Test chemical information (including the solvent control) and their respective abbreviations, CAS RN, and source are presented in Table 1.

All experiments were performed with human lymphoblastoid TK6 (IVTG strain) cells (ECACC #13051501) purchased from Sigma-Aldrich Canada (Oakville, ON). Cells were grown in RPMI1640 cell culture medium (ThermoFisher Scientific, Ottawa, ON) supplemented with 10% horse serum (ThermoFisher Scientific, Ottawa, ON), 2 mM L-glutamine (ThermoFisher Scientific, Ottawa, ON), 1.8 mM sodium pyruvate (ThermoFisher Scientific, Ottawa, ON), and 100 U/mL of penicillin and streptomycin (ThermoFisher Scientific, Ottawa, ON). Cells were incubated at 37°C with 5% CO₂ and maintained below 1 x 10⁶/mL and >90% viability.

Cells were adjusted to a density of 1.5 x 10⁵/mL, aliquoted in 96-well plates, and exposed to 1 µL of test chemical (1% v/v, final volume 100 µL) solubilized in DMSO at a range of concentrations. For test chemicals requiring metabolic activation the medium was supplemented with 10 µL of 5% S9 Mix (0.5% v/v) containing Aroclor-1254 Induced Mutazyme^® (Moltox, Boone, NC). Treated cells were then incubated at 37°C with 5% CO₂ for 4H. Cells were collected via centrifugation at 300 x g for 5 min, washed with Dulbecco’s phosphate buffered saline (DPBS) (ThermoFisher Scientific, Ottawa, ON), resuspended in fresh media, and incubated at 37°C with 5% CO₂ for 20 h. The viable cell count was quantified via propidium iodide fluorescence analysis with the Miltenyi Biotec MACSQuant ^® Analyzer 10 flow cytometer with integrated 96-well MiniSampler device. Instrument settings: autolabel PI. The fluidics parameters were as follows: sample volume of 100 μL at a medium mix rate, followed by a sample uptake volume of 25 µL analyzed at a medium flow rate. The screen mode was used to rinse the probe between samples. Top concentrations were selected (Table 2) up to 10 mM or lower due to solubility (OECD, 2016). If cytotoxic, top concentrations were selected for each assay that induce: approximately 40% viability (i.e., 60% cytotoxicity) for the TGx-DDI assay and in vitro MicroFlow^® assay (Bryce et al., 2008; OECD, 2016), and 20% viability (i.e., 80% cytotoxicity) for the MultiFlow^® DNA Damage assay (Bryce et al., 2016).

Cells were adjusted to a density of 1.5 x 10⁵/mL, aliquoted in 96-well plates, and exposed to 1 µL of test chemical (1% v/v, final volume 100 µL/well) solubilized in DMSO in a six-point concentration range on a half-log distribution scale in duplicate (Table 2). Six DMSO solvent controls were included per plate. For test chemicals requiring metabolic activation, the medium was supplemented with 10 µL of 5% S9 Mix (0.5% v/v) containing Aroclor-1254 Induced Mutazyme^® (Moltox, Boone, NC). Treated cells were incubated at 37°C with 5% CO₂ for 4H. Cells were collected via centrifugation at 300 x g for 5 min, washed with DPBS, and resuspended in 1X TempO-Seq™ Enhanced Lysis buffer (BioSpyder Technologies, Carlsbad, CA, United States) in DPBS (40 µL/well) (ThermoFisher Scientific, Ottawa, ON). If treated with 0.5% rat liver S9, cells were collected via centrifugation at 300 x g for 5 min, washed with DPBS (ThermoFisher Scientific, Ottawa, ON), resuspended in fresh media, and incubated at 37°C with 5% CO₂ for a 3 h recovery period. Based on preliminary data select compounds were only tested in one metabolic condition. Following incubation, cells were collected, washed, and resuspended in 1X TempO-Seq™ Enhanced Lysis buffer (BioSpyder Technologies, Carlsbad, CA, United States) in DPBS (ThermoFisher Scientific, Ottawa, ON) (40 µL/well) and frozen at −80°C.

