Integrated multiomics profiling predicts anti-tuberculosis drug-induced liver injury

Abstract

Background

Anti-tuberculosis drug-induced liver injury (ATB-DILI) severely compromises tuberculosis (TB) treatment. We aimed to identify pretreatment predictors via integrated proteomic, metabolomic, and gut microbiome profiling.

Methods

This prospective multicenter study enrolled 72 adults who were receiving HRZ therapy. Serum, urine, and stool samples were collected before pretreatment. Serum proteins (RBP-4, CHGA, CPB2, ANT3, APOD) were quantified via ELISA. Nontargeted metabolomics (LC-MS) was used to analyze serum/urine, and 16 S rRNA sequencing was used to characterize the fecal microbiota. Liver injury (ALT/TBIL ≥ 2×ULN; RUCAM ≥ 3) was monitored biweekly/monthly. The ATB-DILI (n = 35) and non-ATB-DILI (n = 37) groups were compared statistically. A random forest model was used to integrate significant features (100-fold cross-validation).

Results

ATB-DILI developed at a median of 29 days (IQR:14–30) and was predominantly hepatocellular (54.3%). The pretreatment levels of all five proteins were elevated in ATB-DILI patients (p < 0.0001). Serum metabolomics revealed 163 differentially abundant metabolites (137↑/26↓; OPLS-DA R²Y = 0.692, Q²=0.351), and urine metabolomics revealed 106 (42↑/64↓; R²Y = 0.972, Q²=0.364). Beta diversity differed significantly between groups (Adonis P = 0.004), with Catenibacterium/Lactococcus enriched in ATB-DILI. Strong correlations linked the microbiota, metabolites, and liver enzymes. The integrated multiomics model (serum/urine metabolites, microbiome, proteins) achieved superior prediction (AUC = 0.880), outperforming single-platform models (serum metabolites:0.859; urine:0.803; microbiome:0.691; proteins:0.671).

Conclusion

Pretreatment alterations in serum proteins, host metabolism, and the gut microbiota predict ATB-DILI risk. An integrated multiomics model enables early intervention for personalized TB therapy.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12931-026-03608-3.

Introduction

Prior to the COVID-19 pandemic, tuberculosis (TB) was the leading cause of death from a single infectious agent globally, posing a significant threat to public health [1]. The cornerstone of TB treatment involves combination chemotherapy, which primarily utilizes the first-line drugs isoniazid (H), rifampicin (R), pyrazinamide (Z), and ethambutol (E) [2]. The World Health Organization (WHO) recommends the standardized 2HRZE/4HR regimen for drug-susceptible TB [3]. However, the prolonged duration of treatment and the need for multidrug regimens increase the risk of adverse drug reactions (ADRs), including gastrointestinal disturbances and, most critically, antituberculosis drug-induced liver injury (ATB-DILI) [4–7].

Among the ADRs associated with anti-TB therapy, ATB-DILI is the most common [8]. Its occurrence presents substantial clinical challenges. Mild to moderate ATB-DILI often necessitates dose reduction or discontinuation of hepatotoxic drugs, potentially leading to suboptimal therapy, regimen changes, and prolonged treatment duration [9]. Severe cases can progress to acute liver failure [10], which may require liver transplantation—a costly intervention with limited donor availability [11]. Consequently, ATB-DILI imposes a significant socioeconomic burden and remains a major public health concern. A critical limitation in managing ATB-DILI is the nonspecificity of its clinical presentation and the current lack of validated diagnostic biomarkers for early detection and intervention.

Proteomic and metabolomic approaches have been increasingly applied to investigate drug-induced liver injury (DILI). Previous studies have revealed alterations in various serum proteins and metabolites associated with oxidative stress, inflammation, and bile acid metabolism in DILI patients [12, 13]. However, many of these findings are derived from postinjury samples, limiting their utility for prediction. Additionally, relatively few studies have applied proteomic and metabolomic approaches specifically to ATB-DILI. Furthermore, human studies specifically examining the role of the gut microbiome in ATB-DILI, particularly the fecal microbiome prior to treatment initiation, are notably lacking. Multiomics approaches—integrating proteomics, metabolomics, and microbiome analysis—hold significant promise for identifying biomarkers enabling the early diagnosis of antituberculosis drug-induced liver injury (ATB-DILI) by providing a more comprehensive view of the interconnected pathogenic pathways [14, 15]. Crucially, while animal studies exist, human microbiome research specifically investigating ATB-DILI associated with first-line anti-TB drugs remains scarce, and studies characterizing the fecal microbiome in human ATB-DILI are currently unreported.

Therefore, this study pursued two primary objectives. First, we aimed to validate a panel of candidate serum proteins—retinol-binding protein 4 (RBP-4), chromogranin-A (CHGA), carboxypeptidase B2 (CPB2), antithrombin III (ANT3), and apolipoprotein D (APOD)--- which were previously identified through our internal, unpublished proteomic screening via enzyme-linked immunosorbent assay (ELISA). Second, we integrated these validated proteomic data with pretherapeutic serum/urine metabolomic profiles and fecal microbiome compositions to construct a multiomics-based early warning model for ATB-DILI. This integrated approach provides a novel predictive framework for identifying individuals at high risk prior to liver injury onset.

