Hepatotoxicity Prediction and Multi-omics Reveal Mitochondrial and Lipid Metabolic Dysregulation in PM_2.5-Induced Liver Fibrosis

Abstract

Prolonged exposure to fine particulate matter (PM_2.5) has been linked to chronic liver injury and cancer. However, an alternative risk assessment method to prospective longitudinal studies of exposome-metabolome interactions for liver inflammation-associated hepatocellular carcinoma (HCC) is lacking. This study investigates the risk of long-term real-world PM_2.5 exposure in hepatocarcinogenesis through machine learning techniques. Shotgun mass spectrometry (MS) imaging data were acquired from mouse models across a continuum of fibrosis, cirrhosis, and HCC for training a multiclass classification model to identify “No Risk”, “Cancer Risk”, and “Cancer”. Direct infusion-MS data from PM_2.5-exposed mouse livers were analyzed to classify risk. By integrating data-driven and knowledge-based approaches, 14 disease progression biomarkers were identified for modeling. Our results suggest that chronic real-world PM_2.5 exposure can induce liver fibrosis, presenting cancer risk. Incorporating metabolomics, lipidomics, and transcriptomics, we propose PM_2.5 exposure induces mitochondrial dysfunction, activates AMPK signaling, and increases ceramide accumulation, potentially mediating insulin resistance that contributes to nonalcoholic fatty liver disease and HCC progression. This work represents a significant advancement in assessing hepatotoxicity of environmental toxicants by reducing reliance on traditional animal testing methods. It also underscores the potential of emerging technologies in transforming our understanding of PM_2.5 exposure, paving the way for targeted interventions.

Introduction

Liver cancer represents a significant global health burden, causing approximately 830,000 deaths worldwide in 2020. Hepatocellular carcinoma (HCC) is the predominant form, accounting for 75%–85% of all liver cancer cases. Key risk factors for the development of HCC, a process known as hepatocarcinogenesis, include chronic liver injury and resulting fibrosis (FIR). Chronic liver injury triggers a sustained wound-healing response, leading to progressive liver scarring. In addition to established risk factors such as viral hepatitis, excessive alcohol consumption, and metabolic disorders (including diabetes, and nonalcoholic fatty liver disease or NAFLD), emerging evidence suggests that chronic exposure to air pollution contributes to liver FIR and inflammation-associated HCC progression.

Air pollution has been a leading risk factor for death globally over the past 30 years and has been linked to NAFLD-like metabolic liver diseases in various in vivo , and epidemiological studies. , Specifically, chronic exposure to PM_2.5, particles with an aerodynamic diameter of <2.5 μm, has been associated with liver cirrhosis (CIR). An increased odds ratio of NAFLD has been reported. After inhalation, PM_2.5 particles can penetrate the lungs and enter systematic circulation, reaching the liver for detoxification. − It may trigger liver FIR through inflammation-related mechanisms in rodent models, , which can promote liver injury and initiate carcinogenesis. Besides FIR, ambient PM_2.5 is classified as a human Class 1 carcinogen. Its higher levels have been associated with a significant increased incidence rate for HCC in 3 human cohorts. − Therefore, studying the relationship between chronic atmospheric PM_2.5 exposure and the progression from liver FIR to HCC is essential for understanding environmental risk factors and improving prevention strategies.

Hepatotoxicity of environmental pollutants was conventionally assessed in rodents or by association with liver injury markers in human studies. These methods require prospective longitudinal studies along with disease progression. Current studies are limited to hepatic FIR and CIR from cohort studies or in vivo investigations. Assessing HCC risk is challenging because the toxicity of PM_2.5 is relatively lower than that of other carcinogens such as diethylnitrosamine (DEN). Consequently, conducting a large prospective longitudinal study from FIR to HCC is time-consuming and costly. Recently, the United States Food and Drug Administration announced plans to phase out animal testing requirements, advocating for artificial intelligence (AI)-based computational models instead. In this study, we developed a multiclass classification machine learning (ML) model based on our desorption electrospray ionization-mass spectrometry imaging (DESI-MSI) data from DEN-tetrachloromethane (CCl₄)-induced FIR, CIR, and HCC in mice. It can classify hepatotoxicity of environmental pollutants into “No Risk”, “Cancer Risk”, and “Cancer” groups. Another mouse model was developed for chronic atmospheric PM_2.5 exposure, and the corresponding lipophilic liver extracts were analyzed via direct infusion MS (DI-MS) for shotgun MS data acquisition and HCC risk prediction. The prediction was validated by a histological analysis.

