Metabolomics analysis of urine from patients with alcohol-associated liver disease reveals dysregulated caffeine metabolism

Abstract

Excess alcohol intake causes millions of deaths annually worldwide. Asymptomatic early-stage, alcohol-associated liver disease (ALD) is easily overlooked, and ALD is usually only diagnosed in more advanced stages. We explored the possibility of using polar urine metabolites as biomarkers of ALD for early-stage diagnosis and functional assessment of disease severity by quantifying the abundance of polar metabolites in the urine samples of healthy controls (n = 18), patients with mild or moderate liver injury (n = 21), and patients with severe alcohol-associated hepatitis (n = 25). The polar metabolites in human urine were first analyzed by untargeted metabolomics, showing that 209 urine metabolites are significantly changed in patients, and 17 of these are highly correlated with patients’ model for end-stage liver disease (MELD) score. Pathway enrichment analysis reveals that the caffeine metabolic pathway is the most affected in ALD. We then developed a targeted metabolomics method and measured the concentration of caffeine and its metabolites in urine using internal and external standard calibration, respectively. The described method can quantify caffeine and its 14 metabolites in 35 min. The results of targeted metabolomics analysis agree with the results of untargeted metabolomics, showing that 13 caffeine metabolites are significantly decreased in patients. In particular, the concentrations of 1-methylxanthine, paraxanthine, and 5-acetylamino-6-amino-3-methyluracil are markedly decreased with increased disease severity. We suggest that these three metabolites could serve as functional biomarkers for differentiating early-stage ALD from more advanced liver injury.

NEW & NOTEWORTHY Our study using both untargeted and targeted metabolomics reveals the caffeine metabolic pathway is dysregulated in ALD. Three caffeine metabolites, 1-methylxanthine, paraxanthine, and 5-acetylamino-6-amino-3-methyluracil, can differentiate the severity of early-stage ALD.

INTRODUCTION

Alcohol-associated liver disease (ALD) is a spectrum of liver injury that ranges from simple steatosis to more severe forms such as alcohol-associated hepatitis (AH), cirrhosis, and potentially hepatocellular carcinoma. ALD imposes a heavy public health burden. For instance, liver cirrhosis was responsible for more than 34,000 deaths in America in 2016 (1) and ∼170,000 in Europe yearly (2). Liver damage is generally asymptomatic and reversible in the early stages of ALD but can proceed to permanent liver damage with continued alcohol consumption. The accumulation of triglycerides in the liver in the early stages of ALD can be completely reversed by abstinence without any drug intervention (3). Therefore, identifying ALD in its early stages is critical for successfully treating ALD before more clinically serious or irreversible liver damage occurs (4). Unfortunately, current methods for early diagnosis of ALD can be nonspecific and lack sensitivity (5, 6). Moreover, current tests do not measure hepatic functional capacity. Gaps in the overall diagnosis and medical management of ALD remain a challenge. An estimated 165 million individuals worldwide will suffer from ALD by 2050 (7). Consequently, reliable noninvasive methods are needed to diagnose and stage patients with ALD.

Several pathogenic factors are directly related to liver cirrhosis, including excess alcohol intake, metabolic syndrome, hepatitis B and C virus infection, etc. (8–11). During the past few years, studies on the metabolic signature of ALD showed that alcohol consumption dysregulates the metabolism of lipids and polar metabolites (12–18). 3-Hydroxytetradecanedioic acid, isocitric acid, and sebacic acid were reported to distinguish patients with ALD from healthy subjects with an accuracy of >95% (7). N-Lauroglycine was reported to be the best marker for identifying cirrhosis (100% sensitivity and 90% negative predictive value, NPV), and decatrienoic acid was the best for assessing disease severity (i.e., differentiating patients with cirrhosis from noncirrhotic patients) with 100% sensitivity and 100% NPV (19). However, these studies did not show whether or not these metabolites could identify patients with early stages of ALD.

Our previous study showed that the metabolites involved in glycolysis and the tricarboxylic acid (TCA) cycle are significantly decreased in the feces of mice fed with alcohol (20). Some amino acids (arginine, glutamic acid, and histidine) and modified nucleosides/bases (inosine and 7-methylguanine) are also changed in the feces of alcohol-fed mice (20). We recently found that the abundance levels of bile acids are significantly altered in a stepwise fashion in the urine of patients with different stages of AH (21). All of these data show the potential of metabolites as diagnostic markers for the development and progression of ALD. In addition to diagnosis, specific metabolic features in ALD help elucidate ALD’s pathological mechanisms.

Urine is an ideal sample type for disease biomarker study because it is collected noninvasively, readily available, and relatively less complex than other samples (22, 23). Furthermore, patients can collect urine in large volumes, which is especially important when patients cannot easily go to clinical laboratories, such as during the COVID-19 pandemic. This study aimed to discover the metabolite biomarkers in human urine to find early-stage ALD and to assess disease severity. We collected urine samples from healthy controls and patients with various stages of ALD. We first performed untargeted metabolomics to study the metabolic profile differences between sample groups. Afterward, we developed an analytical method of targeted metabolomics to investigate the concentration changes of caffeine and its metabolites in urine samples.

METHODS

Chemicals and Reagents

A total of 290 authentic standards of polar metabolites were purchased from Sigma-Aldrich Corp. (St. Louis, MO), Fisher Scientific (Loughborough, UK), and Cayman Chemical (Ann Arbor, MI). Caffeine metabolite standards, 1-methylxanthine (1X), 3-methylxanthine (3X), and 1,3-dimethyluric acid (13 U), were purchased from Biosynth Carbosynth (San Diego, CA). 3-Methyluric acid (3 U) and 5-acetylamino-6-amino-3-methyluracil (AAMU) were purchased from Toronto Research Chemicals (Toronto, ON, Canada). Caffeine (1,3,7-trimethylxanthine,137X), 3,7-dimethyluric acid (37 U), 1-methyluric acid (1 U), and 7-methylxanthine (7X) were obtained from Sigma Aldrich (St. Louis, MO). Theophylline (13X), paraxanthine (17X), theobromine (37X), 1,3,7-trimethyluric acid (137 U), 1,7-dimethyluric acid (17 U), 7-methyluric acid (7 U), [¹³C₃]caffeine ([¹³C₃]137X) and [²H₉]1,3,7-trimethyluric acid ([²H₉]137U) were from Cayman Chemical Company (Ann Arbor, MI). LC-MS grade acetonitrile and formic acid were from Fisher Scientific (Waltham, MA). The analytical grade water was purified using a Millipore Synergy UV system (Burlington, MA).