Methods

Study population and sample collection

This prospective multicenter study enrolled patients between May 2022 and June 2023. Sex- and age-matched serum, urine, and stool samples were collected from all participants prior to initiating anti-TB treatment. Patients with active tuberculosis (TB) admitted to the tuberculosis departments of West China Hospital of Sichuan University, the Affiliated Hospital of Zunyi Medical University, and Guangyuan People’s Hospital between May 2022 and June 2023 were enrolled.

The sample size was estimated via G*Power software (version 3.1.9.7). On the basis of preliminary data suggesting an effect size (Cohen’s d) of approximately 0.8 for the preliminary differences observed in key metabolite levels between groups, with an alpha error of 0.05 and a power of 0.8, a minimum total sample size of 52 was needed. We enrolled 72 patients to account for potential dropouts and to ensure robust subgroup analysis.

Inclusion and exclusion criteria

The inclusion criteria were as follows: (1) age ≥ 18 years; (2) diagnosis of active TB (clinically or laboratory-confirmed); (3) prescribed an anti-TB regimen including isoniazid, rifampicin, and pyrazinamide and ethambutol (E), i.e., the standard HRZE regimen (dosages provided in Supplementary Table 2); (4) no antibiotics, probiotics, or gastrointestinal motility medications within the preceding month; and (5) good medication adherence, defined as self-reported missed doses of ≤ 5% of total prescribed doses during follow-up and follow-up capability. The exclusion criteria were as follows: (1) preexisting liver disease (e.g., HBV, cirrhosis) or baseline liver dysfunction (ALT ≥ 2x ULN or TBIL ≥ 2x ULN); (2) concurrent use of hepatotoxic drugs (e.g., acetaminophen, herbal medicines); (3) severe cardiac, pulmonary, or renal insufficiency; (4) comorbid autoimmune diseases, malignancies, HIV infection, or urinary system diseases; (5) gastrointestinal diseases (e.g., intestinal TB, Crohn’s disease); and (6) pregnancy.

Definitions

TB diagnosis followed Chinese national standards (WS 288–2017) [16]. Anti-TB drug-induced liver injury (ATB-DILI) was defined as ALT ≥ 2x ULN or TBIL ≥ 2x ULN during treatment [17]. Causality was assessed via the RUCAM simplified scoring system (a score ≥ 3 indicates probable ATB-DILI) [18]. The injury type (hepatocellular, cholestatic, or mixed) was classified according to the R value (R = [ALT/ALT ULN]/[ALP/ALP ULN]; Supplementary Table 1). All patients provided informed consent. The standard regimen was a 2-month intensive phase (HRZE) followed by a ≥ 4-month consolidation phase (HRE), which was administered once daily (doses in Supplementary Table 2).

Data collection and follow-up

Baseline demographics (sex, age, ethnicity, height, weight, alcohol/smoking history) and routine blood/liver/renal function tests were recorded before treatment. During the first 2 months of treatment, patients underwent biweekly clinical assessments and blood tests, followed by monthly assessments thereafter (with ad hoc reviews for clinical changes) until treatment completion. Data were extracted from electronic medical records. Liver function impairment was managed with treatment adjustment and/or hepatoprotective therapy. The follow-up methods included outpatient visits, telephone follow-ups, and telephone contact.

Specimen processing

Fasting peripheral blood (3 ml) was collected into serum separator tubes, morning midstream urine was collected into sterile containers, and fresh stool was collected into sterile tubes. The serum was separated via centrifugation (4000 rpm, 10 min, 4 °C) within 2 h of collection. Serum, urine (500 µL), and stool aliquots (250 mg) were stored at -80 °C.

ELISA

Serum concentrations of candidate proteins (RBP-4, CHGA, CPB2, ANT3, APOD) were quantified via commercial ELISA kits from Jingmei BioTech (Jiangsu, China) (RBP-4, catalog # JM-04291H1; CHGA, catalog # JM-1002H1; CPB2, catalog # JM-05366H1; ANT3, catalog # JM-6577H1; APOD, catalog # JM-6573H1) according to the manufacturer’s protocols.

Nontargeted metabolomics (serum/urine)

Serum

Thawed serum (80 µL) was mixed with precooled methanol (400 µL), vortexed, sonicated (ice bath, 20 min), and centrifuged (16000 × g, 20 min, 4 °C). The supernatant was vacuum-dried, reconstituted in precooled 50% methanol (80 µL), centrifuged (20000 × g, 15 min, 4 °C), and analyzed via LC‒MS (4 µL injection).

Urine

The urine (100 µL) was mixed with precooled methanol (400 µL), sonicated (ice bath, 20 min), incubated (-20 °C, 1 h), and centrifuged (16000 × g, 20 min, 4 °C). The supernatant was vacuum-dried, reconstituted in methanol: water (1:1, v/v; 100 µL), centrifuged (20000 × g, 15 min, 4 °C), and analyzed via LC‒MS (4 µL injection).