In addition to HCC risk prediction, we conducted untargeted metabolomics, targeted metabolomics, and lipidomics and analyzed 2 external transcriptome data sets from fossil fuel diesel organic extract-treated HepG2 cells to investigate nicotinamide adenine dinucleotide (NAD⁺)-related metabolic pathways and disrupted lipid homeostasis. Our results indicated that chronic PM_2.5 exposure induced mitochondrial dysfunction, activated AMPK signaling, and increased ceramide accumulation in hepatocytes, potentially mediating insulin resistance (IR) that can lead to NAFLD and ultimately HCC, aligning with the “two-hit hypothesis”. The insights gained may contribute to improved early detection and intervention from PM_2.5-derived chronic liver disease to HCC upon the effort of biological validation.

Materials and Methods

Chemical and Reagents

HPLC-grade dimethylformamide (DMF) was purchased from RCI Labscan (Bangkok, Thailand). Chloroform, formic acid (99% to 100%), LC/MS-grade acetonitrile (ACN), methanol (MeOH), and isopropanol (IPA) were obtained from VWR Chemicals BDH (Poole, UK). 2-Oxo-glutaric acid, 5-hydroxy-L-tryptophan, (S)-malic acid, anthranilic acid, citric acid, fumaric acid, indoleacetic acid, isocitric acid, l-asparagine, l-aspartic acid, N-formyl-kynurenine, l-glutamic acid, l-glutamine, L-kynurenine, NAD⁺, NADP⁺, NADPH, nicotinamide (NAM), serine, serotonin, succinic acid, tryptophan, and quinolinic acid were obtained from Sigma-Aldrich (St. Louis, MO, USA). 5-Hydroxy-indoleacetate was purchased from Yuanye (Guangzhou, China). 3-Hydroxyanthranilic acid and indole-3-pyruvic acid were collected from TargetMol (Boston, MA, USA). NADH was obtained from Aladdin (Shanghai, China). Water was purified using a Milli-Q water system (Millipore, Billerica, MA, USA).

Liver Cryosections of the Rodent Model for Fibrosis, Cirrhosis, and Inflammation-Associated Hepatocellular Carcinoma

Mouse liver cryosections were prepared for DESI-MSI. To study a continuum of FIR, CIR, and HCC, DEN-CCl₄-induced mouse model was developed. The technical details of animal model development and the discussions on liver morphology and histological analysis are listed in Supporting Information. All animal experiments and study protocols were approved by the University of Hong Kong’s Committee on the Use of Live Animals in Teaching and Research (CULATR), in accordance with the Animals (Control of Experiments) Ordinance of Hong Kong. Mice were sacrificed if body weight loss exceeded 20%, and the maximal tumor burden allowed was not exceeded. C57BL/6J mice from the Chinese University of Hong Kong were acclimatized for 1 week at 22 ± 2 °C and 60% ± 5% relative humidity under a 12-h light/dark cycle.

Liver samples of 12 FIR mice, 12 age-matched control (AMC) mice for FIR (FIR-AMC), 12 CIR mice, 12 CIR-AMC mice, 12 HCC mice, and 12 HCC-AMC mice were harvested and stored at – 80 °C. The left liver tissues were sliced into 10 μm thick cryosections using a cryostat (CryoStar NX70, Thermo Scientific, USA) and stored at – 80 °C until MSI. Technical details of DESI-MSI are in Supporting Information.

