Biomarkers of DNA damage response improve in vitro micronucleus assays by revealing genotoxic mode of action and reducing the occurrence of irrelevant positive results

Abstract

We have previously described two flow cytometry-based in vitro genotoxicity tests: micronucleus (MN) scoring (MicroFlow^®) and a multiplexed DNA damage response biomarker assay (MultiFlow^®). Here, we describe a strategy for combining the assays in order to efficiently supplement MN analyses with a panel of biomarkers that comment on cytotoxicity (i.e. relative nuclei count, relative increased nuclei count, cleaved PARP-positive chromatin and ethidium monoazide-positive chromatin) and genotoxic mode of action (MoA; i.e. γH2AX, phospho-histone H3, p53 activation and polyploidy). For these experiments, human TK6 cells were exposed to each of 32 well-studied reference chemicals in 96-well plates for 24 continuous hours. The test chemicals were evaluated over a range of concentrations in the presence and absence of a rat liver S9-based metabolic activation system. MultiFlow assay data were acquired at 4 and 24 h, and micronuclei were scored at 24 h. Testing 32 chemicals in two metabolic activation arms translated into 64 a priori calls: 42 genotoxicants and 22 non-genotoxicants. The MN assay showed high sensitivity and moderate specificity (90% and 68%, respectively). When a genotoxic call required significant MN and MultiFlow responses, specificity increased to 95% without adversely affecting sensitivity. The dose–response data were analysed with PROAST Benchmark Dose (BMD) software in order to calculate potency metrics for each endpoint, and ToxPi software was used to synthesise the resulting lower and upper bound 90% confidence intervals into visual profiles. The BMD/ToxPi combination was found to represent a powerful strategy for synthesising multiple BMD confidence intervals, as the software output provided MoA information as well as insights into genotoxic potency.

Introduction

The in vitro micronucleus (MN) test is an important tool for evaluating chemicals’ potential to cause chromosomal damage. Although the chromosome aberration test can detect clastogens, MN formation is sensitive to both clastogenic and aneugenic activity. The broader range of genotoxic effects detected by MN relative to chromosome aberration is one important reason for the former’s increased use (1,2). Other advantageous characteristics include the fact that it requires less training to score via microscopy, and it is more amenable to automated scoring, for instance via image analysis, flow cytometry and imaging flow cytometry (3–5).

In recent years, there have been concerted efforts to supplement the in vitro MN endpoint with biomarkers that comment on genotoxicants’ mode of action (MoA). The MEGA-Screen assay is an image analysis system that combines MN scoring with kinetochore labelling, γH2AX and cell cycle analysis (6). The iScreen assay is a confocal image analysis approach that combines MN scoring with γH2AX, MPM2, phospho-histone H3 (p-H3), centromere protein A (CENPA) and cell cycle analysis (7). Others have used separate, follow-up-type in vitro assays to elucidate MoA (8–11). Recently, Smart et al. described a combination-type assay whereby chemical-exposed TK6 cells were split towards several assays that included in vitro MN analysis via the flow cytometry-based MicroFlow assay and a panel of DNA damage response biomarkers via the flow cytometry-based MultiFlow assay (12).

Given the encouraging results of Smart et al., we investigated approaches for increasing the efficiency of a MicroFlow/MultiFlow combination assay in TK6 cells. This included the evaluation of an optimised, washless exogenous metabolic activation system that was shown to be compatible with MultiFlow biomarkers (13) but as yet was untested in the MicroFlow assay.

Following optimisation, we evaluated the performance of the combination assay by testing 32 well-studied chemicals, with and without metabolic activation (see Figure 1). In addition to considering MN results in isolation, we also investigated the merits of requiring significant MultiFlow DNA damage biomarker response(s) in order to deem apparent MN induction as bona fide genotoxicity. We hypothesised such a requirement would benefit specificity, since the in vitro MN assay has been described as exhibiting poor specificity in the presence of appreciable apoptosis (14–16).

