

Fig. 4. Metabolites discriminates disease severity and association of
mediators with markers of inflammation, glycolipid metabolism and liver
function.

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

   A heatmap of plasma (A) shown the correlation between selected gut
   metabolites with glycolipid metabolism, liver function and inflammatory
   factors indices. Red color represents positive correlation, blue color
   represents negative correlation. Bar chart (B) show the qRT-PCR of
   differentially relative expressed mRNAs of liver function and
   inflammation produced by THLE-3 under different treatments, C control
   group, NAFL nonalcoholic fatty liver, NASH non-alcoholic
   steatohepatitis. p < 0.05 was considered statistically significant.
   *p < 0.05; **p < 0.01.

qRT-PCR verification of differentially expressed mRNAs of liver function and
inflammation

   The mRNA relative expression levels of ALT, CXCL8, IL6, IL17, IL21 and
   TNFα genes in THLE-3 were significantly upregulated after exposure to
   2-Hydroxy-3-methylbutyric acid. In terms of related receptors, the
   level of TLR5 gene relative expression was significantly increased.
   Moreover, the IL1β relative expression was significantly suppressed
   after exposure to Indole, the FXR relative expression was significantly
   increased. The relative expression level of IL-6 gene was significantly
   upregulated exposure to Ribonolactone, relative expression level of
   NFκB gene was significantly upregulated. The relative expression levels
   of IL-6 and IL-17 genes were significantly upregulated exposure to
   cis-Aconitic acid, MYD88, NFκB and TLR5 genes were also significantly
   upregulated (Fig. [83]4B). There was no significant difference in the
   expression level of inflammatory factor genes in THLE-3 after exposure
   to Tauro-w-muricholic acid, L-Thyronine and L-Alanine (Supplementary
   Fig. [84]3).

Molecular docking of metabolites to binding domain of inflammatory receptor
protein

   List of metabolites included in the molecular docking analysis were
   shown in Supplementary Table [85]9. The S value for the TLR5 binding
   cis-Aconitic acid and 2-Hydroxy-3-methylbutyric acid were
   −5.41 kcal/mol and −4.71 kcal/mol. The S value for the NFκB binding
   Ribonolactone and cis-Aconitic acid were −4.61 kcal/mol and
   −4.68 kcal/mol. The S value for the MyD88 binding cis-Aconitic acid was
   −5.39 kcal/mol. The S value for the FXR binding Indole was
   −4.68 kcal/mol (Supplementary Table [86]10). Many hydrophobic amino
   acid residues surrounded the metabolites (Fig. [87]5A, B). These data
   indicated that the binding pocket of inflammatory receptor protein is
   suitable for molecule interactions and metabolites may serve as a
   ligand specifically bound to this ligand-binding domain for modulation
   of inflammatory receptor activity. This study identified three
   inflammatory pathogenic metabolites (2-Hydroxy-3-methylbutyric acid,
   Ribonolactone, and cis-Aconitic acid), may be involved in
   TLR5/MYD88/NFκB pathway.

Fig. 5. Metabolites docks to the ligand-binding domain of inflammatory
receptor.

   [88]Fig. 5
   [89]Open in a new tab

   The superposition of these metabolites bound to the ligand-binding
   domain pocket of inflammatory receptor (A). Two-dimensional interaction
   map of inflammatory receptor with metabolites (B). Hypothesis map of
   inflammatory mechanism pathway is shown in (C).

