Evaluating potential refinements to existing Threshold of Toxicological Concern (TTC) values for environmentally-relevant compounds

Abstract

The Toxic Substances Control Act (TSCA) mandates the US EPA perform risk-based prioritisation of chemicals in commerce and then, for high-priority substances, develop risk evaluations that integrate toxicity data with exposure information. One approach being considered for data poor chemicals is the Threshold of Toxicological Concern (TTC). Here, TTC values derived using oral (sub)chronic No Observable (Adverse) Effect Level (NO(A)EL) data from the EPA’s Toxicity Values database (ToxValDB) were compared with published TTC values from Munro et al. (1996). A total of 4554 chemicals with structures present in ToxValDB were assigned into their respective TTC categories using the Toxtree software tool, of which toxicity data was available for 1304 substances. The TTC values derived from ToxValDB were similar, but not identical to the Munro TTC values: Cramer I ((ToxValDB) 37.3 c.f. (Munro) 30 µg/kg-day), Cramer II (34.6 c.f. 9.1 µg/kg-day) and Cramer III (3.9 c.f. 1.5 µg/kg-day). Cramer III 5^th percentile values were found to be statistically different. Chemical features of the two Cramer III datasets were evaluated to account for the differences. TTC values derived from this expanded dataset substantiated the original TTC values, reaffirming the utility of TTC as a promising tool in a risk-based prioritisation approach.

1. Introduction

The Toxic Substances Control Act (TSCA) mandates the US Environmental Protection Agency (EPA) perform risk-based prioritisation of chemicals in commerce and then, for high-priority substances, develop risk evaluations that integrate toxicity data with exposure information (EPA, 2008). For chemicals with limited chemical-specific toxicity data, one approach being considered is a Threshold of Toxicological Concern (TTC)-to-Exposure ratio. In an earlier manuscript (Patlewicz et al., 2018), a proof of concept study using a dataset of 7986 substances was undertaken to integrate TTC with heuristic high-throughput exposure (HTE) modelling to rank order chemicals for further evaluation. In this study, we sought to evaluate whether the established TTC values that had been used in Patlewicz et al. (2018) were applicable for the types of chemicals of interest to EPA by analysing an expanded toxicity dataset.

The TTC approach establishes different levels of human exposure below which there is expected to be a low probability of risk to human health (Kroes et al., 2004; WHO/EFSA, 2016; EFSA, 2019). Kroes et al. (2004) presented a tiered TTC approach that established several human exposure thresholds over several orders of magnitude, ranging from 0.0025μg/kg-day to 30μg/kg-day. The exposure limit established for each TTC tier was based on an evaluation of existing toxicity data for chemicals in each tier. It should be noted that the TTC approach was initially developed to be used in specific cases where exposure is expected to be low and where no or limited hazard data is available. Moreover, certain chemicals are excluded from the TTC approach because they were not represented in the original toxicity databases supporting TTC (e.g., metals or metal containing compounds, organosilicons, proteins) or because standard risk assessment approaches are more appropriate (e.g. 2,3,7,8-dibenzo-p-dioxin (TCDD) and its analogues, high potency carcinogens such as N-nitroso compounds).

The lowest TTC tier is 0.0025μg/kg-day which is for substances that raise a concern of genotoxicity determined on the basis of structural alerts for genotoxicity/mutagenicity. For substances without structural alerts, there are a series of non-cancer TTC tiers, which are based on the Cramer et al. (1978) decision tree. Derivation of TTC values for each of the three Cramer Classes stems from the work of Munro et al. (1996). Munro and colleagues compiled a database of NOELs for 613 substances that had been tested in repeat-dose oral toxicity studies including subchronic, chronic, reproductive and developmental toxicity. In cases where there were multiple NOELs for a given substance, the lowest one was selected (there were a total of 2941 NOELs for the 613 substances). The substances were then assigned to the appropriate Cramer structural class, and cumulative distributions of the logarithms of NOELs were plotted separately for each structural class. Adjustments were made to extrapolate subchronic NOELs to chronic, and LOELs to NOELs as appropriate. The 5th percentile NOEL was estimated for each structural class, which then was converted into its respective TTC value by applying a safety factor of 100 (10X to account for extrapolation of animals to humans and 10X for human variability). The TTC values established were 30μg/kg-day for Cramer Class I, 9μg/kg-day for Cramer Class II, and 1.5μg/kg-day for Cramer Class III substances. Kroes et al. (2004) evaluated whether chemicals shown to be neurotoxicants, immunotoxicants and teratogens needed to be considered as a separate category. They concluded that with an exception for organophosphate pesticides (OPs) and carbamates, such substances were adequately represented by the TTC approach for systemic toxicity endpoints. A TTC value of 0.3μg/kg-day was derived for organophosphates and carbamates. NOELs for the OPs and carbamates were not subsequently removed from the original Cramer Class III distribution and a number of publications have suggested that this distribution should be re-evaluated without these substances. Indeed, Munro et al. (2008) suggested that the new limit for Cramer Class III would be at least 3μg/kg-day if OPs were removed and an even higher level of 10μg/kg-day if both OPs and organohalogen compounds were removed; however, neither of these refined values have yet been adopted in practice.