Liver Extracts of the Rodent Model for Chronic Real-World PM_2.5 Exposure

Mouse liver extracts were prepared for MS-based lipidomics and metabolomics and DI-MS. To develop the mouse model for chronic real-world PM_2.5 exposure (Figure S1A), 20 4-week-old C57BL/6 female mice were purchased from Beijing Vital River Laboratory Animal Technology Co., Ltd. (Beijing, China). After acclimating them for 7 days, they were placed in an individually ventilated cage (IVC) system (Junsheng Laboratory Animal Equipment Co., Ltd., Jiangsu, China) that was maintained at 25 °C and 40%–70% room humidity on a 12 h light-dark cycle. The IVC system has 2 chambers, 1 fitted with a high-efficiency filter to deliver ultrafiltered air yet another allowing airborne particles to pass through to deliver PM_2.5. Ten mice were randomly selected as control and housed in IVC with filtered air, and the other 10 mice were exposed to real-ambient PM_2.5 inhalation. All mice were fed a normal diet, and the experimental exposure lasted 6 months from first of November in 2020 to 30th of April in 2021 in Taiyuan, China. Water was freely available during the experiment. The mice were sacrificed on first of May 2021, and liver samples were dissected, individually wrapped in labeled aluminum foils and snap-frozen, stored in steel bottle with dry ice, and shipped to Hong Kong immediately. All samples received in Hong Kong were frozen at – 80 °C until further analysis.

For lipidomics and DI-MS, liver samples (25 mg) were collected in individual 1.5 mL centrifuge tubes with 600 μL of ice-chilled ACN/DMF (1:1, v/v) and 3 stainless steel beads. After homogenization and centrifugation at 12000g for 15 min at 25 °C, each 400 μL of supernatant was dried with an IR concentrator (NB-504CIR; N-Biotek MAX-UP, Korea), reconstituted in 200 μL of ACN/IPA/H₂O (65:30:5, v/v/v) for lipidomics, and diluted 10 times for DI-MS. Technical details of lipidomics for 9 case and 8 control samples and of DI-MS for 8 case-control pairs are in Supporting Information.

For untargeted and targeted metabolomics, 1 mL of iced ice-cold ACN/water (8:2, v/v) and 3 stainless steel beads were added for homogenization and extraction. The homogenate was centrifugated at 17,000g for 10 min at 4 °C and moved into the new 1.5 mL centrifuge tube. The aqueous extracts were dried, reconstituted with 200 μL of MeOH/water (1:1, v/v), and mixed well thoroughly by vortex for 1 min. The pooled quality control (QC) sample was prepared by pooling 80 μL of each sample solution together. Technical details of untargeted metabolomics for 9 case and 8 control samples and of targeted metabolomics for 10 case and 8 control samples are in Supporting Information.

Feature Selection and Chemical Identification

Detailed discussions on feature selection procedures are listed in Supporting Information. In brief, data-driven method was applied to FIB/FIB-AMC and HCC/HCC-AMC groups. After denoising, eXtreme Gradient Boosting (XGB) model was implemented to rank feature importances and to mine the most sensitive features. Pearson correlation was then used to eliminate redundant features from the mined features. Features selected from the CIR/CIR-AMC group were based on prior research that identified a set of dysregulated lipids associated with disease progression along the same trajectory. As some features overlapped across groups, 14 features were selected eventually for hepatotoxicity modeling (Table ).

1. Summary of the 14 Identified Molecular Features for Liver Cancer Staging Modeling and PM_2.5 Hepatic Carcinogenicity Prediction.