The TempO-Seq™ Human Surrogate + Tox Panel (S1500+) v2.0 (BioSpyder Technologies, Carlsbad, CA, United States) assay was completed following the manufacturer’s instructions in a 96-well plate format. Each 96-well plate included three assay controls in duplicate: a negative no-cell lysate (1X TempO-Seq™ Enhanced Lysis buffer only), Human Reference Total RNA (Takara Bio, CA, United States), and Human Brain Reference Total RNA (Takara Bio, CA, United States). Briefly, for each treatment, 2 µL of cell lysate in 1X TempO-Seq™ Enhanced Lysis buffer was hybridized to the Human S1500 + Surrogate detector oligo (DO) probe mix (V2.0), incubated at 70°C for 10 min, then ramped down to 45°C over 50 min (0.5°C/min). A nuclease digestion followed to remove excess or incorrectly bound DOs at 37°C for 1 h. The bound DO pairs were then ligated together at 37°C for 1 h and 15 min at 80°C to generate templates for amplification. A 10 µL aliquot of amplification template was transferred to its respective well in the TempO-Seq™ PCR Pre-Mix and Primers plate and amplified on the CFX96 thermocycler (Bio-Rad, Mississauga, ON, Canada) with the following program: 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; 68°C for 2 min. For library building and purification, three 96-well plates (288 sample libraries) were pooled together (5 µL per sample) and purified using the Macherey-Nagel NucleoSpin^® Gel and PCR Cleanup Kit (Clontech Laboratories Inc., Bethlehem, PA, United States) with the adjustments specified by the TempO-Seq™ Assay User Guide. The pooled purified TempO-Seq™ libraries were then diluted, quantified, and assessed for quality using the Agilent High Sensitivity D1000 TapeStation (Agilent Technologies, Santa Clara, CA, United States) and the qPCR KAPA Library Quantification Kit (Universal qPCR Master Mix) for Illumina NextSeq 500. Samples were sequenced on a total of two NextSeq^® 500/550 High Output (75-cycle) flow cells using an Illumina NextSeq^® 500 Sequencing Platform (Illumina, San Diego, CA, United States). A separate pool was completed for a fourth plate and an additional 96 sample libraries were added to the sequencing data for subsequent analysis.

Sequencing data are accessible in the National Centre for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO) database under accession number GSE213454. Briefly, bcl2fastq (v2.20.0.422) was used to demultiplex the raw sequencing data and assign each data set to its respective sample files. The data were filtered using fastp (v0.20.0) to eliminate reads shorter than 50 bp and remove low-quality sequences. The BioSpyder TempO-SeqR v3.0 analysis script, implemented as part of our transcriptomics data processing pipeline (https://github.com/R-ODAF/R-ODAF_Health_Canada), was used to align the FASTQ files for each sample to the reference sequences for the TempO-Seq™ Human Surrogate + Tox Panel (S1500+) v2.0 probes producing a table of counts per probe per sample. A study-wide quality control analysis workflow adapted from Harrill et al. (2021) was applied to the count matrix; 7/276 samples (i.e., S1_pool2, S2_pool2, S31, S79, S95_pool2, S109, S119) were removed due to low quality.

The count matrix underwent log2(CPM +1) normalization to account for read-depth variability between samples. TGx-DDI genes with multiple probes were averaged. For each compound, sample replicates were averaged and the log2 fold changes were estimated for each concentration. Samples that were cytotoxic (viability <40%), or did not achieve a minimum read depth, were removed from the analysis.

Detailed information regarding the TGx-DDI classification statistical analyses has been described previously (Yauk et al., 2016; Buick et al., 2017; Buick et al., 2021). Briefly, to classify chemicals as DDI or non-DDI, a three-pronged statistical approach was applied. Each analysis compares the sample biomarker responses to those of training set compounds with known DDI and non-DDI mechanisms. The three analyses are listed below.

