Early microRNA responses in rodent liver mediated by furan exposure establish dose thresholds for later adverse outcomes

Abstract

To reduce reliance on long-term in vivo studies, short-term data linking early molecular-based measurements to later adverse health effects is needed. Although transcriptional-based benchmark dose (BMD_T) modeling has been used to estimate potencies and stratify chemicals based on potential to induce later-life effects, dose-responsive epigenetic alterations have not been routinely considered. Here, we evaluated the utility of microRNA (miRNA) profiling in mouse liver and blood, as well as in mouse primary hepatocytes in vitro, to indicate mechanisms of liver perturbation due to short-term exposure of the known rodent liver hepatotoxicant and carcinogen, furan. Benchmark dose modeling of miRNA measurements (BMD_miR) were compared to the referent transcriptional (BMD_T) and apical (BMD_A) estimates. These analyses indicate a robust dose response for 34 miRNAs to furan and involvement of p53-linked pathways in furan-mediated hepatotoxicity, supporting mRNA and apical measurements. Liver-sourced miRNAs were also altered in the blood and primary hepatocytes. Overall, these results indicate mechanistic involvement of miRNA in furan carcinogenicity and provide evidence of their potential utility as accessible biomarkers of exposure and disease.

1. Introduction

Short-term toxicogenomic studies can provide mechanistic data that predict later adverse outcomes in a more efficient and cost-effective manner than conventional toxicity testing methods, such as the rodent two-year cancer bioassay. An important quantitative tool to derive predictions of thresholds of chemical hazard is the benchmark dose (BMD) estimate, which is the dose at which there is a measurable change in biological response compared to normal, background levels (1). Multiple studies have demonstrated that transcriptional-based BMD (BMD_T) values using short-term in vivo chemical exposures are similarly or more sensitive than traditional longer-term apical endpoint BMDs and can inform the mode-of-action (MoA) by which the exposure imparts adverse outcomes (1-5). More recently, the potential of in vitro high-throughput transcriptomics (HTTr) to accelerate the pace of chemical risk assessment is being investigated (6, 7). However, the relevance of specific gene expression perturbations in vitro to future adverse effects is often unclear, as the changes may sometimes represent an adaptive rather than adverse response (8, 9).

Very few studies to date have examined the utility of epigenetic measurements in quantitative risk assessment. However, many epigenetic targets are regulatory in nature, which can offer several potential advantages. MicroRNAs (miRNAs) are small non-coding RNA molecules that regulate gene expression post-transcriptionally by binding to complementary sequences on messenger RNA (mRNA) resulting in mRNA degradation or suppression of gene translation (10, 11). MiRNAs are responsive to environmental chemical exposures and are linked to activation of nuclear receptors and other transcription factors representing key events in adverse outcomes of regulatory concern (12-14). Because a single miRNA has multiple gene targets, the dysregulation of miRNAs may represent a node of gene expression changes (15); therefore, miRNA profiling may yield information of functional consequence. In addition, miRNAs are released into accessible biofluids in response to cellular perturbation, where they are stable due to association with proteins, exosomes and apoptotic bodies and can be measured non-invasively and non-destructively (16). For these reasons, miRNAs are being actively investigated for their utility as early biomarkers of chemical exposure that are indicative of primary MoA leading to later adverse health outcomes. A previous study in our laboratory exposed mice for 7-days to di(2-ethylhexyl)phthalate (DEHP) (17) and identified dose-responsive mouse liver miRNAs associated with the PPARα signaling pathway, which is known to be the molecular initiating event for DEHP-induced mouse liver carcinogenesis (18). In contrast, transcriptional profiling was less MoA-specific but provided point of departure (POD) estimates closer to those of the apical endpoints. This example demonstrated that miRNA could be useful in indicating chemical-mediated MoA, but additional studies are warranted to establish the utility of miRNAs for quantitative, predictive toxicology for other exposures.

In this study, we hypothesize that miRNA profiles will similarly predict carcinogenic MoAs in rodents exposed to furan. Furan, a heterocyclic organic compound that is highly volatile, is a rodent hepatocarcinogen at high doses with a postulated MoA of chronic cytotoxicity followed by regenerative proliferation. Previous studies in B6C3F1 mice used global gene expression data to examine the molecular basis of furan’s action through cytotoxicity, activation of oxidative stress response, inflammation, and regenerative proliferation (19). In addition, BMD_T were consistent across genomic platforms and highly predictive of published cancer bioassay point of departure values (3, 19). Using archived liver and blood samples from these studies on B6C3F1 mice exposed to non-carcinogenic and carcinogenic doses of furan, the current study evaluates the utility of miRNAs as early and quantitative biomarkers of adverse outcomes.