Caffeine Metabolites Stock Solution Preparation

Stock solutions of caffeine (1 mg/mL), [¹³C₃]caffeine (1 mg/mL), and 13X (2.5 mM) were prepared using 100% methanol. Stock solutions of AAMU (10 mM), 137 U (10 mM), [²H₉]137U (10 mM), 7 U (10 mM), 17 U (10 mM), 37 U (20 mM), 1 U (40 mM), 1X (40 mM), 17X (40 mM), 3 U (40 mM), 3X (40 mM), and 37X (80 mM) were prepared using 0.1 M NaOH. A stock solution of 13 U (20 mM) was prepared using 100% ammonium hydroxide, and a stock solution of 7X (40 mM) was prepared using 100% dimethyl sulfoxide (DMSO). All standards and stock solutions were stored at −20°C.

Patient Recruitment and Sample Collection

The clinical study was approved by the University of Louisville Institutional Review Board, and information on patient definitions and recruitment are detailed in another publication (21). All study participants provided written informed consent before the study, including appropriate authorization for data and sample collection. Eighteen healthy controls (HC, n = 18) and 46 patients with ALD were enrolled. All study participants had a complete history, physical examination, and laboratory evaluation upon enrollment. Urine samples from 46 patients were categorized into two groups by MELD score. Patients with moderate AH and patients with an elevated ALT in an alcohol detox program were grouped together as nonsevere ALD (n = 21). This represents a spectrum of alcohol-induced liver injury that does not reach the stage of severity where pharmacological therapy with corticosteroids is indicated. Patients with severe alcohol-associated hepatitis (severe AH, n = 25) comprised a second group. Patient and group information are listed in Table 1.

Table 1.

Demographic, drinking, and liver indicators of study participants by group

Variables	HC (n = 18)	Alcohol-Associated Liver Disease (ALD)				P Valuea
		Nonsevere ALD (n = 21)
		AUD (n = 8)	Moderate AH (n = 13)	Severe AH (n = 25)	Total Patients (n = 46)
Age, yr	36 (24–60)	51 (39–67)	50 (34–65)	47 (27–66)	49 (27–67)
T. Bilirubin, mg/dL	0.7 (0.4–1.3)	1.5 (0.8–2.6)	4.2 (1.2–18.2)	12.9 (3.7–34.2)	6.7 (0.8–34.2)	<0.001
Male (Female)	12 (6)	4 (4)	7 (6)	20 (5)	31 (15)
INR	N/A	1.2 (1.1–1.7)	1.5 (1.2–2.8)	2.0 (1.0–3.2)	1.7 (1.0–3.2)
BMI	N/A	31.1 (27.8–33.4)	25.9 (20.6–41.1)	29.7 (22.4–50.5)	29.7 (20.6–50.5)
AST, U/L	27 (19–66)	59 (21–120)	119 (53–347)	88 (16–190)	90 (16–347)	<0.001
ALT, U/L	25 (16–109)	36.8 (14.0–60.0)	48 (18–194)	35 (16–66)	39 (14–194)	0.015
Alkaline phosphatase, IU/L	52 (37–62)	124 (89–232)	173 (80–518)	144 (71–336)	148 (71–518)	<0.001
Albumin, g/dL	4.2 (3.8–4.3)	3.9 (2.6–4.9)	2.8 (1.9–4.5)	2.4 (1.4–4.3)	2.7 (1.4–4.9)	<0.001
Creatinine, mg/dL	0.88 (0.69–1.07)	0.69 (0.36–1.40)	0.68 (0.32–1.30)	0.89 (0.39–5.68)	0.79 (0.32–5.68)	0.082
MELD scores	N/A	9.2 (6.0–11.0)	16 (12–19)	24 (20–39)	18 (6–39)

Values are presented as mean with ranges. AH, alcohol-associated hepatitis; ALD, alcohol-associated liver disease; ALT, alanine aminotransferase; AST, aspartate aminotransferase; AUD, alcohol use disorder; AUDIT, alcohol use disorders identification test; BMI, body mass index; HC, healthy controls; INR, international normalized ratio; MDF, Maddrey’s discriminant function; MELD, model for end-stage liver disease. Patients with non-severe ALD were made up of patients with AUD having some abnormality in liver testing, as well as AUD patients with liver disease advanced enough to meet criteria of moderate AH (12 < MELD < 20). Severe AH had a MELD > 20. ^aMann–Whitney U test between healthy controls and patients with ALD/AH.

The last alcohol intake was variable, but no participant had consumed alcohol within the last 48 h. All participants’ specimens were collected in the morning after overnight fasting and stored at −80°C until use. All deidentified data from participants who provided urine samples were collected at baseline and information on subsequent death, if available, was also obtained. Clinical data include participant demographics (age, sex), drinking history, medical assessments at admission, and medical history. Confirmatory tests for AH (laboratory and imaging) and markers of liver disease severity (MELD score) were collected and analyzed. A laboratory panel specific for this study was composed of a comprehensive metabolic panel (CMP, including liver injury panel), coagulation measures, and complete blood count tests. All data were analyzed at the University of Louisville.

Untargeted Metabolomics for Polar Metabolite Profiling

Polar metabolite extraction.

All sample processing procedures were performed at 4°C to minimize the loss of volatile metabolites unless stated otherwise. Urine samples were thawed on ice. One hundred microliters of urine were transferred to a fresh tube, and metabolites were extracted by adding methanol at a ratio of 1:4 (vol: vol). After being vortexed for 2–3 min, the mixture was centrifugated at 14,000 rpm for 20 min. Then, 100 µL of supernatant was mixed with 100 µL of H₂O for 2DLC-MS analysis. Group-based pooled urine samples were also prepared by mixing a small portion of the supernatant of samples in the same group.

Untargeted metabolomics for metabolite relative quantification.

All samples were analyzed by parallel two-dimensional liquid chromatography-mass spectrometry (2DLC-MS) composed of a Thermo Q Exactive HF Hybrid Quadrupole-Orbitrap Mass Spectrometer coupled with a Thermo DIONEX Ultimate 3000 HPLC system (Thermo Fisher Scientific, Waltham) (20). Briefly, the LC system was configured in parallel mode with a SeQuant ZIC-cHILIC column (2.1 × 150 mm, 3 µm) as the hydrophilic interaction liquid chromatography (HILIC) column and a Waters Acquity UPLC HSS T3 column (2.1 × 150 mm, 1.8 µm) as thereversed-phasee chromatography (RPC) column. Each column was connected with a 2-μL sample loop. A sample was injected and separated on each of the two columns simultaneously. The eluant of the two columns was then mixed and delivered to the mass spectrometer. The temperature was set to 40°C for the two columns. For separation on the HILIC column, 10 mM ammonium acetate (pH 3.25) was used as mobile phase A and 100% acetonitrile was used as mobile phase B. The flow rate was 0.3 mL/min. For separation on the RPC column, water with 0.1% formic acid was used as mobile phase A and acetonitrile with 0.1% formic acid was used as mobile phase B. The flow rate was 0.4 mL/min. The parameters for mass spectrometry were the same as those we reported previously (24).