1) The Nearest Shrunken Centroids Probability Analysis (NSC-PA) method was conducted (Tibshirani et al., 2002) using the pamr() function in R. This summarized the training dataset by calculating a standardized centroid (SC) for the DDI and non-DDI chemicals in the training set. The SC represents the average expression for each gene in a class relative to its within-class standard deviation. To create the NSC, the SCs were shrunken in the direction of the overall centroid. To classify each concentration for each chemical, the expression profiles were compared to the training set NSCs (Li et al., 2017) and assigned to a class (DDI or non-DDI) based on the probability that class membership was >0.90 (else it was not classifiable). This analysis was visualized as a heatmap.

If a chemical had a positive call in one of the three statistical analyses (NSC-PA, PCA, 2-DC) at any concentration, the overall call assigned was DDI. Conversely, if none of the three analyses produced a positive call, the overall call assigned was non-DDI.

Cells were adjusted to a density of 2.0 x 10⁵/mL and cells were aliquoted in 96-well plates. Cells were exposed to 2 µL of solubilized test chemical (1% v/v, final volume 200 µL/well) in a ten-point concentration range on a half-log distribution scale in (Table 2). All chemicals were also tested in the presence of S9 applying the conditions described previously. Treated cells were incubated at 37°C with 5% CO₂ for 4H. At the 4H timepoint, aliquots of cells were collected for the MultiFlow^® DNA Damage assay (Litron Laboratories, Rochester, NY, United States). The remaining cells were collected via centrifugation at 300 x g for 5 min, washed with DPBS (ThermoFisher Scientific, Ottawa, ON), resuspended in fresh media, and incubated at 37°C with 5% CO₂. These cells were incubated for an additional 20H for the second time point of the MultiFlow^® DNA Damage assay and for the in vitro MicroFlow^® assay (Litron Laboratories, Rochester, NY, United States).

At 24-h post-exposure cells were collected via centrifugation at 300 x g for 6 min. The supernatant was carefully removed and freshly prepared Complete Nucleic Acid Dye A (50 µL/well) was added. The plate was then placed under a visible light source on ice for 30 min, 1X Buffer Solution (150 µL/well) was added, cells were collected via centrifugation at 300 x g for 6 min, and the supernatant was carefully removed. Cells were resuspended in Complete Lysis Solution 1 (100 µL/well), mixed thoroughly, and incubated at room temperature for 1 h. Completed Lysis Solution 2 was added (100 µL/well) and the plate was rocked gently to mix. The plate was then analyzed via flow cytometry with the Miltenyi Biotec MACSQuant^® Analyzer 10 flow cytometer with an integrated 96-well MiniSampler. The mixing and fluidics parameters were as follows: sample mixing of 50 μL at a medium mix rate, followed by a sample uptake volume of 50 µL that was analyzed at a medium flow rate. The fast mode was used to rinse the probe between samples. Instrument settings followed instructions stated in the MicroFlow^® MicroNucleus Analysis kit (In Vitro, 96 well) (Litron Laboratories, Rochester, NY) to detect fluorochromes SYTOX Green^® in the FITC channel and ethidium monoazide (EMA), in the PerCP-Cy5.5 channel. An analysis stop gate of 5,000 EMA-negative nuclei per well was applied.