LC‒MS analysis

Nontargeted metabolomic analysis was performed via an ultrahigh-performance liquid chromatography (UPLC) system (Thermo Scientific) coupled to a Q Exactive Plus mass spectrometer with a heated electrospray ionization (HESI) source.

Chromatographic separation was achieved on an ACQUITY UPLC^® HSS T3 column (1.8 μm, 2.1 mm × 100 mm; Waters) maintained at 40 °C. The mobile phase consisted of solvent A (0.1% formic acid in water) and solvent B (acetonitrile). The detailed gradient elution program is provided in Supplementary Tables 3 & 4. The flow rate was 0.3 mL/min and the injection volume was 4 µL.

Mass spectrometric detection was performed in both positive and negative ionization modes. The HESI source conditions are detailed in Supplementary Tables 5, and the mass analyzer settings are listed in Supplementary Table 6.

Quality control (QC) samples, prepared by pooling equal aliquots from all study samples, were analyzed intermittently throughout the analytical sequence (specifically, one QC sample was inserted for every 10 experimental samples) to monitor instrument stability and data quality. The data were considered acceptable when the relative standard deviation (RSD%) of the peak areas for metabolic features in the QC samples was less than 30%.

Data processing and metabolite identification were performed via MSDIAL software (ver. 4.9) for peak extraction, alignment, and normalization. Metabolites were identified by matching the acquired MS and MS/MS spectra against the MassBank public database and an in-house database, using matching criteria of mass accuracy < 10 ppm and an MS/MS spectral similarity score > 0.7.

Fecal 16 S rRNA sequencing

Microbial DNA was extracted (OMEGA Soil DNA Kit), quantified, and quality-checked (agarose gel). The V3-V4 region was amplified (Pfu polymerase), and libraries were prepared (TruSeq Nano DNA LT Kit). Sequencing was performed on an Illumina MiSeq platform. QIIME2 processed sequences: primer removal (cutadapt trim-pairs), quality control, denoising, merging, and chimera removal (DADA2). Operational taxonomic units (OTUs) were clustered at 97% similarity.

Statistical analysis

Demographics/proteins

The data were analyzed via SPSS 26.0 and GraphPad Prism 8.0. Continuous variables (means ± SDs or medians [IQRs]) and categorical variables (%) were compared between the ATB-DILI and non-ATB-DILI groups via t tests, Mann‒Whitney U tests, chi‒square tests, or Fisher’s exact tests (p < 0.05 significant).

Metabolomics

Multivariate statistical analyses, including principal component analysis (PCA) and orthogonal partial least squares-discriminant analysis (OPLS‑DA), were performed in Python. Differentially abundant metabolites were identified on the basis of the following criteria: variable importance in projection (VIP) ≥ 1, fold change (FC) ≥ 1.5 or ≤ 1/1.5, and p < 0.05 using Student’s t‑test with false discovery rate (FDR) correction for multiple comparisons. The data were preprocessed via log-transformation and Pareto scaling prior to multivariate analysis.

Microbiomics

Species annotation (Greengenes database). Alpha diversity (Chao1, Shannon, Pielou’s evenness, Faith’s PD, Good’s coverage; QIIME2) and beta diversity (weighted UniFrac; PCoA/NMDS in R/QIIME2) were assessed. The statistical significance of the overall group differences was confirmed by a permutational multivariate analysis of variance (Adonis test) (P = 0.004). Differentially abundant taxa were then identified via linear discriminant analysis effect size (LEfSe) analysis, with a linear discriminant analysis (LDA) score threshold of > 2 and a p-value < 0.05.

Integrative analysis

Spearman correlation was used to assess links between significant fecal genera, metabolites, and biochemical indices.

Predictive modeling

Significant features identified from univariate and multivariate analyses of each omics dataset (differential serum proteins, serum/urine metabolites, and fecal genera) were selected for integration. The MetaboAnalystR package implements a random forest classifier to build an ATB-DILI early warning model using these significantly different features from fecal genera, serum/urine metabolites, and candidate proteins. Model performance was assessed via the area under the receiver operating characteristic curve (ROC AUC), which was evaluated via 100-fold cross-validation.

Results

Patient characteristics

Seventy-two eligible TB patients were enrolled: 35 developed ATB-DILI (ATB-DILI group), and 37 did not (non-ATB-DILI group). The baseline characteristics were comparable except for significantly lower lymphocyte counts and higher total bilirubin in the ATB-DILI group (p < 0.05; Table 1).

Table 1.