Feature Selection Approach	m/z (-ve mode)	Compound Subclass	Annotation	MSn	Grade/Level	Adduct	Formula	Delta ppm	Retention Time Error (s)	Source
Data-Driven Features for Fibrosis	250.1448	Diacyl-glycero-phospho-serines	PS 34:5	1	3	M-3H	C40 H68 N1 O10 P1	–2		LMGP03010922
										LMGP03010131
										LMGP03010899
										LMGP03010432
										LMGP03010623
										LMGP03010650
	303.2329	Unsaturated fatty acids	FA(20:4)	2	C/2B	M-H	C20 H31 O2	–1 × 10–1		LipidSearch
	295.2278	Unsaturated fatty acids	FA(18:1)+OX:(s)	2	C/2B	M-H	C18 H32 O3	3 × 10–1		LipidSearch
	480.3099	monoacyl	18:0 LYSO-PE	2	1	M-H	C23 H47 N1 O7 P1	4 × 10–1	7.44	HMDB0011130
			LPC(16:0)	2	C/2B	M-CH3	C23 H47 N1 O7 P1	1 × 10–1		LipidSearch
	215.0328	Fatty acids and conjugates	Undecane-dioic acid	2	1	M-H	C11 H20 O4	2 × 10–2	1.65	HMDB0000888
Knowledge-Based Features for the Continuum of Fibrosis, Cirrhosis, and Cancer	255.2330	Straight chain fatty acids	FA(16:0)	2	C/2B	M-H	C16 H31 O2	1 × 10–1		LipidSearch
	327.2330	Unsaturated fatty acids	FA(22:6)	2	C/2B	M-H	C22 H31 O2	–2 × 10–1		LipidSearch
	619.2894	monoacyl	LPI(20:4)	2	A/2B	M-H	C29 H48 O12 P1	–4 × 10–1		LipidSearch
	865.5025	diacyl	PG(22:6_22:6)	2	A/2B	M-H	C50 H74 O10 P1	–8 × 10–1		LipidSearch
Data-Driven Features for Cancer	311.1688	Linoleic acids and derivatives	13(S)-HpODE	2	1	M-H	C18 H31 O4	2 × 10–1	0.41	HMDB0003871
			9(S)-HpODE	2	1	M-H	C18 H31 O4	2 × 10–1	0.41	HMDB0062434
	269.2486	Straight chain fatty acids	FA(17:0)	2	C/2B	M-H	C17 H33 O2	–5 × 10–1		LipidSearch
	738.5062	diacyl	PE(16:0_20:4)	2	A/2B	M-H	C41 H74 N1 O8 P1	5 × 10–02		LipidSearch
			PE(18:2_18:2)	2	A/2B	M-H	C41 H74 N1 O8 P1	5 × 10–2		LipidSearch
	435.2964	Monoradyl-glycerols	2,3-Dihydroxy-propyl octanoate	1	3	2M-H	C11 H22 O4	1		HMDB0254850
	514.2848	Bile acids, alcohols and derivatives	Tauro-alpha-muricholic acid	2	1	M-H	C26 H45 N1 O7 S1	2 × 10–1	5.32	HMDB0258742
			Tauro-b-muricholic acid	2	1	M-H	C26 H45 N1 O7 S1	2 × 10–1	11.92	HMDB0000932
			Tauro-cholic acid	2	1	M-H	C26 H45 N1 O7 S1	6 × 10–1	15.75	HMDB0000036

^aThe selected feature is not only mined for fibrosis but also for hepatocellular carcinoma by data-driven approach.

^bThe selected feature is mined by both data-driven approach and knowledge-based approach.

The annotation for the 14 features follows the Metabolomics Standards Initiative. For level 1 identification confidence, the software TidyMass was employed to match the m/z values of molecular ion and fragments, as well as retention times with referenced values recorded in an in-house database. Level 2B annotations of lipids were achieved using LipidSearch (Thermo Fisher Scientific, San Jose, CA, USA). In LipidSearch, a lipid defined as grade A indicates that all fatty acid (FA) chains belonging to a given lipid were completely identified, while grade C indicates the detection of lipid class-specific ions or FA-derived product ions. For level 3 confidence, a bulk search of MS1 was performed using the Human Metabolome Database (HMDB) and LIPID MAPS. Candidates with the least absolute mass error were selected and reported.

Digital Liver Disease Modeling for Risk Assessment of Chronic PM_2.5 Exposure

A 3-class (“No Risk”, “Cancer Risk”, and “Cancer”) classification model was trained and validated for the risk assessment of chronic PM_2.5 exposure. To establish this model, 14 significant biomarkers were extracted from different liver disease stages. Technical details of feature selection are in Supporting Information. In brief, an MSI training data set was prepared by integrating 3 independent batches (FIB/FIB-AMC, CIR/CIR-AMC, and HCC/HCC-AMC) of cryosection MSI data. For the class “No Risk” with the label “0”, 261,298 pixels were included from 36 cryosections of AMCs (train: 198,536 pixels, 27 cryosections; test: 62,762 pixels, 9 cryosections); in the class of “Cancer Risk” with the label “1”, 228,046 pixels (112,010 pixels were from FIB and 116,036 pixels were from CIR livers) were collected from 24 cryosections (train: 171,018 pixels, 18 cryosections; test: 57,028 pixels, 6 cryosections); and for the “Cancer” class with the label “2”, only the cryosection MSI data of HCC were included (104,747 pixels, train: 76,819 pixels, 9 cryosections; test: 27,928 pixels, 3 cryosections).

The predictive performances of decision tree (DT), gradient boosting (GB), and random fortress (RF) modeling were compared primary by weighted F1 score and secondary by Matthews correlation coefficient (MCC) for the training, validation, and the ingested data sets (see Data sets S1–S3 at 10.5281/zenodo.15831391 for the training records).

