Reconstructed human epidermis‐based testing strategy of skin sensitization potential and potency classification using epidermal sensitization assay and in silico data

The hazards and potency of skin sensitizers are traditionally determined using animal tests such as the local lymph node assay (LLNA); however, significant progress has been made in the development of non‐animal test methods addressing the first three mechanistic key events of adverse outcome pathway in skin sensitization. We developed the epidermal sensitization assay (EpiSensA), which is a reconstructed human epidermis‐based assay, by measuring four genes related to critical keratinocyte responses during skin sensitization. Four in vitro skin sensitization test methods (EpiSensA, direct peptide reactivity assay [DPRA], KeratinoSens™, and human cell line activation test [h‐CLAT]) were systematically evaluated using 136 chemicals including lipophilic chemicals and pre/pro‐haptens, which may be related to assay‐specific limitations. The constructed database included existing and newly generated data. The EpiSensA showed a broader applicability domain and predicted the hazards with 82.4% and 78.8% accuracy than LLNA and human data. The EpiSensA could detect 76 out of 88 sensitizers at lower concentrations than the LLNA, indicating that the EpiSensA has higher sensitivity for the detection of minor sensitizing constituents. These results confirmed the potential use of the EpiSensA in evaluating a mixture of unknown compositions that can be evaluated by animal tests. To combine different information sources, the reconstructed human epidermis‐based testing strategy (RTS) was developed based on weighted multiple information from the EpiSensA and TImes MEtabolism Simulator platform for predicting Skin Sensitization (TIMES‐SS; RTSv1) or Organization for Economic Cooperation and Development (OECD) QSAR Toolbox automated workflow (RTSv2). The predictivities of the hazards and Globally Harmonized System (GHS) subcategories were equal to or better than the defined approaches (2 out of 3, integrated testing strategy [ITS]v1, and ITSv2) adopted as OECD Guideline 497.

These results confirmed the potential use of the EpiSensA in evaluating a mixture of unknown compositions that can be evaluated by animal tests.To combine different information sources, the reconstructed human epidermis-based testing strategy (RTS) was developed based on weighted multiple information from the EpiSensA and TImes MEtabolism Simulator platform for predicting Skin Sensitization (TIMES-SS; RTSv1) or Organization for Economic Cooperation and Development (OECD) QSAR Toolbox automated workflow (RTSv2).The predictivities of the hazards and Globally Harmonized System (GHS) subcategories were equal to or better than the defined approaches (2 out of 3, integrated testing strategy [ITS]v1, and ITSv2) adopted as OECD Guideline 497.