Baseline characteristics of 72 tuberculosis patients

Characteristics	ATB-DILI (n = 35)	Non-ATB-DILI(n = 37)	p-value
Gender, n (%)			0.3292
Female	10(28.57)	15(40.54)
Male	25(71.43)	22(59.46)
Ethnicity, n (%)			0.7111
Han	32(91.43)	32(86.49)
Nonhan	3(8.57)	5(13.51)
Age (years)	46.20 ± 13.06	46.43 ± 16.58	0.9533
BMI (kg/m2)	21.30 ± 3.40	20.81 ± 2.81	0.5898
BMI general division, n (%)			> 0.9999
< 18.5	5(14.29)	5(13.51)
≥ 18.5	30(85.71)	32(86.49)
Smoking, n (%)	5(14.29)	5(13.51)	> 0.9999
Alcohol consumption, n (%)	2(5.71)	0(0.00)	0.2328
Therapeutic category, n (%)			0.4861
Initial treatment	34(97.14)	37(100.00)
Retreatment	1(2.86)	0(0.00)
Location of tuberculosis, n (%)			> 0.9999
Intrapulmonary	34(97.14)	35(94.59)
Extrapulmonary#	1(2.86)	2(5.41)
Arterial hypertension, n (%)	1(2.86)	0(0.00)	0.4861
Diabetes mellitus, n (%)	0(0.00)	0(0.00)	> 0.9999
White blood cell count [×10^9/L]	6.19(4.76–8.21)	6.59(5.20–7.75)	0.9557
Number of neutrophil [×10^9/L]	3.8(2.8–5.2)	4.18(3.15–5.64)	0.7471
Monocyte count [×10^9/L]	0.47(0.36–0.67)	0.46(0.39–0.63)	0.9127
Total lymphocyte count [×10^9/L]	1.04 ± 0.49	1.43 ± 0.50	0.0047
Platelet count [×10^9/L]	267.2 ± 127.69	282.24 ± 80.15	0.5248
Hemoglobin [g/L]	125.70 ± 19.87	128.30 ± 22.29	0.6259
Red blood cell count [×10^12/L]	4.39 ± 0.63	4.32 ± 0.93	0.4844
ALT [U/L]	21.00(15.00-28.50)	17.00(13.00–24.00)	0.1111
AST [U/L]	22.00(19.50–25.00)	24.00(17.50–29.50)	0.5836
ALP[U/L]	88.00 ± 13.87	86.00 ± 20.48	0.8321
Total bilirubin [µmol/L]	12.25(8.90–17.80)	9.90(7.20-13.85)	0.0486
Albumin [g/L]	38.55 ± 6.32	40.87 ± 4.66	0.2237
Uric acid [µmol/L]	331.11 ± 104.28	333.38 ± 83.21	0.1424
Creatinine [µmol/L]	65.80 ± 16.75	64.61 ± 13.04	0.5483
Urea [mmol/L]	4.49 ± 1.06	4.52 ± 1.10	0.9470

ATB-DILI anti-tuberculosis drug-induced liver injury, n number, BMI body mass index, # The “extrapulmonary TB” category includes only cases of pure extrapulmonary tuberculosis. Cases of pulmonary tuberculosis combined with extrapulmonary tuberculosis are classified under the “Intrapulmonary TB” group, ALT alanine aminotransferase, AST aspartate transaminase, ALP alkaline phosphatase, significant results (p < 0.05) are highlighted in bold

ATB-DILI characteristics

Liver injury manifested at a median of 29 days (IQR: 14–30). The injury types included hepatocellular (54.29%), mixed (37.14%), and cholestatic (8.57%) injuries. The median R value was 4.07 (IQR: 2.44–5.06). At diagnosis, the ALT level was 168.00 U/L (IQR: 98.00–233.00), the AST level was 144.80 U/L (IQR: 84.00–227.00), the TBIL level was 20.17 ± 8.70 µmol/L, and the international normalized ratio (INR) was 1.29 ± 0.25 (Table 2).

Table 2.

Clinical characteristics of 35 ATB-DILI patients at the onset of ATB-DILI

Variables	ATB-DILI at Onset (n = 35)	Corresponding Baseline Comparison	p-value (Baseline: ATB-DILI vs. Non-ATB-DILI)
Days from initial TB treatment to DILI recognition (days)	29(14–30)	N/A	N/A
Pattern of liver injury, n (%)		N/A	N/A
Hepatocellular	19(54.29)	N/A	N/A
Mixed	13(37.14)	N/A	N/A
Cholestatic	3(8.57)	N/A	N/A
R value	4.07(2.44–5.06)	N/A	N/A
Lab test at ATB-DILI onset
ALT (U/L)	168.00(98.00-233.00)	21.00 (15.00–28.50) vs. 17.00 (13.00–24.00)	0.1111
AST (U/L)	144.80(84.00-227.00)	22.00 (19.50–25.00) vs. 24.00 (17.50–29.50)	0.5836
ALP (U/L)	123.83 ± 46.73	88.00 ± 13.87 vs. 86.00 ± 20.48	0.8321
GGT (U/L)	61.00(47.25–129.30)	Not measured	N/A
Total bilirubin [µmol/L]	20.17 ± 8.70	12.25 (8.90–17.80) vs. 9.90 (7.20–13.85)	0.0486
Albumin [g/L]	40.71 ± 6.86	38.55 ± 6.32 vs. 40.87 ± 4.66	0.2237
INR	1.29 ± 0.25	Not measured	N/A
White blood cell count [×10^9/L]	5.58(4.18–7.76)	6.19 (4.76–8.21) vs. 6.59 (5.20–7.75)	0.9557
Hemoglobin [g/L]	135.52 ± 20.11	125.70 ± 19.87 vs. 128.30 ± 22.29	0.6259
Platelet count [×10^9/L]	247.79 ± 121.33	267.2 ± 127.69 vs. 282.24 ± 80.15	0.5248
Uric acid [µmol/L]	582.00(223.70–710.00)	331.11 ± 104.28 vs. 333.38 ± 83.21	0.1424
Creatinine [µmol/L]	69.31 ± 12.71	65.80 ± 16.75 vs. 64.61 ± 13.04	0.5483
Urea [mmol/L]	4.21 ± 1.02	4.49 ± 1.06 vs. 4.52 ± 1.10	0.9470

