Unlocking the Power of Transcriptomic Biomarkers in Qualitative and Quantitative Genotoxicity Assessment of Chemicals

Abstract

To modernize genotoxicity assessment and reduce reliance on experimental animals, new approach methodologies (NAMs) that provide human-relevant dose–response data are needed. Two transcriptomic biomarkers, GENOMARK and TGx-DDI, have shown a high classification accuracy for genotoxicity. As these biomarkers were extracted from different training sets, we investigated whether combining the two biomarkers in a human-derived metabolically competent cell line (i.e., HepaRG) provides complementary information for the classification of genotoxic hazard identification and potency ranking. First, the applicability of GENOMARK to TempO-Seq, a high-throughput transcriptomic technology, was evaluated. HepaRG cells were exposed for 72 h to increasing concentrations of 10 chemicals (i.e., eight known in vivo genotoxicants and two in vivo nongenotoxicants). Gene expression data were generated using the TempO-Seq technology. We found a prediction performance of 100%, confirming the applicability of GENOMARK to TempO-Seq. Classification using TGx-DDI was then compared to GENOMARK. For the chemicals identified as genotoxic, benchmark concentration modeling was conducted to perform potency ranking. The high concordance observed for both hazard classification and potency ranking by GENOMARK and TGx-DDI highlights the value of integrating these NAMs in a weight of evidence evaluation of genotoxicity.

1. Introduction

Evaluating a chemical’s potential to damage genetic material is vital for safety assessments, as genotoxicity can lead to severe health outcomes like cancer and/or heritable genetic disorders. Traditional genotoxicity assessment begins with in vitro tests for gene mutations and chromosome aberrations. These in vitro assays are robust, relatively simple, and highly sensitive methods for identifying potential genotoxicants. However, the tests also have limitations, including a low specificity leading to unnecessary in vivo testing.^1,2 Several modifications to the assays have been made to address the issue of low specificity (e.g., decreasing the top concentration, using metabolically competent and human-relevant cells, taking into account the level of cytotoxicity), but some studies suggest that the number of misleading positives remains high.³ Additionally, to stimulate the shift from qualitative to quantitative genotoxicity assessment, the throughput of the tests needs to be increased to collect sufficient data to characterize concentration–response relationships.^4–8 Technologies such as flow cytometry and (semi)automated imaging systems that score, for example, the presence of micronuclei have been developed to address this need.

In addition to improving traditional tests, new approach methodologies (NAMs) are needed to modernize genotoxicity assessment by enabling integrated testing that more broadly captures biological effects and reduces reliance on animal assays. Such NAMs can inform mechanisms of action and efficiently generate concentration–response data.⁹ Ideally, NAMs and traditional tests will eventually be fully integrated to identify and characterize the genotoxic potential of agents without the need for further in vivo testing. Methods applying transcriptomics in human-derived cell systems are recognized as important NAMs because they provide the capacity for high-throughput, high-content and human-relevant assessments of DNA damage.^2,10,11 Transcriptomic profiling detects gene expression changes resulting from chemical exposure, indicating adverse or adaptive biological responses.¹² One approach to simplify whole transcriptome analysis is the development of transcriptional signatures, which are subsets of genes predicting specific toxicological effects like genotoxicity for hazard identification.¹³ When combined with machine learning algorithms, these gene signatures, referred to as “transcriptomic biomarkers”, enable rapid analysis of large transcriptomic data sets.

Two notable transcriptomic biomarkers for genotoxicity are TGx-DDI and GENOMARK. The TGx-DDI biomarker, initially developed in TK6 cells with a 4 h exposure time, distinguishes DNA damage-inducing (DDI) from non-DDI chemicals by analyzing the gene expression of 64 biomarker genes.¹⁴ TGx-DDI has also demonstrated high predictivity in human HepaRG cells.^15,16 A variety of studies have confirmed the robustness of TGx-DDI for genotoxicity assessment using different gene expression technologies^15–20 and demonstrated its utility in integrated testing for hazard identification and concentration–response modeling for potency evaluation.^15,21 Another biomarker, GENOMARK, consists of 84 genes that identify genotoxicants in human HepaRG cells.²² HepaRG cells are becoming more widely used for in vitro testing as this p53-competent, human-relevant cell culture model expresses relevant levels of phase 1 and phase 2 metabolic enzymes, transporters, and nuclear receptors that closely resemble those of primary human hepatocytes.^23,24 The GENOMARK biomarker shows high predictivity in classifying chemicals as (non)genotoxic based on gene expression data collected using microarrays or RT-qPCR after 72 h exposure.^22,25 Similar to TGx-DDI, GENOMARK accurately classifies “misleading positives” (i.e., chemicals positive in one of the traditional in vitro tests but negative in the associated in vivo follow-up test) as nongenotoxic, illustrating its utility for derisking of these chemicals.^{17 ,19 ,25} Although a pilot study using a publicly available data set demonstrated the applicability of GENOMARK on a RNA-sequencing (RNA-seq) data set,²⁵ additional studies are needed to confirm the potential application of GENOMARK across other high-throughput platforms.

While both TGx-DDI and GENOMARK show promise for hazard identification and chemical prioritization, their concordance and integrated use have not been studied. Since GENOMARK and TGx-DDI were extracted from training sets encompassing different reference chemicals, cell lines, and exposure times, we hypothesized that integration of these biomarkers would increase confidence in hazard calls and potency ranking. To test this hypothesis, we first investigated the applicability of GENOMARK across different platforms using Templated Oligo assay with Sequencing readout (TempO-Seq, a high-throughput technology) to generate gene expression data for either a targeted gene set (i.e., a customized version of the Human + S1500 panel²⁶) or a panel representing the entire transcriptome. TempO-Seq specifically targets unpurified RNA in cellular lysates by hybridizing sequencing and sample-specific adapters to the original probes for sequencing.²⁷ This allows a high-throughput assessment of differentially expressed transcripts and concentration–response transcriptomic profiling. We then compared the concordance of GENOMARK against TGx-DDI for hazard classification and potency ranking of chemicals. Our findings highlight the high value of integrating transcriptomic biomarkers in a weight of evidence genotoxicity assessment of chemicals, particularly when utilizing human-relevant cell lines such as HepaRG cells.