| INTRODUCTION
Allergic contact dermatitis is a dermal disease resulting from skin sensitization to causative chemicals.Skin sensitization potential is traditionally assessed using animal tests, such as the murine local lymph node assay (LLNA) (Organization for Economic Cooperation and Development [OECD], 2010).Conversely, the development of nonanimal test methods for identifying skin sensitization has recently been stimulated because of animal welfare and regulatory requirements that ban animal testing to predict this endpoint (Environmental Protection Agency, 2018;European Union, 2009).
The skin sensitization process has been described in the report published by the OECD on "The Adverse Outcome Pathway (AOP) for Skin Sensitization Initiated by Covalent Binding to Proteins" (OECD, 2012).Briefly, the process includes four key events (KEs): protein binding of small reactive chemicals (haptens; KE1); inflammatory response and induction of cytoprotective gene pathways in keratinocytes (KE2); activation and induction of surface molecules in dendritic cells (DCs; KE3); and activation, differentiation, and proliferation of naive T cells (KE4).To assess the skin sensitization potential of various chemicals, several kinds of in chemico and in vitro assays focusing on these KEs have been developed.Recently, the direct peptide reactivity assay (DPRA), the amino acid derivative reactivity assay, and the kinetic direct peptide reactivity assay (KE1), KeratinoSens™ and LuSens (KE2), and human cell line activation test (h-CLAT), U-SENS™, IL-8 luc assay, and genomic allergen rapid detection (GARD™) and GARD™ skin (KE3) have been adopted as OECD test guidelines (OECD, 2022a(OECD, , 2022b(OECD, , 2022c)).
Currently, replacing assessment of sensitization with animals using an available single alternative method is not possible.Thus, testing strategies, known as defined approaches (DAs), combining different information from multiple AOP-based assays have been developed and evaluated (OECD, 2016a(OECD, , 2016b)).The DAs provide information on hazards and potency (strong, weak sensitizer, and non-sensitizer [NS]) to allow for the United Nations Globally Harmonized System (UN GHS) classification and labeling (GHS, 2017); however, some limitations remain because of the applicability domain for each assay being used.For example, cell-based assays are considered less predictive for highly lipophilic chemicals because of their aqueous-phase test system.When a test chemical is very lipophilic and fails to form a stable dispersion, the testing cannot be conducted.Moreover, the assays have a weakness in detecting indirectly acting sensitizers known as pre-and prohaptens, which need to undergo abiological oxidation and metabolic conversion to become sensitizing haptens (Aptula et al., 2007).
Therefore, the chemical space on which each assay is applicable needs to be thoroughly expanded to ultimately provide predictions with high confidence using the testing strategies.Thus, we developed the epidermal sensitization assay (EpiSensA) using a reconstructed human epidermis (RhE) model that mimics skin structure, has a differentiated epidermis and stratum corneum, allows topical application, and incorporates metabolic systems.The EpiSensA is an in vitro test method based on the commercial human skin model LabCyte EPI-MODEL 24 in which gene expressions are measured of four mechanistically relevant markers, namely, encoding activating transcription factor 3 (ATF3); glutamate-cysteine ligase, modifier subunit (GCLM); DnaJ (Hsp40) homolog, subfamily B, member 4 (DNAJB4); and interleukin-8 (IL-8).
The expression of these genes addresses KE2 (keratinocyte responses) in skin sensitization AOP (OECD, 2012).Moreover, the EpiSensA is also considered to cover KE1, which is a covalent modification of epidermal proteins.For a panel of 72 test chemicals, the EpiSensA provided 94% sensitivity and 90% accuracy compared with LLNA results (Saito et al., 2017).Moreover, the EpiSensA was formally validated by the Japanese Center for the Validation of Alternative Methods for assessing skin sensitization hazards and then demonstrated sufficient transferability and reproducibility (OECD, 2023).
This study aimed to evaluate the predictive performance of hazard (sensitizer/NS) and potency categories for the EpiSensA using a large, partly newly generated dataset of 136 chemicals based on the availability of LLNA and human skin sensitization data.This evaluation also provided insight into the applicability and predictivity of challenging chemicals that are deemed to be difficult to test with existing validated alternative methods, such as lipophilic chemicals and pre/pro-haptens.Moreover, we directly com-

| Prediction model for hazard identification and potency identification
The mean value (three tissue units per concentration) of maximum fold induction (I max ) was obtained using data from concentrations with >80% cell viability.When the mean I max of at least one out of the four marker genes exceeded the respective cut-off value (ATF3, >15-fold; GCLM, >2-fold; DNAJB4, >2-fold; and IL-8, >4-fold), the chemical was judged as positive.When the mean I max of all four marker genes does not exceed the respective cut-off values, and if at least one mean cell viability at the tested concentrations is <80% or if the test chemical does not indicate <80% mean cell viability at the highest soluble concentration, the chemical was judged as negative.
Data were accepted when the following criteria were fulfilled:

| DPRA
DPRA was adopted in the OECD Testing Guideline TG 442C in 2015 (OECD, 2022a).Briefly, the assay uses cysteine peptide (Ac-RFAACAACOOH) and lysine peptide (Ac-RFAAKAA-COOH), and the binding to a test chemical was measured by high-performance liquid chromatography with UV detection at 220 nm by determining the concentration of free peptide that is available after incubating the test chemical for 24 h at 25 C.As a cut-off value, the mean cysteine and lysine peptide percent depletion value of 6.38% was used to classify test chemicals as positive.Of the 136 chemicals including 28 newly tested ones, zinc mercaptobenzothiazole with logKow = 5.02 was not applicable in this assay because of solubility issues.DPRA data on the remaining 135 chemicals are presented here.

| KeratinoSens™
KeratinoSens™ was adopted in the OECD Testing Guideline TG

| h-CLAT
h-CLAT was adopted in the OECD Testing Guideline TG 442E in 2016 (OECD, 2022c

| TIMES-SS
TIMES-SS (v.2.31.2) was developed by the OASIS Laboratory of Mathematical Chemistry, Bourgas, Bulgaria, and was run using the skin sensitization metabolism-activated toxicity model (Patlewicz et al., 2014).Thus, the model evaluated the skin sensitization potential of the parent compound and its potential metabolites and estimates skin sensitization potency classes (NS, weak/NS, weak, and strong).
The TIMES-SS can also provide a prediction for a chemical regardless of logP.