The in vitro MicroFlow^® results were analyzed using the analysis template provided by Litron Laboratories (Rochester, NY). The % MN was calculated using the count of MN events relative to the count of nucleated events. Results for graphical and statistical representations were expressed as a relative fold-change from the average plate-specific DMSO solvent control normalized to 1. %   M N = M N   E v e n t s N u c l e a t e d   E v e n t s   x   100

Cytotoxicity was determined using the Relative Survival (RS) equation. Treatments with RS values less than 40% and %EMA-positive nuclei values greater than 4-fold over solvent controls were removed from analysis. R S = N u c l e i   i n   t r e a t e d   c u l t u r e / m L N u c l e i   i n   m e a n   s o l v e n t   c o n t r o l / m L   x   100 F o l d   E M A = %   A p o p t o t i c / N e c r o t i c M e a n   %   A p o p t o t i c / N e c r o t i c   o f   s o l v e n t   c o n t r o l s   x   100

A positive call was indicated by a 2.50-fold increase in %MN and a statistically significant (p < 0.05) %MN increase relative to solvent controls in at least one non-cytotoxic concentration (Avlasevich et al., 2021). Statistical significance was determined using a Poisson regression with a Holm-Sidak multiple correction procedure.

At 4H and 24H sampling times, fresh complete labeling solution was prepared; solution containing nuclei release solution, DNA stain, RNase solution, ɣH2AX Alex Fluor^® 647, phosphor-histone antibody PE, and p53 antibody conjugated to fluorescein isothiocyanate (FITC) (Litron Laboratories, Rochester, NY) was added (50 µL/well) to a fresh 96-well plate. Treated cells were gently resuspended and 25 µL/well was mixed thoroughly with the labeling solution. The plate was incubated for 30 min at room temperature shielded from light. The cells were then analyzed via flow cytometry with the Miltenyi Biotec MACSQuant^® Analyzer 10 flow cytometer with an integrated 96-well MiniSampler. Instrument settings followed the MultiFlow^® Smart Start Guide (Litron Laboratories, Rochester, NY) to detect fluorescence emissions from fluorochromes FITC (in the B1 channel), PE (in the B2 channel), propidium iodide (in the B3 channel) and Alexa Fluor^® 647 (in the R1 channel). The mixing and fluidics parameters were as follows: sample mixing of 40 μL at a medium mix rate, followed by a sample uptake volume of 20 μL, which was analyzed at a medium flow rate. The fast mode was used to rinse the probe between samples (Bernacki et al., 2016; Bryce et al., 2016).

The MultiFlow ^® DNA Damage assay results were analyzed using the analysis template provided by Litron Laboratories (Rochester, NY). The γH2AX and p53 endpoints were measured based on their median fluorescence intensity (i.e., mean channel fluorescence or MCF) relative to the DMSO solvent control. The p-H3 and polyploidy endpoints were measured based on their frequencies among all events (i.e., all 2n–4n and polyploidy (8n) DNA content). The equations for calculating γH2AX shift, nuclear p53 shift, % polyploidy, and % p-H3 positive events are shown below. Results for graphical and statistical representations were expressed as a relative fold-change from the average plate-specific DMSO solvent control normalized to 1 (Bernacki et al., 2016; Bryce et al., 2016). γ H 2 A X   s h i f t = M C F   o f   t r e a t e d   c u l t u r e M e a n   M C F   o f   s o l v e n t   c o n t r o l N u c l e a r   p 53   s h i f t = M C F   o f   t r e a t e d   c u l t u r e M e a n   M C F   o f   s o l v e n t   c o n t r o l %   P o l y p l o i d y = #   e v e n t s   p o l y p l o i d y   8 n   D N A   c o n t e n t #   e v e n t s   2 n ,   4 n   a n d   p o l y p l o i d y   D N A   c o n t e n t     x   100 %   p H 3   e v e n t s = #   e v e n t s   p H 3   p o s i t i v e   4 n   a n d   g r e a t e r   D N A   c o n t e n t #   e v e n t s   2 n ,   4 n   a n d   p o l y p l o i d y   D N A   c o n t e n t     x   100

Latex microspheres were used as counting beads to determine nuclear density. Cytotoxicity was determined using the Relative Nuclei Count (RNC) equation, with cytotoxicity = 100%—RNC at 24H (Bernacki et al., 2016; Bryce et al., 2016): R N C = D e n s i t y   o f   n u c l e i   i n   t r e a t e d   c u l t u r e D e n s i t y   o f   n u c l e i   i n   m e a n   s o v e n t   c o n t r o l   x   100

Litron developed Global Evaluation Factors (GEF) to identify a biologically significant increase in select biomarker levels. The determined GEFs, expressed as fold-change (FC) increase over solvent control, are as follows (Bryce et al., 2017).