2. Materials and Methods

2.1. Chemicals

Table 1 summarizes the test chemicals, CAS numbers, previously published information on genotoxicity, and test concentrations. The 10 chemicals include (i) 5 chemicals previously tested by GENOMARK using microarray and/or qPCR to assess the cross-platform performance of GENOMARK using TempO-Seq and (ii) 5 additional test chemicals to compare the predictive performance and concordance of GENOMARK and TGx-DDI. Overall, eight known in vivo genotoxic chemicals were tested: glycidol (GLY), methylmethanesulfonate (MMS), nitrosodimethylamine (NDMA), 4-nitroquinoline-N-oxide (4NQO), aflatoxin B1 (AFB1), colchicine (COL), cyclophosphamide (CPA), and mitomycin C (MMC). Two known in vivo nongenotoxic chemicals were analyzed: 2-chloro-ethyltrimethylammonium chloride (2CTAC) and eugenol (EUG). The latter is a known “misleading positive” in the traditional in vitro genotoxicity test battery. Test chemicals were purchased from Merck/Sigma-Aldrich (Hoeilaart, Belgium). They were dissolved in dimethyl sulfoxide (DMSO) and further diluted in a medium with a final concentration of 0.5% (v/v) DMSO except for 2CTAC, which was directly dissolved in the medium.

Table 1. Overview of the Test Chemicals^a.

chemical and CAS number	in vitro GTX		in vivo GTX	requires metabolicbioactivation	mechanism of GTX	previously tested by GENOMARK	concentrations tested (μM)					data setb
	Ames	MN/CA					C1	C2	C3	C4	C5
MMC—(50-07-7)	+	+	+	no	mono and bifunctional alkylation	no	0.17	0.5	1.5	4.5	NA	1
4NQO—(56-57-5)	+	+	+	no	formation of DNA adduct and ROS	no	0.9	2.7	8	12*	NA	1
AFB1—(1162-65-8)	+	+	+	yes	formation of DNA adduct and ROS	yes22	0.025	0.1	0.5	1.5	NA	1
CPA—(6055-19-2)	+	+	+	yes	alkylation	yes22	78	233	700	2100	NA	1
NDMA—(62-75-9)	+	+	+	yes	alkylation and DNA adduct formation	yes22	80	240	720	2160	5454**	2
COL—(64-86-8)	–	+	+	no	aneugen	no	0.1	0.3	1	3	NA	1
GLY—(556-52-5)	+	+	+	no	alkylation	no	97	291	873	1200	1402*	2
MMS—(66-27-3)	+	+	+	no	mono functional alkylation	yes22	27	81	243.3	400	500	2
EUG—(97-53-0)	–	+	–	?	NA	yes	60	170	500	1000*	NA	1
2CTAC—(999-81-5)	–	–	–	no	NA	no	370	1100	3300	10000	NA	1

^aChemicals are shown with names and abbreviations, CAS number, previously published in vitro and in vivo genotoxicity data, mechanisms of action, whether previously tested by GENOMARK or not and the concentrations tested (μM). GTX denotes genotoxicity, defined as positive in the Ames test, the micronucleus assay (MN) or via analysis of chromosomal aberrations (CA). NA = not available, only four concentrations tested. *Maximum concentration due to cytotoxicity (IC₅₀). **The highest stock solution possible was prepared at a concentration of 1090 mM in DMSO. ROS = reactive oxygen species; ? = not clear.

^bFor data set 1, the TempO-Seq analysis was conducted at BioClavis (Glasgow, UK) using a customized version of the human S1500+ panel, whereas for data set 2, the analysis was done at the University of Ottawa (Ottawa, Canada) using the Human Whole transcriptome panel.

2.2. HepaRG Cell Culture and Chemical Exposures

Human HepaRG cell culture and chemical exposures were performed as described previously.²⁵ Every experiment was performed in triplicate using different batches of HepaRG cells. In brief, cryopreserved differentiated HepaRG cells were purchased from Biopredic International and cultivated according to the manufacturer’s protocola. Differentiated HepaRG cells were thawed and seeded into collagen-coated wells at a density of approximately 0.072 × 10⁶ viable cells per well (i.e., 100 μL of the cell suspension per well) in 96-well plates, using HepaRG Thawing/Plating/General Purpose Medium 670. After 24 h, the media were changed to HepaRG Maintenance Medium 620 for cell maintenance or HepaRG Induction Medium 640 for cell treatment. Cells were incubated for 7 days at 37 °C, 5% CO₂, and saturating humidity. At day 7, cells were treated with either 4 or 5 concentrations of test chemical (see Table 1) and refreshed with test chemical in media daily for 72 h (i.e., 3 day repeat exposures; 0, 24, and 48 h). After 72 h of exposure, cells were either used for cell viability (n = 3 per treatment group alongside matched solvent control, see further in Section 2.3) or used to generate cell lysates for gene expression analysis (n = 3 per treatment group per assay for TempO-Seq). For the TempO-Seq experiments, concentrations were based on the results of the cell viability assays (Table 1). The IC₁₀ value was selected as a guide (C3, the third concentration within the concentration range) to derive the concentration range to be tested. For each chemical, lower and higher concentrations around the IC₁₀ value were included, resulting in concentration intervals of 3-fold, with the exception of AFB1, for which the concentration range was based on previous qPCR experiments. Top concentrations induced a maximum of 50% cytotoxicity (i.e., IC₅₀) as per the established protocol used for TGx-DDI.^15,16 If there was no cytotoxicity, then the top concentration was 10 mM, which is the limit of solubility. After exposure, the media were aspirated, and the cells were washed with phosphate-buffered saline (PBS) prior to adding 1× TempO-seq lysis buffer (BioSpyder Technologies, Carlsbad, CA, USA) to each well for 10 min at 37 °C. Cell lysates were then frozen and stored at −80 °C for subsequent transcriptome profiling as described below (Section 2.4).