| Derek Nexus
Derek Nexus (v.6.2.0 using DEREK Knowledge Base 2020 1.0) from Lhasa Limited is an expert knowledge-based system for toxicity predictions containing structural alerts derived from both public and proprietary data and was used to make in silico assessment of the skin sensitizing potential of a chemical (Macmillan et al., 2016).Derek mainly addresses structural features and predicts lipophilic chemicals and pre/pro-haptens.
2.8 | OECD QSAR Toolbox v4.5 The OECD QSAR Toolbox Skin sensitization automated workflow for DASS is a software application for assessing the chemical hazards (https://qsartoolbox.org/).In addition, the automated workflow for EC3 from LLNA predicted an EC3 value of a target chemical by readacross prediction.

| RTSv1 and RTSv2
RTS is the DA that requires weighted multiple information from the EpiSensA and TIMES-SS (RTSv1) or OECD automated workflow (RTSv2).For the EpiSensA, the Min EC values were converted to a score of 3 or 1 based on 0.098 w/v%, and the negative outcome was assigned to 0. For the in silico prediction, the potency class prediction was converted to 2 (strong), 1 (weak or weak/NS), or 0 (NS).Then, the summed score, ranging from 0 to 5, can be used to provide hazard identification and potency category information according to UN GHS Cat.1A (4-5), Cat.1B (2-3), and NC (0-1) (Table 1).

| 2 out of 3
The 2 out of 3 requires information from the DPRA, KeratinoSens™, and h-CLAT and can be used to evaluate the skin sensitization hazard potential of chemicals.The overall result relies on two concordant results.

| Data analysis
The hazard predictive performance for each assay was calculated using the standard prediction model.Sensitivity is the rate that sensitizers predicted as positive.Specificity is the rate that NS predicted as negative.Accuracy is the overall rate of correct predictions.Balanced accuracy is the mean of sensitivity and specificity.
The prediction of potency identification was also calculated based on 3 Â 3 contingency table counting, and the concordance of potency prediction to GHS subcategorization and human potency classification was analyzed.The GHS subcategorization of sensitizers into Classes  92.7% accuracy.The EpiSensA showed 89.1% accuracy, and its performance was greater than or similar to those of other test methods, with accuracy ranging from 76.4% to 81.8%.For pre/pro-haptens (n = 23), 100% sensitivities were observed for the EpiSensA and LLNA.In the overall 80 chemicals, the EpiSensA showed similar predictive performance to the LLNA and h-CLAT, whereas the DPRA and KeratinoSens™ had lower sensitivity and accuracy but higher specificity.

| Contribution of respective marker genes in the EpiSensA on hazard prediction
In the EpiSensA, information on the expression of four marker genes involved in the antioxidant and inflammatory responses is necessary to identify skin sensitizers with high confidence.However, this does not imply that a sensitizer induces the expressions of all four genes.As presented in Table 4, some of the sensitizers upregulated the expression in only one of the four marker genes.

| Sensitivity for detection between the EpiSensA and LLNA
The EpiSensA allows the direct application of undiluted test chemicals onto the surface of the RhE.This testing condition is very close to that of animal tests such as the LLNA.For 89 sensitizers classified as

