Performance of the GHS Mixtures Equation for Predicting Acute Oral Toxicity

Abstract

Acute oral toxicity classifications are based on the estimated chemical dose causing lethality in 50% of laboratory animals tested (LD₅₀). Given the large number of pesticide registration applications that require acute toxicity data, an alternative to the in vivo test could greatly reduce animal testing. The United Nations Globally Harmonized System of Classification and Labelling of Chemicals (GHS) Mixtures Equation estimates the acute toxicity of mixtures using the toxicities of mixture components. The goal of this study was to evaluate the concordance of LD₅₀s predicted using the GHS Mixtures Equation and LD₅₀s from the in vivo test results. Using the EPA classification system, concordance was 55% for the full dataset (N = 671), 52% for agrochemical formulations (N=620), and 84% for antimicrobial cleaning products (N=51). Most discordant results were from substances LD₅₀ >2000 mg/kg (limit test) or 2000<LD₅₀<5000 mg/kg that were predicted as LD₅₀>5000 mg/kg. A supplementary analysis combining all formulations with an LD₅₀ >500mg/kg produced a concordance of 82%. The lack of more toxic formulations in this dataset prevented a thorough evaluation of the GHS equation for such substances. Accordingly, our results suggest the GHS equation is helpful to predict the toxicity of mixtures, particularly those with lower toxicity.

1. Introduction

Acute systemic toxicity testing identifies the potential for a chemical to cause toxicity in humans following ingestion (oral toxicity tests), absorption through the skin (dermal toxicity tests), or inhalation (inhalation toxicity tests). These tests are usually conducted in rodents or rabbits. The primary endpoint is lethality, expressed as LD₅₀ or LC₅₀ depending on the route of administration, with LD₅₀ used to represent the dose of a test substance that would be expected to kill 50% of the animals tested via oral route. The U.S. Environmental Protection Agency (EPA) is one of several U.S. federal agencies that require acute oral toxicity data (Strickland et al. 2018) to assign substances to toxicity categories that, in turn, determine the hazard warnings displayed on product labels and the types of personal protective equipment required to handle the substance.

The Organisation for Economic Co-operation and Development (OECD) has issued several test guidelines for the evaluation of acute oral toxicity, which have evolved over time to use fewer animals. While current test designs (OECD, 2001, 2002a, 2002b, 2008) use fewer animals than the traditional LD₅₀/LC₅₀ methods, they can still require up to 15 animals per test. Regulators such as EPA require that acute oral toxicity be assessed for both pure substances and mixtures. The focus of this evaluation is on formulations that are regulated as pesticides by EPA, such as herbicides, fungicides, insecticides, and cleaning products that make antimicrobial claims. Pesticide formulations are typically a mixture of active and inert ingredients, with inert ingredients optimized to increase the safety or effectiveness of the formulation. Importantly, “inert” ingredients are not necessarily “nontoxic”, but simply do not have any activity against the target organism. The EPA Office of Pesticide Programs (OPP) requires acute toxicity data on all active ingredients as well as end-use formulations, with data generated using U.S. or internationally adopted test guidelines (EPA, 1998a, 1998b, 2002a, 2002b). While the number of new active ingredients registered each year is typically between 10–20, the number of new pesticide formulations can be over 250 per year (US EPA, 2018a). Accordingly, an alternative to conducting the in vivo acute oral toxicity test for formulations has the potential to reduce animals tested per year by the thousands.

The Globally Harmonized System of Classification and Labelling of Chemicals (GHS) is used internationally for hazard classification of substances. A mathematical approach to calculating toxicity of mixtures has been developed within the context of the GHS that is based on the weighted contribution of each formulation component [UN Guidance, (United Nations, 2019)]. The acute toxicity estimate (ATE) of a mixture (in this case, the estimated LD₅₀) is calculated by summation of the ATE values for all relevant ingredients using the following equation:

$$\frac{100}{ATEmix}={\displaystyle \sum }_n\frac{Ci}{ATEi}$$

where Ci = concentration of the ingredient; n ingredients and i is running from 1 to n; ATEi = acute toxicity estimate of ingredient i. Mixture components present at 1% or greater in the formulation with a known acute toxicity value are included in the calculation while ingredients that are presumed not acutely toxic (e.g., water, sugar) and ingredients with an oral limit test that does not show acute toxicity at 2,000 mg/kg bw (body weight) are ignored. Using an assumption of additive toxicity, this approach, referred to as the GHS Mixtures Equation (hereafter GHS equation), allows calculation of an estimate of acute oral toxicity without additional in vivo testing if all the components have been previously tested.