Each biological sample was analyzed in random order in positive mode (+) and negative mode (−) to obtain the full MS data for metabolite relative quantification. The pooled samples were analyzed to acquire MS/MS spectra at different collision energies (20, 40, and 60 eV) for metabolite identification. The parameters for MS data acquisition were a full scan from 60 to 900 (m/z), maximum injection time 50 ms, full MS resolution 30,000, and MS/MS resolution 15,000.

Targeted Metabolomics for Absolute Quantification of Caffeine and Its Metabolites

Caffeine metabolite extraction from human urine.

Owing to the large concentration differences of the caffeine metabolites in urine, we prepared two sets of samples using the same extraction method. Initially, 200 µL urine was mixed with 800 µL methanol to precipitate proteins. After 2 min vortex, the mixture was centrifugated at 14,000 rpm for 15 min. To prepare the low-concentration set of samples, 200 µL of supernatant was transferred to a 2-mL Eppendorf tube. The sample was lyophilized to dryness overnight. The dried extract was reconstituted in 400 µL 20% acetonitrile that contains 5 µM [¹³C₃]caffeine and [²H₉]137U. To prepare a high-concentration set of samples, 400-µL supernatant was transferred to a 2-mL Eppendorf tube, and the samples were prepared under the same conditions as the low-concentration set to remove proteins. After overnight lyophilization, the dried extract was reconstituted into 40 µL 20% acetonitrile. Then, all samples were transferred to LC vials for LC-QqQ MS analysis.

LC-QqQ MS/MS analysis.

Analysis of caffeine and its metabolites was performed on a Waters Acquity H-class UHPLC system coupled with a Waters Xevo TQ-S micro triple quadrupole mass spectrometer (Waters, Milford, MA). The chromatographic separation was carried out on a Waters Acquity UPLC HSS T3 column (2.1 × 150 mm, 1.8 µm), and the column temperature was held at 40°C. The mobile phases were A: 0.1% formic acid in water (vol/vol) and B: 0.1% formic acid in acetonitrile (vol/vol). The flow rate was set to 0.3 mL/min. The multislope LC gradient was as follows: kept 3% B during 0–2 min, increased to 5% B during 2–3 min, 5% B during 3–13.6 min, 12% B during 13.6–19.5 min, and then stayed at 12% B during 19.5–24.5 min. The mobile phase B was dropped to 3% B at 24.6 min to re-equilibrate the column. The overall run time was 35 min per injection.

MS detection was performed under positive and negative electrospray ionization (ESI+ and ESI−) modes. The quantification of each metabolite was achieved by multiple reaction monitoring (MRM). The specific MRM transition and detection mode of each metabolite is presented in Supplemental Table S1. Nitrogen was used as the desolvation gas and argon as the collision gas. In ESI+ mode, the capillary voltage was set to 3.70 kV. The cone voltage was compound-specific ranging from 4 V to 44 V. Desolvation temperature was 350°C with a desolvation gas flow of 650 L/h. The collision energy was compound-specific, ranging from 16 eV to 26 eV. In ESI− mode, the capillary voltage was set to 2.6 kV. The other parameters remained the same as in ESI+ mode.

MRM method development and evaluation.

The optimal MRM transition for each metabolite was determined using three types of samples, including the diluted stock solutions of caffeine and its 14 metabolites, a blank, and a mixture of all 15 metabolites (caffeine and its 14 metabolites). The top five abundant transitions were first identified for each metabolite by direct infusion mass spectrometry using a 5 μM standard solution diluted from its stock solution using 20% acetonitrile. The blank was 20% acetonitrile. The mixture of 15 metabolites contained 20 μM of each metabolite prepared from their stock solutions using 20% acetonitrile. Each metabolite in the mixture was identified by its retention time, parent ion m/z, and MS/MS spectrum. The best transition of a metabolite was determined based on the uniqueness of the transition, abundance, and m/z value of the daughter ion. Specifically, the best transition must be unique for the metabolite at its retention time; the abundant transition was chosen if multiple transitions were unique for the metabolite, and the transition with a larger m/z value of daughter ion was selected if multiple transitions had similar abundance.

The precision was evaluated using intraday and interday variations by analyzing three mixtures of 15 metabolites with three different concentrations (0.5, 8, and 40 μM). For the intraday variation, each of the three mixtures was analyzed four times. For the interday variation analysis, each mixture was analyzed one time per day for four consecutive days.

Analytical recovery.

The recovery rate was obtained using stable isotope-labeled (SIL) caffeine and 137 U as internal standards by spiking 5 µM [¹³C₃]caffeine and [²H₉]137U to urine samples before the steps of protein precipitation and after the lyophilization, respectively.

Matrix effect.

SIL internal standard of 5 µM ([¹³C₃]caffeine and [²H₉]137U) was spiked in a blank sample (n = 3) and urine samples after extraction and lyophilization (n = 3). The matrix effect was determined by:

$$\text{matrix\hspace{0.17em}effect}\hspace{0.17em}\left(\%\right)=\frac{A_{\text{s}}}{A_{\text{b}}}\times 100\%$$

where A_s is the peak area of SIL standards in urine sample matrices, A_b is the peak area of SIL standards in blank sample.

Internal calibration.

To obtain the concentration of caffeine metabolites in the urine sample, SIL [¹³C₃]caffeine and [²H₉]137U were added to samples as internal standards. The response factor (RF) for each metabolite related to either [¹³C₃]caffeine or [²H₉]137U was assessed by analyzing 10 µM caffeine metabolites standard mixed with both 10 µM [¹³C₃]caffeine and [²H₉]137U. The RF for each caffeine metabolite is calculated by:

$$\text{RF}=\frac{A_A^{\circ }·C_{\text{SIL}}}{A_{\text{SIL}}·C_A^{\circ }}$$

where $A_A^{\circ }$ is the peak area of analyte A, A_SIL is the peak area of a SIL internal standard, $C_A^{\circ }$ is the concentration of analyte A, and C_SIL is the concentration of a SIL internal standard (25).

External calibration.