Discussion

   For the first time in this study, we systematically performed an
   in-depth and comprehensive gut targeted metabolomic profiling of plasma
   and fecal samples to identify NAFLD-associated metabolites in children
   with obesity, NAFL, and NASH. Targeted metabolomics revealed extensive
   inflammatory mediators from gut into plasma in children with NAFLD. We
   found 22 significantly increased metabolites and 2 significantly
   decreased metabolites in fecal and plasma samples of the NAFL or NASH
   children than in the controls. Moreover, 9 metabolites
   (2-Hydroxy-3-methylbutyric acid, cis-Aconitic acid, w-TMCA, w-MCA,
   indole, Ribonolactone, L-Isoleucine, L-Alanine and L-Thyronine) were
   significantly correlated with inflammatory factors and associated with
   inflammatory metabolic pathways. These metabolites were also involved
   in amino acid metabolism, carbohydrate metabolism, digestive system,
   tricarboxylic acid (TCA) cycle, energy metabolism and metabolism of
   cofactors and vitamins. Regarding disease status, metabolite
   fingerprinting by machine learning revealed 7 metabolites
   (2-Hydroxy-3-methylbutyric acid, w-TMCA, indole, Ribonolactone,
   cis-Aconitic acid, L-Alanine and L-Thyronine) that discriminated
   different stage of NAFLD. Meanwhile, one of the advantages of our
   research was that targeted metabolomics, could be used for absolute
   quantification rather than relative quantification. Moreover, by
   analyzing the 24 differentially expressed metabolites screened by
   machine learning, we found that 95.8% of the metabolites had fecal
   concentrations higher than plasma concentrations, except for
   cis-Aconitic acid, which may be due to its involvement in energy
   metabolism leading to higher endogenous production concentrations than
   in fecal. Future animal experiments are needed to compare the
   concentrations of metabolites in the gut, liver and circulation to
   further determine whether they are produced by liver metabolism.

   Although the inflammatory metabolite, indole, was previously reported
   in NAFLD, the following inflammatory metabolites in NAFLD are being
   reported for the first time: 2-Hydroxy-3-methylbutyric acid, L-Alanine,
   cis-Aconitic acid, Ribonolactone, L-Thyronine, and w-TMCA. Gut
   microbiota and metabolites have been considered to be key factors
   affecting hepatic lipid accumulation and inflammation during the
   development of liver diseases, such as NAFLD and hepatocellular
   carcinoma^[90]12,[91]13. The following discussion focuses on how the
   metabolites, found by machine learning in this study, produce
   inflammatory effects through different metabolic pathways. Overall,
   these metabolites seem to form a TCA cycle metabolic network in amino
   acid, lipid, and carbohydrate metabolism. First, in amino acid
   metabolism, branched-chain amino acids are used to synthesize
   glutathione in response to oxidative stress and
   inflammation^[92]14,[93]15. L-Thyronine seems to have similar functions
   with branched-chain amino acids, which can inhibit inflammation.
   However, in this study, L-Alanine, L-Valine, Sarcosine and L-Isoleucine
   seem to have antagonistic effects with branched-chain amino acids.
   These amino acid metabolites may inhibit glutathione production and
   activate inflammatory response. Second, in lipid metabolism, liver
   steatosis is an important part of NAFL and NASH, which is mainly caused
   by the imbalance of triglyceride production and secretion in the
   liver^[94]16. Hepatic de novo lipogenesis is stimulated by carbohydrate
   and insulin, and the metabolism of fatty acids produces oxylipins,
   which usually increase the risk of inflammatory-related diseases such
   as insulin resistance, diabetes, and obesity^[95]17,[96]18. In this
   study, Hydroxy-3-methylbutyric acid, a lipid metabolite, may be
   involved in the metabolism of fatty acids and oxylipins, thus promoting
   inflammation and NASH. Bile acids are secreted in response to the meal
   to facilitate the digestion of fats^[97]19. Disturbance of bile acid
   metabolism leads to underactivation of bile acid receptors FXR and
   TGR5, causal for decreased energy expenditure, increased lipogenesis,
   bile acid synthesis and hepatic inflammation^[98]20. This phenomenon
   was also observed in this study. The levels of bile acid metabolites
   w-MCA and w-TMCA were significantly increased in NASH group, which may
   promote the occurrence and development of inflammation and NASH.
   Finally, in carbohydrate metabolism, excessive hepatic mitochondrial
   TCA cycle and gluconeogenesis in patients with NAFLD^[99]21,[100]22. In
   this study, Ribonolactone, a metabolite of carbohydrates, and
   cis-Aconitic acid, an intermediate of TCA cycle, may participate in the
   energy metabolism in the occurrence and development of NASH, and may
   promote inflammation. These findings raise the possibility that
   dysregulated TCA cycle metabolism is central in the NAFLD, providing a
   potential link between amino acid, lipid, and carbohydrate metabolism.