• For clastogenic biomarkers: 4H γH2AX, 1.51 FC; 24H γH2AX, 2.11 FC; 4H p53, 1.40 FC; 24H p53, 1.45 FC.

A clastogenic or aneugenic call required two successive concentrations that met or exceeded the GEFs in at least 2 of the 4 respective MoA biomarkers. Chemicals that met or exceeded two GEFs in a single non-cytotoxic concentration were designated as “weak” responses. In cases where both clastogen and aneugen biomarkers exceeded the GEF the MoA was considered pan-genotoxic. In cases where less than two biomarkers met the GEFs, the call was non-genotoxic. Exposure concentrations that resulted in greater than 80% cytotoxicity were excluded from MoA classification.

Gene expression data were used for BMC modeling as described by Buick et al. (2021). Briefly, the read counts were log₂ normalized. Groups where N = 1 were removed and filtered to exclude probes with fewer than 5 reads. The log₂ transformed data were run with the BMDExpress V2.3 software following the guidelines outlined by the US National Toxicology Program (NTP) Approach to Genomic Dose-Response Modeling report (USNTP, 2018). The log2 data were prefiltered using a Williams Trend Test applying a permutation p-value cutoff <0.05 with 500 permutations and linear fold change of ≥1.5. The pre-filtered data were analyzed using EPA BMDS parametric models to derive BMCs. This included: the Exponential 2, 3, 4, and 5 models, Linear, 2° Polynomial, and the Restricted Power (≥1). The benchmark response factor (BMR) was set to 1 standard deviation. The “Best BMC” was selected based on the best-fit model [i.e., model with the lowest Akaike Information Criterion (AIC)]. The upper (BMCU) and lower (BMCL) 95% confidence limits of the BMCs were calculated. The calculated gene BMCs were filtered and removed if: the model fit was insufficient (p-value <0.1), BMC/BMCL ratio ≥20, BMCU/BMC ≥20, or BMCU/BMCL ≥40. Applying a bootstrapping method described in detail by Buick et al. (2021), the median BMC of the TGx-DDI biomarker genes was derived to represent an overall TGx-DDI BMC for each compound. The 95% confidence intervals were calculated from the bootstrapped distribution of the BMC median.

BMC modeling was done on test chemicals with positive MultiFlow^® and/or MicroFlow^® calls. The modeling was performed with the PROAST R package (V70.3) developed by the RIVM (i.e., Dutch National Institute for Public Health and the Environment). To model the continuous concentration-response data, a single 5-parameter exponential model (y = a*[c^{(1−exp(−x/b)^d)}]) was applied (White et al., 2020). For the MicroFlow^® assay, the % induction of MN relative to test chemical concentration was modeled; a BMR of 1.0 (i.e., 100% increase in response relative to control) was applied (Long et al., 2018; White et al., 2020; Avlasevich et al., 2021). For the MultiFlow assay, the fold change in response for each biomarker (e.g., 4H and 24H p53; 4H and 24H ɣH2AX) relative to test chemical concentration was modeled. A BMR of 0.5 (i.e., 50% increase) was applied (Avlasevich et al., 2021). The lowest BMC of the MultiFlow biomarkers was selected as the overall BMC.