In 2016, OPP initiated the “Mixtures Equation Pilot Program to Reduce Animal Testing” (US EPA, 2016) to evaluate the performance of the GHS equation. As part of this effort, OPP accepted submissions of acute oral toxicity data paired with calculations using the GHS equation to support evaluations of pesticide formulations. The goal of the pilot program was to evaluate the utility and acceptability of the GHS equation as an alternative to animal oral toxicity studies for pesticide formulations.

The purpose of the current project discussed in this paper was to compare LD₅₀s predicted for formulations based on the GHS equation to LD₅₀s determined from in vivo results with the complete formulation. Paired LD₅₀s based on in vivo results and calculated using the GHS equation were collected through a pilot program administered by the EPA (US EPA, 2016) as well as data provided directly to the National Toxicology Program Interagency Center for the Evaluation of Alternative Toxicological Methods (NICEATM). The concordance of the paired LD₅₀ values and their associated hazard classifications based on the EPA and GHS systems was determined.

2. Materials and methods

2.1. Data collection

NICEATM obtained data for a total of 671 formulations including both in vivo measured LD₅₀s coupled with LD₅₀s calculated using the GHS equation. Formulations were provided by BASF (N=209) Clorox (N=19), Control Solutions (N=4), Dow (N=63), DuPont (N=58), Ecolab (N=15), Procter & Gamble (N=13), and Syngenta (N=290). Data were collected from studies submitted for pesticide registration. Five companies submitted data under the EPA OPP pilot program, which OPP then provided to NICEATM. BASF, Dow, and Dupont provided data directly to NICEATM. Companies also provided the ATE calculations along with the EPA and GHS classifications for the LD₅₀ values. Once formulations were compiled, they were assigned a number one to 671. Formulations were evaluated based on two separate categories due to the substantive differences in composition between the two types of formulations: antimicrobial cleaning products (AMCPs, N=51) provided by Clorox, Control Solutions, Ecolab, and Procter & Gamble; and agrochemicals (N = 620) provided by BASF, Dow, Dupont, and Syngenta. EPA and GHS hazard categories (see Table 1) were assigned to each substance based on both measured LD₅₀ (i.e., the LD₅₀ derived from in vivo acute oral toxicity test) and the estimated LD₅₀ (i.e., the LD₅₀ calculated using the GHS equation).

Table 1:

EPA and GHS Hazard Categorization Schemes

EPA Category	LD50 range	GHS Category	LD50 range
I	≤ 50 mg/kg	1	≤ 5 mg/kg
		2	> 5 ≤ 50 mg/kg
II	> 50 ≤ 500 mg/kg	3	> 50 ≤ 300 mg/kg
III	> 500 ≤ 5,000 mg/kg	4	> 300 ≤ 2,000 mg/kg
IV	> 5,000 mg/kg	5	> 2,000 mg/kg
		Not Classified	>5000 mg/kg

2.1. Analysis

2.2.1. Concordance analysis

A qualitative assessment was conducted on the agreement (concordance) between hazard categories based on either the EPA or GHS classification schemes, derived using in vivo measured LD₅₀ values and LD₅₀s calculated using the GHS equation. To calculate concordance of formulations under the GHS classification system, formulations classified as Category 5 or Not Classified (NC) were considered equivalent and were combined. Concordance was calculated by determining the percentage of formulations for which classifications derived from in vivo data agreed with classifications derived from GHS equation calculations.

2.2.2. Dataset processing

A portion of the in vivo LD₅₀ values (N=186) were provided as limit dose values ranging from LD₅₀>300 mg/kg to LD₅₀>3000 mg/kg for which multiple hazard classifications were possible. For example, for a substance with an LD₅₀ >2000 mg/kg, it is not clear if the actual LD₅₀ is between 2000–5000 mg/kg (which would be classified as EPA Category III), or if it in fact it is >5000 mg/kg (which would be classified as EPA Category IV).