   Specifically, among these inflammatory mediators,
   2-Hydroxy-3-methylbutyric acid belongs to the class of hydroxy fatty
   acids and is mainly involved in lipid metabolism pathways. The
   2-Hydroxy-3-methylbutyric acid showed good ability to distinguish
   NAFL/NASH from controls in a machine learning model, and it has been
   identified in the urine of patients with maple syrup urine disease and
   lactic acidosis^[101]23,[102]24. In another study, altered
   2-Hydroxy-3-methylbutyric acid was correlated with severity of
   inflammation, which has important diagnostic value for
   community-acquired pneumonia^[103]25. These findings are particularly
   consistent with our finding that the level of 2-Hydroxy-3-methylbutyric
   acid was significantly upregulated in NAFLD patients, and associated
   with impaired inflammation (IL1β, IL6, IL17, IL21, IL32, TNFα, and
   CXCL8), glycolipid metabolism, and liver function.

   Lipid metabolism in the liver is mainly involved in the occurrence and
   development of NAFLD through two metabolic pathways, one of which is
   oxidized to produce energy and ketones, such as acetoacetate and
   β-hydroxybutyric acid^[104]26. The other metabolic pathway is
   re-esterified to glycerol esters, such as diglycerides and
   triglycerides. Primary bile acids and cholesterol esters are
   synthesized from cholesterol. Bile acid metabolism is involved in the
   progression of NAFLD’s response to a high-fat diet to facilitate the
   digestion of fats^[105]27,[106]28. Glyceraldehyde 3-phosphate, serine
   and diglycerides are also involved in the synthesis of sphingomyelins
   and ceramides^[107]29,[108]30. Newly synthesized triglycerides can be
   secreted as VLDL or stored as lipid droplets. These metabolic pathways
   have the effect of promoting liver inflammation and oxidative
   stress^[109]4.

   Cis-Aconitic acid belongs to the class of tricarboxylic acids (TCA) and
   derivatives, and is mainly involved in the TCA cycle pathway. In a
   machine learning model, cis-Aconitic acid showed good ability in the
   discrimination of the NASH patients from the controls. Also,
   cis-Aconitic acid was found to be significantly correlated with
   inflammatory response in an animal study^[110]31. The pathogenesis of
   NAFLD metabolism has been reported to be accompanied by these
   functional and phenotypic changes in liver pro-inflammatory
   macrophages. Pro-inflammatory M1 macrophages rely primarily on
   glycolysis and experience disruption in the mitochondrial TCA cycle,
   leading to the accumulation of itaconic acid, cis-Aconitic acid, and
   succinic acid^[111]32. Upregulation of glycolysis and disruption of the
   TCA cycle underpin this switch to activated liver pro-inflammatory
   macrophages. Moreover, excess cis-Aconitic acid may lead to the
   accumulation of HIF1α and activate the transcription of glycolytic
   genes, thus sustaining the glycolytic metabolism of liver
   pro-inflammatory M1 macrophages^[112]32,[113]33. Our findings also
   indicated that cis-Aconitic acid was associated with impaired
   inflammation (IL1β, IL6, IL21, IL12, TNFα and CXCL8) and glycolipid
   metabolism.

   Ribonolactone belongs to the class of carbohydrates, peptides, and
   analogues, and is mainly involved in the carbohydrate metabolism
   pathway. In the machine learning model, Ribonolactone showed good
   ability to discriminate the NASH patients from the controls. Recent
   evidence suggests that carbohydrate metabolism and TCA cycle metabolic
   pathway may be closely combined to play a synergistic role in the
   pathogenesis of NAFLD^[114]32,[115]34. Specifically, concerning
   carbohydrate metabolism, glucose of amino acids, lactate, or glycerol
   is produced by glycogenolysis or gluconeogenesis, and De novo
   lipogenesis (DNL) is stimulated by carbohydrate and Acetyl-CoA^[116]35.
   Amino acid metabolism and carbohydrate metabolism are connected through
   the key metabolite pyruvate and participate in the progression of NAFLD
   through TCA cycle, a key signal pathway^[117]32,[118]34. TCA cycle
   intermediates (Acetyl-CoA and Succinyl-CoA) are also involved in the
   progression of NAFLD^[119]34,[120]36.