Although the Munro et al. (1996) dataset was intended to cover a broad chemical domain, the dataset is now over 20 years old, and a question that could be raised is whether the TTC values that had been derived ought to be updated if an expanded dataset were to be used. Indeed, there have been a wealth of studies which have built upon the work by Munro et al. (1996), including identifying groups of chemicals for which the TTC approach is not appropriate, proposing additional TTC values for specific endpoints, or utilising additional datasets to re-evaluate the original TTC values. Many of these studies have been cited in the WHO/EFSA (2016) and EFSA (2019) reports. Examples of studies include that from Cheeseman et al. (1999), Munro et al. (1999), Blackburn et al. (2005), van Ravenzwaay et al. (2011), Kalkhof et al. (2012), Dewhurst and Renwick, (2013), Leeman et al. (2014), Boobis et al. (2017) and Yang et al. (2017). One evaluation of particular interest, and which inspired our own case study was that undertaken by Yang et al. (2017) who enriched the original Munro et al (1996) dataset to capture cosmetics-related substances. The case study here was structured to consider an expanded dataset that was more representative of the chemicals of interest to the EPA. Specifically, the objectives were as follows:

1. Verify TTC values from Munro et al. (1996)

2. Extract data from US EPA’s Toxicity Values database (ToxValDB) available via the US EPA’s CompTox Chemicals Dashboard (https://comptox.epa.gov/dashboard; Williams et al. (2017))

3. Use the Kroes et al. (2004) workflow implemented in Toxtree to assign substances present in ToxValDB into their respective TTC categories

4. Derive TTC values using the toxicity data extracted from ToxValDB for Cramer class chemicals

5. Evaluate whether the newly derived TTC values were statistically equivalent to those derived from the Munro et al. (1996) dataset

6. Derive confidence intervals for the 5^th percentile values underpinning the newly derived TTC values

7. Compare and contrast the chemistry of the two datasets to rationalise any (dis)similarities in the TTC values

8. Profile a large inventory of ~45,000 chemicals taking into account insights gained from the preceding objectives

2. Materials and Methods

2.1. Toxicity Data Sources

Two sources of toxicity data were utilised in this study: 1) the TTC dataset from Munro et al. (1996) referred to as the ‘Munro dataset’ and 2) the US EPA’s Toxicity Values (ToxVal) database (version 7) referred to as ToxValDB.

The Munro dataset was downloaded as an Excel file from the European Food Safety Authority (EFSA) website (http://www.efsa.europa.eu/en/supporting/pub/en-159). This was converted to a comma separated value (csv) file to facilitate use within the R scripting environment (https://www.r-project.org) (R Core Team, 2018).

ToxValDB consists of a collection of summary level in vivo test data from a variety of study types typically used in risk assessments. It comprises point of departure (POD) values such as no-observed (adverse) effect levels and lowest-observed (adverse) effect levels (NO(A)ELs and LO(A)ELs). These data have been aggregated from over 40 publicly available sources including US Federal and State agencies (e.g. US EPA, US Food and Drug Administration (FDA), and California EPA) alongside international organisations (e.g. World Health Organisation (WHO)), as well as data submitted under regulatory frameworks such as the European Union’s REACH regulation (e.g. non-confidential registration data submitted to the European Chemicals Agency (ECHA) by industry registrants). The entire ToxValDB was downloaded for subsequent filtering and processing (Supplementary Table 1).

2.2. Chemical structure data

2.2.1. Profiling of substances through the Kroes et al. (2004) workflow within Toxtree

Chemicals with defined structures (such as SMILES: Simplified Molecular-Input Line-Entry System) were needed for profiling through the TTC decision tree within Toxtree.

QSAR-ready SMILES strings were extracted through a batch search using the US EPA’s CompTox Chemicals Dashboard (https://comptox.epa.gov/dashboard; Williams et al. (2017)) for the chemicals present in ToxValDB. Of the 15,960 unique substances present in ToxValDB, QSAR-ready SMILES were available for 4,554 chemicals. These were subsequently profiled through Toxtree (v3.1.0) (IdeaConsult Ltd) using two of the original modules, namely the Cramer rules (Patlewicz et al., 2007) and Kroes TTC decision tree as well as 3 custom modules developed ad hoc by Patlewicz et al. (2018) intended to identify carbamates, organophosphates (OPs), and steroids.

SMILES strings provided in the Munro dataset from the EFSA website were converted to their corresponding Kekule form using the ChemAxon Standardizer (v17.13.0) software. The Munro Cramer class chemicals were also processed through Toxtree to address objective 6.