Transcriptomics by RNA Sequencing

Two external RNA sequencing (RNA-Seq) data sets were used to validate our metabolomics data in this study. In brief, PM_2.5 derived from low-sulfur fossil diesel and ethanol/diesel fuel were collected with stainless steel fibers and extracted by organic solvent. The liver epithelial cell line HepG2 cells were exposed to 2 membrane-filtered organic extracts separately for 24 h, together with the control (extract of clean stainless-steel fibers). The paired-end libraries (2 × 150 bp) were sequenced in an Illumina HiSeq 4000 plataform for the extracted total RNA. In our study, the obtained RNA-Seq data were preprocessed and reanalyzed on RStudio 2023.12.1 + 402 (R version 4.3.3). The scripts can be found at https://github.com/TommyNHL/liverCancerStagingProj.

Statistical Analysis and Data Visualization

Python v.3.11 on Jupyter Notebook (Anaconda3), R version 4.3.3 on RStudio 2023.12.1 + 402, and Microsoft Excel 2024 were used for statistical analyses and data visualization. Results were expressed as mean ± standard deviation (SD). Student’s t test was used to assess statistically significant differences between the disease group and its AMC group, by which a value of P ≤ 0.10 was considered as statistically significant.

Results and Discussion

In this study, we first develop a predictive model for the risk assessment of long-term PM_2.5 exposure. After the hepatic carcinogenicity of PM_2.5 was predicted, we performed untargeted metabolomics, targeted metabolomics, and lipidomics to examine the major metabolic pathways and signaling networks that were found to be dysregulated along the liver injury. These pathways generally are NAD⁺-related based on our previous investigations. Finally, we selected 2 transcriptome data sets that are the most relevant to our study to further validate the metabolic alterations by the differentially expressed genes (DEGs) on HepG2 cells.

Fourteen Significant Markers for Orbitrap Mass Spectrometry-Based Hepatic Carcinogenicity Prediction

The identifications of these 14 discriminative features are annotated and reported in Table . The discriminative features that were selected for HCC/HCC-AMC classification by a data-driven approach included 13­(S)- or 9­(S)-HpODE, FAs (such as FA 17:0, FA 18:1;O, FA 17:4;O3, FA 11:1;O2), diacyl lipids [phosphatidylethanolamine (PE), and phosphatidylserine (PS)], and a bile acid. Among these features, some were also found to be significant for FIB/FIB-AMC classification by our data-driven approach, such as PS, FA 18:1;O, and FA 11:1;O2. Besides these, a polyunsaturated FA, FA 20:4 (arachidonic acid), was found to be elevated in metabolic dysfunction-associated steatohepatitis (MASH), a fibrotic liver disease that can cause CIR. Although the role of 18:0-lyso PE in NAFLD is rare reported in the literature, its structural lipid isomer lysophosphatidylocholine was proposed to be involved in the pathogenesis of NAFLD.

As de novo lipogenesis (DNL) was previously reported to be upregulated by the overexpression of the sterol regulatory element-binding protein (SREBP)-related genes Srebf1 and Srebf2 across a continuum of liver disease from FIB to CIR, and eventually to HCC, we further incorporated the lipids that were detectable in negative ESI mode with consistent increases or decreases as a part of the features. The straight-chain saturated FA palmitate that is the first lipid end-product in DNL was included. Besides palmitate, the downstream polyunsaturated FA, FA 22:6 (docosahexaenoic acid, DHA), was found to be negatively correlated with FA 20:4, was also selected. Interestingly, FA 20:4 is also one of the significant features mined by the knowledge-based approach. The monoacyl lipid lysophasphatidylinositol 20:4, which was also positively associated with NAFLD-associated fibrosis and IR, and the diacyl lipid phosphatidylglycerol 22:6_22:6, which was identified in steatosis, were also identified.

Predictive Power of Machine Learning Models

Predictions made using DT, GB, and RF. The optimal RF models demonstrated moderate weighted F1 score (>0.60) and MCC score (>0.50), while performed superiorly for the training, validation, and ingested data sets (weighted F1 scores are 0.87, 0.64, and 0.72, respectively, Table ). Other optimal tree-based models, including DT and GB models, were overfit to the training data as their weighted F1 scores approximate to 1 while <0.50 for the validation and ingested data sets. Therefore, we selected our RF model that was trained using 35 trees, a minimum of 25 leaves per tree with the depth of 3, and a minimum sample split of 2, to proceed for the PM_2.5 in silico toxicity analysis.