2. Materials and methods

2.1. Animal study and sample collection

Archived samples generated from a previously conducted study were used for this evaluation and the animal methods are detailed elsewhere (19). Briefly, 6–7-week-old female B6C3F1 mice were housed five per cage in polycarbonate cages in a specific pathogen free (SPF) and Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC) accredited facility. Female mice were chosen for these studies to replicate previous studies with furan; cytotoxic responses to furan were similar for male and female B6C3F1 (20-22). Mice were exposed by oral gavage to non-carcinogenic (0, 1 or 2 mg/kg bodyweight per day) or carcinogenic (4 or 8 mg/kg bodyweight per day) doses of furan in corn oil for 21 days. Four hours after the final dosing mice were anesthetized by CO₂ inhalation prior to euthanasia by exsanguination and livers were removed. Blood and liver samples were flash frozen and stored at or below −70°C until used for molecular measurements, described below.

2.2. Primary hepatocyte culture conditions, exposure, and lysis

Fresh hepatocytes isolated from female B6C3F1 mice were obtained from Triangle Research Laboratories (TRL, Research Triangle Park, NC). Isolated hepatocytes were seeded into 25 mL Erlenmeyer flasks at 2 X 10⁶ cells/mL in 3 mL per flask, using Williams’ Medium E containing 5% fetal bovine serum, 2 mM L-glutamine, 10 ng/ml dexamethasone, 50 μg/ml gentamycin, 5 μg/mL insulin, 5 μg/ml transferrin, and 5 ng/ml selenium as the culture medium. Cells were exposed to furan at 0.5 and 2 μM, or to the vehicle control (DMSO), in duplicate flasks per exposure. The flasks were incubated at 37°C and 5% CO₂ with gentle shaking for 4 hours. Following removal of the treatment media, the hepatocytes were washed with phosphate-buffered saline and plated in the culture medium at 3-4 x 10⁵ cells/well in collagen-coated 12-well plates Cells were incubated for 24, 48, or 72 hours at 37 °C and 5% CO₂ in air (duplicate plates, each with duplicate wells per plate). Cells were then washed once with PBS, and 350 μL lysis solution (Exiqon miRCURY^™ RNA Isolation kit) was added to each well. The lysed cells (pooled 2 wells per treatment plate) were collected into a clean, labelled microfuge tube, then flash frozen and stored at approximately −80°C until RNA extraction.

2.3. RNA extraction

Total RNA was extracted from frozen liver tissue (n=4-5/treatment) using a ceramic bead-based homogenization combining an initial RNAzol RT^® (MRC, Cincinnati, OH) extraction with subsequent purification using the RNeasy MinElute kit (Qiagen, Germantown, MD). Clean-up used the modifications for RNA binding indicated in the MinElute kit protocol supplement to retain the small RNA fractions. Total RNA from whole blood (n= 4-5/treatment) was isolated using the mirVANA RNA isolation kit (ThermoFisher Scientific, Waltham, MA) and from in vitro samples (n=2/treatment) using the miRCURY RNA isolation kit (Exiqon Inc., Woburn, MA). RNA was quantified by Nanodrop spectrophotometer (NanoDrop Technologies, Wilmington, DE) and/or Qubit fluorometric assay (ThermoFisher Scientific). All liver and cell culture-derived RNA exhibited high purity with A260/280 ratios >2.0. RNA isolated from archived blood samples exhibited slightly poorer quality, ranging from 1.63-2.09 with all but three above 1.8.

2.4. Gene and miRNA expression analysis

Agilent gene expression microarrays (GEO # GSE48644) (19, 23, 24) and RNA-Seq (GEO # GSE64371) (3) were previously used to measure changes in liver gene and long non-coding RNA expression for these samples. Count data for RNA-sequencing (RNA-seq) was generated as previously described (3) and differentially expressed genes (DEGs) were subsequently determined by DESeq2 in the Partek Flow version 6.0 (Partek, St. Louis, MO) environment using Wald hypothesis testing, independent filtering, and parametric fit (25). We measured liver miRNA by next generation sequencing on the Illumina HiScanSQ platform (Illumina Inc., San Diego, CA) using the TruSeq Small RNA Library Prep Kit following manufacturer’s instructions. Quality and size range of prepped libraries were checked using Agilent Bioanalyzer before loading on flow cell. FASTQ files were generated using CASAVA v1.8.2 (Illumina). FASTQ and normalized data are available in Gene Expression Omnibus (GEO # GSE224206). The processed FASTQ files were analyzed in Partek Flow version 8.0.19.1027 (Partek, St. Louis, MO) with minimum read length set to 15 base pairs and bases trimmed of adapter sequence from the 3’ end (TGGAATTCTCGGGTGCCAAGG) and according to Quality Score from both 5’ and 3’ ends. Contaminant reads (rRNA, tRNA, and mtDNA) were first filtered out using Bowtie 2 v2.2.5 (26). Samples were aligned with Bowtie (27) to the mm10 reference index using up to a 1 base pair mismatch and a seed length of 10. Aligned sequences were then annotated to miRBase mature miRNA v21 with the Partek E/M (expectation maximum) algorithm and a minimum overlap with feature of 90% or greater. Features (miRNAs) were excluded if the geometric mean of unnormalized counts were less than or equal to 10. Differentially expressed miRNAs (DEmiRS) were then determined by DESeq2. Ingenuity Pathway Analysis (IPA; Qiagen) was used to identify molecular pathways affected by furan treatment and to predict miRNA-mRNA interactions.