Fig. 1.

Graphical depiction of the combination MicroFlow/MultiFlow assay. Human TK6 cells are exposed to a series of finely spaced test article concentrations in 96-well plates, in the presence and absence of a rat liver S9 activation system. At 4 and 24 h, cells are removed for MultiFlow biomarker measurements. At 24 h, the remaining cells are used for MicroFlow analyses. The resulting data provide a comprehensive assessment of test article-induced cytotoxicity and genotoxicity.

Finally, given the numerous biomarker/timepoint combinations evaluated, we developed an approach for synthesising multiple benchmark dose (BMD) results for each chemical into an aggregate potency score and associated graphic. This was performed with PROAST software (17) in combination with ToxPi software (18), and the merits and current limitations of this strategy for describing chemicals’ relative genotoxic potency are discussed.

Materials and Methods

Chemicals, cells and culture conditions

The identity and source of 32 chemicals, as well as information about presumed genotoxic MoA, are provided in Table 1.

Table 1.

Test chemicals, source and biological effects

Chemical (abbreviation)	CAS no., source	Notes about MoA, biotransformation	References
Alosetron hydrochloride	122852-69-1, Sigma-Aldrich	5-HT3 receptor antagonist	(19)
AMG900	945595-80-2, Selleckchem	Aurora kinase inhibitor	(20)
2-Aminoanthracene	613-13-8, Sigma-Aldrich	Aromatic amine, requires metabolic activation (CYP1B1, 2A family)	(21)
Benzo[a]pyrene	50-32-8, Sigma-Aldrich	Polycyclic aromatic hydrocarbon, requires metabolic activation (CYP1A1, 1B1 and epoxide hydrolase)	(19)
Brefeldin A	20350-15-6, Sigma-Aldrich	ER-golgi transporter inhibitor, ER stress- induced apoptosis	(22)
Carbonyl cyanide m-chlorophenyl hydrazone (CCCP)	555-60-2, Sigma-Aldrich	Uncouples oxidative phosphorylation, potent apoptogen	(23)
Cisplatin	15663-27-1, Sigma-Aldrich	Atypical alkylator	(19)
Chlorambucil	305-03-3, Sigma-Aldrich	Nitrogen mustard-type alkylator	(24)
Colchicine	64-86-8, Sigma-Aldrich	Tubulin binder	(19)
Cycloheximide	66-81-9, Sigma-Aldrich	Protein synthesis inhibitor	(25)
Cyclophosphamide monohydrate	6055-19-2, Sigma-Aldrich	Nitrogen mustard, requires metabolic activation (CYP2B6, CYP2C19, CYP2C9 and CYP3A4/5)	(19, 26)
Dexamethasone	50-02-2, Sigma-Aldrich	Glucocorticoid receptor agonist	(27)
Dibenzo[a,l]pyrene	191-30-0, Sigma-Aldrich	Polycyclic aromatic hydrocarbon, requires metabolic activation (thought to be primarily activated by CYP1A1)	(28)
1,2-Dimethylhydrazine	540-73-8, Sigma-Aldrich	DNA methylating agent, requires metabolic activation	(29)
Ethyl methanesulfonate	62-50-0, Sigma-Aldrich	DNA ethylating agent	(30)
Etoposide	33419-42-0, Sigma-Aldrich	Topoisomerase 2 inhibitor	(19)
Genistein	446-72-0, Sigma-Aldrich	Topoisomerase 2 inhibitor	(31)
Hydroxyurea	127-07-1, Sigma-Aldrich	Ribonucleotide reductase inhibitor; non- mutagenic clastogen	(32)
Ixabepilone	219989-84-1, Selleckchem	Tubulin binder	(33)
d-Mannitol	69-65-8, Sigma-Aldrich	Polyol	(19)
Melphalan	142-82-3, Sigma-Aldrich	Nitrogen mustard-type alkylator	(24)
Methotrexate	59-05-2, Sigma-Aldrich	Anti-metabolite	(34)
Methyl methanesulfonate	66-27-3, Sigma-Aldrich	Alkylator	(19)
Paclitaxel	33069-62-4, Sigma-Aldrich	Tubulin binder	(19)
2-amino-1-methyl-6-phenylimidazo [4,5-b]pyridine (PhIP)	105650-23-5, Toronto Research Chemicals	Heterocyclic amine, requires metabolic activation (CYP1A family)	(19, 35)
PF03814735	942487-16-3, Selleckchem	Aurora kinase inhibitor	(36)
Temozolomide	85622-93-1, Sigma-Aldrich	DNA alkylator	(37)
Thapsigargin	67526-95-8, Sigma-Aldrich	Cytotoxicant, ER stress-induced apoptosis	(38)
Thiotepa	52-24-4, Sigma-Aldrich	Alkylator	(24)
Vinblastine sulfate	143-67-9, Sigma-Aldrich	Tubulin binder	(19)
VX680	639089-54-6, Selleckchem	Aurora kinase inhibitor	(39)