ATB-DILI anti-tuberculosis drug-induced liver injury, R value = [serum ALT/ALT ULN]/[serum ALP/ALP ULN], ALT alanine aminotransferase, AST aspartate transaminase, ALP alkaline phosphatase, GGT gamma glutamyl transferase, INR international normalized ratio

Serum proteins

The concentrations of RBP-4, CHGA, CPB2, ANT3, and APOD were significantly greater in the ATB-DILI group than in the non-ATB-DILI group (p < 0.0001; Fig. 1), a finding that is consistent with our preliminary proteomic screening. Despite these significant differences, the predictive power of the combined five-protein panel, as assessed by the area under the curve (AUC), was 0.671. This suggests that while these proteins are associated with ATB-DILI risk, their baseline levels alone may exhibit substantial interindividual variation or are insufficient to capture the full complexity of the injury process, highlighting the need for a multifactorial prediction approach.

Fig. 1.

Comparison of the serum RBP-4, CHGA, CPB2, ANT3 and APOD concentrations between the ATB-DILI group and the non-ATB-DILI group. DILI, anti-tuberculosis drug-induced liver injury; NoDILI: non-anti-tuberculosis drug-induced liver injury

Serum metabolomics

A total of 1195 metabolites were identified. PCA and OPLS-DA revealed distinct metabolic profiles between the groups (Fig. 2A, B, with the QC samples clustered tightly, indicating good analytical stability). The OPLS-DA model (R2Y = 0.692, Q2 = 0.351) demonstrated good predictive ability without overfitting (permutation test Q2 intercept < 0; Fig. 2C). A total of 163 metabolites differed significantly (137 upregulated and 26 downregulated in ATB-DILI; Fig. 2D, Supp Table 7).

Fig. 2.

Serum metabolic changes associated with ATB-DILI. A. PCA score plot comparing patients with ATB-DILI (blue) and non-ATB-DILI (green). B. Score plot of the OPLS-DA showing patients with ATB-DILI (blue) and non-ATB-DILI patients (green). C. Permutation test OPLS-DA. The upper dot represents the R2Y value of the permutation test, and the lower triangle represents the Q2 value of the permutation test. The two dashed lines represent the regression lines of R2Y and Q2. D. Volcano plot showing the important metabolites included in the OPLS-DA model

Urine metabolomics

A total of 1952 metabolites were identified. PCA revealed overlap, but OPLS-DA revealed separate groups (Fig. 3A, B, with QC samples clustered tightly). The robust model (R2Y = 0.972, Q2 = 0.364) was validated (permutation test; Fig. 3C). A total of 106 metabolites differed significantly (42 upregulated and 64 downregulated in ATB-DILI; Fig. 3D, Supp Table 8).

Fig. 3.

Urinary metabolic changes associated with ATB-DILI. A. PCA score plot comparing patients with ATB-DILI (blue) and patients without ATB-DILI (green). B. Score plot of the OPLS-DA showing patients with ATB-DILI (blue) and non-ATB-DILI patients (green). C. Permutation test OPLS-DA. The upper dot represents the R2Y value of the permutation test, and the lower triangle represents the Q2 value of the permutation test. The two dashed lines represent the regression lines of R2Y and Q2. D. Volcano plot showing the important metabolites included in the OPLS-DA model