2.5. nCounter miRGE assay

The top dose-responsive liver miRNA candidates were measured in blood and in vitro samples using nCounter miRGE assay (NanoString Technologies), with total RNA concentrations normalized to 33 ng/μl and submitted to NanoString for processing. The resultant data files were analyzed with nSolver Analysis Software, version 4.0. For each dataset counts were normalized to the top few most stable genes chosen in RefFinder (http://blooge.cn/RefFinder/), with stability rankings <3.

2.6. Dose-response modeling

BMDExpress version 2.3 was used to perform BMD analysis on RNA-seq, small RNA-seq, and NanoString datasets. Prior to modeling, the datasets were log2 transformed and pre-filtered for differential expression by ANOVA p ≤ 0.05 with multiple test correction and fold change ≥ 1.3 (miRNA) or ≥ 2.0 (mRNA). Pre-filtering was relaxed with no multiple test correction for blood and cell culture miRNA samples with the NanoString measurements. Best BMD/BMD lower-limit (BMDL) ratio was set at less than 20, best BMD upper-limit (BMDU)/BMDL ratio was set at less than 40 and best fit required a p-value of greater than 0.1. Hill, power, linear, poly 2, exponential 2, exponential 3, exponential 4, and exponential 5 models were used for modeling. For each gene the software chose the best fitting model using the nested Chi Square (p-value cut off 0.05) to select the best poly model followed by lowest AIC. Other parameter settings included maximum iterations = 250, confidence level = 0.95, constant variance = 1, BMR type = standard deviation, BMR Factor = 1.349, restrict power = 0, flag Hill model with ‘k’ parameter less than 1/3 the lowest possible dose, best model selection with flagged Hill model = select next best model with p-value > 0.05. Functional category analyses (Gene Ontology Analyses for miRNA and Signaling Pathway Analyses for mRNA) were filtered at a minimum of 3 expressed genes per category with a minimum percentage of the total at 5% and a maximum Fisher’s Exact Two-tail p-value of 0.001.

3. Results

3.1. Gene and miRNA expression analysis

Principal component analysis (PCA; Fig. 1A) of gene expression in furan treated mice (2, 4 and 8 mg/kg bw for 3 weeks) demonstrated a clear dose-dependent clustering of hepatic expression profiles based on RNA-seq. The 8 mg/kg bw dose had over 4.5 times the number of DEGs than the 2 mg/kg bw dose (Fig. 1B). The number of DEGs were similar for 4 mg/kg and 2 mg/kg doses; however, approximately 90% of the 4 mg/kg DEGs overlapped with the 8 mg/kg dose group, whereas only 44% of the 2 mg/kg DEGs were differentially expressed in the 8 mg/kg group (Fig. 1C). There was minimal overlap of DEGs noted between the 2 and 4 mg/kg dose groups. The 1 mg/kg bw dose level was not included in the archived RNA-sequencing data, but these samples were included for the miRNA assessment.

Figure 1.

(A) PCA analysis of log₂ normalized data shows increasing separation from control with dose. (B) Differentially expressed genes (DEGs), filtered at FDR ≤ 0.05, absolute fold change > 2, number of DEGs by treatment group, and (C) Venn diagram of DEG overlap between treatment groups.

PCA of the miRNA data (Fig. 2A) indicated a distinction between carcinogenic (4 and 8 mg/kg bw) and non-carcinogenic (1 and 2 mg/kg bw) doses, but individual dose levels were not as distinct as the hepatic gene expression profiles. The number of DEmiRs increased from 0 at the 1 mg/kg bw dose to 53 at the 8 mg/kg dose (Fig. 2B). Similar to mRNA expression data, the greatest overlap observed was between the 4 and 8 mg/kg bw (carcinogenic) dose groups. The 35 miRNAs in this overlap are listed in Table 1. Superscript ^a indicates a quantifiable dose-response that could be modelled to derive a BMD (BMDExpress). Using miRNA target filter analysis, we predicted target genes for these miRNAs (using DEmiRs with FDR ≤ 0.05 and absolute fold-change > 2.0 cut-offs in the 8 mg/kg bw dose group) that were also significantly altered in our RNA-seq analysis (Supplemental Table 1).