TK6 cells were purchased from ATCC^® (cat. no. CRL-8015). Cells were grown in a humidified atmosphere at 37°C with 5% CO₂ and maintained at or below 1 × 10⁶ cells/ml. The culture medium consisted of RPMI 1640 with 200 µg/ml sodium pyruvate (both from Sigma-Aldrich, St. Louis, MO), 200 µM l-glutamine, 50 units/ml penicillin and 50 µg/ml streptomycin (from Mediatech Inc., Manassas, VA) and 10% v/v heat-inactivated horse serum (Gibco^®, a Thermo Fisher Scientific Company, Waltham, MA).

Cell treatments

The optimised S9 (low concentration, continuous exposure) activation approach has been described in detail elsewhere (12). Briefly, phenobarbital-/β-naphthoflavone-induced rat liver S9 was prepared in a cofactor mix at a concentration of 2.5% v/v. This 10x solution was maintained on ice until it was added at a 1:9 ratio to TK6 cells adjusted to 2 × 10⁵/ml in culture medium for a final S9 concentration of 0.25%. The 0% S9 cultures were TK6 cells at 2 × 10⁵/ml in culture medium without cofactors or S9.

Dose range-finding experiments were performed to generate 24 h relative nuclei count (RNC) and relative increased nuclei count (RINC) data for each chemical and to evaluate test substances for contributions to background fluorescence. Chemical treatments occurred in U-bottom 96-well plates, with 198 µl TK6 cell suspensions with and without S9/cofactor mix as described above. Test chemicals prepared in dimethyl sulfoxide (DMSO) were added at 2 µl/well for final DMSO concentrations of 1% v/v. The highest test chemical concentration was 10 mM unless solubility or previous experience with a chemical indicated that this would be overly cytotoxic. Testing occurred at 20 concentrations in single wells, and each concentration differed from the one above by a factor of 70.71%. Solvent was tested in at least 10 replicate wells spread throughout the plate. Upon addition of test chemicals, plates were immediately incubated in a humidified atmosphere at 37°C with 5% CO₂ for 24 h.

Definitive experiments with the 32 test chemicals were conducted in 96 well plates, with and without S9, as described above. As with the dose range-finding experiments, test chemicals were prepared in DMSO and delivered at 2 µl/well for 1% v/v. The highest concentration was derived from the dose range-finding experiment. In the case of non-cytotoxic freely soluble chemicals, the top concentration was 10 mM. When precipitate was evident, the lowest precipitating concentration was evaluated. Otherwise, the highest concentration tested was chosen with the goal of achieving approximately 80% reduction to RNC at 24 h. Nine additional lower concentrations were tested using the 70.71% dilution scheme described above. Each concentration was tested in triplicate wells, whereas DMSO controls were evaluated in 12 replicate wells spread throughout the plate. Upon addition of test chemicals, plates were immediately incubated in a humidified atmosphere at 37°C with 5% CO₂ for 24 h.