   This study identified three novel elevated inflammatory pathogenic
   metabolites (2-Hydroxy-3-methylbutyric acid, Ribonolactone, and
   cis-Aconitic acid) and the relationship between increased inflammation
   (CXCL8, IL17, IL6 and TNFα), may be involved in TLR5/MYD88/NFκB
   pathway, as shown in Fig. [121]5C. Specifically,
   2-Hydroxy-3-methylbutyric acid, Ribonolactone, and cis-Aconitic acid
   mediated TLR5 signaling induces MYD88/IRAK4/TRAF6 by TAB1/TAB2/TAK1,
   and activation of NFκB by IKKs complex, and promoted the secretion of
   IL6 and TNFα^[122]37,[123]38.

   Indole belongs to the class of aromatic heterocyclic organic compound,
   and is mainly involved in amino acid metabolism pathway. In the machine
   learning model, indole showed good ability in discriminating NASH
   patients from the controls. Indole and its derivatives have been
   indicated to have beneficial effects in mitigating liver inflammatory
   responses and preserving tight junctions of epithelial
   integrity^[124]7,[125]39. An animal study showed that mice receiving
   indole displayed resistance to metabolic alternations of cholesterol
   and liver inflammation induced by lipopolysaccharide^[126]40. Some
   evidence also suggests that a certain dose of indole and its
   derivatives exert a protective effect, as it reduces the induction of
   pro-inflammatory cytokines including MCP-1, TNFα, and IL1β; and indole
   and its derivatives can be stimulated by ligand for pregnane X receptor
   and toll-like receptor 4 to act on both macrophages and hepatocytes
   against NAFLD^[127]41,[128]42. These findings are consistent with our
   finding that the level of indole was significantly downregulated in
   NAFLD patients and associated with the effect of anti-inflammation
   (IL1β, IL6, IL17, TNFα, and CXCL8). The increase in indole-3-lactic
   acid consumes indole, leading to pro-inflammatory effects (IL1β, IL6,
   IL17, and IL21).

   Collectively, our results capture a specific gut metabolite
   fingerprints in the clinical course of the NAFLD, and these metabolic
   reprogramming may provide potential insights and drivers into the
   pathogenesis of NAFLD in children. Pathway analysis of these
   differential gut metabolites in children with NAFLD demonstrated
   significant enrichment of several processes including amino acid
   metabolism, carbohydrate metabolism, digestive system, TCA cycle,
   energy metabolism and metabolism of cofactors and vitamins. These
   pathways have been reported to contribute to oxidative stress and
   systemic inflammation, which further lead to disorder in liver function
   and glycolipid metabolism^[129]15,[130]43,[131]44. The binding
   coefficients of the docking of metabolites to the ligand-binding
   regions of inflammatory receptors has the following practical
   significance. Firstly, by combining spatial and energy matching, the
   interaction force between inflammatory receptors and metabolites was
   qualitatively represented by the coefficient S value. When the S value
   is close to 0, it indicates that metabolite was unstable in combination
   with inflammatory receptor. In that case, this pair of inflammatory
   receptor and metabolite will not be considered for future in vitro
   experiments. This will greatly enhance the true positive rate of in
   vitro animal experiments, and optimize the efficiency of the use of
   research funds, especially when there are many metabolites that need to
   be screened. The S value of the six receptors and metabolites in this
   study were all < −4 kcal/mol, indicating that metabolites were unstable
   in combination with inflammatory receptors in this study. Secondly, the
   smaller the binding coefficient S value, the more stable the
   conformation, indicating a stronger binding between the receptor and
   the metabolite. In the future vitro experimental validation, we will
   prioritize selecting pair of inflammatory receptor and metabolite with
   lower binding coefficients S value for animal experiments, such as pair
   of TLR5 and cis-Aconitic acid, pair of MyD88 and cis-Aconitic acid, and
   pair of TLR5 and 2-Hydroxy-3-methylbutyric acid, et al.