2.3. Cell Viability Measurement

A CellTiter-Glo Luminescence Viability Assay kit (cat. no. G7571; Promega Corporation, Madison, USA) was used to measure cell viability according to the manufacturer’s protocol. Briefly, wells containing 100 μL of cell of sample were equilibrated at room temperature for 30 min prior to the addition of 100 μL of CellTiter-Glo Reagent. The contents were mixed for 2 min on an orbital shaker to induce cell lysis prior to incubation at room temperature for 10 min to stabilize the luminescent signal. Luminescence was measured on a GloMax Explorer microplate reader (Promega Corporation, Madison, USA). Each assay was performed three times, every time with a more precise and narrower range, to find a concentration range with a top concentration inducing a maximum of 50% cytotoxicity. For each chemical, the IC₁₀- and IC₅₀-values and their R² values were evaluated using best-fit dose–response inhibition curves with GraphPad Prism v9.0 Software (GraphPad Software, Inc.). Cell viability was calculated as the percentage of viable cells in the treatment condition compared with the vehicle control. The cell viability assay was performed for each chemical except for CPA, MMS, and NDMA as the IC₁₀ of these three chemicals was based on IC₁₀-values in HepaRG cells previously published by our group.²²

Observations of cytotoxicity from the ATP assays were paired with read counts from TempO-Seq analyses to further identify cytotoxic concentrations and remove outliers from subsequent transcriptomic analyses. Overtly cytotoxic concentrations cause a large reduction in reads recovered during sequencing that are subsequently flagged in the Omics data analysis framework for regulatory application (R-ODAF) pipeline for removal during quality assurance and quality control (QA/QC) as described under Section 2.5.²⁸

2.4. TempO-Seq Assay: Library Building for Next-Generation Sequencing

The lysates of the samples of data set 1 were shipped to BioClavis (Glasgow, UK), a spinout company of BioSpyder Technologies, on dry ice where the TempO-Seq assay was conducted. As the TempO-Seq Human 1500+ Surrogate Panel did not contain all 84 GENOMARK genes, BioSpyder Inc. (BioSpyder Technologies, Carlsbad, CA, USA) designed a customized version of this panel comprising 3456 probes targeting 2792 genes. Bioclavis used the customized targeted transcriptome panel to carry out the proprietary BioSpyder TempO-Seq assay following the User Manualb, as described in detail below, to prepare the libraries. For data set 2, the TempO-Seq analysis was conducted at the University of Ottawa (Ottawa, Canada). The TempO-Seq Human Whole Transcriptome Reagent Kit (v2.1) (BioSpyder Technologies, Carlsbad, CA, USA) was used to prepare libraries in a 96-well plate format from exposed and control HepaRG cell lysates according to the manufacturer’s instructions and as previously described.²⁷

Although a different probe set was used for data sets 1 and 2, the TempO-Seq library building protocol was identical for both data sets and was completed by following the TempO-Seq User Manual. All concentrations of each test chemical were used for gene expression analysis. Assay controls included a negative no-lysate control (1× TempOSeq Lysis Buffer only) and two positive controls: qPCR Human Reference Total RNA and Human Brain Total RNA (Takara Bio, CA, USA; four replicates per control). Briefly, 2 μL of each sample (cell lysate and controls) in 1× TempO-Seq Lysis Buffer was hybridized to the detector oligo (DO) Pool using an annealing kit for the custom panel (data set 1) or whole human transcriptome panel (data set 2) supplied by BioSpyder for 10 min at 70 °C. This was followed by a temperature gradient with a ramp rate of 0.5 °C/min to 45 °C over 50 min with a 16 h hold at 45 °C and then cooled to 25 °C. Next, nuclease digestion was employed to remove excess, unbound, or incorrectly bound DOs enzymatically at 37 °C for 90 min. The DO pairs bound to adjacent target sequences were then ligated (60 min at 37 °C followed by 15 min of enzyme denaturation at 80 °C) to generate a pool of amplification templates. Each amplification template (10 μL of ligated DOs) was transferred into its respective well of the 96-well PCR plate containing PCR Pre-Mix and primers, supplied by BioSpyder. Amplification was conducted using a QuantStudio 3 Real-Time PCR System (ThermoFisher, ON, Canada) to add a sequence tag unique to each sample and the sequencing adaptors using the following PCR program settings: 37 °C for 10 min, 95 °C for 1 min; 25 cycles of 95 °C for 10 s, 65 °C for 30 s, 68 °C for 30 s (with fluorescence read for visual sample QC); 68 °C for 2 min, followed by a hold at 25 °C prior to library pooling and purification. NucleoSpin Gel and PCR cleanup kits (Macherey-Nagel, PA, USA) were used to pool and purify labeled amplicons. Libraries were quantified by using a KAPA Library Quantification Kit (Roche, QC, Canada). Libraries were sequenced using a NextSeq 2000 High-Throughput Sequencing System (Illumina, San Diego, CA, USA) using 50 cycles from a 75-cycle high-throughput flow cell to achieve a median read depth of 2 million reads per sample.