| DISCUSSION
Existing dataset curated for the validated alternative methods is highly biased to hydrophilic chemicals that are amenable to these test methods.This is a well-recognized limitation that may arise from the aqueous condition in the test system.In addition, a limited metabolic capability of test methods is one of the technical issues because some of sensitizers may be pro-haptens (Aptula et al., 2007) Lu et al., 2014).DNAJB4 suppresses protein misfolding by oxidative stress as chaperone (Pesce et al., 2016).The induction of the DNAJB4 gene is regulated by both the Nrf2-ARE pathway and heat shock transcription factor 1 (HSF-1)/heat shock factor response element pathway (Satoh et al., 2011).IL-8 serves as a potent chemotactic peptide for neutrophils, which are critically involved in both the sensitization and elicitation phases of contact hypersensitivity (Weber et al., 2015).Sensitizers induce adenosine triphosphate (ATP) release from keratinocytes that express ATP receptor P2X 7 in response to ATP (Inoue et al., 2005;Onami et al., 2014).The induction of the IL-8 T A B L E 6 Hazard identification and potency categorization performance of LLNA, RTSv1, RTSv2, 2 out of 3, ITSv1, and ITSv2 in comparison with human data.gene is regulated by the ATP/P2X 7 pathway (Montreekachon et al., 2011;Qiu et al., 2014) and the p38 MAPK pathway (Mitjans et al., 2008).ATF3 functions as a hub of cellular adaptive-response network by negatively modulating inflammatory cytokines and chemokines (Hai et al., 2010).The induction of the ATF3 gene is regulated by ATP (Ohara et al., 2010) and nuclear factor kappa B (NF-κB) (Hai et al., 2010).The upregulation of DNAJB4 and GCLM, or ATF3 and IL-8, in human keratinocytes by a skin sensitizer was partly controlled by Nrf2 or P2X 7 (Saito et al., 2017).Thus, the four marker genes of the EpiSensA could be mechanistically relevant for skin sensitization.
The combination of four genes related to inflammatory (ATF3 and IL-8) and cytoprotective (DNAJB4 and GCLM) responses provided better sensitivity and accuracy than each single marker (Saito et al., 2017).
Moreover, sensitizers with only one positive marker gene are listed in Table 3.Each marker gene included strong sensitizers in the LLNA with EC3 < 1% or lipophilic sensitizers with logKow > 3.5, indicating that a sensitizer may independently interact with the respective marker genes and lead to discordant results.Thus, the combination of these four marker genes in the EpiSensA is critical to detect various sensitizers.
The EpiSensA successfully assesses the sensitizing potential of complex mixtures such as silicone-based compounds, surfactant mixtures, and crop-protection formulations (Mizumachi et al., 2021) by the RTS when compared with the LLNA.However, this LLNA result was reported to be probably a false positive, as confirmed by the human repeat insult patch test and human maximization test results (Natsch et al., 2023;Urbisch et al., 2015).
Furthermore, RTSv1 and RTSv2 are more predictive of human sensitization data rather than the LLNA for both hazards and potency.
RTSv1 and RTSv2 predicted human hazards with 84.1% and 78.8% balanced accuracy, respectively, whereas the LLNA showed 75.4% balanced accuracy.The 2 out of 3, ITSv1, and ITSv2 had balanced accuracy ranging from 78.6% to 80.3% in predicting human hazards.
pared the estimated concentration of a chemical required to produce positive reactions in the EpiSensA and LLNA to determine which test method is more sensitive to detect minor sensitizing constituents.Finally, we developed the RhE-based testing strategy (RTS) combining the EpiSensA with TImes MEtabolism Simulator platform for predicting Skin Sensitization (TIMES-SS; RTSv1) or OECD QSAR Toolbox automated workflow (RTSv2), taking into account the broad applicability domain of the EpiSensA and in silico prediction.The score-based information transferred from the EpiSensA and in silico prediction data were used in the RTS to predict the hazard and potency (GHS subcategorization) of a given chemical and compared with the DAs (2 out of 3, integrated testing strategy [ITS]v1, and ITSv2) adopted by the OECD (2021) using the dataset of 136 chemicals.
(i) Each cell viability of at least two epidermises of the vehicle control should be ≥95%; (ii) the mean cell viability of both positive controls should be ≥80%; (iii) in the positive control of 0.78 w/v% clotrimazole, the mean values of fold induction for ATF3 and IL-8 should exceed the cut-off value; and (iv) in the positive control of 0.10 w/v % 4NBB, the mean values of fold induction for GCLM and DNAJB4 should exceed the cut-off value.For positive chemicals in the EpiSensA, the minimum estimated concentration (Min EC value) at which any of the four marker genes exceeded the respective cut-off was used for potency classification.A test chemical was classified as having strong or weak potency if the lowest Min EC value was at ≤0.098 or >0.098 w/v %, respectively.If negative at concentrations with >80% cell viability, a test chemical was considered an NS.The potency classification of the EpiSensA (strong, weak, and NS) was used to predict GSH subcategories (UN GHS Cat.1A, Cat.1B, and not classified [NC]).