The external calibration curve for each caffeine metabolite was generated using a set of mixtures with different concentrations of caffeine and its metabolite standards, ranging from 24 nM to 200,000 nM. The sensitivity of each metabolite was assessed using the limit of detection (LOD), the lower limit of quantification (LLOQ), and the upper limit of quantification (ULOQ). The LOD of a metabolite was defined as its peak intensity in extracted ion chromatogram with a signal-to-noise ratio (S/N) ≥ 3.0. The peak area versus its concentration was plotted for each metabolite to generate a calibration curve, where the weighting factor was 1/X. The LLOQ and the ULOQ values were, respectively, set at a concentration where the deviation of a calculated concentration and the true concentration is less than 20% (24).

Data Analysis

2DLC-MS for untargeted metabolomic profiling.

2DLC-MS data were first converted into mzXML format by the MSConvert (64-bit) software (Proteowizard, Palo Alto, CA). XCMS software was used for spectrum deconvolution (26). The MS/MS spectra of 290 metabolite standards were generated as an in-house database and imported to mzVault software, a component of Compound Discover software (version 3.2, Thermo Fisher Scientific, Inc., Germany). Metabolite identification was then performed using Compound Discover as described previously (20). Briefly, the experimental 2DLC-MS/MS data were first matched to the information of the 290 authentic standards (i.e., parent ion m/z, retention time, and tandem MS spectrum). The experimental data without a match to the authentic standards were then matched to the MS/MS spectra in the Compound Discovery software. The threshold of spectrum matching was set to m/z variation ≤ 5 ppm, retention time variation ≤ 0.15 min, and spectrum similarity score ≥ 0.5. MetSign software was used to process 2DLC-MS data for alignment and normalization (27).

LC-QqQ-MS/MS for metabolite absolute quantification.

MassLynx (v4.2) software (Waters, Milford, MA) was used to acquire and process LC-MS/MS spectral data. The peak integration and external standard calibration curve were generated using its built-in TargetLynx application.

Statistical data analysis.

Statistical analyses were performed using SPSS software (version 25, IBM Corporation, Armonk, NY) and R (version 4.1.0, https://www.r-project.org/). Distributional assumptions of continuous outcomes were checked, and, if needed, a data transformation (e.g., log-transformation) was applied to meet the normality assumption. Partial least squares-discriminant analysis (PLS-DA), a supervised technique that uses the PLS algorithm to explain and predict the membership of samples to groups, was performed to give an overview of the metabolic profile differences among groups. Univariate analysis of metabolite abundance among groups was conducted using one-way ANOVA with Bonferroni’s post test. The decision tree algorithm was used for hierarchical classification to select metabolites that differentiate groups. Group cross linear-by-linear association test (also known as Mantel-Haenszel test for trend) was used for linear trend analysis. Receiver operating characteristic (ROC) analysis was used to classify patients based on the abundance of metabolites in the urine samples. Spearman’s rank correlation was used to measure the association of each metabolite with the clinical parameter MELD score. The thresholds of statistical significance tests were set as one-way ANOVA P ≤ 0.05, areas under the ROC curve (AUC) > 0.7 or < 0.3, and Spearman’s rank correlation test coefficient $\left|\rho \right|$ ≥ 0.4. PLS-DA and quantitative pathway enrichment analysis were performed using Metaboanalyst (version 5.0, https://www.metaboanalyst.ca/). The error bars in histogram plots are the standard errors of the mean (SEM).

RESULTS

Untargeted Metabolic Profiling of Polar Metabolites

We categorized all 64 urine samples into three groups in the current study: healthy controls (HC, n = 18), nonsevere ALD (MELD ≤ 19, n = 21), and severe AH (MELD ≥ 20, n = 25). Polar metabolites extracted from all urine samples were analyzed by 2DLC-MS via untargeted metabolomics. About 8,000 features (i.e., isotopic peaks) were detected, and 380 metabolites were identified. Among the identified metabolites, 63 were identified from our in-house database by parent ion m/z, retention time, and spectrum matching. In contrast, 317 metabolites were identified from the public database by spectrum matching. PLS-DA was performed to study the metabolic profile difference among the three groups using the abundance of the identified metabolites. Figure 1 shows that the metabolic profiles differ among the three groups, with 112 metabolites having variable importance projection (VIP) scores greater than 1.0. The very large values of R² = 0.92 and Q² = 0.82 indicate good discrimination and predictability of the PLS-DA model.

Figure 1.

PLS-DA analysis using metabolites detected by 2DLC-MS. All metabolites detected under positive mode and negative mode were merged for PLS-DA analysis, with R²= 0.92 and Q²= 0.82. Eighteen urine samples were collected from healthy controls (HC, n = 18). Patient urine samples were categorized into nonsevere ALD (MELD ≤ 19, n = 21) and severe AH (MELD ≥ 20, n = 25). AH, alcohol-associated hepatitis; ALD, alcohol-associated liver disease; MELD, model for end-stage liver disease; PLS-DA, partial least squares-discriminant analysis; 2DLC-MS, two-dimensional liquid chromatography-mass spectrometry n, number of subjects in a group.

One-way ANOVA was performed to study the abundance changes of the identified metabolites. A total of 209 metabolites have significant abundance changes (P < 0.05) and fold-change > 20%, of which 92 metabolites have VIP > 1.0 in PLS-DA. Receiver operating characteristic (ROC) analysis was used to evaluate the diagnostic performance of the metabolite abundance changes for differentiating severe AH from nonsevere ALD. The result showed that 157 metabolites are significantly changed, and 112 have an area under the ROC curve (AUC) larger than 0.7 or smaller than 0.3. Table 2 lists the metabolites whose abundance (peak area of 2DLC-MS data) and MELD score are significantly correlated with Spearman’s rank correlation coefficient |ρ| ≥ 0.4.

Table 2.