2.5. Data Processing and QC

The reads of data set 1 were aligned to the customized BioSpyder TempO-Seq Human Targeted Transcriptome probe set (3456 probes, targeting 2792 genes). For data set 2, the reads were aligned to the BioSpyder TempO-Seq Human Whole Transcriptome probe set (22,537 probes over 19,687 genes). Sequencing data have been deposited in the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) database under accession number GSE241069. Reads were demultiplexed (i.e., assigned to respective sample files) from the BCL files and processed into FASTQ files using bcl2fastq v2.20.0.422. The FASTQ files of both data sets were processed with the R-ODAF analysis pipeline in R (version 4.2.2), which is available at https://github.com/R-ODAF/R-ODAF_Health_Canada, to ensure that the data processing of both data sets was streamlined. Briefly, this pipeline used STAR version 2.7.10b to align reads to the reference FASTA file for the TempO-Seq probe sequences of the human whole transcriptome platform. Read counts were quantified from the BAM files by using the qCount function from the QuasR R package (version 1.38.0). A study-wide QC was applied to exclude samples of low quality based on established metrics.²⁹ Finally, we applied the gene-level filtering criteria described in R-ODAF, which is designed to reduce false positive differentially expressed gene detection.²⁸

2.6. Gene Expression Biomarker Analysis: GENOMARK and TGx-DDI Classification

Statistical modeling and bioinformatics tools were used to classify chemicals as (i) genotoxic (GTX) or non-GTX (NGTX) using the GENOMARK biomarker or (ii) DDI or non-DDI (NDDI) using the TGx-DDI genomic biomarker. Therefore, for both the GENOMARK and TGx-DDI biomarker genes, CPM-normalized read counts were log₂ transformed.³⁰ Gene symbols for either the GENOMARK or TGx-DDI biomarker genes that had multiple probes were summed prior to log₂ normalization. Details on the biomarker analyses and classification approaches for GENOMARK and TGx-DDI were published previously.^15,25 Briefly, in the case of GENOMARK (84 genes), two classifiers based on supervised machine learning algorithms were applied: (1) the support vector machine (SVM) using R packages e1071 and caret and (2) the random forest (RF) algorithm using the randomForest package in R Cran, Version 4.0.4 as described previously.²⁵ The output of both the SVM and RF classifier is a probability value between 0 and 1 for genotoxicity. A chemical is classified overall as GTX when the probability is >0.55 in both classifiers and as nongenotoxic when the probability is <0.45 in both classifiers. When the probability values are between 0.45 and 0.55 or when the chemical produced different results in both classifiers, the chemical is marked as inconclusive. For TGx-DDI (64 genes), three separate analyses were conducted to classify the chemicals: (1) the Nearest Shrunken Centroids (NSCs) using the pamr function of R; probability analysis (PA; visualized using heatmaps); (2) principal component analysis (PCA) using the prcomp() function in R; and (3) hierarchical clustering (HC) using the hclust() function in R (www.r-project.org). Classification was completed as follows: if a chemical yielded a DDI call in any one of three classification analyses (NSC PA heatmaps, PCA, or HC), it was classified as DDI; whereas, a chemical was classified as non-DDI if it did not have any DDI calls in any of the aforementioned analyses.^15,16

2.7. Benchmark Concentration Modeling

Log₂ normalized read counts were uploaded into BMDExpress v3 to conduct the benchmark concentration (BMC) analysis, in accordance with recommendations outlined in the US National Toxicology Program Approach to Genomic Dose–response Modeling report.^31,32 Test chemicals that were classified as GTX or DDI for GENOMARK or TGx-DDI, respectively, were subjected to BMC analysis for potency ranking. The highest concentration (C4) of EUG was eliminated because of overt cytotoxicity. Biomarker genes were analyzed and filtered using the Williams trend test retaining features with a permutation p-value <0.01 (with 500 permutations) with fold changes >1.5 for at least one concentration relative to matched solvent controls. To derive the BMC confidence intervals (CIs), the biomarker genes that passed the prefilters were fit to the following models: linear, exponential (3 and 5), polynomial 2, and the restricted power (power restricted to ≥1). A best fit model was selected with the lowest Akaike Information Criterion (AIC) value (lowest complexity). The benchmark response (BMR) was set to 1 standard deviation (SD) based on the recommendations from the expert panel held by the US National Toxicology Program.³¹ BMCs were filtered based on the goodness of fit (p-value >0.1), a BMC/BMCL ratio <20, a BMCU/BMCL ratio <40, and the BMC < the highest concentration.

Two approaches were used for the transcriptional biomarker PoD (tPoD) generation for TGx-DDI and GENOMARK:

• (a)Calculation of the bootstrap median BMC: a secondary analysis was conducted to generate CIs for the BMC_TGx-DDI and BMC_GENOMARK values using the bootstrap method. For each gene, 500 bootstrap samples were generated using the residual bootstrap method.³³ These data were then imported into BMDExpress v3 with the same filtering criteria and model selection as in the BMC analysis. As bootstrap samples are independent, the BMDExpress results were then used to simulate 2000 experiments, where each gene in the biomarker has a probability for inclusion into a bootstrap sample based on the relative frequency of that gene estimated as the total number of BMCs for that gene that passed all the filtering criteria divided by 500. For each simulated experiment, the median of the biomarker genes with BMCs was estimated. From the resulting 2000 simulated experiments, the distribution of the median BMC was estimated and the 95th percentile CIs were obtained for the median BMC_TGx-DDI and BMC_GENOMARK (bootstrap). GraphPad Prism 9.3.0 (GraphPad Software Inc.) was used to log₁₀ transform the BMCL and BMCU values and to visualize the CIs of each chemical in graphs.
• (b)Calculation of the BMD normalized gene set enrichment scores (NES): per sample GSEA was conducted for the TGx-DDI and GENOMARK gene sets. For each chemical, data were normalized to controls by subtracting the mean of the controls from the log2 counts per million. Fast GSEA using the fgsea R package³⁴ was applied to each sample to estimate the NES scores using 500 permutations for both biomarkers. NES scores were then imported into BMDExpress v3 for BMC analysis applying the same filtering criteria and model selection as in the BMC analysis. For chemicals that did not meet the filtering criteria, a second analysis relaxing the p-value and fold change cut-offs was conducted to identify which prefiltering criteria were not achieved and to obtain a BMC value.