Figure 2.

(A) PCA analysis of log₂ normalized data shows two clusters: (0 and 1 mg/kg doses) and (4 and 8 mg/kg doses) with the 2 mg/kg dose overlapping both clusters. (B) Differentially expressed miRNAs (DEmiRs), filtered at FDR ≤ 0.05, absolute fold change > 1.3, number of DEmiRs by treatment group, and (C) Venn diagram of DEmiRs overlap between treatment groups.

Table 1.

Fold change of liver miRNAs significantly altered by furan exposure.

13 miRNAs significantly altered, 3 top doses	1mg/kg	2 mg/kg	4 mg/kg	8 mg/kg
mmu-miR-125a-5pa	−1.01	−1.75	−3.30	−2.74
mmu-miR-125b-5pa	1.03	−1.68	−2.39	−2.14
mmu-miR-30c-1-3p	−1.03	−1.37	−1.36	−1.58
mmu-miR-676-3pa	−1.01	−1.34	−1.47	−1.58
mmu-miR-532-5pa	−1.10	−1.35	−1.34	−1.49
mmu-miR-671-3p	−1.01	−1.41	−1.33	−1.36
mmu-miR-378d	−1.05	−1.34	−1.35	−1.35
mmu-miR-26b-5p	1.01	1.54	1.65	1.40
mmu-miR-17-5pa	1.13	1.43	1.67	1.70
mmu-miR-203-3pa	1.04	1.10	1.24	1.19
mmu-miR-324-5pa	1.19	1.66	1.87	2.00
mmu-miR-146b-5p	1.22	2.48	2.72	2.37
mmu-miR-34a-5pa	1.35	3.22	6.03	8.05
22 miRNAs significantly altered at top 2 doses
mmu-miR-501-3pa	1.19	−1.58	−6.38	−5.71
mmu-miR-99b-5pa	−1.02	−1.99	−7.35	−5.39
mmu-miR-100-5pa	1.00	−1.86	−5.85	−4.86
mmu-miR-3071-5p	−1.26	−2.14	−3.91	−3.30
mmu-miR-221-5pa	1.05	−1.62	−4.55	−3.25
mmu-miR-96-5pa	1.02	−1.32	−2.99	−2.84
mmu-miR-145a-3pa	1.19	−1.26	−2.30	−2.40
mmu-miR-99a-5pa	1.03	−1.50	−2.85	−2.33
mmu-miR-183-5pa	1.16	−1.21	−3.67	−2.09
mmu-miR-362-5p	1.15	−1.36	−2.58	−2.07
mmu-miR-10a-5p	1.26	−1.12	−2.04	−1.8
mmu-miR-1948-5p	−1.29	−1.59	−1.86	−1.67
mmu-miR-192-5p	1.08	−1.20	−1.51	−1.45
mmu-miR-378c	1.00	−1.35	−1.54	−1.42
mmu-miR-106b-5pa	−1.14	1. 07	1.37	1.34
mmu-miR-16-5pa	1.12	1.18	1.43	1.34
mmu-miR-181b-5p	−1.01	1.18	1.33	1.36
mmu-miR-3073b-3p	1.27	1.23	1.52	1.40
mmu-miR-20a-5pa	1.02	1.23	1.31	1.42
mmu-miR-421-3pa	1.16	1.09	1.37	1.70
mmu-miR-3073a-3p	1.25	1.46	1.99	2.14
mmu-miR-744-5pa	1.14	1.24	1.50	1.48

^amiRNAs exhibiting a robust dose-response trend using BMDExpress version 2.3.

3.2. Pathway Analysis

IPA Core pathway analyses were performed on two datasets: 1) DEGs at the 8 mg/kg bw dose and 2) the IPA-predicted gene targets of DEmiRs at 4 and 8 mg/kg bw doses. Canonical pathways that were significantly enriched (p ≤ 0.01 with at least 3 representative genes) for mRNA DEGs and predicted miRNA gene targets are shown in Figs. 3A and 3B, respectively. The top 4 pathways for both datasets were G2/M DNA damage checkpoint regulation, ATM signaling, p53 signaling, and aryl hydrocarbon receptor signaling. Pathways with predicted activation in the mRNA dataset were the mitotic roles of polo-like kinase, NRF2-mediated oxidative stress response, and nicotine and melatonin degradation pathways (Fig. 3A, asterisks). Increased hepatocyte proliferation and decreased hepatocyte apoptosis were indicated for both datasets (Figs. 4A & 4B) which is consistent with observations of increased liver cell proliferation following furan treatment based on mice treated with BrdU in parallel (19) and in other studies (21, 28). Predicted upstream regulators are shown (Fig. 5). These analyses further support the involvement of NRF2-mediated oxidative stress response and p53 signaling with carcinogenic doses of furan exposure in mouse liver and support previous findings regarding MoA.