MultiFlow DNA damage assay: sample processing and flow cytometric analysis

TK6 cells were prepared for flow cytometric analysis using reagents and instructions included in the MultiFlow^® DNA Damage Kit—p53, γH2AX, Phospho-Histone H3, Cleaved-PARP (Litron Laboratories, Rochester, NY). Components and preparation of the MultiFlow working solution have been described in detail elsewhere (9,40). At the 4 and 24 h sampling times, cells were resuspended with pipetting, and then 25 µl were removed from each well and added to a new 96-well plate containing 50 µl/well of pre-aliquoted working MultiFlow reagent solution. Although the definitive experiments’ working reagent solution contained each of the kit-supplied fluorescent antibodies (anti-p53,anti-γH2AX, anti-p-H3 and anti-Cleaved PARP), they were omitted for the dose range-finding experiment to facilitate assessments of background fluorescence as described in detail by Tian et al. (13). Mixing was accomplished by pipetting the contents of each well several times. After incubation at room temperature for 30 min, samples were analysed via flow cytometry.

Flow cytometric analysis was carried out using a Miltenyi Biotec MACSQuant^® Analyzer 10 flow cytometer with integrated 96-well MiniSampler device. Stock photomultiplier tube detectors and associated optical filter sets were used to detect fluorescence emissions associated with the fluorochromes: Brilliant Violet™ 421, FITC, PE, propidium iodide and Alexa Fluor^® 647 (detected in the V1, B1, B2, B3 and R1 channels, respectively). Flow cytometry data were analysed using FlowJo software v10.5.0 (Ashland, OR, USA).

Representative bivariate graphs, gating logic and position of regions have been described in detail elsewhere (9, 40). γH2AX and p53 responses were based on the shift in median channel fluorescence intensity relative to same-plate solvent controls, whereas p-H3-positive, cleaved PARP-positive and polyploidy data were based on their frequency among other nuclei. Nuclei to counting bead ratios were calculated for each sample, and these ratios were used to determine absolute nuclei counts (those with 2n and greater DNA-associated propidium iodide fluorescence). Nuclei counts were used to derive RNC, and for the MultiFlow assay, %cytotoxicity was calculated as 100 minus %RNC at 24 h.

MultiFlow DNA damage assay: data analysis

Data analyses described herein were restricted to those concentrations that did not exceed the MultiFlow assay’s cytotoxicity limit. That is, as described by Dertinger et al. (41), the top concentration of each chemical had to exhibit ≤80% cytotoxicity at the 24-h time point, and no more than two concentrations within the cytotoxicity range 70–80% were permitted.

Test substances were corrected for background fluorescence as necessary, as described by Tian et al. (13). MultiFlow data were then converted to fold-change values relative to mean solvent control. This was performed for every biomarker, well, and time point for those test article concentrations that were not excluded due to excessive cytotoxicity or other limits described above.

An ensemble of three machine learning models—multinomial logistic regression, artificial neural network and random forest—has been described in detail previously (42). Briefly, these models utilise 4 and 24 h MultiFlow data fold-change values to predict whether a chemical is clastogenic and/or aneugenic.

Micronucleus (MicroFlow) assay: sample processing and flow cytometric analysis

After 24 h of exposure, cells were processed for MN frequencies via flow cytometric analysis using In Vitro MicroFlow^® Kit reagents (Litron Laboratories, Rochester, NY, USA). These methods have been reported in detail elsewhere (43–45) and utilised all remaining cells that were left over from the two 25 µl samples that went towards MultiFlow analyses as described above.

Flow cytometric analyses were carried out using a BD FASCANTO^™ II flow cytometer equipped with a BD High Throughput Sampler device (both from Becton, Dickinson and Company, San Diego, CA, USA). Stock photomultiplier tube detectors and associated optical filter sets were used to detect fluorescence emissions associated with the MicroFlow kit-provided fluorochromes: SYTOX Green^® (detected in the FITC channel) and ethidium monoazide (EMA, PerCP-Cy5.5 channel). The stop mode for these analyses was the acquisition of 10 000 EMA-negative nuclei per replicate well. Additional details, including region placement and gating logic, have been described in detail elsewhere (43–45).