3. Results

3.1. Cell Viability

Cell viability studies were performed to select a top concentration for the TempO-Seq experiments that induced a maximum of 50% cytotoxicity. This was done for all chemicals (Figure S1) except for CPA, MMS, and NDMA; the top concentrations for these chemicals were based on previously published studies.^22,23 MMC, 4NQO, AFB1, EUG, COL, and GLY caused a decline in viability, whereas 2CTAC showed no cytotoxicity up to 10 mM. In the TempO-Seq experiments, cytotoxicity was also evaluated during QA/QC. EUG (C4) was overtly cytotoxic and was removed from subsequent transcriptome analyses.

3.2. GENOMARK Biomarker Classification

To differentiate test chemicals as either GTX or NGTX utilizing the GENOMARK biomarker, two individual prediction models based on SVM and RF were applied to the transcriptional profiles generated through TempO-Seq sequencing (Table 2). The two NGTX chemicals, namely, EUG and 2CTAC, were accurately classified as NGTX. Notably, a concentration-dependent increase in the probability of genotoxicity was observed in both models for the eight GTX chemicals. Across various concentrations, all of the GTX chemicals were correctly classified as GTX, even when an inconclusive outcome was conservatively interpreted as positive. This yields a sensitivity of 100%, specificity of 100%, and predictive accuracy of 100% using the TempO-Seq technology.

Table 2. Overview of the GENOMARK and TGx-DDI Hazard Classifications Based on Gene Expression Data Obtained with TempO-Seq^a.

^aChemicals (GLY, MMS, NDMA, 4NQO, AFB1, COL, CPA, MMC, 2CTAC, and EUG) are listed in increasing test concentrations (C1–C4/C5) with their corresponding in vivo genotoxicity data. For GENOMARK, the individual prediction scores of the RF and SVM model are given together with the overall classification result for genotoxicity. A probability result <0.45 is considered nongenotoxic (NGTX) (green), ≥0.45 and ≤0.55 inconclusive (INCONCL) (yellow), and >0.55 genotoxic (GTX) (red). For TGx-DDI, classification, the results of the individual statistical analyses (NSC-PA, PCA, and HC) for direct DNA damaging (DDI, in red) or nondirect DNA damaging (non-DDI, in green) and the overall classification result are provided. *Excluded for transcriptome analysis due to overtly cytotoxic in QA/QC.

3.3. TGx-DDI Classification

To classify test chemicals as either DDI or non-DDI based on the TGx-DDI biomarker’s transcriptional profiles, three independent statistical analyses, namely, HC, PCA, and NSC-PA, were used to derive the overall classifications (Table 2). Again, the two NGTX chemicals, EUG and 2CTAC, were classified correctly as non-DDI and the eight GTX chemicals were classified as DDI. COL, an aneugen and therefore non-DDI, produced a “DDI” call for all concentrations in the NSC-PA and PCA models and an “inconclusive” call in the HC analysis for the three highest concentrations, but its overall classification remained DDI. Overall, TGx-DDI resulted in a sensitivity of 100%, specificity of 100%, and predictive accuracy of 100% for genotoxicity hazard classifications using the TempO-Seq technology.

3.4. BMC Analysis of the TempO-Seq Data for the GENOMARK and TGx-DDI Biomarker Genes for Potency Ranking

The TempO-Seq data for the chemicals that produced GTX or DDI calls in GENOMARK or TGx-DDI, respectively (i.e., all eight genotoxic chemicals), were further used for BMC modeling. A BMR of 1 SD was selected to derive BMCs for GENOMARK (BMC_GENOMARK) and TGx-DDI (BMC_TGx-DDI). The bootstrap method was used to calculate transcriptomic BMCs as this method allows for the calculation of the 95% CIs for the gene set (CIs; i.e., the interval spanning the BMCL and BMCU). The calculated median bootstrap BMC values for GENOMARK and TGx-DDI and the corresponding CIs are shown in Figure 1 and Table 3. Furthermore, the number of biomarker genes that could be modeled and the ratio between the BMCs obtained in both biomarkers (BMC_GENOMARK/BMC_TGx-DDI) of the transcriptomic BMC values were determined (Table 3).

Figure 1.

Comparison of potency rankings for the chemicals classified as genotoxic using the GENOMARK or TGx-DDI biomarker genes with bootstrap median BMCs. Median BMCs are expressed as log₁₀ concentration (μM) on the x-axis with their lower and upper 95% confidence limits. The potency ranking was from the TGx-DDI transcriptomic biomarker (in red). The potency ranking was from the GENOMARK transcriptomic biomarker (in green).