The calculated endpoint BMCs were integrated and converted into a single visual profile for each compound using the Toxicological Prioritization Index (ToxPi) software (v2.3) (Reif et al., 2013; Marvel et al., 2018). The BMC, BMCL, and BMCU values for each compound were formatted in Excel using the templates provided by ToxPi and saved as a comma-separated value (csv) file. A value of 10,000 µM (i.e., 10 mM) was entered when an infinite upper confidence interval or a BMC could not be calculated for a particular endpoint; or if a calculated BMC value was greater than the actual top passing/valid concentration. Differential weighting was assigned to the slices. TGx-DDI and MN were assigned 1/3 of the profile each; the four p53 and ɣH2AX timepoints were assigned 1/12 of the profile each to make up 1/3. The BMCL and BMCU values were assigned to their respective slices and were −log₁₀ transformed. The ToxPi analysis was run to derive ToxPi scores. First, the BMCL and BMCU values for each endpoint were summed and transformed into a single slice score for each compound. This was achieved by normalizing the summed BMC values to a [0, 1] interval by dividing by the slice maximum (USNRC, 2014). Thus, in this case, values closer to 1 (i.e., maximum unit score) denoted a higher potency. Alternatively, values closer to 0 (closer to the origin) denoted a lower potency; and values that did not extend from the origin were denoted as inactive. The overall ToxPi score was calculated by combining all weighted slice scores for each compound and the 95% confidence intervals were calculated. The ToxPi hierarchical clustering algorithm was applied to cluster like-chemical ToxPi profiles. ToxPi analysis figures (i.e., profiles, rank plot, and hierarchical clustering) were downloaded.

To classify test compounds as DDI or non-DDI using the TGx-DDI biomarker, transcriptional profiles were generated using TempO-Seq™ S1500 + sequencing. A trio of independent statistical analyses, 2-DC, PCA, and NSC-PA, were used to derive the overall TGx-DDI classifications (Figure 1; Supplementary Figures S2, S3). To ensure a conservative assessment of genotoxic potential, if one of the analyses classified a test compound as DDI, the overall call given was DDI. The biomarker correctly classified the three control chemicals: DMBA and EPEG were classified as DDI; the negative control DMANN was classified as non-DDI.

Eight out of ten NSACB compounds were identified as DDI, with varying potencies. NSACB compounds #5 and #7 showed the strongest responses. The trio of statistical analyses identified NSACB #5 as DDI across all concentrations (2-DC was inconclusive at the highest four concentrations); and NSACB #7 was DDI across all concentrations except the lowest. The strength of this response can be visualized in Figure 1, i.e., when comparing the similarity of the heatmaps of NSACB #5 and #7 to those DDI compounds in the training set. NSACB #9 had a moderate DDI response; the 2-DC statistical analysis identified #9 as DDI for the majority of concentrations, but the NSC-PA and PCA analyses only identified this compound as DDI at the highest non-cytotoxic concentration. NSACB compounds #1, 2, 3, 4, and 8 were also classified as DDI. However, these were considered weaker DDI responses as these compounds were only positive at the highest non-cytotoxic concentrations and the DDI calls were not consistent across the statistical analyses (with the exception of #3). NSACB compounds #6 and 10 were classified as non-DDI; all three analyses yielded non-DDI calls across all concentrations.

Chromosome damage was assessed using the in vitro MicroFlow^® assay for the test chemicals (Figure 2, Supplementary Figure S4). A positive result was designated by at least one concentration that induced >2.5 fold-increase in %MN and was statistically significant (p < 0.05) compared to vehicle controls. Both positive controls, EPEG and DMBA, exhibited significant fold-increases in %MN, whereas this was not observed with the negative control DMANN (Supplementary Figure S4).

All ten NSACB chemicals met the classification criteria for a positive call in at least one S9 condition (i.e., with/without S9) (Figure 2; Supplementary Figure S4). The majority were positive without S9, but #8 and #10 only tested positive with S9 conditions. NSACB #1, 5, 6, 7, 8, and 9 showed a robust fold-increase in %MN with a clear concentration-response relationship. NSACB #2 and #3 displayed variable responses with a significant increase observed only at the highest concentration. NSACB #4 and #10 exhibited a concentration-response pattern at the lowest concentrations before decreasing or plateauing at the higher concentrations.