2. Summary of the Selected Machine Learning Algorithms’ Performance.

	data set
	training		validation		ingested
	wF1	MCC	wF1	MCC	wF1	MCC	weighted recall
DT	0.96	0.94	0.38	0.20	0.49	0.38	0.56
GB	0.98	0.97	0.51	0.35	0.66	0.57	0.69
RF	0.87	0.82	0.64	0.50	0.72	0.61	0.73

^aTraining data set was prepared from the desorption electrospray ionization mass spectrometry imaging (DESI-MSI) data of 9 cryosection tissue samples for each liver disease stage (fibrosis, cirrhosis, and hepatocellular carcinoma) and their corresponding age-matched controls (AMCs) (total: 54 subjects and 446,373 MS1 pixels of shotgun analysis).

^bValidation data set was used for model validation and was prepared from the DESI-MSI data of 3 cryosection tissue samples for each liver disease stage and their corresponding AMCs (total: 18 subjects and 147,718 MS1 pixels).

^cIngested data set was prepared from combining training and validation data sets (total: 72 animal subjects and 594,091 MS1 pixels).

^dWeighted F1.

^eDecision Tree.

^fGradient Boosting.

^gRandom Forest.

In Silico Analysis for the Hepatotoxicity Prediction of Long-Term Real World PM_2.5 Exposure

The optimized multiclass model, that is used to predict a MS1 scan belonging to “No Risk” (label “0”), “Cancer Risk” (label “1”), or “Cancer” (label “2”) was deployed for the DI-MS data from 8 case/control pairs. For each DI-MS acquisition of the PM_2.5-exposed liver extract, there are ∼200 MS1 scans. Our results indicated that the PM_2.5-exposed mice have the highest (76%) probability in “Cancer Risk” on average (Figure ). Histological study confirmed that the mice were suffering from hepatic fibrosis (Figure S1B−F), showing long-term exposure to atmospheric PM_2.5 may have the inflammation-associated HCC risk.

1.

Prediction for hepatic carcinogenicity of chronic atmospheric PM ₂ .₅ exposure. Predictions for 8 direct infusion mass spectrometry data of mouse livers exposed to filtered air or real world PM_2.5 for 6 months. The classes of “No Risk”, “Cancer Risk”, and “Cancer” are represented by labels “0”, “1”, and “2”, respectively.

Untargeted Metabolomics Identifies Chronic Real World PM_2.5 Exposure Dysregulated NAD⁺-Related Metabolism

To investigate the underlying mechanism of PM_2.5 exposure leading to the HCC risk, we initiated our mechanistic study by using untargeted LC/MS-based metabolomics. Metabolite set enrichment analyses on the untargeted metabolomics data revealed that steroidogenesis and nicotinate and NAM metabolism were both the significant metabolic pathways for the enrichment utilized pathway databases Kyoto Encyclopedia of Genes and Genomes (Figure S2A) and the Small Molecule Pathway Database (Figure S2B). From the untargeted metabolomics results (Figure S3A), the relative MS signal abundance of deoxycortisol (a precursor in cytoplasm for cortisol biosynthesis) was upregulated, while that of tetrahydrocortisone (urocortisone, a breakdown product of cortisol in endoplasmic reticulum) was downregulated upon PM_2.5 exposure. As the synthesis of steroid hormone cortisol from deoxycortisol takes place in mitochondria and required NADH consumption, we hypothesized that the alterations in steroidogenesis intermediates were due to mitochondrial dysfunction and insufficient NAD⁺ supply. Therefore, we performed targeted metabolomics for the NAD⁺- or energy-metabolism-related metabolic pathways, including tryptophan metabolism, nicotinate/NAM metabolism, fatty acid metabolism, alanine/aspartate/glutamate (AAG) metabolism, and citrate cycle.