   This study has a few limitations. First, some confounding factors, such
   as dietary habit and ethnicity, may influence the gut microbiota and
   metabolomic profiles. Second, although we successfully recruited 145
   patients, the relatively small sample size may limit the
   generalizability and statistical significance of the results. However,
   high consistency of the gut and blood circulation results consolidated
   our findings. Future multicenter studies should consider expanding the
   sample size to enhance the robustness and external validity of these
   findings. Third, our study may not reflect changes in liver metabolites
   due to the poor accessibility of liver biopsy. Future studies in
   patients with liver biopsy are needed. Fourth, samples from the case
   cohort were collected after diagnosis of NAFLD, although most of them
   were collected closer to the date of diagnosis. Future studies of
   metabolomic profiles from more time points are required to explore
   molecular dynamics during temporal analysis of NAFLD progression.
   Finally, the investigation into the underlying mechanisms is still in
   its early stages. Future studies of in-depth in vitro experiments can
   better elucidate these mechanisms are needed, such as through the
   knockout or overexpression of TLR5 signaling pathway genes. Moreover,
   future studies need to consider incorporating animal or liver organoid
   model studies, which would greatly enhance the credibility and
   scientific rigor of the findings. This combined approach would provide
   a solid foundation for future drug development and clinical
   application.

   We identified 9 metabolites that played a key role in metabolic
   reprogramming of lipids, carbohydrates, amino acid metabolism and TCA
   cycle, hence signifying their critical role in the inflammation
   pathogenesis of NAFLD, especially NASH. Machine learning analysis found
   that 7 metabolites play an important role in identifying disease states
   and verified by cell experiments in vitro and molecular-docking
   analysis. Last, this study identified three novel elevated inflammatory
   pathogenic metabolites (2-Hydroxy-3-methylbutyric acid, Ribonolactone,
   and cis-Aconitic acid) and the relationship between increased
   inflammation (CXCL8, IL17, IL6, and TNFα), may be involved in
   TLR5/MYD88/NFκB pathway.

Methods

Study population

   This study was performed in 290 fecal samples and 290 plasma samples
   from 92 children with NAFLD (53 NAFL patients and 39 NASH patients) and
   53 control children from the Institute of Child Health, Hunan
   children’s Hospital (Changsha, China). All legal guardians of these
   children and the Hospital Ethics Research Committee (XYGW-2018-04) gave
   informed and written consent to the clinical and biomaterial omics
   investigations. The study was undertaken according to Helsinki
   Declaration II from January 2020 to September 2021. We recruited
   children with/without NAFLD with strict inclusion and exclusion
   criteria to avoid the influence on the targeted metabolites. A detailed
   selection flow chart of the children with/without NAFLD included in
   this study is shown in Fig. [132]6. According to the standard of
   “Screening for overweight and obesity among school-age children and
   adolescents” (WS/T 586-2018), the BMI status of children was classified
   into 3 categories, normal, overweight and obesity (Supplementary Table
   [133]11). Demographic, clinical data and inflammatory factors for each
   child were obtained at the Institute of Child Health, Hunan children’s
   Hospital.

Fig. 6. Flow chart showing the methods used for metabolomics analysis and
bioinformatics function prediction.

   [134]Fig. 6
   [135]Open in a new tab

   Patient recruitment and inclusion and exclusion criteria. Various
   clinical samples were obtained from NAFLD patients and controls.
   Integrative analysis for potential harmful and beneficial gut
   metabolites. Targeted metabolomics (LC-MS/MS) offers a plethora of
   information on metabolites. NAFL, nonalcoholic fatty liver, NASH
   non-alcoholic steatohepatitis, LC-MS/MS liquid chromatography coupled
   to tandem mass spectrometry, RF random forest.