442D in 2015(OECD, 2022b).Briefly, the assay is a reporter cellbased assay measuring luciferase gene induction as an indicator of the activation of the Kelch-like ECH-associated protein 1 (keap1)nuclear erythroid 2-related factor 2 (Nrf2)-antioxidant response element (ARE) pathway in HaCaT cells.Cells were incubated with the test chemical for 48 h, and luciferase induction was then measured by luminometer.The test chemicals were classified as positive if the luciferase gene induction showed a statistically significant increase greater than 1.5-fold over the vehicle control at a concentration of <1000 μM, with cell viability of >70%.When a negative result was obtained in a test with a maximal concentration of <1000 μM and no cytotoxicity was reached, the result should be considered inconclusive.Of the 136 chemicals, 5 with high log-Kow values such as clotrimazole, tocopherol, squalene, oxyfluorfen, and carbonic acid dioctyl ester were classified as inconclusive because of negative results even without cytotoxicity observed at the tested concentrations.Of the 136 chemicals including 29 newly tested ones, KeratinoSens™ data on 131 chemicals are presented here.
1A and 1B aims to discriminate between strong and weak sensitizers based on the LLNA EC3 threshold of 2%.Six classes of human sensitizing potency reported byBasketter et al. (2014) andApi et al. (2017) were split into three classes such as Categories 1 and 2 (Cats. 1 and 2), Categories 3 and 4 (Cats.3 and 4), and Categories 5 and 6 (Cats.5 and 6).Chemicals in human Cats. 1 and 2 were considered higher potency sensitizers, Cats. 3 and 4 as lower potency sensitizers, and Cats. 5 and 6 as NS.3 | RESULTS3.1 | Hazard prediction of the EpiSensA and individual non-animal test methods compared with LLNA and human dataThe predictive performances of the EpiSensA, DPRA, KeratinoSens™, and h-CLAT were evaluated for the ability to discriminate sensitizers from NS when compared with LLNA and human data.
and therefore require metabolic activation to produce a positive response.In this study, we compiled the dataset containing 136 chemicals (136 with LLNA data; 80 with human data), which included 69 lipophilic chemicals with logKow > 3.5 and 37 pre/pro-haptens.This unbiased dataset allows the meaningful comparative evaluation of four nonanimal test methods.For the DPRA, KeratinoSens™, and h-CLAT, obtaining conclusive predictions has been difficult in several chemicals under the conditions of the methods.The results on 24% (32/136) of the chemicals assessed were inconclusive for the h-CLAT, whereas all results of 136 chemicals were conclusive for the EpiSensA, indicating that the EpiSensA exhibited a wider applicability domain.The overall sensitivity and accuracy of the EpiSensA were clearly superior to those of DPRA and KeratinoSens™ for predicting sensitizers in LLNA and human data, whereas they were nearly comparable with those of the h-CLAT in terms of chemicals with conclusive predictions.Indeed, the overall sensitivity, specificity, and accuracy of the EpiSensA were proved to be very similar to those of the LLNA when compared with human data.In contrast to the majority of validated alternative methods for the detection of sensitizers, which are based on a single or a few of the markers, the EpiSensA measures the induced expression of four genes (GCLM, DNAJB4, ATF3, and IL-8) involved in antioxidant and inflammatory response.GCLM serves as key determinant of biosynthesis for glutathione (GSH), which can regulate redox signaling and antioxidant response (S.C.Lu, 2013).The induction of the GCLM gene is regulated by both Nrf2-ARE and activating protein 1 (AP-1) pathways (C.Y. Abbreviations: EpiSensA, epidermal sensitization assay; GHS, Globally Harmonized System; Inc., inconclusive; ITS, integrated testing strategy; LLNA, local lymph node assay; NC, not classified; RTS, reconstructed human epidermis-based testing strategy; TIMES, TImes MEtabolism Simulator platform.T A B L E 6 (Continued) RTSv1 and RTSv2 predicted three human potency classes with 72.4% and 68.0%accuracy, respectively, whereas the LLNA showed 65.8% accuracy.ITSv1 and ITSv2 predicted human potency classes with 65.2% and 62.3% accuracy.RTSv1 obtained the highest accuracy for human hazards and potency prediction, which relies on the EpiSensA and TIMES-SS prediction.The limitations of the RTS are related to the limitations of the EpiSensA and TIMES-SS or OECD QSAR Toolbox.First, the applicability of DA must be considered for mono-constituent substances and defined mixtures because the TIMES-SS and OECD QSAR Toolbox require a structure as input.The EpiSensA has been evaluated using mono-constituent substances rather than mixtures.Second, the TIMES-SS and OECD QSAR Toolbox are regularly