2.3. Verification of Munro et al. (1996) TTC values

The column with the header “NOEL_calculated_Munro_mg/kg/day” in the Munro dataset was used to calculate the 5^th percentile values associated with each Cramer class. However, when calculating the thresholds using this dataset, it became clear that there were discrepancies between the published 5^th percentile values in Munro et al. (1996) and those calculated using the Munro dataset retrieved from the EFSA website. Upon investigation, the following adjustments discussed in Munro et al. (1996) needed to be applied, namely, the 3-fold safety factor for being either: 1) a subchronic study or 2) one of several ad hoc reproductive/teratology studies. After applying these adjustments, the published and calculated thresholds were equal confirming the validity of the Munro dataset.

2.4. New threshold calculations for Cramer class substances using ToxValDB

Data that met the study criteria outlined in Munro et al. (1996) were extracted from ToxValDB for each chemical assigned to the 3 Cramer classes. ToxValDB study criteria were as follows: 1) study types that were included were subchronic, chronic, reproductive, developmental, or multigenerational. Short term and acute studies were not considered; 2) route of exposure – oral, other routes were excluded; 3) species – rodents; 4) point of departure – no observed (adverse) effect level (NO(A)EL); and 5) point of departure units – mg/kg-day. As per Munro et al. (1996), all NO(A)ELs from subchronic studies were divided by a factor of 3 to calculate an approximation of the NO(A)EL that would likely be generated by a chronic study. For chemicals with multiple NO(A)EL values, the minimum NO(A)EL was used once extreme outliers were identified and removed. Tukey’s fences (Tukey, 1977) were calculated for each chemical using the following method:

$$lower bound=Q1-1.5\left(Q3-Q1\right)$$

$$upper bound=Q3+1.5\left(Q3-Q1\right)$$

where Q1 is the lower quartile value and Q3 is the upper quartile value. Extreme outliers were identified as NO(A)EL values that were either less than the lower bound value or greater than the upper bound value.

The empirical cumulative distributions of the (minimum) NO(A)ELs for every chemical were plotted and fitted with a lognormal distribution for each Cramer class. The 5^th percentile NO(A)EL values were calculated and converted to their corresponding TTC values by applying a safety factor of 100 as discussed earlier.

2.5. Pairwise Comparison of the TTC values derived from ToxValDB and Munro datasets

2.5.1. Comparison of NO(A)ELs distributions and their fifth percentile values

Pairwise comparisons of the NO(A)EL distributions of the Cramer classes from the ToxValDB dataset were performed using the non-parametric, pair-wise Kolmogorov-Smirnov (K-S) test (Conover, 1999). The K-S test was also used to compare the distributions of each Cramer class between the ToxValDB and Munro datasets (i.e. were the Cramer class distributions statistically different between the two datasets). The 5^th percentiles derived for each Cramer class were also compared between the two datasets to calculate whether there was a statistically significant difference. This was performed using the qcomhd function from the WRS2 package available in R, which utilises a Harrell-Davis estimator, in conjunction with bootstrapping (i.e. random sampling with replacement) (Mair and Wilcox, 2018). Briefly, two groups (e.g. Cramer class I from ToxValDB and Munro) were independently bootstrapped n-times. For each bootstrap sample, the 5^th percentile for each dataset and the difference between the 5^th percentiles were calculated. Once the bootstrapping was complete, the 95% confidence intervals of the difference between the two samples was utilised to identify whether the 5^th percentiles were statistically different (i.e. is the 5^th percentile difference between the two datasets significantly different from zero). In this study, 5000 bootstrap samples were run to calculate the difference between the 5^th percentiles. Bootstrap sampling using 5000 samples was further used to calculate the 95% confidence intervals around the 5^th percentile NO(A)EL values for each dataset and Cramer class.

2.6. Characterisation of the chemical landscape

2.6.1. Investigation of differences in Cramer class chemical landscapes

To provide an overview of the differences in the chemical landscape between the Munro and ToxValDB datasets, bar graphs were generated for each Cramer class which compared the frequency of ToxPrint chemotypes. First, a binary molecular fingerprint was generated for each chemical in the Munro and ToxValDB datasets utilising the publicly available ToxPrint chemotype feature set (https://toxprint.org) and the ChemoTyper software (https://chemotyper.org/). ToxPrint chemotypes consist of a predefined library of 729 sub-structural features designed to encapsulate a broad range of chemical atoms and scaffolds, which were developed by Altamira and Molecular Networks under contract by the US Food and Drug Administration (Yang et al., 2015). Next, the full 729-bit ToxPrint fingerprint was condensed to a length of 70-bits so that any differences between the two datasets could be more readily visualised. To do this, the ToxPrints were condensed based upon the root of the ToxPrint name. For example, the ToxPrints “bond:C#N_cyano_cyanamide”, “bond:C#N_nitro_isonitrile”, and “bond:C#N_nitrile_generic” were concatenated to form the more generalised name “bond:C#N” that is common amongst these ToxPrints. The frequency of these condensed fingerprints in each dataset were then calculated and plotted.