Pathway analysis

To elucidate the potential biological mechanisms underlying the observed metabolic perturbations, pathway enrichment analysis was performed on the differentially abundant metabolites identified in serum and urine via MetaboAnalyst on the basis of the KEGG database. In the serum of patients prone to ATB-DILI, the most significantly perturbed pathways were associated primarily with membrane transport, signal transduction, and core metabolic processes (Fig. 4). The most prominent pathway was ABC transporters (n = 7, p = 1.89E-05), which play a critical role in the hepatic excretion of drugs and metabolites. This finding suggests a preexisting baseline alteration in the cellular efflux machinery, potentially leading to increased intracellular accumulation of hepatotoxic drugs or their metabolites. Significant enrichment was also detected in key signaling pathways, including the cAMP signaling pathway (n = 5, p = 3.95E-04) and the sphingolipid signaling pathway (n = 3, p = 5.26E-04), which are central to regulating cell survival, inflammation, and stress responses. Concurrently, widespread disturbances were detected in several amino acid metabolism pathways, most notably alanine, aspartate and glutamate metabolism (n = 7, p = 7.00E-04). This pathway is crucial for ammonia detoxification (urea cycle) and gluconeogenesis and provides precursors for glutathione synthesis, indicating potential vulnerabilities in hepatic detoxification and antioxidant defense capacities even before drug exposure. Furthermore, the enrichment of the bile secretion pathway (n = 3, p = 5.03E-03) points to a potential predisposition to cholestatic injury. The urinary metabolic profile revealed enrichment in pathways related to translation, specific disease states, and systemic regulation (Fig. 5). The most significant pathway was aminoacyl-tRNA biosynthesis (n = 3, p = 3.26E-03), which is fundamental to protein synthesis. Its perturbation may reflect a systemic shift or stress in cellular biosynthetic activity. Like in serum, the sphingolipid signaling pathway (n = 3, p = 5.31E-03) was also enriched, reinforcing the notion of altered lipid-mediated signaling. Several pathways related to substance dependence (e.g., morphine addiction, alcoholism) and neurodegenerative diseases (e.g., Parkinson’s disease) were identified, likely reflecting shared metabolic nodes with neurological and stress‒response systems rather than direct etiology. Notably, pathways involved in digestive and systemic homeostasis were also affected, such as protein digestion and absorption (n = 4, p = 4.79E-03) and mineral absorption (n = 4, p = 1.48E-02), potentially indicating alterations in nutrient processing or gut‒kidney‒axis communication.

Fig. 4.

KEGG enriched pathways of serum metabolites

Fig. 5.

KEGG enriched pathways of urine metabolites

Overall, the pathway analysis revealed multifaceted pretreatment metabolic vulnerability in individuals susceptible to ATB-DILI. The concurrent enrichment of ABC transporters (impaired efflux), glutamate/amino acid metabolism (compromised detoxification and redox balance), and bile acid metabolism highlights potential intrinsic weaknesses in the liver’s core defensive and excretory functions. Simultaneously, disturbances in key signaling pathways (cAMP, sphingolipid) suggest a dysregulated cellular stress response network. These findings imply that future ATB-DILI patients should be treated with a metabolic landscape already primed for hepatotoxicity, where drug exposure may overwhelm preexisting limitations in detoxification, efflux, and cellular resilience.

Fecal microbiomics

Sequencing revealed that 4,067,802 high-quality sequences clustered into 5486 OTUs (ATB-DILI: 1889 unique; Non-ATB-DILI: 2662 unique; Fig. 6B). The sequencing depth was sufficient (Shannon rarefaction curves plateaued; Supp Fig. 1A). The alpha diversity indices were not significantly different (Fig. 6A). Beta diversity analysis (PCoA, NMDS; Fig. 6C, Supp Fig. 1B) and the Adonis test (P = 0.004) confirmed distinct microbial community structures (NMDS stress = 0.195).

Fig. 6.

Comparison of fecal flora richness and diversity between patients with ATB-DILI and those without ATB-DILI. A. α diversity was evaluated via the Chao1, Shannon and Pielou’s evenness indices and the PD and Good’s coverage indices, and the results revealed that there was no significant difference between the ATB-DILI group and the non-ATB-DILI group. B. Venn diagram showing the shared and unique operational taxonomic units (OTUs) in the two groups. C. PCoA based on weighted UniFrac distances revealed that the distribution of microbial communities in the ATB-DILI group was significantly different from that in the non-ATB-DILI group (P < 0.01). PCoA: principal coordinate analysis; DILI, anti-tuberculosis drug-induced liver injury; NoDILI: non-anti-tuberculosis drug-induced liver injury

Differential microbiota

LEfSe identified 14 differentially abundant taxa (Fig. 7A). Catenibacterium and Lactococcus were enriched in the ATB-DILI group. The genera Gilliamella, Moryella, Leuconostoc, RF39, Dorea, and Eubacterium siraeum were enriched in the Non-ATB-DILI group. The circumferential cladogram (Fig. 7B) visually summarizes the phylogenetic distribution of these differentially abundant taxa across different taxonomic levels.

Fig. 7.

Taxonomic characteristics of the fecal microbiota of patients in the ATB-DILI group and patients in the non-ATB-DILI group. A LEfSe analysis identified taxa with different levels between the ATB-DILI and non-ATB-DILI groups. Only taxa with LDA thresholds greater than 2.0 are shown. B Circumferential cladograms derived from LEfSe analyses showing the taxa that differed most between ATB-DILI patients and non-ATB-DILI patients. LDA: linear discriminant analysis. DILI, anti-tuberculosis drug-induced liver injury; NoDILI: non-anti-tuberculosis drug-induced liver injury

Correlations

Enterococcus abundance was positively correlated with ALT, whereas Faecalibacterium, Adlercreutzia, and Parabacteroides abundances were negatively correlated with ALT (Fig. 8). Significant correlations existed between fecal genera and serum/urine metabolites (Figs. 9 and 10). Veillonella was positively correlated with 19 serum metabolites, most strongly with rosmarinic acid (r = 0.55, p < 0.001). Ruminococcus and Blautia were strongly positively correlated with urinary metabolites.