Table 3. Comparison of Benchmark Concentrations for the GENOMARK (BMC_GENOMARK) and TGx-DDI Biomarker (BMC_TGx-DDI) Genes^a.

chemical	GENOMARK bootstrap median BMC			TGx-DDI bootstrap median BMC			ratio BMCGENOMARK/BMCTGx-DDI
	#GENOMARK genes modeled	median BMCGENOMARK (BMCL–BMCU) (μM)	ratio (BMCU/BMCL)	#TGx-DDI genes modeled	median BMCTGx-DDI (BMCL–BMCU) (μM)	ratio (BMCU/BMCL)
COL	77	0.07 (0.02–0.15)	6.25	57	0.08 (0.01–0.52)	307.66	0.82
AFB1	78	0.07 (0.05–0.09)	1.87	59	0.06 (0.03–0.9)	2.87	1.24
MMC	78	0.15 (0.11–0.20)	1.85	61	0.15 (0.09–0.24)	2.75	0.99
4NQO	79	1.06 (0.73–1.51)	2.07	61	0.89 (0.50–1.62)	3.23	1.19
MMS	76	32.57 (23.28–45.31)	1.95	47	34.55 (20.76–62.78)	3.02	0.94
GLY	75	95.84 (55.57–178.36)	3.21	52	147.76 (75.43–395.99)	5.25	0.65
CPA	75	182.03 (118.74–318.66)	2.68	59	327.52 (133.83–567.30)	4.24	0.56
NDMA	70	310.37 (96.75–540.58)	5.59	41	234.21 (36.79–529.52)	14.39	1.33

^aChemicals that were classified as genotoxic (GLY, MMS, NDMA, 4NQO, AFB1, COL, CPA, and MMC) were modeled to derive BMCs. The number of biomarker genes that passed the prefilter criteria are listed. Median BMCs are given together with their lower and upper CIs (BMCL and BMCU, respectively).

The number of GENOMARK biomarker genes (84 in total) that fit models ranged from 70 (NDMA) to 79 (4NQO) using the BMC approach. In GENOMARK, COL, and AFB1 were the most potent chemicals, both with median GENOMARK BMCs of 0.7 μM. These were followed by MMC and 4NQO, with BMCs of 0.15 and 1.06 μM, respectively. The BMCs for MMS and GLY were 32.57 and 95.84 μM, respectively, and 182.03 μM for CPA. The least potent GTX chemical based on the GENOMARK BMC in HepaRG was NDMA, with a median gene BMC of 310.37 μM. The potency ranking for genotoxicity using the GENOMARK biomarker was thus as follows: COL = AFB1 > MMC > 4NQO > MMS > GLY > CPA > NDMA. However, CIs overlapped for COL, AFB1, and MMC as well as for GLY, CPA, and NDMA (Table 3).

The number of TGx-DDI biomarker genes (64 in total) that fit models ranged from 41 (NDMA) to 61 (4NQO and MMC). AFB1 and COL were the most potent chemicals, with median TGx-DDI BMCs of 0.06 and 0.08 μM, respectively, followed by MMC and 4NQO, with median BMCs of 0.15 and 0.89 μM, respectively. MMS, GLY, and NDMA had higher BMC values of 34.55, 14.76, and 234.21 μM, respectively. CPA was the least potent chemical in HepaRG based on TGx-DDI BMC modeling, with a median BMC of 327.52 μM. The median BMC_TGx-DDI potency ranking was as follows: AFB1 > COL > MMC > 4NQO > MMS > GLY > NDMA > CPA. However, the CIs of AFB1, COL, MMC, and 4NQO overlapped, as well as the CIs of MMS, MMC, GLY, NDMA, and CPA.

A highly similar potency ranking was obtained by using the GENOMARK and TGx-DDI biomarker BMCs (Figure 1). One notable difference was the width of the CI of COL, which was much larger, with a ratio of 307.66 using TGx-DDI compared to the width of the CI obtained with GENOMARK (ratio of 6.25). Overall, the width of the CIs (ratio of BMCU/BMCL) of the median BMCs of the eight chemicals was slightly larger in TGx-DDI than the CIs obtained in GENOMARK (Table 3). The BMC_GENOMARK/BMC_TGx-DDI ratio ranged from 0.56 to 1.33 (Table 3). Overall, the median BMC_GENOMARK and BMC_TGx-DDI were highly correlated (Figure 1).

A second analysis based on the BMC from NES scores resulted in nearly similar potency rankings as the bootstrap approach; however, a higher consistency was seen in terms of the median bootstrap BMCs and bootstrap confidence between both biomarkers (Figure S2 and Table S1).

4. Discussion

Transcriptomic biomarkers have emerged as important NAMs to predict molecular initiating events and key events in toxicity pathways. For genotoxicity, TGx-DDI and GENOMARK are the most documented transcriptomic biomarkers. Initially, the gene expression data collected with these transcriptomic biomarkers were only analyzed qualitatively, i.e., for hazard identification, as until recently regulatory testing and decision-making have been based on qualitative outcomes of in vitro and/or in vivo genotoxicity tests. However, it is now widely accepted that genotoxicity data can be used quantitatively to rank chemicals based on their genotoxic potency or to derive PoDs for risk assessment.^4,7,8,35,36 Such a quantitative analysis requires the use of high-throughput technologies, such as genomic tools, to generate data for more test chemical concentrations.¹² TempO-Seq, a targeted sequencing approach, is a particularly interesting high-throughput technology, as it directly targets the RNA contained in small amounts of raw cellular lysates while eliminating additional purification steps. This not only results in lower costs and analysis time but also circumvents increased variability associated with these steps.²⁷ Whereas the applicability of TGx-DDI to data generated with TempO-Seq has previously been demonstrated, herein, we confirm that the GENOMARK biomarker also correctly predicts genotoxic potential using TempO-Seq gene expression data. This compatibility offers an integrated testing approach with both biomarkers on the same data set, allowing for a more comprehensive genotoxicity assessment. Moreover, the two transcriptomic biomarkers can be combined with other biomarkers to identify biological perturbations after chemical exposure in a high-throughput workflow.^26,29

Interestingly, despite their different biomarker gene compositions, both biomarkers correctly classified the 10 test chemicals, yielding a predictive accuracy of 100%. GENOMARK and TGx-DDI share only two biomarker genes, i.e., DUSP14 and E2F7, which are both p53-regulated.^37,38 In fact, both biomarkers contain a substantial number of p53-regulated genes. Ates et al. demonstrated that the p53 pathway together with the EGF signaling pathway are the most represented in GENOMARK,²² whereas up to 75% of the genes in the TGx-DDI biomarker are associated with p53-regulation.³⁹ The difference in biomarker gene composition might be related to differences in the cell systems, exposure times, and reference chemicals used during their development.