2.6.2. Chemotype enrichment analysis

A chemotype enrichment analysis was conducted to further investigate what, if any, impact the difference in chemical landscape between the two datasets had on the differences in 5^th percentile NO(A)EL values. A more comprehensive explanation of the chemotype enrichment analysis workflow is available in Wang et al. (2019). Briefly, chemotype enrichment analysis identifies sub-structural features that are over-represented with respect to a given endpoint. Typically, this endpoint may be activity in a particular (suite of) assay(s); however, in this study the “endpoint” was the presence/absence of the chemical in the given Munro Cramer class chemical list. Chemicals present in the Munro set were indicated by a value of 1, whilst chemicals originating from the corresponding Cramer class in the ToxValDB dataset were indicated by a value of 0.

To conduct this analysis, the full 729-bit ToxPrint fingerprints generated above were annotated with an additional binary column representing which data set the chemical originated from: either Munro (1) or ToxValDB (0). Chemicals present in both datasets were retained as duplicates to avoid modifying the chemical space of either dataset. This combined dataset was subsequently processed using the chemotype enrichment workflow developed by NCCT researchers and implemented in Python. The odds ratio (OR) and p-values generated by the workflow were utilised to identify those ToxPrints that were more highly enriched in the Munro dataset compared to the ToxValDB dataset. For the purposes of the analysis, enriched ToxPrints were defined as having an OR ≥3, a p value ≤0.05 and number of true positives (TP) ≥3. For this study, only ToxPrints with an OR of infinity were carried forward as this signified that the ToxPrint was not present in any chemical in the ToxValDB dataset. Chemicals within the Munro set that contained any one or more, of the ToxPrints with an OR of infinity were excluded and the 5^th percentile NO(A)EL values were re-calculated. This enabled an exploration of the impact chemicals containing these chemotypes had on the 5^th percentile NO(A)EL value.

2.7. Software

Data processing and analysis was conducted in the R scripting language v3.5.2 (R Core Team, 2018) unless otherwise stated. Code and datasets are available as supplementary information.

3.0. Results and Discussion

3.1. Verification of previously published fifth percentile NO(A)EL values for Cramer class assigned chemicals

To ensure the previously published Cramer class TTC values could be reproduced, the 3-fold adjustment factor was applied to relevant chemicals in the Munro dataset and the 5^th percentile NO(A)EL values were then calculated. The number of chemicals present in the overall Munro dataset, as well as the individual Cramer classes, provided from the EFSA website were the same as those published within the Munro et al. (1996) article. Minor discrepancies were found between the 5^th percentile values calculated and those originally published in Munro et al. (1996) (Table 1). These discrepancies were most likely due to rounding differences in terms of the number of significant figures used in the calculations of the 5^th percentile NO(A)ELs. The 5^th percentile NO(A)ELs calculated were equivalent to those reported by Munro et al. (1996), thus confirming the validity of the Munro dataset.

Table 1.

Verification of Munro dataset provided by EFSA. Comparing the original 5^th percentile NOEL values published by Munro et al (1996) to the recalculated 5^th percentile values.

	Number of chemicals	Original 5th percentile NOEL from Munro (mg/kg bw/day)	Recalculated 5th percentile NOEL from Munro (mg/kg bw/day)
Cramer class I	137	3.0	2.9
Cramer class II	28	0.91	0.95
Cramer class III	448	0.15	0.15

3.2. Calculation of fifth percentile NO(A)ELs from ToxValDB for Cramer class assigned chemicals

Of the 4,554 chemicals present in ToxValDB with QSAR-ready SMILES, 1,241 (27%) were excluded because the chemical either was: 1) not applicable for TTC, i.e. compound-specific toxicity data were required (114 chemicals); 2) considered a potential genotoxicant based upon the presence of a structural alert, thus, requiring the use of the TTC of 0.0025µg/kg-day (1025 chemicals); or 3) considered to be an organophosphate (OP) or carbamate (102 chemicals). Two additional chemicals were excluded from further analysis since they could not be properly profiled through the Cramer workflow implemented in Toxtree. As the decision tree laid out in Kroes et al. (2004) was being followed, substances that presented a structural alert for genotoxicity or were considered to be an OP or carbamate were excluded from further analysis. They will be evaluated separately as part of ongoing work. Of the remaining 3,311 chemicals, 1,476 were assigned to Cramer class I, 162 were assigned to Cramer class II, and 1,673 were assigned to Cramer class III (Table 2).

Table 2.

Number of chemicals from ToxValDB with QSAR-ready SMILES assigned to each TTC category. Two chemicals could not be properly profiled through the Cramer (original) module in Toxtree and were additionally removed.

	Number of chemicals profiled	Number of chemicals with toxicity data
Not applicable for TTC	114	NA
Presence of genotoxicity alert	1025	NA
OPs/Carbamates	102	NA
Cramer class I	1476	565
Cramer class II	162	39
Cramer class III	1673	700
Could not be profiled	2	NA
Total	4554	1304

Upon separating the chemicals into their respective Cramer class, associated toxicity data that satisfied the study criteria set out in Munro et al. (1996) were extracted from ToxValDB - i.e., only subchronic, chronic, reproductive, developmental, or multigenerational studies conducted in rodents with an oral route of exposure and generating a NO(A)EL in mg/kg-day were utilised. This decreased the number of chemicals down to 565, 39, and 700 for Cramer class I, II, and III respectively (Table 2).