Figure 3.

Significantly altered (p≤0.01, n ≥ 3) canonical pathways using IPA of DEGs at the 8 mg/kg bw furan treatment (A), and for the 8 + 4 mg/kg bw miRNA target analysis (B). Asterisks indicate IPA predicted pathway activation (calculated z-score of 2 or greater).

Figure 4.

Significantly altered (p≤0.05) “Diseases and BioFunctions” gene lists in IPA with z-scores for the 8 mg/kg furan DEGs (A) and carcinogenic 4 + 8 mg/kg (B) miRNA target analysis datasets.

Figure 5.

(A) Predicted upstream regulators in IPA at the 8 mg/kg bw furan dose. MYC and FOXM1 were also predicted upstream regulators in the 4 + 8 mg/kg miRNA target analysis. (B). Graphical summary of the 8 mg/kg bw furan data showing predicted activation that highlights involvement of the p53 and Nrf2 pathways.

3.3. Benchmark Dose Modeling

Thirty-four miRNA and 145 mRNA passed the BMD filters (ANOVA p < 0.05 with MTC, best BMD/BMDL < 20, best BMDU/BMDL < 40, best fit P-value > 0.1) indicating they exhibited a robust furan dose-response. Mean and median BMD and BMDL values are reported in Table 2. The Hill model was the best fit for 61.8% of the miRNA. Many dose-responsive miRNA overlapped with the list of miRNAs significantly altered at both 4 and 8 mg/kg furan (Table 1, indicated by ^a). The majority have been associated with liver cancer or specifically hepatocellular carcinoma (HCC) in the literature (Supplementary Fig. 1). The linear model was the predominant best fit for the mRNA (33.1%). Seventy of these dose-responsive genes were identified by IPA as targets of significantly altered miRNAs, and many were relevant to the furan MoA (Supplementary Table 1). The BMD values for the dose-response modeled miRNA were significantly lower than those for the dose-response modeled mRNA (Table 2). BMD_miR values were more conservative than BMD_a calculated for the adverse outcomes of interest two years after exposure (HCC BMD(L): 5.1 (4.2); hepatocellular adenoma (HCA) BMD(L): 2.6 (1.6); HCC or HCA BMD(L): 2.3 (1.3)) (19,21).

Table 2.

BMD analyses

Transcriptional BMD(BMDL)s	miRNA	mRNA**	miRNA target mRNAs only*
Best BMD(L) Mean	1.99 (1.36)	2.85 (1.93)	2.77 (1.89)
Best BMD(L) Median	1.90 (1.15)	2.70 (1.92)	2.65 (1.86)
Best BMD(L) 95% CI	1.71-2.27 (1.13-1.60)	2.56-3.15 (1.73-2.13)	2.36-3.19 (1.61-2.17)
Lowest median gene set:	1.53 (1.00)	0.43 (0.30)	0.47 (0.30)
GO Pathways	RISC complex, RNAi effector complex	Intrinsic apoptotic signaling pathway in response to DNA damage by p53 class mediator	Intrinsic apoptotic signaling pathway in response to DNA damage by p53 class mediator

^*Significantly different (p≤0.05 Student’s two-tailed t-test) from miRNA dataset

^**Significantly different (p≤0.01 Student’s two-tailed t-test) from miRNA dataset

We also assessed BMD changes from a gene pathway level to determine which pathways exhibited the most conservative PODs. Functional categorization of the dose-responsive genes using Reactome (https://reactome.org/) and BioPlanet (https://tripod.nih.gov/bioplanet/) databases (Fig. 6) supported findings observed using IPA, with p53 signaling, cell cycle checkpoints and ATM pathways heavily represented. The pathway with the lowest median BMD was The role of GTSE1 in G2/M progression after G2 checkpoint (0.48 mg/kg bw), and the lowest median GO category (Table 2) was Intrinsic apoptotic signaling pathway in response to DNA damage by p53 class mediator (0.43 mg/kg bw). Reactome and BioPlanet pathways were not available for miRNA analysis, but the lowest median GO category of the dose-responsive miRNA (Supplementary Fig. 2, Table 2) was RNAi effector complex and RISC with a POD estimate of 1.53 mg/kg bw. Functional categorization of IPA-identified miRNA target genes (Fig. 7) again illustrated the probable involvement of miRNA regulation in these pathways, especially the p53, cell cycle and DNA damage response pathways. In particular, the significance of miRNA regulation in p53 signaling was demonstrated by an increased -log (p-value) when examining both IPA canonical pathways and Reactome/BioPlanet pathways for the miRNA target genes.