We noted that TGx-DDI tended to classify the test chemicals as DDI at lower concentrations than GENOMARK. Several lower test concentrations were classified as “inconclusive” by GENOMARK, reflecting the different classification approaches of both biomarkers. GENOMARK uses two classifiers based on supervised machine learning algorithms to generate probability scores for genotoxicity, resulting in an ‘inconclusive’ category when results diverge between the SVM and RF models or when the probability falls between 0.45 and 0.55 in one or both models. Therefore, GENOMARK introduces a gray zone of “inconclusive” results that, from a statistical perspective, can help to characterize regions of the predictions with poor discriminatory ability.⁴⁰ In contrast, TGx-DDI uses three statistical algorithms to provide binary (DDI vs NDDI) results in each of the classifiers and classifies a chemical as ‘DDI’ even if one of the classifiers yields a ‘DDI’ call, leading to a more “protective” approach in the context of regulatory hazard classification. Therefore, integrating the outcome of both biomarkers can increase the certainty and/or confidence in the hazard calls obtained.

Our results indicate that both TGx-DDI and GENOMARK produce positive responses for genotoxic agents even at low, noncytotoxic concentrations. We observed concentration-dependent increases in positive hazard calls for genotoxicity. We tested up to a 50% cytotoxicity threshold. We recommend continuing to test up to these higher concentrations, in keeping with the OECD test guidelines for in vitro mammalian genotoxicity tests,⁴¹ to maintain the high sensitivity of the assay. Specifically, we suggest a maximum of 50% cytotoxicity (IC₅₀) or 10 mM when no cytotoxicity occurs. We caution that transcriptomic signatures may be confounded by overt cytotoxicity beyond this limit, although in many cases, the data processing pipeline would identify these concentrations as outliers because of the lower numbers of reads recovered. This was seen for EUG, which was used as a negative control in the current study; however, it is known that EUG induces toxicity and has pro-oxidative properties at high concentrations.^42–44 EUG was classified as nongenotoxic by both biomarkers at all concentrations tested in our study. Importantly, the highest concentration of EUG was removed, as it did not pass the QC criteria due to the low number of read counts. This is likely due to high levels of cytotoxicity, which were also visually observed under the microscope. Thus, the use of the transcriptomic QA/QC data processing filtering parameters provides a built-in methodology to flag and remove concentrations that are overtly toxic to the cells.

Our results further confirm that the TGx-DDI biomarker, although originally developed in TK6 cells with a 4 h exposure period, is also applicable under the GENOMARK conditions, i.e., in HepaRG following 72 h of exposure. Previous work has shown the value of integrating the high-throughput comet and the MN as integrated tests with TGx-DDI in HepaRG cells.^15,16 Our study extends this to provide additional critical validation of the value of integrating TGx-DDI with GENOMARK, both comprising very different gene sets, in the weight of evidence evaluation of genotoxicity in HepaRG cells.

An intriguing finding in this study was the classification of COL, an aneugen, as DDI in the TGx-DDI biomarker. This was contrary to expectations based on its mechanism of action, i.e., the inhibition of tubulin polymerization. The main difference between the two biomarkers is the classification of aneugens, since aneugens in the reference chemicals of GENOMARK are grouped as genotoxic but as non-DDI in TGx-DDI. In addition to COL, other aneugens were used as non-DDI agents in the training set to develop TGx-DDI in TK6 cells.¹⁴ COL induces mitotic arrest at specific concentrations, and therefore, the aneugenic effect would be concentration-dependent.^45,46 In previous studies using the TGx-DDI biomarker, COL tested in an equivalent concentration range to this study, was classified as NDDI using either TK6 cells or HepaRG cells after 4 and 55 h of exposure, respectively (Buick et al. 2020). A possible explanation for the positive hazard call of COL by TGx-DDI in the present study is that we used the experimental design developed for GENOMARK that applies a longer exposure time of 72 h. In a previous study by Bernacki et al., COL caused marked phospho-histone H3 increases after 4 and 24 h in TK6 cells, accompanied by delayed p53 responses (at 24 h) in the in vitro MultiFlow assay, highlighting the importance of the biomarker/time point for detecting aneugenicity.⁴⁷ As TGx-DDI is enriched in p53-responsive genes, the 72 h exposure period may have induced a robust change in the transcription of p53-regulated genes following COL treatment, resulting in a DDI classification.¹⁷ The data suggest that TGx-DDI also identifies aneugens as DDI using this design, but more work is necessary to support this application.