Metabolites with Spearman’s rank correlation coefficient |ρ| > 0.4 with MELD score

Compound Name	ρ	Compound Name	ρ
7,9-Dimethylguanine	0.71	N-Phenylacetylglutamine	−0.57
β-d-Glucopyranuronic acid	0.65	Lysylvaline	0.55
Saccharopine	0.60	Glucosylgalactosyl hydroxylysine	0.55
Ureidoisobutyric acid	−0.58	Citramalic acid	0.55
3,7-Dimethyluric acid	−0.57	Hydantoin-5-propionic acid	0.54
m1I	0.57	Paraxanthine	−0.54
4-Acetamidobutanoic acid	0.56	Cystathionine	0.53
3-Hydroxyanthranilic acid	0.55	Ethanolamine	0.51
1,3,7-Trimethyluric acid	−0.40	l-Pyroglutamic acid	−0.49
3-Indoxyl sulfate	−0.51	m2,2G	0.48
N3,N4-Dimethyl-l-arginine	0.40	n-Ribosylhistidine	0.48
3-Methylcrotonylglycine	0.48	m3C /m5C	−0.48
Benzoic acid	−0.48	Acetoacetic acid	0.40
Sucrose	0.47	Tyr-lys	−0.47
m1A	0.47	Hypaphorine	−0.47
Am	0.47	p-Cresylsulfate	−0.46
Acetoacetic acid	0.40	DHU	0.45
3-Hydroxysuberic acid	0.45	N6-Acetyl-l-lysine	0.45
4-Hydroxy-3-methoxyphenylglycol sulfate	0.45	Leucine	0.45
Anserine	0.44	Hypoxanthine	−0.45
But-1-ene-1,2,4-tricarboxylic acid	0.40	Cytosine	0.40
1,7-Dimethyluric acid	−0.43	Creatinine	0.44
3-Methyluric acid	−0.42	5-Aminovaleric acid	0.41
1-(β-d-Ribofuranosyl)-1,4-dihydronicotinamide	0.41	N-Acetyl-d-galactosamine 4-sulfate	0.42

DHU, dihydrouridine; ρ, Spearman’s rank correlation coefficient; m¹I, N1-methylinosine; m¹A, N1-methyladenosine; Am, 2′-O-methyladenosine; m^2,2G, N2,N2-dimethylguanosine; m³C/m⁵C, 3-methylcytidine/5-methylcytidine, these two nucleosides co-eluted in 2DLC-MS.

The abundance of the 209 significantly changed metabolites were then used for pathway enrichment analysis, showing that the caffeine metabolic pathway is the most affected in the urine of patients with ALD (Fig. 2A). Supplemental Fig. S1 shows the metabolic pathway of caffeine and the chemical structures of its metabolites (28). Among the 209 metabolites, 13 are in the caffeine metabolic pathway (caffeine and its 12 metabolites). Figure 2B displays metabolites in the caffeine metabolic pathway detected by 2DLC-MS, and the abundance of these metabolites is inversely correlated with MELD score. Five metabolites have Spearman’s rank correlation coefficient |ρ| > 0.4, of which 3,7-dimethyluric acid has the best correlation with a ρ = −0.57, and paraxanthine has the second-best correlation score ρ = −0.54. However, caffeine has a very weak correlation with MELD score ρ = −0.09 (Fig. 2B), and its abundance change is not statistically significant (Fig. 2C). Figure 2D depicts the relative quantification of metabolites in the caffeine metabolic pathway by untargeted metabolomics. Statistical significance tests show that most of these metabolites have statistically significant abundance changes between the two groups. The linear-by-linear association analysis also indicates that all caffeine metabolites have a statistically significant decreasing trend with disease severity (Fig. 2D).

Figure 2.

Metabolites in the caffeine metabolism pathway with significant abundance decrease in patient urine. A: results of quantitative pathway enrichment analysis using all significantly changed metabolites detected by 2DLC-MS. The caffeine metabolism pathway was the most affected pathway by alcohol consumption, followed by the arginine biosynthesis pathway, aminoacyl-tRNA biosynthesis pathway, and histidine pathway. B: metabolites in the caffeine metabolism pathway detected in urine samples by 2DLC-MS and their Spearman’s rank correlation coefficients with MELD score. C: quantification results of caffeine in the urine samples by 2DLC-MS. D: quantification of other 12 metabolites in the caffeine metabolism pathway by 2DLC-MS. One-way ANOVA was used for statistical significance test and Bonferroni correction was used for the post hoc analysis: *P < 0.05; **P < 0.01; ***P < 0.001. The P values of linear-by-linear association test for linear trend are indicated as #P < 0.05, ##P < 0.01, ###P < 0.001. MELD, model for end-stage liver disease; 2DLC-MS, two-dimensional liquid chromatography-mass spectrometry.

Quantification of Caffeine Metabolites by Targeted Metabolomics

To confirm the results of caffeine metabolites measured by untargeted metabolomics, we further developed a targeted metabolomics method, LC-QqQ MS using multiple reaction monitoring (MRM). We noticed that the concentrations of 1X, 3X, 7X, 17X, 37X, and 13X (Supplemental Fig. S1) are much higher than the concentrations of other metabolites of caffeine pathway in the urine samples, and their concentrations are all beyond the ULOQ of the calibration curves. Therefore, we prepared two sets of urine samples for LC-QqQ MS analysis. The metabolite concentration in the high-concentration set of samples was 20 times higher than in the low-concentration set. We also prepared a pooled sample for quality control (QC) that contained the same volume of each sample. The QC sample was analyzed once after every 10 injections of urine samples, and the data from the QC sample were used to inspect the variations of LC-QqQ MS during the analysis of urine samples. The relative standard deviation (RSD) of the median for the calculated concentrations of the analytes in the QC sample was 3.6% (Supplemental Table S2), which agrees with the data of interday variation (Supplemental Table S9). AAMU and 3 U had large RSD (AAMU 9.7% and 3 U 10.9%) owing to their ultralow concentration in human urine.

The concentrations of caffeine and all of its 14 metabolites were quantified by targeted metabolomics. Figure 3 depicts the concentrations of caffeine and its metabolites between the three groups determined by external standard calibration. One-way ANOVA shows that 17X and 1X have significant concentration differences between any two groups, i.e., HC versus nonsevere ALD, HC versus severe AH, and nonsevere ALD versus severe AH (Fig. 3, Supplemental Table S4). Figure 4 depicts the same information determined by internal standard calibration, and one-way ANOVA also shows the concentrations of 1X, 17X, and AAMU to have significant differences between any two groups (Fig. 4, Supplemental Table S5). Although the caffeine concentration varied among the three groups, such variation is not statistically significant (Figs. 3 and 4), which agrees with the untargeted metabolomics data (Fig. 2C). Overall, the concentrations of caffeine and its 14 metabolites calculated by internal and external standard calibration have no significant differences. The Spearman’s rank correlation coefficients (ρ value) are greater than 0.96 (Supplemental Table S3), showing the two calibration methods provide very similar results even though they are not identical. Therefore, further data analysis is based on the concentration calculated by the internal standard calibration method.

Figure 3.

Intergroup concentration comparison of caffeine and its 14 metabolites in human urine determined by external standard calibration among healthy control (HC), nonsevere ALD (N-S), and severe AH (S). *P < 0.05; **P < 0.01; ***P < 0.001 by one-way ANOVA with Bonferroni’s posttest. #P < 0.05; ##P < 0.01; ###P < 0.001 by linear-by-linear association test. AH, alcohol-associated hepatitis; ALD, alcohol-associated liver disease.