Targeted Metabolomics and Lipidomics Reveal All NAD⁺-Related Pathways are Downregulated for the Chronic PM_2.5-Exposed Mice

To further explore the disturbances in NAD⁺ metabolism identified in the untargeted analyses, we performed targeted RPLC/tandem MS on rodent liver tissue samples. Targeted metabolomics revealed that almost all the metabolites in the NAD⁺-related pathways were significantly decreased in the PM_2.5-exposed livers (Figure A–F, see Data set S4 at 10.5281/zenodo.15831391). Citric cycle intermediates such as citrate, isocitrate, and succinate are significantly reduced (Figure F). These metabolic alternations were observed in both the HepG2 cells that were treated with the organic extract of PM_2.5 and an in vivo study with similar experimental setting for 6-month ambient PM_2.5 exposure. An isotopic tracing experiment was further confirmed by these findings on citric cycle.

2.

Metabolomic and lipidomic changes upon long-term atmospheric PM_2.5 exposure. (A) overview of the energy metabolism-related pathways. Upregulated genes and metabolites are red-colored. Downregulated genes and metabolites are blue-colored. Nonmeasured metabolites are gray-colored. Bold text indicated statistical significance with P-value <0.05. Stage-sensitive alterations in metabolites and lipids in (B) tryptophan metabolism, (C) nicotinate and nicotinamide metabolism, (D) fatty acid metabolism, (E) alanine, aspartate, glutamate metabolism, and (F) citrate cycle. Abbreviations: ∧, P < 0.1; *, P < 0.05; **, P < 0.01; ***, P < 0.001.

Our lipidomics analysis also indicated a significant decrease in hepatic acylcarnitine levels in the PM_2.5-exposed livers (Figure A and D, see Data set S5 at 10.5281/zenodo.15831391), aligning to the finding in a previous study with similar settings. Interestingly, the levels of hepatic ceramides were significantly elevated. The observation of upregulated ceramide levels was reported by a recent study on a fatty liver mouse model for 12-week low-dose traffic-related PM_2.5 exposure.

The metabolic profile of tryptophan pathway downregulation, incomplete FA oxidation (FAO), and the accumulation of ceramides were suggested to be attributed by the mitochondrial damage in the hepatocytes upon the exposure of the lipophilic components in PM_2.5. The organic extract of PM_2.5 also exhibited similar metabolic alternation of energy metabolism in HepG2 cell model and in the rodent study with same experimental setting for 6-month ambient PM_2.5 exposure. Therefore, we selected 2 external RNA-Seq data sets of the HepG2 cells that were incubated with the organic extracts of PM_2.5 derived from fossil fuel diesel and ethanol/diesel fuel. Moreover, the dose–response toxicodynamics by in vivo methods could be tailored to mimic in vitro methods based on Haber’s rule, which considers the mathematical product of exposure concentration and duration. In our study, the 181-day (6-month) PM_2.5 exposure yielded an average regional concentration of 64.14 μg/m³ (data obtained online). Given that a 70-kg human breathes approximately 20–30 m³ of air daily, a 25-g mouse would inhale about 0.00714–0.01286 m³ of air, translating to 18–33 ng/g and 3–6 μg/g of PM_2.5 for 1 day and 181 days, respectively. In comparison, HepG2 cells were exposed to 3.2 μg of PM_2.5 extract over a 24-h period, which is comparable to the overall dose from our 181-day in vivo study.

Chronic Atmospheric PM_2.5 Exposure May Cause High Cortisol Stress-Induced AMPK Activation, Leading to Ceramide Accumulation-Associated Mitochondrial Dysfunction

Transcriptomics analyses revealed that the HepG2 cells exhibited similar patterns upon treatment with the organic extracts of PM_2.5 derived from low-sulfur fossil diesel and ethanol/diesel fuel (Figure and Table S1). As the metabolic characteristics show from our targeted analysis, RNA-Seq data also indicated that the genes regulating the citric acid cycle were downregulated, as well as the gene encoding the rate-limiting enzyme of FAO, carnitine palmitoyltransferase 1A (Cpt1a). Besides citrate cycle and FAO, reduced energy supply was also observed from oxidative phosphorylation. Reduced levels of adenosine triphosphate and increased levels of adenosine monophosphate (AMP) were expected, leading to the upregulation of AMP-activated protein kinase (AMPK) signaling pathway.

3.

Differentially expressed genes in HepG2 cells treated with the organic extracts of PM_2.5 derived from diesel fuel and ethanol/diesel fuel. The genes exhibited overexpression in the same direction, with at least one group having a magnitude of log2 fold change exceeding 0.1. The selection was made regardless of statistical significance due to the limited sample size (only triplicate experiments). In vitro data were adapted from a study done by Orozco-Jiménez et al. and reanalyzed for this study.