updated, resulting in updated predictions that may influence the RTS output.Thus, new versions must be carefully tracked.In addition, the use of outof-domain in silico prediction might have a certain effect on the overall RTS outcome.Even if chemicals with out-of-domain prediction were considered inconclusive, the overall predictive performance of RTSv1 and RTSv2 for the LLNA and human data remained unchanged in this dataset (data not shown).However, more uncertainty may be associated with DA predictions when the DA outcomes are nonsensitizing compared with when the outcomes are sensitizing.Overall, the EpiSensA can be applied to various chemicals with varying functional groups and physicochemical properties and can provide high predictivity of skin sensitizing potential and potency class for chemicals that fall out of the applicability domain or show poor predictivity in validated alternative methods, for example, lipophilic chemicals and pre/pro-haptens.When TIMES-SS prediction is used in combination with the EpiSensA within the DA, the RTS can provide skin sensitization hazards or GHS subcategories and can be used as a part of integrated approaches to the testing and assessment of skin sensitization.AUTHOR CONTRIBUTIONS Hideyuki Mizumachi: Data curation; conceptualization; writingreview and editing.Sho Suzuki: Data curation; writing-review and editing.Megumi Sakuma: Data curation; writing-review and editing.Midori Natsui: Data curation.Noriyasu Imai: Data curation; writingreview and editing.Masaaki Miyazawa: Investigation; conceptualization; writing-original draft; supervision.
Schematic of the RTSv1 and RTSv2 defined approaches.EpiSensA, epidermal sensitization assay; LLNA, local lymph node assay; Min EC, minimum estimated concentration; NC, not classified; NS, non-sensitizer; OECD, Organization for Economic Cooperation and Development; RTS, reconstructed human epidermis-based testing strategy; TIMES-SS, TImes MEtabolism Simulator platform for predicting Skin Sensitization.
TableS1shows the expanded dataset of 136 chemicals including data obtained by the LLNA, human, EpiSensA, DPRA, KeratinoSens™, and h-CLAT.As shown in Table2, the predictive performance of each assay against the LLNA was quantitatively assessed for the maximal subset of chemicals with conclusive predictions out of 136 chemicals.Table2summarizes data on lipophilic chemicals (logKow > 3.5, n = 69), hydrophilic chemicals (logKow ≤ 3.5, n = 67), pre/pro-haptens (n = 37), and overall chemicals (n = 136).For lipophilic chemicals, the LLNA data accuracy ranged from 48.4% to 78.3%, being the highest for the EpiSensA.For hydrophilic chemicals, the accuracy with LLNA data ranged from 73.1% to 86.6%, being higher for the EpiSensA and h-CLAT than for the DPRA and KeratinoSens™.As negative h-CLAT sensitivity and accuracy than the other two test methods.These data indicated that the EpiSensA can obtain a higher sensitivity for various chemicals.The predictive performance of the LLNA and non-animal test methods for human data is shown in Table3.For lipophilic chemicals (logKow > 3.5, n = 25), the EpiSensA predicted human hazard with a T A B L E 1 sensitivity of 92.3%.Its performance was comparable with that of the LLNA and greater than those of the DPRA and KeratinoSens™.By contrast, the EpiSensA and LLNA showed quite low specificities of 16.7% and 0%.Thus, all false positive chemicals of the EpiSensA against human data were rated as positive in the LLNA.Regarding hydrophilic chemicals (logKow ≤ 3.5, n = 55), the LLNA showed Abbreviations: DPRA, direct peptide reactivity assay; EpiSensA, epidermal sensitization assay; h-CLAT, human cell line activation test; LLNA, local lymph node assay.T A B L E 3 Hazard identification performance of LLNA of each test method for analyzed chemicals in comparison with human data.Abbreviations: DPRA, direct peptide reactivity assay; EpiSensA, epidermal sensitization assay; h-CLAT, human cell line activation test; LLNA, local lymph node assay.
Note: I max in bold indicated values greater than the cut-off values for the marker gene.

Table 5 )
. For RTSv1, cobalt chloride of 136 chemicals was excluded from the assessment of this DA because of the inability to generate TIMES-SS prediction.The balanced accuracy of RTSv1 was 86.3% for NC) with accuracy rates of 74.1% and 70.8%, respectively.For the same set of chemicals, the 2 out of 3 could not provide conclusive results for 19/136 chemicals (14.0%) and correctly predicted LLNA hazards with 70.1% balanced accuracy.Likewise, ITSv1 and ITSv2 could not provide conclusive results for 20/136 (14.7%) and 22/136 chemicals (16.1%) and correctly predicted LLNA hazards with 75.3% and 78.0%balanced accuracy and the GHS subcategorization with 72.0% and 72.0%accuracy, respectively.Therefore, RTSv1 and RTSv2 demonstrated equivalent or superior performance to 2 out of 3, ITSv1, and ITSv2 in predicting LLNA hazard and GHS subcategorization.Hexyl salicylate (GHS Cat.1A) was mispredicted as an NS