Figure 4.

Intergroup concentration comparison of caffeine and its 14 metabolites in human urine among healthy control (HC), nonsevere ALD (N-S), and severe AH (S). Metabolite concentration was determined by internal standard calibration using isotope labeled standards. *P < 0.05; **P < 0.01; ***P < 0.001 by one-way ANOVA with Bonferroni’s posttest. #P < 0.05; ##P < 0.01; ###P < 0.001 by linear-by-linear association test. AH, alcohol-associated hepatitis; ALD, alcohol-associated liver disease.

The receiver operating characteristic analysis (ROC) was also applied to evaluate the diagnostic performance of the concentration changes of caffeine metabolites for differentiating HC from nonsevere ALD and nonsevere ALD from severe AH. The ROC analysis indicates that 1X, 17X, and AAMU are significantly changed. The area under the curve (AUC) of these three metabolites are 0.81, 0.81, and 0.79 between HC and nonsevere AH, and 0.72, 0.80, and 0.81 between nonsevere ALD and severe AH, respectively (Supplemental Table S6). Overall, different bioanalytical methods (internal standard and external standard calibrations) and data analysis methods (one-way ANOVA, Spearman’s rank correlation, and ROC) reveal that 1X, 17X, and AAMU can be used to assess the severity of ALD.

We recategorized patients into alcohol use disorder with mild liver injury (AUD, n = 8) and all patients with alcohol-associated hepatitis, with both moderate and severe stages (AH, n = 38), and performed the same study. Supplemental Fig. S3 depicts the concentration differences of caffeine metabolites among HC, AUD, and AH. The concentration of 1X shows a significant change between any two groups, i.e., HC versus AUD, AUD versus AH, and HC versus AH. However, 17X and AAMU cannot differentiate HC and AUD groups. The ROC result in Supplemental Table S14 also supports that 1X can distinguish HC, AUD, and AH, with an AUC of 0.86 between HC and AUD and 0.90 between AUD and AH.

We further categorized patients into AUD (n = 8), patients with moderate alcohol-associated hepatitis (M-AH, n = 13), and patients with severe alcohol-associated hepatitis (S-AH, n = 25). Supplemental Fig. S4 depicts the concentration differences of caffeine metabolites among HC, AUD, M-AH, and S-AH. The linear-by-linear association test shows that some of the caffeine metabolites, including 1X, 17X, and AAMU, have a significant decreasing trend with the increase of disease severity. However, one-way ANOVA indicates that no single caffeine metabolite can differentiate all four groups. Compared with other metabolites, 1X performs the best in one-way ANOVA in differentiating groups. But it still cannot differentiate M-AH from S-AH and AUD from M-AH (Supplemental Fig. S4). ROC analysis has the same results (Supplemental Table S15).

Metabolites in the 17X Subpathway Serve as Diagnostic Markers for ALD

Previous studies reported that more than 80% of caffeine in the human body is metabolized into 17X, and only 3% or less of caffeine intake is excreted unchanged in urine (29–32). The current study did not control for caffeine intake, and we do not know whether ALD altered the percentage of excreted caffeine in urine. As an alternative approach to investigating the alcohol/ALD effect on the caffeine metabolites, we normalized the concentration of every metabolite in a participant sample by the concentration of the excreted caffeine as $\left\{\raisebox{1ex}{${\text{C}}_i$}\!\left/ \!\raisebox{-1ex}{${\text{C}}_{\text{c}}$}\right.\right\}$, where C_i is the concentration of a caffeine metabolite, and C_c is the concentration of excreted caffeine in urine. Figure 5 depicts the results of statistical significance tests, showing that the concentration ratio of seven metabolites to caffeine has significant intergroup alteration. It is interesting that all of these seven metabolites are in the 17X subpathway (Supplemental Fig. S1). Among the seven metabolites, the concentration ratios of 17 U, 1X, AAMU, and 17X to caffeine differ significantly among the three groups and could differentiate any two groups with P < 0.05. The linear-by-linear association test also shows that all seven metabolites in 17X subpathway have a significant linear decreasing trend from HC to nonsevere ALD and reach the minimum in severe AH (P < 0.001). Compared with the results depicted in Fig. 4 using the concentration of each metabolite for one-way ANOVA, 1X, 17X, and AAMU are the only three metabolites that can differentiate the ALD stages using either the concentration or the concentration ratio of a metabolite to excreted caffeine. It is noteworthy that 1X can be metabolized from 13X and 17X (Supplemental Fig. S1). Our data show that 1X and 17X are significantly decreased from HC to nonsevere ALD and reach their minimum in severe AH, using either concentration or concentration ratio for analysis (Figs. 4 and 5). However, 13X does not have significant changes among HC, nonsevere ALD, and severe AH groups. These data strongly suggest that the concentration alteration of 1X derives mostly from the 17X subpathway.

Figure 5.

Concentration ratios of metabolites in 17X subpathway to excreted caffeine in human urine among healthy control (HC), nonsevere ALD (N-S), and severe AH (S). Metabolite concentration was determined by internal standard calibration using isotope-labeled standards. *P < 0.05; **P < 0.01; ***P < 0.001 by one-way ANOVA with Bonferroni’s posttest; ###P < 0.001 by linear-by-linear association test. AH, alcohol-associated hepatitis; ALD, alcohol-associated liver disease.

Table 3 lists the results of ROC analyses using the concentration ratios. The AUCs of most of these metabolites are larger than 0.7 when comparing HC and nonsevere ALD, whereas the AUCs of the majority are small when comparing nonsevere ALD and severe AH, except 1X, 17X, and AAMU. The AUCs of 1X, 17X, and AAMU are 0.795, 0.901, and 0.841 in HC versus nonsevere ALD, and 0.733, 0.737, and 0.819 in nonsevere ALD versus severe AH. The large value of AUCs of these three metabolites indicates that the concentration ratio of these three metabolites to the excreted caffeine can simultaneously differentiate different stages of ALD, which agrees with the results of one-way ANOVA.

Table 3.