Fig. 8.

Correlation analysis between fecal microbiological and biochemical parameters. Red indicates a positive correlation, blue indicates a negative correlation in the heatmaps, and darker colors indicate stronger correlations; *0. 01 < P < 0.05; **0. 001 < P ≤ 0.01; ***P ≤ 0.001

Fig. 9.

Correlation analysis between fecal microbiological and serum metabolites. Red indicates a positive correlation, blue indicates a negative correlation in the heatmaps, and darker colors indicate stronger correlations; *0. 01 < P < 0.05; **0. 001 < P ≤ 0.01; ***P ≤ 0.001

Fig. 10.

Correlation analysis between fecal microbiological and urinary metabolites. Red indicates a positive correlation, blue indicates a negative correlation in the heatmaps, and darker colors indicate stronger correlations; *0. 01 < P < 0.05; **0. 001 < P ≤ 0.01; ***P ≤ 0.001

ATB-DILI early warning model

The random forest model using pretreatment serum metabolites demonstrated the highest individual predictive power for ATB-DILI (AUC = 0.859; Fig. 11; Table 3). Models based on urine metabolites (AUC = 0.803), the fecal microbiome (AUC = 0.691), and serum proteins (AUC = 0.671) performed poorly. Integration of all pretreatment omics data (serum/urine metabolites, microbiome, proteins) yielded the best predictive model (AUC = 0.880; Table 3).

Fig. 11.

Predictive Model Performance and Feature Importance Across Multi-omics Data. A Performance evaluation of the integrated model. Panel A shows the receiver operating characteristic (ROC) curve of the predictive model constructed by integrating all omics features (candidate proteins, fecal microbiota, serum metabolites, and urine metabolites). The area under the curve (AUC) is 0.91, indicating excellent discriminatory performance of the model. The grey shaded area represents the 95% confidence interval. The diagonal dashed line (AUC = 0.5) corresponds to the random chance level. B Contribution of key biomarkers. Panel B displays the most important features driving the aforementioned predictive model. The bar chart ranks the top biomarkers in descending order of importance, with their names annotated, including: 3-(3-Methoxy-4,5-methylenedioxyphenyl)-2-propen-1-ol, Retinol, Anticopalic Acid, N-Methoxycarbonylanonaine, N-Acetylgalactosamine (GalNAc), Gallic acid, 3-Phenyl-3-(pyridin-2-yl)propanoic acid, among others. The importance scores were calculated by the machine learning model (Random Forest), with higher values indicating a greater contribution of the feature to the predictive model. DILI, ATB-DILI group; NoDILI, non-ATB-DILI group

Table 3.

Performance of random forest models in discriminating ATB-DILI patients from those without ATB-DILI on the basis of 100-fold cross-validation

Subjects	100-fold Cross-Validation
	AUC (95%CI)	Average accuracy
Serum metabolites	0.859(0.679–0.963)	0.695769231
Urine metabolites	0.803(0.622–0.948)	0.780769231
Fecal microbiota	0.691(0.469–0.868)	0.624615385
ATB-DILI early diagnostic candidate protein	0.671(0.485–0.805)	0.616538462
All	0.88(0.754–0.911)	0.786153846

AUC area under the curve, 95% CI confidence interval

Discussion

This study represents the first integration of pretherapeutic fecal microbiome composition, serum and urine metabolomic profiles, and validated protein biomarkers to construct an early warning model for antituberculosis drug-induced liver injury (ATB-DILI) via a random forest classifier. Our findings suggest that multiomics profiling prior to treatment initiation has significant potential for identifying patients at high risk for ATB-DILI.

Our targeted analysis confirmed that the levels of five candidate serum proteins (RBP-4, CHGA, CPB2, ANT3 and APOD) were significantly elevated pretreatment in patients who later developed ATB-DILI. These proteins may contribute to or reflect processes that increase susceptibility to liver injury. For example, RBP-4 is involved in retinol metabolism, which can be disrupted in liver disease [19]. CHGA, a precursor of bioactive peptides, may reflect neuroendocrine stress responses associated with drug toxicity [20]. The observed elevations suggest underlying metabolic or inflammatory perturbations that are predisposed to hepatotoxicity.

Metabolomics offers a powerful approach for identifying diagnostic markers and elucidating pathogenic pathways in ATB-DILI [15, 21]. Importantly, pretreatment metabolic signatures may predict susceptibility to liver injury [22]. Consistent with this premise, our analysis revealed significant differences in the serum (n = 163) and urine (n = 106) metabolite profiles between patients who subsequently developed ATB-DILI and those who did not prior to the start of anti-TB therapy. Our pathway analysis further revealed potential disruptions in fatty acid metabolism, bile acid synthesis, and amino acid metabolism, which are processes central to liver function. These findings are partially consistent with a previous metabolomics study by Wang et al. [15], which also reported alterations in lipid and bile acid metabolism in ATB-DILI patients, although their analysis was conducted postinjury. Our study adds value by demonstrating that these metabolic disturbances are present beforehand, suggesting a predisposing metabolic state. These distinct metabolic alterations represent potential biological indicators for early ATB-DILI risk stratification.