Analysis of gut metabolites by targeted metabolomics

   Fecal and plasma levels of gut metabolites were determined by G350 Kit,
   measured using the liquid chromatography coupled to tandem mass
   spectrometry (LC-MS/MS). The G350 Kit has coverage of up to 346
   metabolites and 40 biochemical classes, including the differential
   metabolites detected by our research group using untargeted screening
   and key metabolic pathways involved in the pathogenesis of NAFLD such
   as fatty acids, amino acids, and carbohydrates, as well as other
   metabolites such as indoles, bile acids and benzenoids^[136]45–[137]47.
   Common nomenclature and biochemical classes of these gut metabolites
   are detailed in Supplementary Table [138]12.

Metabolite extractions

   To extract metabolites from the fecal samples, 800 μL of cold
   methanol/acetonitrile/water (2:2:1) extraction solvent was added to
   100 mg fecal sample, and adequately vortexed. Moreover, to extract
   metabolites from plasma samples, 400 μL of cold methanol/acetonitrile
   (1:1, v/v) extraction solvent was added to 100 μL plasma sample remove
   the protein and extract the metabolites, then adequately vortexed. For
   absolute quantification of the metabolites, stock solutions of stable-
   isotope internal standards were added to the extraction solvent
   simultaneously, and then the elution was collected and dried in a
   vacuum centrifuge at 4 °C. For LC-MS analysis, the samples were
   re-dissolved in 100 μL acetonitrile/water (1:1) solvent and transferred
   to LC vials.

LC-MS/MS analysis

   Analyses were performed using an UHPLC (1290 Infinity LC, Agilent
   Technologies, USA) coupled to a QTRAP MS (6500, Sciex, USA). The
   analytes were separated on HILIC (Waters UPLC BEH Amide column,
   2.1 mm × 100 mm, 1.7 µm) and C18 columns (Waters UPLC BEH
   C18-2.1 × 100 mm, 1.7 μm). For HILIC separation, the column temperature
   was set at 35 °C; and the injection volume was 2 μL. Mobile phase A:
   100 mM ammonium acetate and 1.2% Ammonium hydroxide in water, mobile
   phase B: acetonitrile. A gradient (85% B at 0–1 min, 80% B at 3–4 min,
   70% B at 6 min, 50% B at 10–12.5 min, 85% B at 12.6–18 min) was then
   initiated at a flow rate of 300 μL/min. For RPLC separation, the column
   temperature was set at 40 °C, and the injection volume was 2 μL. Mobile
   phase A: 50 mM ammonium formate and 0.4% formic acid in water, mobile
   phase B: methanol. A gradient (5% B at 0 min, 60% B at 5 min, 100% B at
   11–13 min, 5% B at 13.1–16 min) was then initiated at a flow rate of
   400 μL/min. The sample was placed at 4 °C during the whole analysis
   process. 6500 QTRAP (AB SCIEX, USA) was performed in positive and
   negative switch mode. The ESI positive source conditions were as
   follows: source temperature: 550 °C; ion Source Gas1 (Gas1): 55; Ion
   Source Gas2 (Gas2): 55; Curtain gas (CUR): 40; ion Sapary Voltage
   Floating (ISVF): +4500 V; the ESI negative source conditions were as
   follows: source temperature: 550 °C; Gas1: 55; Gas2: 55; CUR: 40; ISVF:
   −4500 V. MRM method was used for mass spectrometry quantitative data
   acquisition. A polled quality control (QC) samples were set in the
   sample queue to evaluate the stability and repeatability of the system.
   MultiQuant (version 3.0) was used for quantitative data processing. The
   ratio of the peak area (area of the peak for substance/area of the peak
   of the internal standard) was used for obtaining the absolute
   quantitation for each substance according to the calibration curve.
   Peakview (version 1.2) was used for the metabolomic total ion
   chromatograms (TIC) analysis of feces and plasma.