ROC analysis of concentration ratio of metabolites/caffeine

	Healthy Control vs. Nonsevere ALD			Nonsevere ALD vs. Severe AH
	P Value	AUC	95% CI	P Value	AUC	95% CI
137U	<0.001	0.865	0.735–0.994	0.181	0.625	0.438–0.812
1X	0.002	0.795	0.646–0.944	0.009	0.733	0.580–0.885
1U	0.001	0.833	0.678–0.989	0.575	0.553	0.371–0.734
17U	<0.001	0.900	0.789–1.000	0.013	0.736	0.573–0.899
3U	0.006	0.796	0.626–0.965	0.800	0.526	0.316–0.736
7U	0.001	0.837	0.699–0.976	0.038	0.703	0.541–0.865
17X	<0.001	0.901	0.803–0.999	0.008	0.737	0.582–0.892
37U	0.003	0.821	0.665–0.976	0.603	0.556	0.347–0.764
AAMU	0.001	0.841	0.695–0.988	0.005	0.819	0.655–0.982
13U	0.001	0.830	0.682–0.978	0.252	0.612	0.422–0.802
7X	0.003	0.804	0.658–0.949	0.068	0.679	0.503–0.854
3X	0.015	0.752	0.585–0.918	0.482	0.569	0.389–0.748
37X	0.032	0.709	0.537–0.881	0.848	0.518	0.337–0.698
13X	0.261	0.607	0.423–0.791	0.353	0.583	0.405–0.761

ALD, alcohol-associated liver disease; AH, alcohol-associated hepatitis; AUC, area under the curve; CI, confidence interval; ROC, receiver operating characteristic.

The alternative approach to evaluate the alcohol/ALD effect using the concentration ratio of a metabolite and the excreted caffeine was also applied to different patient grouping methods for early-stage ALD diagnosis. When patients were categorized as AUD and AH, the three metabolites, 1X, 17X, and AAMU, could not differentiate HC from AUD (Supplemental Fig. S5). When patients were further grouped as HC, AUD, M-AH, and S-AH, none of the three metabolites can differ M-AH from S-AH either (Supplemental Fig. S6). However, the linear-by-linear association analysis shows statistically significant stepwise descending trends of the concentration ratio of these metabolites with the ascending severity of ALD (Supplemental Fig. S6).

DISCUSSION

Caffeine is mainly absorbed in the intestine and protects the liver from steatosis by improving fat homeostasis (33). It reduces intrahepatic lipid content and stimulates β-oxidation in the liver by an autophagy-lysosomal pathway (34). The mobilization and hydrolysis of triglycerides to free fatty acids (FFAs) through the autophagy-lysosomal pathway leads to increased delivery of FFAs to the mitochondria, and in turn, increases the β-oxidation of fatty acids and oxidative phosphorylation (34). Caffeine has also been reported to prevent hepatic fibrosis by reducing oxidative stress (33) and functioning as an antagonist of the adenosine receptor (35). In ALD, caffeine is thought to protect the liver against alcohol-associated liver injury by attenuating the inflammatory response and oxidative stress (36). Caffeine treatment decreases serum and tissue inflammatory cytokines levels and tissue lipid peroxidation and inhibits the necrosis of hepatocytes (36). In a model of alcohol-induced liver fibrosis, caffeine protected against liver fibrosis by dampening the cAMP/PKA/CREB pathway in rat hepatic stellate cells through adenosine A2A receptors (37).

The decrease of caffeine metabolites in the urine of patients with ALD was first reported more than 20 years ago (38). The quantification results of the most caffeine metabolites in the current study are consistent with the literature report. Our study shows that the abundance level of caffeine is not changed in the urine of patients, suggesting that caffeine transport (i.e., caffeine excreting into the urine) may not be disrupted by alcohol. All caffeine metabolites decrease with the increasing severity of ALD, although the theophylline (13X) decrease is not statistically significant in the targeted metabolomics data (Figs. 3 and 4), and metabolites 1X, 17X, and AAMU can differentiate different stages of ALD. In contrast, previous studies concluded that the decrease of metabolites in the caffeine metabolism pathway does not reflect the ALD disease severity. This apparent discrepancy might be partly explained by different patient categorization methods used in the studies. A Child-Pugh score was used for patient categorization in the previous publication (38), but MELD score is used in the current study. The Child-Pugh scoring system categorizes patients with chronic liver disease into stages ranging from well-compensated cirrhosis (A) to decompensated cirrhosis (C) and is not designed to evaluate less severe forms of liver disease such as mild or moderate AH. The MELD score is a newer assessment for the severity of chronic liver disease. It has been demonstrated that MELD scores better assess disease stage than the Maddrey discriminant function (MDF) score in alcohol-associated hepatitis (39). Our previous study on quantifying bile acids in the urine of patients with ALD also showed that the MELD score performs the best for patient sample grouping when compared with MDF and the ratio of AST to ALT (21).

Another reason for the discrepancy between the current study and the literature report is the analysis methods and instruments used. An HPLC system coupled with a UV detector was used in a previous report to identify and quantify caffeine metabolites in ALD (38, 40). HPLC has limited separation power, resulting in many metabolites coeluting from the column and entering the UV detector together. Any UV-absorbing compounds that coelute with the target analyte in HPLC could affect the quantitation results. With the dramatic rapid development of chromatography and mass spectrometry, we coupled UPLC and MS to greatly increase the peak capacity of chromatography and further separate the coeluting compounds by their m/z values in MS. Therefore, the analysis methods employed in the current study have much-improved sensitivity, accuracy, and resolving power. Furthermore, we first analyzed the polar metabolite abundance differences between the urine of HC and patients with ALD using a parallel 2DLC-MS platform to achieve a much-increased metabolite coverage and high confidence in metabolite identification. We then developed a targeted metabolomics method, LC-QqQ MS using MRM, to validate the results of untargeted metabolomics. The lower limit of quantification (LLOQ) of our targeted metabolomics approach is 0.049–0.391 µM, interday and intraday variation is less than 10%, sample preparation recovery rate is higher than 84%, and the matrix effect is between 84% and 110%. The detailed LC-QqQ MS method development and assessment are described in the Supplemental material.

Eighteen healthy controls and 46 patients with ALD were enrolled in this study. Based on MELD scores, patients can be grouped into alcohol use disorder with mild liver injury (AUD, n = 8), patients with moderate alcohol-associated hepatitis (M-AH, n = 13), and patients with severe alcohol-associated hepatitis (S-AH, n = 25). Grouping patients using different criteria clearly affects the analysis results. For instance, no metabolite can differentiate AUD from M-AH and M-AH from S-AH if the patients are grouped into three groups (Supplemental Fig. S6). If we merge AUD and M-AH to form a nonsevere ALD (n = 21) group, three metabolites in the 17X subpathway, 1X, 17X, and AAMU, can differentiate all three groups, i.e., HC, nonsevere ALD and severe AH (Figs. 3, 4, and 5). Many factors, including biomedical reasons and statistical power, contribute to such a difference. Statistical power is reduced when multiple groups are formed from a limited sample size. Therefore, we chose to group the patients into nonsevere ALD and severe AH.