The 5^th percentiles calculated for each Cramer class are provided in Table 3. The expected trend in TTC values with more conservative values for Cramer III relative to Cramer I was observed. However, there was only a minimal separation in 5^th percentile values between the Cramer I and Cramer II chemicals.

Table 3.

Comparison of the 5^th percentile NO(A)EL and TTC values calculated using the ToxValDB and Munro data for Cramer class I-III. Note that the 95% confidence intervals calculated using bootstrapping are in parentheses.

Cramer Class	ToxValDB		Munro
	5th percentile (mg/kg-day)	TTC value (µg/kg-day)	5th percentile (mg/kg-day)	TTC value (µg/kg-day)
Class I	3.73 (2.97 – 4.79)	37.3 (29.7 – 47.9)	3.0 (1.71 – 5.31)	30 (17.1 – 53.1)
Class II	3.46 (1.5 – 8.63)	34.6 (15 – 86.3)	0.91 (0.32 – 3.02)	9.1 (3.2 – 30.2)
Class III	0.39 (0.3 – 0.53)	3.9 (3 – 5.3)	0.15 (0.11 – 0.22)	1.5 (1.1 – 2.2)

Figure 1 shows how the lognormal distributions and empirical cumulative distribution functions (CDFs) for the ToxValDB class I and II chemicals are poorly separated; this is especially true below the 10% quantile where the distributions overlap. The K-S test was utilised to evaluate whether the distributions were significantly different between the Cramer classes. The distributions between Cramer classes I and II (n = 604) and Cramer classes II and III (n = 739) were not statistically different at a significance level of 0.05; in contrast, the difference between the Cramer class I and III (n = 1,265) distributions was significant. Given that the K-S test tends to be more sensitive near the centre of the distribution and, looking at Figure 1, we can see that the main differences in the distributions are at the lower quantiles, the fact that the Cramer class II distribution was not significantly different from either Cramer class I or III may not be wholly surprising especially given the few substances it contains.

Figure 1.

Cumulative distribution function and fitted lognormal distribution of NO(A)EL values from ToxValDB for chemicals in Cramer class I (in green), II (in orange), and III (in red). The distributions for Cramer classes I and III were seen to differ significantly, whilst the distributions for classes I and II and classes II and III did not differ significantly (p > 0.05).

3.3. Pairwise Comparison between ToxValDB and Munro et al. (1996)

For each Cramer class, the distribution for the ToxValDB dataset was compared to its corresponding distribution for the Munro dataset (e.g. the ToxValDB Cramer class I distribution was compared to the Munro Cramer class I distribution). As can be seen in Figure 2, the empirical CDFs and fitted distributions for Cramer class II and III between the ToxValDB and Munro datasets are visually more distinct than those for Cramer class I: where the lognormal distributions intersect below the 20% quantile. To statistically investigate whether the distributions were significantly different, the K-S test was employed. Both the ToxValDB and Munro Cramer class I (n = 702) and II (n = 67) distributions were not statistically different (at a p-value of 0.05), whilst the Cramer class III (n = 1,148) distributions were observed to be significantly different. Furthermore, the level of overlap between the two datasets was examined and a total of 219 chemicals were found to be in common. Of these, 82 chemicals (37%) had a difference in NO(A)EL between the two datasets of at least ±0.5 log units, whereas 32 chemicals (15%) had a difference in NO(A)EL of at least ±1 log unit (Supplementary Table 2, Supplementary Figure 1). Therefore, for the vast majority of the overlapping chemicals, any discrepancy in NO(A)EL values was considered to be captured by the variability that is inherent to in vivo studies (Pham et al., 2019; Pham et al. in prep). Additionally, the overall NO(A)EL distributions for the intersecting chemicals was assessed; whilst the distribution of ToxValDB NO(A)ELs was marginally left-shifted compared to that of the Munro NO(A)ELs (Supplementary Figure 2), the distributions were not significantly different according to the K-S test (n = 438).

Figure 2.

Comparison of the cumulative and fitted lognormal distributions for the ToxValDB and Munro NO(A)EL data for each Cramer class. The ToxValDB and Munro Cramer class I and II distributions were not significantly different (p > 0.05). Meanwhile, the Cramer class III distributions were significantly different between the two datasets (p < 0.05)

The ToxValDB 5^th percentile NO(A)EL value was greater than that for the Munro dataset across the three Cramer classes (Table 3). This was especially true for Cramer class II, where the ToxValDB 5^th percentile value was almost 4-fold larger than the Munro value. However, there were relatively few chemicals present in the Cramer class II category for both datasets: 28 and 39 chemicals for Munro and ToxValDB respectively. Thus, a small shift towards less potent NO(A)ELs could have a strong impact on the resulting 5^th percentile value. Indeed, this seemed to be the case for the Cramer class II datasets. Only 6 of the 28 chemicals (21%) in the Munro Cramer class II set had a NO(A)EL ≥100mg/kg-day, whereas 19 of the 39 chemicals (49%) in the ToxValDB set had a NO(A)EL ≥100mg/kg-day. Furthermore, the Munro Cramer class II set contained only one chemical with a NO(A)EL ≥1000mg/kg-day, whilst the ToxValDB Cramer class II set has five chemicals with a NO(A)EL ≥1000mg/kg-day. In addition, 39% of the chemicals in the Munro set had a NO(A)EL ≤10mg/kg-day, whilst this was the case for only 18% of the ToxValDB set.