Initially, data (N = 671) were analyzed according to EPA and GHS classification systems and if the data provided did not allow for classification into a single category, the more conservative classification was used for analysis. In the example above (LD₅₀ >2000 mg/kg), the more conservative classification (EPA Category III) would be used for this substance. A second analysis was conducted, hereafter referred to as “supplementary analysis,” that considered all substances with an LD₅₀ > 500 mg/kg together (note that one formulation with an in vivo LD₅₀ reported as a range from 300 to 2000 mg/kg was excluded as indeterminate). Because substances with LD₅₀ >2000 would default to EPA Category III for precautionary labeling purposes, Category III and Category IV substances were combined (>500 mg/kg).

2.2.3. Examining discordant LD₅₀ Values for the proximity of paired values

When there was disagreement between classifications resulting from in vivo and calculated LD₅₀s, the proximity of the paired LD₅₀ values was examined. Specifically, both in vivo and calculated LD₅₀ values were log transformed and used ±0.25 log mg/kg as a confidence interval, which was established based on an evaluation of observed experimental variability in replicate in vivo acute oral toxicity data (Karmaus, Mansouri, Blake, et al. submitted). Discordant LD₅₀ pairs (in vivo and corresponding calculated LD₅₀) ±0.25 log mg/kg were plotted on a log10 scale to determine whether the ranges of the LD₅₀ obtained via the different approaches overlapped. To provide a measure of how well the calculated LD₅₀s approximate the in vivo values, the root mean squared error (RMSE) was calculated using the metrics package in R (Hamner and Frasco, 2018). Root Mean Square Error (RMSE) is the standard deviation of the residuals. Whereas correlation coefficient (r2) is a relative measure of fit (how tight data are to the regression line), RMSE is an absolute measure of fit (the spread of how far residuals are from the regression line; i.e., how close are predicted values to experimental values). RMSE is a good measure of how accurately values are predicted, and it is the most important criterion for fit when the main purpose of the model is prediction. There is a direct relationship between RMSE and r2; if the r2 correlation coefficient is 1, the RMSE will be 0, because all of the points lie on the regression line (and therefore there are no errors or spread away from the regression line). Lower values of RMSE indicate better fit (Kenney and Keeping, 1962).

3. Results

3.1. Full data set

Figure 1 provides a breakdown of the dataset by EPA and GHS classification systems. It should be noted that the data set is heavily weighted towards less toxic formulations.

Figure 1. Distribution of 671 formulations by hazard classification category.

*Hazard category for each formulation provided by the data submitter and based on the LD₅₀ determined from an in vivo study. The formulations are presented by EPA hazard classification system (Figure 1A) and by the GHS hazard classification system (Figure 1B).

3.2. Concordance analysis

The dataset of formulations was analyzed for concordance between in vivo LD₅₀s and LD₅₀s calculated using the GHS equation. Concordance was analyzed for both the EPA and GHS hazard classification systems.

3.2.1. Concordance using the EPA categories

EPA hazard classifications assigned based on the GHS equation were 55% (367/671) concordant with those assigned based on in vivo data (see Figure 2A). Most misclassified formulations (37%; 245/671) were under classified, with the majority of these (163/245) being formulations classified as EPA Category III based on in vivo data but EPA Category IV based on the GHS equation. Most of these “discordant” substances (79%; 128/163) had in vivo LD₅₀s values measured between 2000 and 5000 mg/kg or limit test LD₅₀ > 2000 mg/kg. Similarly, of the 20 substances that were classified as EPA Category IV based on in vivo data but over classified based on the GHS equation, most (70%; 14/20) had calculated LD₅₀ values that were between 2000 and 5000 mg/kg. When the subset of formulations identified as AMCP (N = 51) were considered separately, overall concordance was greater (84%; 43/51; see Figure 2E) than for those formulations identified as agrochemicals (52%; 324/620; see Figure 2C).

Figure 2. Concordance Tables using the EPA Classification System.

The dataset was analyzed using the EPA Classification System (Figures 2A, 2C, and 2E) and using a supplemental concordance analysis where > 500 mg/kg were combined together (Figures 2B, 2D, and 2F). Figures 2A and 2B show the concordance of the entire dataset; Figures 2C and 2D show the concordance of agrochemical formulations; Figures 2E and 2F show the concordance of AMCP formulations.