Notably, the serum and urinary metabolomic profiles showed a high degree of consistency. Several key pathways—including sphingolipid signaling and amino acid metabolism—were similarly dysregulated in both biofluids, reflecting systemic metabolic disturbances detectable across distinct sample matrices. This concordance between serum and urine reinforces the robustness of the identified metabolic signatures as potential predictive biomarkers for ATB-DILI risk.

Gut dysbiosis is increasingly implicated in the pathogenesis and progression of diverse liver diseases [23–25], potentially serving as an indirect mechanism in drug-induced hepatotoxicity [26]. First-line anti-TB drugs are known to disrupt the gut environment, reduce beneficial bacteria, enrich potentially pathogenic taxa [25], compromise intestinal barrier integrity, and increase permeability [27, 28]. Our LEfSe analysis revealed significant preexisting compositional differences: the genera Catenibacterium (a producer of short-chain fatty acids but potentially proinflammatory in certain contexts [29]) and Lactococcus (generally considered beneficial, but some species may have context-dependent effects [30]) were enriched in the ATB-DILI group, whereas potential beneficial taxa such as Gilliamella, Moryella, Leuconostoc, and others were more abundant in the Non-ATB-DILI group. These findings suggest that inherent interindividual variations in the gut microbiota composition may predispose patients to ATB-DILI, potentially exacerbated by subsequent drug-induced dysbiosis and barrier dysfunction contributing to injury development. The strong correlations we observed between specific gut bacteria (e.g., Enterococcus positively correlated with ALT; Faecalibacterium negatively correlated with ALT) and host metabolites further support the existence of a gut‒liver axis interaction in ATB-DILI susceptibility.

The random forest algorithm has demonstrated utility in integrating multiomics data for disease prediction, such as stratifying chronic kidney disease severity via metagenomics and metabolomics [31]. With this approach, we developed the first ATB-DILI early warning model that combines the pretreatment fecal microbiome, serum/urine metabolome, and serum proteome data. This integrated model achieved high predictive accuracy (AUC = 0.880), demonstrating robust potential for identifying patients at risk of ATB-DILI prior to initiating first-line anti-TB treatment. Notably, the model outperformed models based on single-omics datasets, underscoring the complementary nature of information derived from different biological layers.

While the comprehensive multiomics approach presented here shows excellent predictive performance, its direct translation into routine clinical practice, especially in resource-limited TB-endemic settings, requires consideration of cost and logistical complexity. Future work should focus on validating a simplified prediction panel derived from the most influential features identified in this study, potentially focusing on a smaller set of key metabolites and proteins measurable by more accessible platforms such as targeted mass spectrometry or automated immunoassays. The economic and practical implications of implementing such a panel would need careful evaluation against the potential benefits of preventing ATB-DILI, including reduced hospitalization costs and improved treatment outcomes.

Several limitations warrant consideration: (1) Lack of longitudinal data: the absence of serial metabolomic and microbiome sampling during treatment precludes analysis of dynamic changes associated with ATB-DILI onset and progression. (2) Confounding factors: The gut microbiome composition is influenced by host genetics, health status, diet, and environmental exposures. While baseline characteristics were generally comparable, unmeasured confounders could influence the observed differentially abundant metabolites and microbial taxa. (3) Biomarker validation: The identified candidate biomarkers require further functional characterization and validation in larger, independent cohorts. (4) External validation: The model’s performance needs confirmation in external populations because of the current multicenter but geographically limited design and moderate sample size.

Conclusion

Our findings demonstrate that patients destined to develop ATB-DILI exhibit distinct pretherapeutic signatures across multiple biological layers, encompassing serum proteins (RBP-4, CHGA, CPB2, ANT3, APOD), serum and urine metabolites, and fecal microbiota composition. Integrating these features via a random forest classifier successfully established a high-accuracy early warning model for ATB-DILI risk prediction following first-line anti-TB therapy initiation. This multiomics approach provides a novel strategy for personalized risk assessment and proactive management of ATB-DILI.

Authors’ contributions

QXL and JQH conceptualized and designed the study; CZW and GS collected the data; JQH coordinated and supervised the data collection; QXL drafted the initial manuscript; and CZW and GS carried out the initial analyses. All the authors critically reviewed the data and revised the manuscript. All authors approved the final submitted manuscript and agreed to be accountable for all aspects of the work.

Funding

This work was supported by the Sichuan Science and Technology Program Key R&D Project (Grant No. 2023YFS0236).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of West China Hospital of Sichuan University, the Affiliated Hospital of Zunyi Medical University and Guangyuan People's Hospital [Approval No. KLL-2022-735]. The study was conducted with the written informed consent of each subject. The confidentiality of the data is maintained, and no identifiable information is collected. The study was carried out in accordance with the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.