Figure 6.

Using significantly altered dose-responsive transcripts (FDR≤0.05), enriched Reactome and BioPlanet pathways were examined. Bubble size equals the percentage of genes in the gene set that pass the BMD filters (ranging from 5-20%).

Figure 7.

Using significantly altered dose-responsive miRNA-target transcripts, Reactome and BioPlanet pathways. Bubble size equals the percentage of genes in the gene set that pass the BMD filters (ranging from 5-17%).

3.4. Comparison to Microarray Data

Genes relevant to the furan mode of action that are significantly different from controls in response to 8 mg/kg bw furan in both the current study and the original analysis by Jackson et al. (19) are listed in Supplemental Table 2 and show good concordance in both directionality and significance.

3.5. Comparison of furan miRNA signature to human hepatocellular cancer (HCC) miRNA signature

Many of the mouse liver miRNA altered by a 3-wk exposure to carcinogenic doses of furan are also dysregulated in human hepatocellular cancer (Supplementary Fig. 1). For a few of these miRs, the direction of response to furan exposure was opposite from that reported in HCC tissue (miRs-221-5p and -183-5p). However, more often alterations were concordant: miRs-100-5p, -125a-5p, -125b-5p, -99a-5p, -199a-3p, -150-5p, -200b-3p and -101a-3p, which were down-regulated in liver by furan in our study, were also reported to be down-regulated in HCC; miRs-93-5p, -301a-3p, 18a-5p, 181b-5p and 106b-5p were upregulated by both furan treatment and in HCC.

3.6. Blood and primary cell culture miRNA measurements

Selected miRNA biomarker candidates were measured in RNA isolated from archived blood of furan-treated mice and furan-treated primary hepatocyte cell lysates using NanoString with BMD modeling. Several miRNAs that were significantly altered by furan in the liver demonstrated a significant dose response in the blood in the same (miR-122-5p, increased in liver only at 2mg/kg dose) or opposite (miRs-10-5p, 221-5p, 501-3p, 96-5p and 99a-5p decreased in liver, Table 1) direction (Fig. 8). None of the miRNAs measured in cell lysates were statistically altered by dose in our small in vitro pilot study (n=2/treatment); however, three did fit a dose response model at both 48 and 72h (Supplementary Fig. 3), decreasing with dose 48h post-treatment but increasing with dose at 72h post-treatment.

Figure 8.

Dose-responsive miRNAs in the blood measured by NanoString. Data points represent data mean +/− SD, fit to a dose-responsive model (see Methods), and extrapolated BMD(L) values based on 10% BMR.

4. Discussion

MiRNAs are critical for the regulation of gene expression and translation. The intracellular expression of these molecules can change rapidly with exogenous perturbation and these alterations can impact entire gene networks, partially driven by the ability of a given miRNA to target multiple genes. MiRNAs are also released and stable in biofluids and alterations in these matrices can reflect cellular exposure, adversity, and disease. These hallmarks indicate these molecular regulators could serve as powerful toxicologic biomarkers, yet use of miRNAs in this context is limited. In this study, we measured miRNA alterations in the livers of B6C3F1 mice exposed short-term to the known rodent liver carcinogen, furan. We found that the liver miRNA alterations, like what was observed with gene expression, were robust and overlapping at the higher carcinogenic furan doses, compared to the lower non-carcinogenic doses. In contrast to the gene expression changes, calculated BMD values were more conservative on average than for gene expression measured by multiple methods. Significant dose-responsive miRNA changes were predicted to target genes involved in the DNA damage response, in particular activation of the p53-ATM axis. In addition, some of these alterations were measured in the blood. Overall, our measured miRNA alterations primarily linked to response to DNA damage and cell death processes, and BMD estimates of these alterations occur at lower doses than calculated PODs for cancer and, on average, gene expression.