3.2.2. Concordance based on the supplemental analysis

When formulations with LD₅₀s >500 mg/kg were considered as a single group, overall concordance between calculated and measured in vivo classifications improved from 55% to 82% (see Figure 2B). As in the initial analysis, when the subset of formulations identified as AMCP (N = 51) were considered separately overall concordance was greater (100%; 51/51; see Figure 2F) than for those formulations identified as agrochemicals (80%; 497/619; see Figure 2D).

3.2.3. Concordance using the GHS categories

Overall concordance using the GHS classification system to analyze the full dataset was 72% (484/671) (see Figure 3A). However, as noted above, the available dataset is heavily skewed towards less toxic substances (i.e., GHS Category 4 and 5/NC). In general, misclassified substances (N= 184) were under classified regardless of formulation type, with a large portion of these (45%; 85/187) being in vivo GHS Category 4 substances predicted to be GHS Category 5/NC by the GHS equation.

Figure 3. Concordance Tables Using the GHS Classification System.

. Figure 3A shows the concordance of the entire dataset; Figure 3B shows the concordance of agrochemical formulations; Figure 3C shows the concordance of AMCP formulations.

As with the EPA classification system, when the subset of formulations identified as AMCP (N = 51) were considered separately using the GHS classification system, overall concordance was much greater (98%; 50/51; see Figure 3C) than for those formulations identified as agrochemicals (70%; 434/620; see Figure 3B). The lone discordant AMCP formulation was under classified by the GHS equation; the in vivo LD₅₀ was 1098 mg/kg, but the LD₅₀ based on the GHS equation was 2092 mg/kg.

3.3. Analysis of discordant formulations

Most of the substances exhibiting discordance between in vivo classifications and calculated classifications were under classified by one category, regardless of the classification system used. Among the under classified formulations, most cases were either EPA Category III (67%) or GHS Category 4 (70%) substances predicted as EPA Category IV or GHS Category 5/NC systems, respectively.

The difference between measured and calculated values was further evaluated by calculating the RMSE for the dataset. Analyzing the full dataset (N = 671) produced a RMSE of 0.77. Removing limit values and LD₅₀s reported as a range analysis yielded an RMSE of 0.67 for the remaining formulations (N = 261). In a final analysis step, nine calculated values that ranged from 100,000 to 1,000,000, 21 that were between 10,000 to 100,000, and 16 from 5,001 to 10,000 were artificially set to a value of 5000 which reduced the RMSE to 0.49. Although artificial data with limited ranges improved the RMSE, in general these results suggest that calculated values still vary from experimental values.

3.3.1. Analysis of overlapping log10±0.25 ranges

To examine discordant formulations the ranges of the in vivo and calculated log10 LD₅₀ ± 0.25 (based on observed variability in the rodent LD₅₀ test) were examined for overlap. To limit the analysis to LD₅₀s with a single numerical value, LD₅₀ values that were expressed as ranges, those based on a limit dose (i.e., greater than a certain value), and calculated LD₅₀s that had no classified ingredients (and therefore no numerical LD₅₀ values) were excluded from this analysis. These restrictions left 261 formulations with numerical values for both the in vivo and calculated LD₅₀s for which the log10 LD₅₀ ± 0.25 ranges could easily be compared (see Figure 4). Using the EPA classification system, roughly half of the discordant formulations had a calculated log10 LD₅₀ ± 0.25 range that overlapped the in vivo LD₅₀ ± 0.25 log10 range.

Figure 4.

Comparison of in vivo and calculated LD₅₀ values. Log transformed LD₅₀ values for discordant formulations were compared using ±0.25 log mg/kg as a confidence interval. Figure 4A – 4D display EPA under classified formulations, EPA over classified formulations, GHS under classified formulations, and GHS over classified formulations, respectively. Horizontal lines represent classification cut-offs.

The analysis of all EPA discordant formulations revealed:

• Of the 61 substances with an in vivo LD₅₀ between 50 to 500 mg/kg that were predicted as 500 to 5000 mg/kg based on the GHS equation:

  • ⚪33 had an in vivo LD₅₀ based on a limit dose or range.
  • ⚪13 of the remaining 28 had overlapping log10 ± 0.25 ranges for in vivo versus predicted.