To identify whether the 5^th percentile values between the two datasets were statistically different, bootstrapping was utilised to calculate the 95% confidence intervals for each Cramer class (Figure 3). As shown in both Figure 3 and Table 3, the 5^th percentile values for Cramer class I and II were not statistically different. Therefore, even though the unprocessed 5^th percentile values for Cramer class II appeared different, there was actually a relatively large overlap in their 95% confidence intervals, likely due to the small number of chemicals assigned to this class. In contrast, the 5^th percentile values for Cramer class III differed significantly between the two datasets. The TTC values are shown in Table 3 for completeness.

Figure 3.

Fifth percentile values and associated 95% confidence intervals calculated using 5000 bootstrap samples for each Cramer class from ToxValDB (in red) and Munro (in blue). Only the 5^th percentile values for the Cramer class III chemicals from the two datasets were seen to be significantly different (p < 0.05).

3.4. Investigation of Cramer class III

Since the Cramer class III 5^th percentile values differed significantly, we sought to investigate whether this was due to an underlying difference in the types of chemicals represented in the ToxValDB and Munro Cramer class III datasets. Comparison of the frequency of ToxPrints present in chemicals in both ToxValDB and Munro Cramer class III datasets provided an initial overview of the differences in chemical landscape. Figure 4 illustrates that the Munro dataset contained a higher frequency of certain ToxPrints, including, but not limited to: aromatic and heterocyclic ring structures, phosphate and phosphonate bonds, (amino) carbonyl structures, and halogen containing chemicals. On the other hand, the ToxValDB dataset contained a higher frequency of chemicals possessing a linear alkane chain.

Figure 4.

Comparison of the frequency of the ToxPrints (after being condensed to the 70 level 2 ToxPrints) for the chemicals in Cramer class III for both ToxValDB (in red) and Munro (in blue).

A chemotype enrichment analysis was utilised to provide a more detailed assessment of which specific ToxPrints differed between the two datasets. A total of 63 chemotypes were calculated to be enriched (OR ≥3 and p-value ≤0.05) in the Munro Cramer class III set compared to corresponding ToxValDB set. Of these, 29 ToxPrints were only observed in the Munro Cramer class III set of chemicals (OR “Inf”, Supplementary Table 3). These 29 ToxPrints were used to investigate what impact if any, removal of chemicals containing at least one of these ToxPrints had on the Munro Cramer class III 5^th percentile, specifically whether a re-derived value remained statistically different from the ToxValDB class III 5^th percentile value. After filtering out chemicals that contained any one, or more, of these 29 ToxPrints, 306 chemicals remained with a re-calculated 5^th percentile NO(A)EL of 0.22mg/kg-day (Table 4); meanwhile, the 142 chemicals that contained at least one of the 29 ToxPrints and were removed had a 5^th percentile value of 0.075mg/kg-day. After bootstrapping, the re-calculated 5^th percentile value was not statistically different to that of ToxValDB Cramer class III although the K-S metrics still reflected a difference in the distributions. This suggested that at least some of the potent chemicals contained some of these 29 ToxPrints. Accordingly, the K-S test was utilised to investigate whether these 29 ToxPrints were actually separating out the more potent chemicals from the Munro class III set. The 306 Munro class III chemicals that did not contain any of the 29 ToxPrints were compared with the 142 chemicals from Munro class III that did contain at least one of the 29 ToxPrints. The two distributions were shown not to be statistically different. As observed in Figure 5A, the two distributions are very closely aligned to one another and even intersect at approximately the mean. Therefore, although the 29 ToxPrints used to filter the Munro class III chemicals raised the 5^th percentile NO(A)EL value, the chemicals extracted were comparable in potency to chemicals that did not contain one of these ToxPrints. The 29 ToxPrints were not able to account the difference in 5^th values between the 2 datasets.

Table 4.

Comparison of the 5^th percentile values for the Munro Cramer class III chemicals that were retained and removed after utilising different methods. This investigation was undertaken to identify potential reasons behind the discrepancy in 5^th percentile values between the Munro and ToxValDB class III chemicals.

Method used for separation	Number of chemicals	Re-derived 5th percentile (mg/kg-day)	Number of chemicals removed	Removed chemical 5th percentile (mg/kg-day)	Statistically different from ToxValDB class III 5th percentile
Chemotype enrichment	306	0.22	142	0.075	No
Original SMARTS	386	0.2	62	0.056	No
Updated SMARTS	397	0.23	51	0.037	No

Figure 5.