Gene expression analysis derived from both microarray and RNA-seq data for these archived samples had been previously analyzed (3, 19). Here, we reanalyzed the RNA-seq data to update the pathway analysis and integrate with the miRNA measurements performed. Our new gene expression analysis supported key cellular events in the progression of liver cancer due to carcinogenic furan exposures in mice described previously, especially at the 8 mg/kg bw dose. This progression included furan metabolism to cis-2-butene-1,4-dial (BDA) mediated by CYP2E1 -> oxidative stress and depletion of glutathione reserves -> cytotoxicity and cell death signaling -> chronic NRF2 activation and inflammation -> regenerative proliferation -> cancer (19). CYP2E1 metabolism of furan results in reactive oxygen species production and glutathione depletion (29), which was evidenced by upregulation of glutathione-s-transferases (Gsta1, Gsta2, Gstp2, etc.) and activation of the Nrf2-mediated oxidative stress response (Nqo1, Hmox1, Srxn1, etc.) in our current and previous studies. We observed activation of the DNA damage response/p53 pathway (increased expression of downstream gene targets Cdkn1a, Ccng1), increased death receptor/TNF signaling (increased expression of Tnfrsf10b/DR5 intracellular death domain), apoptosis activator Bax, and increased expression of the JNK-specific phosphatase/inactivator, Dusp8, which inhibits cell proliferation. We also observed some evidence for inflammatory cascades, such as upregulation of the pro-inflammatory chemokine C-X-C motif ligand, Cxcl10. Finally, we observed enrichment of disease pathways that indicated cell growth and proliferation, including increased expression of growth factors Pdgf and Gdf9/15 and predicted activation of HCC-associated c-Myc and Foxm1.

Analysis of the miRNA alterations are somewhat limited compared to gene expression, primarily due to the databases available to map and predict their involvement in cellular pathways and function. To overcome this, we supplemented the limited direct miRNA analysis with predicted gene target estimation of altered miRNAs (based on experimentally validated data from TarBase and miRecords and computationally predicted miRNA-mRNA interactions from TargetScan in the IPA environment). Using this method, profiling of the miRNA target genes indicated more specific involvement in the p53-ATM cell cycle and DNA damage response pathways than what was more globally indicated by the full gene expression analysis. Likewise, results showed BMD_miR estimates based on dose-responsive miRNA to be more sensitive than BMD_T estimates based on dose-responsive mRNA; however, the lowest median BMD(L) estimation for both gene expression and miRNA-target sets using GO were similar and linked to the p53 pathway (0.43 (0.30), 0.47 (0.30) mg/kg bw/day, respectively). The combination of the dataset analyses overwhelmingly highlighted processes of p53-mediated DNA damage response, and tumor-suppressing cell arrest and death mechanisms, in these mice after 3 weeks of furan exposure. This finding was also supported by the previous observation that short-term furan exposure in a Sprague-Dawley rat model augments a miRNA strongly linked to the p53 pathway, miR-34a (30).

Activation of the p53 pathway is a general indication of a cell’s protective response to DNA damage and suggests that if overwhelmed, genotoxicity may be a potential event that contributes to liver cancer in these mice. However, the genotoxic potential of furan has long been a point of contention in the literature. In the Ames assay, furan did not induce mutation with or without S9 metabolism; however, furan was positive in the L5178Y/TK^+/− mouse lymphoma assay without S9 fraction (31). In addition, both positive and negative findings have been reported for chromosomal aberrations in Chinese hamster ovary (CHO), hamster V79 cells, and B6C3F1 mouse bone marrow cells (32-34). Micronucleus assays in mice and rats have been largely negative (35-38). CYP2E1 metabolism of furan results in a highly reactive metabolite, BDA, which may be mediating many of the potential genotoxic and cytotoxic effects. Indeed, DNA adducts, single-strand breaks, and direct mutations in the S. typhimurium TA104 strain and the L5178Y/TK^+/− were observed with BDA exposure (31, 39-42). A recent study suggested that furan, through the action of BDA, is a clastogen based on in vitro and 4-week oral exposures in C57BL6 mice and increased micronuclei in splenocytes, and that this action was independent of an oxidative stress mechanism (28). It has also been previously suggested that oxidative stress generated in part by CYP2E1 metabolism mediates putative genotoxic actions of furan, but that this MoA may be secondary to the cytotoxicity and proliferative generation of the liver, ultimately leading to liver cancer (19, 36, 43). Our analysis supports that miRNA are involved in response to DNA damage, but it is not clear if this ultimately led to genotoxicity, or whether this damage is mediated through BDA, oxidative stress, or other mechanisms. Additionally, this stress is indicated at lower than carcinogenic doses of furan, suggesting these miRNA biomarkers of p53 are indeed conservative for furan in this case study.