• Of the 20 substances with an in vivo LD₅₀ between 50 to 500 mg/kg that were predicted as > 5000 mg/kg based on the GHS equation:

  • ⚪14/20 had an in vivo LD₅₀ based on a limit dose or range.
  • ⚪None of the remaining 6 had overlapping log10 ± 0.25 ranges for in vivo versus predicted.

• Of the 163 substances with an in vivo LD₅₀ between 500 to 5000 mg/kg that were predicted as > 5000 mg/kg based on the GHS equation:

  • ⚪117 had an in vivo LD₅₀ based on a limit dose, range, or had no classifiable ingredients to calculate a predicted LD₅₀ using the GHS equation (and thus would not be classified).
  • ⚪8 of the remaining 46 had overlapping log10 ± 0.25 ranges.

• Of the 34 substances with an in vivo LD₅₀ between 500 to 5000 mg/kg that were predicted as between 50 to 500 mg/kg based on the GHS equation:

  • ⚪3 had an in vivo LD₅₀ based on a limit dose, range, or had no classifiable ingredients to calculate a predicted LD₅₀ using the GHS equation.
  • ⚪15 of the remaining 34 had overlapping log10 ± 0.25 ranges.

• Of the 35 substances with an in vivo LD₅₀ > 500 mg/kg that were predicted as between 50 to 500 mg/kg based on the GHS equation:

  • ⚪2 had an in vivo LD₅₀ based on a limit dose, range, or had no classifiable ingredients to calculate a predicted LD₅₀ using the GHS equation.
  • ⚪15 of the remaining 33 had overlapping log10 ± 0.25 ranges.

• Of the 81 substances with an in vivo LD₅₀ between 50 and 500 mg/kg that were predicted as > 500 mg/kg based on the GHS equation:

  • ⚪47 had an in vivo LD₅₀ based on a limit dose or range.
  • ⚪13 of the remaining 34 had overlapping log10 ± 0.25 ranges.

A similar analysis conducted using the GHS classification system indicated that roughly one-third of the discordant formulations had predicted log10 LD₅₀ ± 0.25 values that overlapped the in vivo values.

The analysis of GHS discordant formulations revealed:

• Of the 26 substances with an in vivo LD₅₀ between 50 to 300 mg/kg that were predicted as between 300 to 2000 mg/kg based on the GHS equation:

  • ⚪10 had an in vivo LD₅₀ based on a limit dose or range.
  • ⚪4 of the remaining 16 had overlapping log10 ± 0.25 LD₅₀ ranges.

• Of the 10 substances with an in vivo LD₅₀ between 50 to 300 mg/kg that were predicted as > 2000 mg/kg based on the GHS equation:

  • ⚪6 had an in vivo LD₅₀ based on a limit dose or range.
  • ⚪None of the remaining 4 had overlapping log10 ± 0.25 LD₅₀ ranges.

• Of the 85 substances with an in vivo LD₅₀ between 300 to 2000 mg/kg that were predicted as > 2000 mg/kg based on the GHS equation:

  • ⚪51 had an in vivo LD₅₀ based on a limit dose, range, or had no classifiable ingredients to calculate a predicted LD₅₀ using the GHS equation.
  • ⚪10 of the remaining 34 had overlapping log10 ± 0.25 LD₅₀ ranges.

• Of the 17 substances with an in vivo LD₅₀ between 300 to 2000 mg/kg that were predicted as between 50 to 300 mg/kg based on the GHS equation:

  • ⚪2 had an in vivo LD₅₀ based on a limit dose.
  • ⚪4 of the remaining 15 had overlapping log10 ± 0.25 LD₅₀ ranges.

Therefore, irrespective of the hazard classification system, many of the formulations evaluated do not have overlapping in vivo and calculated LD₅₀s despite consideration of variability. However, as demonstrated in Figure 4, in most cases the differences observed do not alter the hazard classification by more than one category.

4. Discussion

EPA requires toxicity assessments to be conducted on every pesticide formulation (agrochemicals and AMCPs) proposed for marketing, regardless of whether the active or inert ingredients included in the formulation have been tested previously. These assessments can lead to substantial animal testing. The GHS equation represents a possible alternative approach to reduce animal testing for formulations. When the toxicities of a formulation’s components are known, the computational approach provided by the GHS equation may enable a prediction of formulation toxicity to be made with no additional testing. Further, applications for pesticide products may qualify as identical or substantially similar in composition and labeling to a currently registered pesticide product or differ in ways that would not significantly increase the risk of unreasonable adverse effects on the environment as outlined in FIFRA Sections 3(c)(3)(B) and 3(c)(7)(A). The GHS equation prediction tool may be utilized by EPA under this statute as part of a weight of evidence waiver rationale.