Cell lines

   Human normal hepatocyte cell line hepatocytes (THLE-3) were obtained
   from the American Type Culture Collection (Manassas, Virginia, USA),
   which is non-tumorigenic and alpha fetoprotein expression negative. The
   cells were cultured in RPMI-1640 medium (GE Health Life Sciences, USA)
   supplemented with 10% (vol/ vol) fetal bovine serum (FBS) (Gibco, USA),
   and 1% penicillin/streptomycin (100 IU/ml penicillin and 100 mg/ml
   streptomycin) (Gibco, USA) at 37 °C in a humidified atmosphere of 5%
   CO^2. We terminated with 0.25% trypsin-EDTA (Gibco, USA) and collected
   logarithm growth cells. One million of THLE-3 was seeded in each well
   of a six-well culture plates. Then, cells were exposed to with or
   without indicated metabolites-supplemented for 24 h.

Quantitative RT-PCR (qRT-PCR) for liver function and inflammation assay

   Total cellular RNA was extracted using the RNA isolater Total RNA
   Extraction Reagent Kit (Vazyme, NanJing, China). Reverse transcription
   was conducted using the HiScript ® II Q Select RT SuperMix cDNA
   Synthesis Kit (Vazyme, NanJing, China) on S1000TM Thermal Cycler
   (Bio-Rad, Hercules, CA, USA). mRNAs were assayed qualitatively by ChamQ
   Universal SYBR qPCR Master Mix (Vazyme, NanJing, China) on a
   LightCycler96 Real-Time PCR System (Roche, Switzerland). All the
   procedures were repeated three times. The mRNA relative expression
   levels were calculated by the 2^−∆∆Ct method using GAPDH as an internal
   reference, respectively. The sequences of mRNA primer pairs used were
   in Supplementary Table [139]13.

Molecular docking of metabolites to binding domain of inflammatory receptor
protein

   The DOCK module of the MOE2019.01 software program was used to predict
   the preferable binding sites between inflammatory receptor and the
   respective metabolites. Water molecules in the crystal structure were
   removed with the MOE software, and the energy was minimized before
   molecular docking. The S value scoring function was based on
   GBVI/WSA∆G, that estimates the binding free energy of a metabolites to
   an inflammatory receptor.

Statistical analysis

   All statistical tests were performed using R v3.5.1 statistical
   software. Non-parametric statistical tests (Kruskal-Wallis
   test/Mann–Whitney test) were used to assess the differences of
   demographic and clinical features between cases and controls. Spearman
   correlation was used to assess correlation between demographic and
   clinical features and gut metabolites. The hierarchical cluster
   analysis was used to assess clustering algorithm, and the Euclidean
   distance was used to assess the relationship between samples and gut
   metabolites; and between demographic and clinical features.

   Data of gut metabolites were log transformed and centered before
   conducting the Orthogonal Partial least squares discriminant analysis
   (OPLS-DA), which was used to assess the metabolic alterations between
   cases and controls. The Bonferroni correction was applied to correct
   for multiple-testing. The 7-fold cross-validation and 200 times
   permutation testing were used to assess the risk of overfitting and the
   robustness of the OPLS-DA model. The functions of differential
   metabolites were performed using the Kyoto Encyclopaedia of Genes and
   Genomes (KEGG) database.

   Machine learning with random forest models was used to verify whether
   the metabolite combinations were potential discriminators for
   identifying children with NAFL or NASH. We selected important
   metabolites with mean decrease accuracy using RF analysis. Combined
   with important variables of machine learning screening, multivariate
   logistic regression analysis was used to control confounding factors
   such as age, BMI, and gender, and further account for the impact of
   these plasma metabolites on the children with NAFL or NASH. We further
   combined the results of RF analysis and biological function annotation
   to select potential inflammatory metabolites for the next in vitro cell
   experiment verification. Unless otherwise stated, p < 0.05 was
   considered statistically significant. Completed reporting checklist is
   shown in Supplementary Table [140]14.

Supplementary information

   [141]Supplementary Information^ (1.8MB, pdf)

Acknowledgements