Despite AMPK-induced sirtuin (AMPK-SIRT) activation, the NAD⁺ salvage pathway was suppressed (Figure and Table S1). The upregulated AMPK-SIRT repressed cholesterol biosynthesis pathway and induced ceramide synthase 6 (CerS6) upregulation, resulting in the elevation of lipotoxic ceramide and subsequent mitochondrial damage. Since mitochondria is essential for steroid and steroid hormones metabolism, together with the insights drawn to steroidogenesis in untargeted metabolomics (Figures S2 and S3A), we then determined the relevant DEGs. RNA-Seq data showed downregulation of the energy-consuming cortisol biosynthesis in mitochondria. Cortisol is a stress hormone of which level is proposed to be triggered by PM_2.5 exposure, we observed its breakdown was upregulated through Hsd11b2 overexpression and simultaneously downregulation of Hsd11b11 (Figure and Table S1), undoubtedly in endoplasmic reticulum. In addition, cortisol was evidenced to induce AMPK activation and CerS6 overexpression, boosting the conversion of the intermediates that are generated from the incomplete FAO and cannot transported to mitochondria into ceramides and potentially leading to IR.

Conclusion and Limitations

In this study, we applied the generic biomarkers discovered by the data-driven approach for FIB and HCC, with the significant lipids mined by the knowledge-based approach across the continuum of FIR, CIR, and HCC to develop a ML model for risk assessment of chronic real-world PM_2.5. exposure. We confirmed the prediction of the HCC risk upon long-term exposure to PM_2.5 by histological analysis. Multiomics analysis was performed to investigate the underlying mechanism. Collectively, we assume chronic atmospheric PM_2.5 exposure may increase the cortisol level that promotes IR as a first hit. The cortisol-activated AMPK-SIRT-CerS6 axis may then result in mitochondrial damage or dysfunction and ceramide lipid accumulation as the second hit, potentially mediating NAFLD progression to HCC, requiring additional investigation in our future work.

This study has several limitations. First, there is a distinction between the chronic exposure in the in vivo study and the acute exposure in the in vitro study, complicating direct comparisons and extrapolation of results across species. Additionally, the predictive model trained on multibatch DESI-MSI data sets may include false positives, as not all pixel data correspond to true positives for “Cancer Risk” due to the lack of specific selection for tumor regions, leading to potential contamination in the training data. Moreover, the genes in transcriptomics analysis were selected based on overexpression criteria without strict statistical significance due to the limited sample size of only triplicate experiments.

The developed digital liver model was provided for shotgun MS data is available at https://github.com/TommyNHL/liverCancerStagingProj (JOBLIB). The training records for ML model selection (Data set S1–S3), the processed targeted metabolomics data (Data set S4) and lipidomics data (Data set S5) are available at 10.5281/zenodo.15831391) (XLSX).

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/envhealth.5c00401.

• Technical details of mouse model development; DESI-MSI; MS-based lipidomics and metabolomics; DI-MS; MS data preprocessing; feature selections; RNA-Seq analysis; code availability for model development and application; histological analyses of PM_2.5-exposed mouse livers; pathway enrichment for the PM_2.5-exposed mouse livers; heatmap for untargeted metabolomics; histological analyses of DEN-CCl₄ mouse livers; regression plots of multibatch MS data; plots relevant to feature selection for FIB/FIB-AMC classification; summary of the DEGs in the RNA-Seq analysis; collision energy used in targeted metabolomics; and computational power used in this study (PDF)

Conceptualization, H.-L.N., J.Z., Y.C., Y.S., Z.Y., and Z.C.; resources, H.-L.N., J.Z., Z.Q., Z.Y., and Z.C.; data curation, H.-L.N.; writing-original draft preparation, H.-L.N.; writing-review and editing, H.-L.N., Y.C., Y.S., and H.Y.; visualization, H.-L.N. and H.Y.; supervision, Z.Y., H.Y., and Z.C.; project administration, Z.Y., H.Y., and Z.C.; funding acquisition, Z.C. All authors have read and agreed to the published version of the manuscript.

The authors declare no competing financial interest.

Published as part of Environment & Health special issue “Transforming Environmental Exposure and Health Research through Emerging Technologies”.