Quantitative analysis of the gene expression data from both biomarkers using BMDExpress software revealed highly similar tPoDs and potency rankings for the eight genotoxic test chemicals. There are multiple methods for calculating a tPoD,⁴⁸ but in this study, we examined two different approaches: (a) a bootstrap median BMC and (b) BMC modeling of NES scores. Modeling the NES scores produced in less coherent tPoDs and CIs between the two biomarkers. Thus, we propose that potency ranking for biomarkers having different gene sets be done by using the bootstrap median BMD approach. Using this approach, the potency ranking for both biomarkers was nearly identical. AFB1 was the most potent genotoxic compound. The only exception to potency ranking between the biomarkers was for CPA, which was slightly more potent in the GENOMARK assay, surpassing NDMA. However, since CIs overlap for these compounds, they are not significantly different from each other. In the literature, BMD/BMC values derived from both in vitro and in vivo genotoxicity studies are available for several of the tested chemicals. Nevertheless, due to differences in experimental design (e.g., test system, exposure time, end point measured, limited dose range, choice of BMR, etc.), direct comparison to other studies is challenging. Although the potency rankings were similar between both biomarkers, BMC_TGx-DDI values were associated with larger BMCU/BMCL ratios, especially for COL, compared to BMC_GENOMARK values, indicating a lower precision. The higher uncertainty associated with the BMC_TGx-DDI values may be because the data were produced under the GENOMARK experimental conditions. The high consistency in potency ranking between both biomarkers builds on evidence from previous studies^21,36,49 that in vitro transcriptomic data can be used quantitatively for screening and prioritization of data-poor chemicals since this approach is highly informative of hazard and potency. Studies such as the present quantitative analysis of the concentration–response in a metabolically proficient human cell line are crucial to support the move away from the traditional binary “screen-and-bin” genotoxicity assessment.

To advance the use of tPoDs for regulatory decision-making and risk assessment, standardization of approaches and agreement on best practices should be prioritized. Further studies coupling transcriptomic PoDs to estimates of human in vivo internal exposure via in vitro–in vivo extrapolation models (IVIVE) may enable the determination of health-based guidance values or chemicals that pose a greater risk to humans and thus should be prioritized for further testing. Indeed, Beal et al.⁴ demonstrated that the application of IVIVE to in vitro genotoxicity data yielded administered equivalent doses (AEDs) that are lower or equal to PoDs from animal studies for 65% of the chemicals they tested. Interestingly, seven of the eight genotoxic chemicals from our study, except for GLY, were included in this study. In the study by Beal et al., these chemicals were ranked for their median AED based on all the available in vitro genotoxicity data as follows: AFB1 > COL > MMC > 4NQO > NDMA > MMS > CPA. Afterward, the median AEDs were compared to the median in vivo PoDs; the seven chemicals were ranked based on the available in vivo data as follows: COL> 4NQO > MMC > AFB1 > MMS > NDMA > CPA. While our ranking based on the median BMCs aligns with the rankings obtained with the median AEDs and in vivo PoDs in the study of Beal et al., it remains challenging to directly compare the potency rankings across studies due to variations in experimental design (e.g., a different BMR, conversion to AEDs). A next logical step would be to convert the transcriptomic BMCs into AEDs by applying the IVIVE modeling approach to further calculate the bioactivity exposure ratios (BERs) that could be used in risk assessment.⁵⁰ Given the high consistency in potency ranking between GENOMARK and TGx-DDI, we posit that the AEDs derived from these biomarkers would lead to similar regulatory decisions based on the BERs, although more research is needed to confirm this.

In summary, our study demonstrates how gene expression profiling provides information on the genotoxic interaction of a compound with a biological system both in a qualitative (i.e., potential to induce genotoxicity) and quantitative (i.e., concentration level or potency corresponding to the compound’s effects) manner. Although transcriptomics is not a new technology, the novelty of our research lies in the quantitative analysis of the biomarker gene expression data and the striking agreement between potency estimates produced by these two very different biomarkers. Importantly, there is currently no consensus on how to implement this type of data to support regulatory decision making. One of the outstanding issues is determining how best to quantitatively analyze and interpret these large gene expression data sets. Our study confirms that transcriptomic biomarkers enable robust, objective, and efficient hazard classification in addition to potency estimation. This is an important finding, especially as the two biomarkers were originally developed under different test conditions (different cell systems and exposure times) and have only two genes in common. More research is needed to evaluate how these BMC values compare to BMC values derived in other in vitro and in vivo genotoxicity assays. Overall, this study, together with other ongoing efforts, contributes to the ultimate goal of implementing more accurate and efficient genotoxicity assessment approaches that combine traditional in vitro tests with NAMs to reduce reliance on time-consuming and costly genotoxicity studies using experimental animals.

Data Availability Statement

All Templated Oligo-Sequencing (TempO-Seq) data are publicly available through NCBI Gene Expression. Omnibus (GEO) under the accession number GSE241069.

Supporting Information Available

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.chemrestox.3c00318.

• Viability results of the test chemicals used in this study, comparison of potency rankings based on either the GENOMARK or TGx-DDI biomarker genes using the BMCs based on NES for the chemicals classified as genotoxic, comparison of BMCs for the GENOMARK and TGx-DDI biomarkers using BMC NES, and overview of the pie charts to illustrate the distribution of best-fit models for each chemical selected by BMDExpress for the statistically significant biomarker genes (PDF)

Author Contributions

^⊥ T.V. and B.M. are equally contributing last authors. A.T.: conceptualization, writing—original draft and review and editing, data curation, formal analysis, investigation, and visualization; E.C.: conceptualization, writing—review and editing, and data curation; M.M.: conceptualization, writing—review and editing, data curation, and formal analysis; A.W.: conceptualization, writing—review and editing, data curation, and formal analysis; C.L.Y.: conceptualization, writing—review and editing, funding acquisition, and supervision; V. R.: conceptualization, writing—review and editing, funding acquisition, and supervision; B. M.: conceptualization, writing—review and editing, funding acquisition, and supervision; and T. V.: conceptualization, writing—review and editing, funding acquisition, and supervision.

The authors declare no competing financial interest.