Micronucleus (MicroFlow) assay: data analysis

MN analyses were restricted to those concentrations that resulted in ≤60% reduction to RINC, the lowest precipitating concentration or 10 mM, whichever was lowest. %MN were converted to fold-change values relative to mean solvent control (with and without S9 arms data considered separately). This was performed for every well and test article concentration that was not excluded due to excessive cytotoxicity or other limits described above.

Initially, two criteria had to be met in order to consider test article-induced MN significant: >2.5× over concurrent mean solvent control and a mean %MN that exceeded an upper bound tolerance interval calculated from historical negative control distributions (with and without S9 data were considered separately, two-tailed tolerance interval, alpha 0.05 and coverage 99%). For the 0% S9 condition, the upper bound tolerance interval was 0.63% (n = 439); for the 0.25% S9 condition, the upper bound tolerance interval was 1.39% (n = 429).

In addition to evaluating MN assay results in isolation, we investigated the performance of the combination assay by requiring apparent MN induction to be accompanied by MultiFlow biomarker results that yielded a clastogen or aneugen prediction.

Benchmark dose and ToxPi analyses

BMD analyses (16) were performed with PROAST v69 for each of eight genotoxicity-related biomarker/time point combinations: 4 and 24 h p53, γH2AX and p-H3, and 24 h polyploidy and MN. Critical effect size (CES) values were +0.5 for all the biomarkers except for MN which was set at +1.0, and polyploidy set at +5.0. In all cases, an exponential model was used to fit the dose–response curves. These analyses were conducted for a single chemical and biomarker at a time, with S9 condition and harvest time (when applicable) as covariate(s).

BMD output was then prepared for ToxPi analysis. Whenever a chemical’s BMD value for a particular biomarker was less than or equal to the top passing/valid concentration, the exact corresponding BMD lower limit (BMDL) and BMD upper limit (BMDU) values were entered into a ToxPi-formatted Excel spreadsheet. There were several exceptions to this: (i) when a calculated BMD value was greater than the actual top passing/valid concentration, we entered 10 000 µM for the BMDL and BMDU values; (ii) when there was no evidence of a dose-related increase as indicated by an infinite upper bound, we entered 10 000 µM for the BMDL and BMDU values; (iii) when there was no evidence of a dose-related increase as indicated by a curve fit that depicted a downturn, we entered 10 000 µM for the BMDL and BMDU values; and (iv) when BMDL or BMDU values were less than 0.0001 µM, we entered 0.0001 µM.

The ToxPi-formatted Excel file was converted to a csv file and analysed with ToxPi software, v2.3. A model comprised of eight slices was created, with each slice corresponding to the following biomarker/timepoint combinations: 4 and 24 h p53, γH2AX and p-H3, and 24 h polyploidy and MN. BMDL and BMDU values were assigned to the designated biomarker/timepoint, and they were −log₁₀ transformed. The slices were colour coded and arranged to provide visual cues as to genotoxic MoA. That is, the top-most two slices were colored shades of purple and correspond to pan-genotoxic activity (24 h MN and p53); slices to the right were colored shades of red/orange and correspond to clastogenic activity (4 and 24 h gH2AX and 4 h p53); and slices to the left were colored shades of blue and correspond to aneugenic activity (4 and 24 h p-H3, and 24 h polyploidy). Differential weighting of the biomarkers/slices was not performed in this initial proof-of-concept work. See Figure 2 (note: the resulting model and associated BMDL and BMDU data are available upon request).

Fig. 2.