To provide insight into the MoA, the test compounds were assessed with the MultiFlow^® assay. The assay data were visualized using radar plots that depict the fold changes in the MultiFlow^® biomarkers following chemical exposure (Figure 3, Supplementary Figure S5). If a test chemical induced a significant increase in two out of four clastogen-specific biomarkers (i.e., 4H and 24H p53, and ɣH2AX), a clastogenic MoA was predicted. Similarly, significant responses in two out of four aneugen-specific biomarkers (4H p-H3, 24H p53, p-H3, and polyploidy) predicted an aneugenic MoA. If both of these criteria were met, a pan-genotoxic MoA was predicted. If none of these criteria were met the chemical was classified as non-genotoxic.

The control chemicals produced the expected results (Supplementary Figure S5). Positive controls, EPEG and DMBA, had concentration-dependent increases in the clastogen biomarkers (EPEG: 4H p53, 24H p53, and 4H ɣH2AX; DMBA: 4H & 24H ɣH2AX).

In total, six of the ten NSACB compounds were classified as genotoxic using the MultiFlow^® assay. NSACB compounds #2, 3, 5, 7, and 9 exhibited concentration-dependent increases in MultiFlow^® clastogen-specific biomarkers and were therefore classified as clastogens (Figure 3). NSACB #3, 5, and 7 induced responses in 4H p53 and 4H ɣH2AX; NSACB #2 induced responses in 4H ɣH2AX and 24H p53; and NSACB #9 induced responses in 4H and 24H ɣH2AX. NSACB #2, 5, and 9 had the most robust increases, whereas NSACB #3 and 7 had more moderate responses. NSACB #4 was the only test chemical classified as a pan-genotoxicant. This compound displayed significant increases in clastogen biomarkers 4H p53 and 4H ɣH2AX, as well as aneugen marker 4H p-H3 and pan-genotoxic biomarker 24H p53. The remaining compounds, NSACB #1, 6, 8, and 10 were classified as non-genotoxic. Three of these compounds (#1, 6, and 10) had a moderate to robust response increase in one of the aneugenic biomarkers (i.e., 4H p-H3) at more than one concentration. However, without an increase in any other biomarker, an aneugenic call could not be made.

An overview of the TGx-DDI, MicroFlow^®, and MultiFlow^® classification calls for the test chemicals is depicted in Table 3. The control chemicals performed as expected. EPEG and DMBA were positive across all three assays; DMANN tested negative in all three assays. For the data-poor compounds, eight out of ten (NSACB #1, 2, 3, 4, 5, 7, 8 and 9) were classified as DDI by the TGx-DDI transcriptomic biomarker in at least one S9 condition. In contrast, all ten NSACB compounds were positive for MN induction in at least one S9 condition. Using MultiFlow^®, five were classified as clastogens and one was classified as a pan-genotoxicant in at least one S9 condition.

Overall, six compounds (i.e., NSACB #2, 3, 4, 5, 7, and 9) were positive in all three assays. NSACB #1 and #8 were positive in the TGx-DDI and MicroFlow^® assays, but were classified as non-genotoxicants by the MultiFlow^® assay. The remaining compounds, NSACB #6 and #10, only tested positive in the MicroFlow^® assay, and were classified as non-DDI and non-genotoxicants by the TGx-DDI and MultiFlow^® assays, respectively.