Cumulative distribution function and fitted lognormal distribution for the Munro Cramer class III chemicals after being split using A) ToxPrints identified using chemotype enrichment analysis; and B) the OP/carbamates modules developed by Patlewicz et al (2018). After utilising the enriched ToxPrints to separate the Munro class III chemicals, the two distributions were not statistically different; however, after removing chemicals identified as being AChE inhibitors the distributions were significantly different.

Earlier studies have suggested that the Cramer class III value ought to be re-evaluated excluding OP and carbamate insecticides (EFSA, 2012; Leeman et al. 2014; Kroes et al. 2000; Kroes et al. 2004; Munro et al., 2008). This is due to fact that both of these insecticide classes are neurotoxicants that act via inhibiting acetylcholinesterase (AChE) either reversibly (e.g. carbamates) or irreversibly (e.g. OPs) (Colovic et al., 2013). Inhibition of AChE leads to acetylcholine accumulating in the nerve synapse, culminating in over stimulation of the nicotinic and muscarinic acetylcholine receptors and, therefore, increased neurotransmitter activity (Colovic et al., 2013, Naughton and Terry, 2018). Given the results of these earlier studies, OPs and carbamates were identified using Toxtree, and the modules developed by Patlewicz et al. (2018) and removed to determine whether this would instead account for the differences in 5^th percentile values found for the two datasets.

A total of 62 Cramer class III chemicals from the Munro dataset were re-assigned as OPs/carbamates. Excluding these chemicals from the Munro Cramer class III dataset, resulted in a rederived 5^th percentile value of 0.2mg/kg-day (based on 386 NO(A)ELs) (Table 4). This re-calculated 5^th percentile NO(A)EL value was not statistically different from the ToxValDB Cramer class III chemicals. Again, the cumulative distribution between this filtered Munro Cramer class III and the ToxValDB Cramer class III distribution remained statistically different. The 5^th percentile NO(A)EL value of the OPs/carbamates excluded from the Munro Cramer class III set was 0.056mg/kg-day, corresponding to a TTC value of 0.56μg/kg-day (c.f. reported TTC value of 0.3μg/kg-day). Furthermore, the K-S test and a CDF plot was used to investigate whether the distribution of the AChE inhibitors extracted from Munro class III were distinct from the chemicals that remained in the Munro class III. According to the K-S test, the two distributions were shown to be statistically different (p-value = 4.404 × 10⁻⁵) (Figure 5B). Taken together, these results provide a plausible explanation for the difference in 5^th percentile NO(A)EL values between the original Munro and ToxValDB class III datasets.

Previous work by Leeman et al. (2014) recalculated the Cramer class III 5^th percentile NOEL and associated TTC threshold after manual inspection, and removal, of OP/carbamate insecticides from the original Munro dataset. Leeman et al. (2014) identified a total of 40 chemicals as being either an OP or carbamate insecticide. After removing these chemicals, they reported an increase in Cramer class III 5^th percentile value from 0.15mg/kg-day to 0.22mg/kg-day (i.e. to a TTC value of 2.2µg/kg-day). In our study, the number of chemicals identified as AChE inhibitors by Toxtree was greater than those published by Leeman et al. (2014), thus suggesting that the SMARTS patterns contained within the original Toxtree modules were too broad. Therefore, the chemicals identified as OPs and carbamates by Toxtree were more closely inspected to determine what refinements could be made to the SMARTS patterns to make them more specific.

Preliminary inspection revealed that some of the identified OP/carbamates were not OP/carbamate insecticides. For example, albendazole is an anti-helminthic that contains a carbamate moiety attached to a benzimidazole and whose mode of action is inhibiting the polymerisation of β-tubulin into microtubules rather than AChE activity. Furthermore, there were several chemicals identified by Munro et al. (1999) and/or EFSA (2012) as either an OP or carbamate insecticide with AChE activity yet these were not identified by Toxtree modules, e.g. diethyldithiocarbamate, merphos, and glufosinate-ammonium. The list of 40 OPs and carbamate AChE inhibitors identified by Munro et al. (1999) and EFSA (2012) (as referenced by Leeman et al. (2014)) were utilised to generate more specific SMARTS patterns (Supplementary Table 4) and implemented into updated OP and carbamate Toxtree modules. The entire Munro Cramer class III chemicals were reprocessed through Toxtree using the updated OP and carbamate modules; resulting in 51 chemicals identified as AChE inhibitors. The 5^th percentile NO(A)EL value of the remaining 397 chemicals was 0.23mg/kg-day, which is not statistically different from the ToxValDB Cramer class III 5^th percentile NO(A)EL (Table 4). However, the distributions between the updated Cramer and ToxValDB class III datasets still differed significantly. The updated OP/carbamate chemicals excluded from Munro Cramer class III resulted in a 5^th percentile value of 0.037mg/kg-day, which corresponds to a TTC value of 0.37μg/kg-day (c.f. reported TTC value of 0.3μg/kg-day). The updated 5^th percentile values calculated for 1) the Munro Cramer class III chemicals without AChE inhibitors and 2) the OPs/carbamates with the updated Toxtree modules were comparable with those reported by Leeman et al. (2014).