Left image: Anatomy of a ToxPi profile. This generic profile shows results for one chemical and three biomarkers, each of which is represented by a different coloured pie slice. For each slice, the distance from the origin to the thin white line represents the ‘slice score’ (i.e. longer protrusion = higher genotoxic potency for that particular biomarker). Note that since we chose to portray potency using BMD values, a −log₁₀ transformation was necessary. The radial angle (width) of each slice indicates the statistic’s weight in the overall model. Optional lower and upper bounds confidence intervals (CIs) are indicated as the lighter-shaded area at the boundary of each slice arc. The −log₁₀ transformation explains why the BMD lower confidence interval value is portrayed at the outer edge of each pie slice. Right image: ToxPi key for the eight genotoxicity biomarkers studied in the current report. Note that the arrangement and colouring of the biomarker slices were configured to provide visuals cues as to genotoxic MoA.

After running initial ToxPi analysis and deriving aggregate ‘ToxPi scores’, two other operations were performed. First, ToxPi’s 95% bootstrap confidence interval algorithm was performed. Since each slice contains only two data points (BMDL and BMDU), the algorithm simply returns those two values—that is, the 90% confidence interval derived from PROAST. This provided a means to graphically depict the BMDL and BMDU interval as a lighter-shaded band at the periphery of each ToxPi profile. Secondly, ToxPi’s hierarchical clustering algorithm was executed (clustering options = complete; circular format). An overview of the complete data acquisition/analysis pipeline is shown in Figure 3.

Fig. 3.

Data acquisition and analysis pipeline for the combination MicroFlow/MultiFlow assay.

Results and Discussion

Experimental design and metabolic activation considerations

Several practical aspects regarding the conduct of the combined assay were addressed. Since 25 µl of cell suspension were removed two times from each well of a 96-well plate to accomplish MultiFlow readings, it was important to determine whether there were enough cells remaining to support MicroFlow analyses. We found that we could indeed consistently acquire sufficient EMA-negative nuclei (10 000 per well) for MN frequency determinations. Whenever this was not possible, the test chemicals’ concentrations exceed the cytotoxicity limit and were, therefore, not of interest.

Whereas the MultiFlow assay uses RNC-based cut-off to establish top passing concentration (9), we used RINC to align with current in vitro MN assay Organisation for Economic Cooperation and Development (OECD) Test Guideline recommendations (46). It should also be noted that the maximal cytotoxicity values are different for these assays. Thus, by setting the top MultiFlow concentration as usual, it was conceivable that too many concentrations would be deemed overly cytotoxic for MN analysis. Indeed, a reduction in the number of passing/valid concentrations was typical for the MN assay. However, this was not an issue, as there were always enough analysable test article concentrations (≥3) for an adequate MN evaluation. Temozolomide without S9 serves as an illustrative example. In this case, MultiFlow and MicroFlow assays were both limited by cytotoxicity (as opposed to limit concentration or test article solubility). Given the differences in cytotoxicity metrics and tolerances, MultiFlow effectively evaluated nine concentrations ranging from 1.95 to 500 µM, whereas MicroFlow was restricted to six concentrations ranging from 1.95 to 62.5 µM. See Figure 4.

Fig. 4.

Thirteen biomarker readings are graphed against temozolomide concentration (µM, log₁₀ transformed). For each concentration, every plotted point represents a replicate measurement performed on a separate well. The biomarkers are grouped according to the type of effects they are sensitive to. Top panels = cytotoxicity-associated metrics: RNC, RINC, EMA-positive chromatin and cleaved PARP. Middle panels = pan-genotoxicity biomarkers: MN and p53 activation at 24 h. Bottom left panels = aneugen-responsive biomarkers: phospho-histone H3 (p-H3; increases are a signature of tubulin binders) and polyploidization. Bottom right panels: clastogen-responsive biomarkers: γH2AX, and p53 activation at 4 h.