In addition to hazard calls, quantitative analyses of genotoxicity data can be applied for chemical potency ranking and subsequent prioritization. To assess the relative potency of the NSACB compounds with positive hazard flags, we conducted BMC modeling to derive BMC values for the TGx-DDI biomarker gene set, and the MicroFlow^® and MultiFlow^® assay endpoints (Supplementary Figure S6). To evaluate the consistency of the potency rankings derived from each assay, the NSACB compounds with concordant positive hazard calls (i.e., #2, 3, 4, 5, 7, and 9) were compared (Figures 4A–C). The most to least potent chemicals based on the TGx-DDI BMC ranking was (Figure 4A): EPEG (lowest median gene BMC) > DMBA > NSACB 7 > NSACB 4 > NSACB 2 > NSACB 3 > NSACB 9 > NSACB 5 (greatest median gene BMC). Two main groupings for the NSACB compounds can be observed. The first group, NSACB #7, 4, 3, and 2, have confidence intervals (CIs) that overlap and are thus are not significantly different from each other. The second group, NSACB #9 and #5, also have CIs that overlap, but are in a distinct, less potent, group from NSACB #7, 4, 3, and 2. MicroFlow^® BMC analysis yielded a highly similar potency ranking (Figure 4B): EPEG > DMBA > NSACB 7 > NSACB 3 > NSACB 4 > NSACB 2 > NSACB 9 > NSACB 5. However, due to the overlapping confidence intervals, all the data-poor compounds are not significantly different from each other. Finally, MultiFlow^® BMCs were evaluated and the lowest BMC of the clastogen biomarkers was selected as the point of departure (Figure 4C). This analysis revealed the following potency ranking: EPEG > DMBA > NSACB 7 > NSACB 3 > NSACB 4 > NSACB 2 > NSACB 9 > NSACB 5. As with TGx-DDI, two distinct groups can be observed for the data-poor compounds. 1) NSACB #7, 3, 4, and 2 make up one group with overlapping CIs, and 2) NSABC #9 and 5, also with overlapping CIs, make up a distinct, less potent, group. Overall, BMC modeling of the endpoints in the proposed in vitro testing strategy yielded nearly identical potency rankings for the genotoxic NSACB compounds.

The ToxPi software was used to integrate all endpoint BMC metrics into a single score; the scores were subsequently used to rank the NSACB compounds and visualize mechanistic information regarding their genotoxic hazards (Figure 5; Supplementary Figure S7). With respect to the latter, the analyses generated toxicological profile graphics for each compound. The endpoints are represented by the coloured pie slices; the radius of each slice denotes the relative effect size expressed as −log10 BMC. Thus, a lower (i.e., more potent) BMC is indicated by greater protrusion from the origin. To compare to the independent endpoint BMC rankings displayed in Figure 4, the compounds with concordant positive hazard flags were first ranked based on their overall ToxPi scores for all endpoints (Figure 5A). The higher the overall ToxPi score, the more potent the compound. In this case, the most potent to least potent ranking was: EPEG > DMBA > NSACB 7 > NSACB 4 > NSACB 2 > NSACB 3 > NSACB 9 > NSACB 5. More specifically, based on the presented ToxPi profile (Figure 5A), EPEG is the most potent for five of the six endpoints examined: TGx-DDI, MN, 4H ɣH2AX, 24H p53 and 4H p53. In contrast, except for a small 4H p53 slice, NSACB #5 had almost no slices protruding from the center. The majority of the compounds, with the exception of NSACB #4, 2, and 9, lacked a 24H ɣH2AX slice. Comparisons of multiplexed ToxPi profiles (Figure 5B) showed that compounds #7, 3, 4, and 2 clustered together on a single branch; NSACB #9 and 5 clustered separately. This clustering is similar to the compound groupings observed in Figures 4A, C.

An additional ToxPi analysis was conducted for all NSACB compounds (Supplementary Figure S7), including those with discordant results (i.e., #1, 6, 8 and 10). In this case, the most potent to least potent ranking was: EPEG > DMBA > NSACB 7 > NSACB 4 > NSACB 1 > NSACB 2 > NSACB 3 > NSACB 8 > NSACB 9 > NSACB 6 > NSACB 5 > NSACB 10. The ToxPi profile of compound #1 clustered with #7, 3, 4 and 2, while the profiles of compounds #6, 8, and 10 clustered with #9 and 5.