Previous studies failed to support a definitive conclusion of the utility of the GHS equation. Whereas Corvaro et al. (2016) reported that this approach produced high accuracy and specificity, Van Cott et al. (2018) concluded that the components of the formulations they examined likely interacted to influence the toxicity of the mixtures in a manner not accounted for by the GHS equation, resulting in a high percentage of misclassifications.

To further evaluate the performance of the GHS equation, EPA initiated a pilot program to evaluate its usefulness as an alternative to in vivo acute oral toxicity studies for pesticide formulations. Data on nearly 700 formulations with both in vivo and calculated LD₅₀ values were collected to evaluate their concordance. The data set was comprised of agrochemicals and AMCP formulations, and analysis of the data set by formulation type demonstrated that the performance of the GHS equation may vary by type of formulation. Across the data set, the concordance for AMCP formulations was relatively high 84% using the EPA classification system and 98% using GHS classifications. The higher concordance of the GHS equation with antimicrobials could be used with other data in a weight of evidence approach that has potential benefits including reduced animal testing and likely results in reduced regulatory agency resources for review of new product submissions. Using an alternative method for antimicrobial products could facilitate reduced review times in scenarios where new or amended products are needed in an expedited manner (i.e., for emerging viral pathogens such as SARS-CoV-2). Additionally, given the recent significant influx of new AMCP submissions that EPA OPP has seen, using alternative methods is particularly impactful in maintaining reduced animal testing. While concordance for AMCPs was relatively high in the current dataset, the relatively low number of substances (N = 51) were primarily comprised of EPA Category IV (N = 44) or GHS Category 5/NC (N = 49) formulations, which skewed the overall interpretation of these results on the applicability of the equation more broadly.

The overall concordance of all formulations ranged from 55% (EPA system) to 72% (GHS system), with the majority of “mispredicted” chemicals falling into the range of in vivo LD₅₀ >2000 mg/kg (EPA Category III) but predicted LD₅₀ >5000 mg/kg (EPA Category IV). However, most substances in the data set were LD₅₀ > 500 and were predicted with relatively high concordance of 93%. With respect to the least toxic class, within-class concordance for the in vivo EPA Category IV substances was 87%. The prediction for substances with LD₅₀ < 500 mg/kg was, low, but that was also coupled with a comparatively small data set. This does skew the overall concordance, but it is reflective of the available dataset.

Our initial analysis assigned the more toxic classifications to formulations having multiple possible classifications, such as for a formulation with a reported LD₅₀ of >2000 mg/kg, which could potentially be EPA Category III or IV. If, as in this example, the GHS equation produced a classification of IV, the default to the more toxic in vivo classification likely erroneously inflated the number of under classified formulations. Although the LD₅₀ ranges used for EPA and GHS hazard classifications do not precisely overlap, the increased performance of the GHS system (which as presented combines all substances with LD₅₀ > 2000 mg/kg into one category and has a 72% overall concordance rate) relative to the EPA supports this conclusion.

A supplemental analysis was conducted to evaluate the impact on concordance of combining substances with LD₅₀ > 500 mg/kg. This evaluation provides a practical assessment to establish whether discordance between LD₅₀s derived from the additivity equation versus those determined in vivo would impact precautionary labeling since substances with LD₅₀ >500 or >2000 would both be classified as EPA Toxicity Category III. Additionally, such labeling is markedly different for substances depending on whether the LD₅₀ is > or < 500 mg/kg. Under this scenario, the number of formulations under classified by the GHS equation was substantially decreased (from 163 to 81), while the number of formulations over classified based on the GHS equation increased slightly from 20 to 36, and the overall concordance increased from 55% to 82%. This improvement underscored that limitations of our data set with the majority (550/671) of substances analyzed with LD₅₀ >500.

It should be noted that the current data set has relatively few formulations classified as highly toxic; our data set only included four formulations that were classified as EPA Category I under the EPA classification system (all 4 would be GHS Category 2). One of these four highly toxic formulations was under classified as EPA Category II based on the GHS equation LD₅₀ estimate. Similarly, 81 of the 115 EPA Category II formulations in our data set were under classified as EPA Category III or IV, as were 36 of the 50 GHS Category 3 formulations. Despite the small sample size, these results suggest a tendency of the GHS equation to under classify formulations relative to the in vivo classification for more toxic formulations.