A third aspect of the experimental design that was considered was the optimised (low/continuous) S9 metabolic activation approach. While it was shown to be effective in the MultiFlow assay (13), it was unknown whether this activation system would be appropriate for the MN endpoint. Furthermore, since we were attempting to streamline cell processing by forgoing dedicated washing steps that are usually performed when using conventional S9 activation protocols, we had concerns that small insoluble particles in S9 preparations would interfere with MN scoring. Data presented in Table 1 for four S9-requiring chemicals (2-aminoanthracene, benzo[a]pyrene, cyclophosphamide and dibenzo[a,l]pyrene) provided assurances that the low/continuous S9 method is compatible with automated MN scoring. Whereas each of the chemicals induced MN in the presence of S9, they were observed to be inactive in its absence.

Sensitivity

Diverse genotoxicants were tested in the MicroFlow/MultiFlow combination assay with and without S9 activation. Table 2 provides a summary of the results. When considered in isolation, the MN endpoint showed good sensitivity: 22/24 genotoxicants identified as such, with 38 of 42 metabolic activation arms performing according to a priori expectations. Only 1,2-dimethylhydrazine and methotrexate did not meet our criteria for significant MN induction.

Table II.

Summarized Results

When an apparent MN response required a confirmatory MultiFlow assay result for an overall positive result, assay sensitivity did not change: 22/24 genotoxicants were identified as such. In fact, the MultiFlow results appeared to shed light on the MN results for methotrexate and 1,2-dimethylhydrazine. That is, the contradictory (but correct) MultiFlow clastogen predictions may be pointing to (i) a suboptimal treatment/harvest schedule for MN induction by these chemicals; (ii) the fact that the MultiFlow assay is less dependent on dividing cells relative to MN; or some combination of the two.

Specificity

Diverse non-genotoxicants were tested in the combination assay with and without S9 activation. Since apoptosis is known to confound MN assays, we included several potent apoptogens in this group. Table 2 provides a summary of the results. When considered in isolation, the MN endpoint showed moderate specificity: 5/8 non-genotoxicants identified as such. By considering both metabolic arms, as well as the pro-genotoxicant chemicals, a total of 22 results were anticipated to show no appreciable MN induction. In this case, we observed 15/22 non-genotoxicants as correctly identified according to a priori expectations—again, not an outstanding level of specificity (68%).

When an apparent MN response required a confirmatory MultiFlow assay result for an overall positive result, assay specificity markedly improved: 8/8 non-genotoxicants were identified as such, and 21/22 metabolism arms performed according to expectations. Importantly, this was the case even for the potent apoptogens (e.g. CCCP, brefeldin A and thapsigargin), despite the fact they were tested to higher concentrations and a greater cytotoxicity limit for MultiFlow analyses relative to parallel MicroFlow analyses.

Mode of action

As summarised in Table 2, the MultiFlow assay provided MoA predictions via an established machine learning ensemble. Each of the 32 chemicals and 63/64 MoA predictions were correct. A ToxPi analysis option, hierarchical clustering, provided an additional opportunity to investigate MoA. First, as shown in Figure 5, the clades show good discrimination between clastogens, aneugens and non-genotoxicants, corroborating the machine learning-based MoA predictions. Second, the irrelevant positive MN chemicals were grouped together, and the accompanying visuals effectively highlighted the lack of supporting MultiFlow-associated biomarker responses. Third, the clustering function provided insights into molecular mechanisms. Specifically, the tubulin binder-type aneugens appeared in a different subclade relative to the mitotic kinase inhibitor-type aneugens.

Fig. 5.

ToxPi profiles for 32 chemicals, 2 metabolic activation conditions and 8 genotoxicity-centric biomarkers (i.e. a total of 512 dose–response analyses). This graph was generated by ToxPi’s hierarchical clustering algorithm that automatically groups similar profiles. Several distinct categories are apparent: clastogens and what might be considered weak clastogens (Clast and W Clast), mitotic kinase inhibitor-type aneugens (An: KI), tubulin binding-type aneugens (An: TB), non-genotoxicants (Non-G) and chemicals that induced what might be considered irrelevant positive MN responses (IP).