Several approaches for POD selection have been utilized in the toxicogenomic literature that are predictive of the dose range for adverse apical outcomes, including the mean, median or first mode of transcript BMDs, lowest pathway BMDs, or BMDs based on MoA-relevant pathway(s) only. The Nrf2-mediated oxidative stress response pathway was previously chosen as the MoA-relevant transcriptional BMD and the calculated BMD(L) best predicted the dose at which furan induced later liver cancer (3). However, this targeted approach is appropriate only for chemicals for which there is at least some understanding of the chemical’s biological effects. In situations in which the adverse effect is unknown, a non-targeted approach to determine a transcriptomic POD protective of any downstream apical effect is required. A common method maps all genes with a BMD to defined gene sets. The transcriptomic POD is then chosen from the gene set with the lowest median BMD(L) value. There is growing consensus that BMDT determined in this manner can estimate BMD for apical measures (BMDA) in multiple study designs, including using a subchronic BMDT to estimate a carcinogenic apical endpoint POD (44). In the current study, the multi-omics analysis of both miRNA and mRNA pinpointed the p53-linked pathway BMD(L)s instead of the NRF2 pathway (Figs. 6 & 7). We have observed this narrowing of biological focus using the miRNA data in a previous study of short-term tumorigenic phthalate exposures in a mouse model. In that study, a 7-day DEHP exposure resulted in dose-responsive mouse liver miRNAs associated with the PPARα signaling pathway (17), which is known to be the molecular initiating event for DEHP-induced mouse liver carcinogenesis (18). Interestingly, the BMD_miR estimates were higher than both the BMD_T estimates for the most responsive PPARα target genes and the BMD estimates derived from the 2-year tumor data. Like our current study, the dose-responsive miRNA indicated the primary molecular signal which reflected the most conservative BMD_T values. This multi-omics approach, therefore, built greater evidence for the primary MoA for DEHP-mediated liver cancer in the mouse model, suggesting the utility of this method if the MoA was unknown.

Few studies have examined epigenetic measurements for deriving BMD estimates. Miousse et al. compared a variety of epigenetic and metabolic-based measurements to other apical measurements in response to a 7-day exposure to the hepatotoxicants clofibrate and phenobarbital in male F344 rats (45). Decreased expression was observed using a panel of liver toxicant-associated miRNA, but the small response was not conducive to BMD modelling. Other epigenetic-based measurements for which BMD estimates were calculated were not as conservative as gene expression targets for nuclear receptors linked to known MoA for these hepatotoxicants. In contrast, Rager et al. (46) modeled multiple molecular endpoints from human cord blood following prenatal exposure to inorganic arsenic, including mRNA expression, protein expression, miRNA expression and DNA (CpG) methylation, and found 12 miRNAs significantly associated with exposure (46). The BMD values were 33 μg As/L urine (maternal exposure) for CpG methylation, 42 μg/L for miRNA, 45 μg/L for protein and 64 μg/L for mRNA. Interestingly, it was the epigenetic markers (including miRNAs) that established the most conservative BMD values; however, all endpoints supported the epidemiological evidence that prenatal effects of inorganic arsenicals occur below 100 μg/L urine. Stermer et al. (47) BMD-modeled percent total reads of miRNA and other small RNA molecular measurements of toxicity in rat sperm following exposure to the testicular toxicant ethylene glycol monomethyl ether (EMGE). They found a BMD of 62 mg/kg using RSH (retained spermatid head), whereas the percent miRNA BMD was 59.2 mg/kg, suggesting that the miRNA PODs are sensitive and predictive biomarkers of exposure.

An advantage of miRNAs is their accessibility and stability in biofluids. For this reason, we also measured miRNA expression in the blood of furan-treated mice. We found a significant and dose-dependent increase in miR-122, which is the most abundant miRNA in hepatocytes and has been extensively studied as a liver toxicity biomarker (48). The calculated benchmark dose in the blood for this miRNA and several others was above the carcinogenic 4 mg/kg dose. Identification of biomarkers indicative of furan concentration below the level of carcinogenicity would be desirable, and miR-501-3p is a potential candidate, with a BMD below 2 mg/kg. There are several reasons that our whole blood miRNA measurements may not have been optimal. First, highly abundant miRNAs present in erythrocytes can complicate or mask evaluation of lower-abundance miRNAs in the blood. Therefore, examination of serum or plasma would have been preferable. Second, the blood samples had been archived for over 5 years prior to RNA isolation, which may have reduced the quality of the measurements; and third, a limited number of potential biomarker candidates were examined. Future evaluations could take steps to minimize technical variables. Similarly, the in vitro pilot indicated treatment related effects for three of the same miRNAs altered in the liver and blood, all of which are relevant to HCC and liver cancer (Supplementary Fig. 1) (49). Expansion of the study to include adequate replicates would likely have led to identification of additional biomarkers of effect. Development and validation of in vitro, high-throughput new approach methodologies (NAMs) will be necessary moving forward to accelerate the pace of chemical screening and hazard identification.

5. Conclusion

Our results support the potential utility of miRNA profiling in determining chemical MoA and predictive POD for a later adverse health outcome. More studies are required, examining chemical classes with differing MoA, to establish reliability and reproducibility of miRNA as a biomarker to different toxicological endpoints of regulatory concern. Our findings point to mechanistic involvement of miRNAs in furan carcinogenicity and identify several candidates with potential utility as accessible, dose-responsive biomarkers of chemical-mediated disease outcome.