Initial analysis of discordant formulations demonstrated a high number of LD₅₀s in the range of 2000 to 5000 mg/kg, including a number of formulations tested in vivo using a maximum dose of 2000 or greater. Under the EPA classification system, 66% of under classified formulations were EPA Category III classified by the GHS equation as EPA Category IV. Most of these under classified formulations (128/163) have an LD₅₀ between 2000 and 5000 mg/kg or from a limit test (>2000 mg/kg or above). Whereas the EPA classification system classifies all substances having LD₅₀s between 500 mg/kg and 5000 mg/kg as EPA Category III, the GHS classification system splits LD₅₀s in this range between GHS Categories 4 and 5/Not classified. Therefore, in many jurisdictions, the under classified EPA Category III formulations would be considered minimally toxic (and not classified) and their under classification may thus be of lower significance. Given the demonstrated inherent variability of the animal test (Karmaus, Mansouri, Blake, et al. submitted) it is certainly possible that a similar under classification in vivo would also be observed following a repetition of the animal test. In addition, the current in vivo acute oral toxicity test uses far fewer animals to estimate toxicity, and the fixed-dose (OECD TG 420) and up-and-down (OECD TG 425) procedures produce acute toxicity range values that may represent a less precise estimate of the acute oral LD₅₀ than the point estimate produced by the obsolete OECD TG 401. In addition, the fixed-dose procedure relies on evidence of toxicity and avoids administration of doses expected to be lethal. This raises the larger issue of considerations that should be made when using in vivo animal data as a reference for validation of an alternative approach. Further, as pointed out in Van Cott et al. (2018), ingredients typically are not tested in vivo with the vehicle used for the formulations which may change the bioavailability of the ingredient. This is perhaps highlighted by the GHS system analysis, where the impact on concordance of combining all substances with LD₅₀ >2000 mg/kg into a single group (i.e., GHS Category 5/Not Classified) can be seen with within class concordance of 88%. This difference resulted in much improved overall concordance (from 55% to 72%), relative to that for the EPA classification system. Similarly, further improvement in concordance (to 82% overall) was observed by combining all substances with LD₅₀ >500 mg/kg into a single group. This improvement demonstrates the potential utility of the equation in a weight of evidence rationale when the expected toxicity of a mixture is expected to be low based on related in vivo data. For example, pairing the additivity prediction with in vivo data on a substantially similar product, or determining whether a formulation change could impact toxicity.

Our results suggest that substances predicted by the GHS equation to have low or negligible acute oral toxicity are likely to be adequately identified as such. Additional work could further define a rationale for observed discordances, including a close evaluation of the components of a formulation and its associated LD₅₀ values, as well as the challenges of classification using limit tests. Additionally, if data are available, a larger number of substances with LD₅₀ < 500 mg/kg needs to be evaluated using this approach to facilitate a definitive outcome and recommendation. This could necessitate expanding the analyses beyond pesticide formulations as this data set is considered to represent most available data on such formulations. Key to any future success in evaluation and acceptance of the GHS equation is continued cooperation and transparency among stakeholders from both the government and industry sides. This work represents an important step towards international harmonization by assessing the utility and application of the GHS equation in a US regulatory context.

Highlights.

• The concordance of in vivo acute toxicity results and LD₅₀ values calculated using the GHS Mixtures Equation were evaluated.

• Most mispredictions occurred between the two least toxic categories.

• Concordance improved up to 98% for a small set of antimicrobial cleaning products.

• The GHS Mixtures Equation may be able to identify minimally toxic substances (i.e., LD₅₀ > 2000 mg/kg).

Abbreviations:

AMCP

antimicrobial cleaning products

EPA

U.S. Environmental Protection Agency

GHS

United Nations Globally Harmonized System of Classification and Labelling of Chemicals

NICEATM

National Toxicology Program Interagency Center for the Evaluation of Alternative Toxicological Methods

OECD

Organisation for Economic Co-operation and Development

OPP

Office of Pesticide Programs

NA

not applicable

NC

Not Classified (GHS)