Potency

Another useful ToxPi software feature is generation of aggregate ‘ToxPi scores’ for each chemical/S9 condition. Figure 6 shows the resulting ToxPi score ‘rank plot’ based on 8 biomarkers and 32 chemicals both with and without S9 activation. While the rank plot is interactive within the software program, this static image highlights several chemicals for illustrative purposes. (Note that the fully functional ToxPi model and associated data are available upon request, including results from the 512 total BMD analyses.)

Fig. 6.

ToxPi scores for 32 chemicals and 2 metabolic activation conditions. This is an interactive plot in ToxPi software. Here, the static image is used to highlight several well-studied chemicals and to demonstrate how ToxPi Scores may represent a useful metric that synthesise disparate assay/biomarker results into an aggregate potency score.

As shown in Figure 6, vinblastine generated the highest aggregate ToxPi score, as low sub-micromolar concentrations affected both the pan-genotoxicity biomarkers and each of the aneugen-responsive biomarkers. Dibenzo[a,l]pyrene was also identified as a relatively potent genotoxicant, but in this case, there was a stark difference between scores with and without S9. Hydroxyurea was observed to affect the pan-genotoxicity biomarkers as well as clastogen-responsive biomarkers, but orders of magnitude higher concentrations were required compared with vinblastine and dibenzo[a,l]pyrene with S9, hence the shorter ToxPi slice protrusion and lower overall score. The rank plot also makes it readily apparent that, in the case of hydroxyurea, rat liver S9 did not influence its genotoxic potency.

Conclusions and Future Directions

This work demonstrates that important advantages are realised when MN scoring is supplemented with panel(s) of DNA damage response-type biomarkers. The results show that genotoxic MoA is reliably predicted, and assay specificity is substantially improved, even in the case of challenging apoptogens.

Given the amount of data generated by the combination MicroFlow/MultiFlow assay, we found it necessary to explore new ways of displaying and evaluating the results. The novel combination of BMDL and BMDU metrics with ToxPi software represents a powerful tool for synthesising and visualising the multiple genotoxicity biomarkers. As shown in Figure 5 (hierarchical clustering), MoA information is clearly addressed.

In addition to MoA, the ToxPi score represents an intriguing aggregate potency metric that warrants further study. In order to assess the reliability of this metric, it will be necessary to compare it to ground truth-type metric(s), for example corresponding in vivo BMD values—see, for instance reference (47). While the ToxPi modelling described herein did not assign different weights to any of the biomarkers, this functionality is built-in to the software (see Figure 2). Thus, if empirically derived approaches for differentially weighting biomarkers are achieved, and if the resulting ToxPi scores are indeed predictive of in vivo genotoxic potency, BMD/ToxPi analyses may greatly facilitate prioritisation of chemicals for further study and contribute to regulatory decision-making. While we considered developing differential weighting approaches beyond the scope of the current investigation, this as well as optimisation of biomarkers’ CES values are among the goals of a recently convened subgroup of the Health and Environmental Sciences Institute—Genetic Toxicology Technical Committee.

Acknowledgements

We would like to thank members of the Health and Environmental Sciences Institute—Genetic Toxicology Technical Committee’s Quantitative Workgroup. They recently formed a subgroup to investigate the utility of combining Benchmark Dose analyses with ToxPi software, and their intellectual support is greatly appreciated.

Funding

This work was funded by grants from the National Institute of Health/National Institute of Environmental Health Sciences (NIEHS; grant no. R44ES029014 and R44ES033138). The contents are solely the responsibility of the authors, and do not necessarily represent the official views of the NIEHS.

Conflict of interest statement: Several authors are employed by Litron Laboratories. Litron holds patents for flow cytometry-based analyses described herein and sells the In vitro MicroFlow^® Kit and the MultiFlow^® DNA Damage Kit—p53, γH2AX, Phospho-Histone H3, Cleaved PARP. The contents are solely the responsibility of the authors and do not necessarily represent the official views of the NIEHS.