Our study shows that the concentrations of three metabolites in the caffeine pathway (1X, 17X, and AAMU) strongly correlate with the MELD score, suggesting these metabolites may be potential markers for liver injury. It has been demonstrated that caffeine suppresses transforming growth factor (TGF)-β-dependent and -independent connective tissue growth factor (CTGF) expression in hepatocytes in vitro and in vivo (41). Paraxanthine (17X) is the most effective pharmacological repressor of hepatocellular CTGF expression among the caffeine-derived metabolic methylxanthines and may function as a potential antifibrotic drug (42). A study on narcolepsy showed that paraxanthine significantly promotes wakefulness and proportionally reduces rapid eye movement and nonrapid eye movement sleep in both control and narcoleptic mice. The wake-promoting action of paraxanthine is greater than that of caffeine (43). These data indicate that paraxanthine might be the most important metabolite contributing to caffeine abuse-related effects in the caffeine metabolic pathway. This study shows that 1X, 17X, and AAMU have a high correlation with ALD severity as measured by MELD score, and their concentrations decrease in urine with an increase in disease severity. The same trend is observed using the concentration ratios of those three metabolites to caffeine. These data suggest that the level of these molecules is related to the stage of ALD. We suggest that 1X, 17X, and AAMU from human urine could be used to diagnose early ALD and delineate disease severity.

This study has limitations. First, a relatively small sample size makes it impossible to identify the roles of metabolites with minor effect sizes, and we could not identify any underlying sex effects. Therefore, more extensive studies are needed to precisely elucidate the role of various demographic measures. Second, this study was not designed as a treatment-based investigation; thus, identifying the role of any metabolites in therapeutic-mechanistic pathways was also not within the scope of this study. However, given the potential prognostic value of the candidate metabolites, this study supports the need for longitudinal AH studies where the efficacy of treatment and change in liver function can be measured in relation to the candidate metabolites identified in this study. Third, the pharmacokinetics of caffeine has been substantially studied. The three potential biomarkers found in this study, 1X, 17X, and AAMU, are metabolized from caffeine by enzymes, cytochrome P450 1A2 (CYP1A2) and N-acetyltransferase 2 (NAT2) (40, 44). It has been reported that decreased activity of CYP1A2 is positively correlated to the severity of liver diseases (45). However, other lifestyle choices, such as long-term smoking, may also induce the activity of CYP1A2 (46, 47). The activities of CYP1A2 and NAT2 in each participant is out the scope of the current study. Next, the current study did not attempt to detect metabolites in the patient’s liver tissue where caffeine biotransformation occurs. This study was also not designed to determine whether caffeine or its metabolites provide hepatic protection. Finally, we did not compare ALD to other types of liver disease/injury. Based on these limitations of the current study, we will do a much larger validation assay and determine whether a spot urine is as good as a urine specimen following a caffeine load.

The study also has unique advantages. Previous reports gave oral caffeine “loads” (sometimes using restricted diets) and often evaluated multiple time points. Thus, these tests were not practical for everyday clinical practices. We could determine existing caffeine and metabolite levels without specific diets or caffeine intake. And our results show that a randomly collected spot urine appears to be a sensitive biomarker for the severity of ALD across the spectrum of liver injury. The fact that our studies were not complicated by requirements of specific caffeine intakes and that levels could be measured in a single urine sample makes this very practical. The relation of “endogenous” caffeine to a metabolite provides a markedly simplified functional assay. Caffeine metabolites provide an evaluation of the functional activity of the liver and are not just markers of liver injury, as are AST and ALT. Moreover, while AST and ALT may detect early liver injury in ALD, they do not reflect the severity of injury in ALD. Indeed, AST and ALT levels are often higher in patients with normal bilirubin levels in alcohol detoxification programs than in patients with severe AH with MELD > 20 (48). Finally, if the caffeine metabolites are the therapeutic components, this study suggests that patients with increasing severity of liver injury would accrue decreasing benefits.

Conclusions

We performed untargeted metabolomics by parallel 2DLC-MS to identify and quantify the polar metabolites in human urine. Pathway enrichment analysis of the untargeted metabolomics data indicated that caffeine metabolism was the most highly affected pathway in ALD. To further measure the concentrations of caffeine and its metabolites, we developed a targeted LC-QqQ MS MRM method that can quantify caffeine and its 14 metabolites in 35 min. The targeted metabolomics data agree with the results of untargeted metabolomics, showing that all caffeine metabolites have significant concentration differences between the healthy controls and patients. Three caffeine metabolites, mainly in the paraxanthine subpathway, 1-methylxanthine, paraxanthine, and 5-acetylamino-6-amino-3-methyluracil, are highly correlated to the severity of ALD. Their urine concentrations can differentiate healthy controls from nonsevere ALD and nonsevere ALD from severe AH. These three metabolites might serve as functional biomarkers in diagnosing and staging ALD.

DATA AVAILABILITY

Data will be made available upon reasonable request.

SUPPLEMENTAL DATA

Supplemental Figures S1–S6 and Supplemental Tables S1–S15: https://doi.org/10.6084/m9.figshare.21611196.v4.

GRANTS

This work was supported by National Institutes of Health (NIH) grants S10OD020106 (to X.Z.), K23AA029198 (to V.V.), 1P20GM113226 (to C.J.M.), 1P50AA024337 (to C.J.M.), 1U01AA026934 (to C.J.M.), 1U01AA026926 (to C.J.M.), and 1U01AA026980 (to C.J.M.). This work was also supported by the Department of Veterans Affairs, 1I01BX002996-01A2 (to C.J.M.).

DISCLAIMERS

The content is solely the authors’ responsibility and does not necessarily represent the official views of the National Institutes of Health.

DISCLOSURES

No conflicts of interest, financial or otherwise, are declared by the authors.

AUTHOR CONTRIBUTIONS

C.J.M. and X.Z. conceived and designed research; R.X., L.H., and V.V. performed experiments; R.X., L.H., X.M., and S.K. analyzed data; R.X., L.H., V.V., S.K., W.F., C.J.M., and X.Z. interpreted results of experiments; R.X., L.H., and X.M. prepared figures; R.X. and L.H. drafted manuscript; R.X., L.H., V.V., S.K., W.F., C.J.M., and X.Z. edited and revised manuscript; R.X., L.H., V.V., X.M., S.K., W.F., C.J.M., and X.Z. approved final version of manuscript.