These analyses demonstrate the most plausible explanation for the differences in 5^th percentile NO(A)EL values between the ToxValDB and Munro Cramer class III chemicals was due to the presence of the OP and carbamate insecticides within Munro Cramer class III. Once these substances were excluded from the Munro Cramer class III dataset, the 5^th percentile NO(A)EL values increased from 0.15mg/kg-day to 0.23mg/kg-day. Given that multiple previous studies have also suggested removing the OPs and carbamates from Cramer class III and the Kroes workflow already provides a separate TTC value for these chemicals, based on this study, it does seem appropriate to consider an update to the TTC value for Cramer class III substances (Kroes et al. 2004, Munro et al. 1999, 2008, EFSA 2012, Leeman et al. 2014).

3.5. Practical impact

To illustrate the practical consequence of this change, we processed a publicly-available inventory of chemicals (~45,000 substances) along with their TTC category assignments that were reported by Scitovation as part of ACC LRI supported research (accessed 28^th June 2019). Since the reported assignments were carried out using an earlier version of Toxtree, the substances were re-profiled using the Kroes workflow in the same manner as the ToxValDB chemicals in this study. To investigate what effect the refined OP/carbamates module had on the TTC category assignments, the chemicals were profiled using both the original and the refined OP/carbamates module developed through this study. Of the ~45,000 chemicals in the publicly available list, the vast majority were assigned to the same TTC category irrespective of the OP/carbamate module utilised. However, there were a total of 654 chemicals (0.015%) with a discrepancy in their TTC category assignment between the two OP/carbamate modules (Figure 6, Supplementary Table 5).

Figure 6.

Tile plot comparing the frequency (in log₁₀ space) of TTC assignments for the ~45,000 chemicals in the publicly available dataset using the original OP and carbamate modules and the updated OP and carbamate modules developed in this study. NB: The values present in each tile display the raw number of chemicals.

Between the original and refined OP/carbamate module, 49 chemicals were reassigned from Cramer class I, one chemical (Terbucarb) was reassigned from Cramer class II, and 99 chemicals were reassigned from Cramer class III to an OP/carbamate (Supplementary Table 5). Upon investigation of the chemicals that were reassigned, a number of them were OP/carbamate insecticides that had previously been missed. For example, fenobucarb, formparanate, and propoxur were previously categorised in Cramer class I; terbucarb was previously categorised as Cramer class II, and; aldicarb, carbofuran, carbaryl, and thiodicarb were categorised as Cramer class III. However, all of these chemicals are known insecticides that act via AChE inhibition (Colovic et al. 2013, Knowles and Ahmad 1972).

Meanwhile, the updated OP/carbamate module also identifies some chemicals it perhaps should not, such as bambuterol, ladostigil tartrate, pyridostigmine, and rivastigmine, which are pharmaceuticals; nevertheless, these chemicals act via AChE inhibition (Colovic et al. 2013, Feldman and Karalliedde 1996, Weinstock et al. 2006). For example, rivastigmine is a reversible inhibitor of AChE activity that is used in the treatment of Alzheimer’s disease (Colovic et al. 2013).

Given that only 40 chemicals were utilised in the development of the SMARTS in the updated OP/carbamate module, these false positives are probably to be expected. Additionally, there are still likely some OP/carbamate containing insecticides that are currently being missed because their chemical structure was outside the domain of applicability for the chemicals used in the development of the updated module. Notwithstanding these limitations, the increased identification of OP and carbamate insecticides by the updated module provides an advantage over the original module by further limiting the number of chemicals that were mis-classified. Future work could involve the use of in vitro high-throughput screening data to further refine the SMARTS patterns identified in this study or in the generation of additional SMARTS patterns for OP/carbamate insecticides.

4. Conclusions

Overall, this analysis demonstrates that the TTC values published by Munro et al. (1996) remain consistently below the thresholds derived from the expanded dataset of chemicals of relevance to the EPA (ToxValDB). We were able to utilise bootstrap sampling to calculate 95% confidence intervals surrounding the 5^th percentile NO(A)EL values for the Munro and ToxValDB datasets. Even though the Munro 5^th percentile NO(A)EL values for Cramer class I-III were lower than those using the data from ToxValDB, only the Cramer class III values were significantly different between the two datasets. Chemotype ToxPrint enrichments were explored to identify plausible explanations to account for the variation in 5^th percentile values, but the discrepancies in chemical features were not sufficient to rationalise these differences. Rather, identification and removal of OP and carbamate insecticides from the Munro Cramer class III set, was able to account for the differences, thus, lending further support to previous work by Munro et al. (2008) and others who have proposed updating the TTC value for Cramer class III. The insights were used to refine the existing SMARTS that are used in Toxtree to identify OP and carbamates. The updated module for OPs/carbamates was used to process a large inventory of substances (~45,000) to illustrate the impact these changes had on how substances are assigned and the consequence this has on the TTC values that are applicable. This showed that the updated module was better equipped to identify OP/carbamate insecticides that act via AChE inhibition than the original module. That said the TTC values derived from this expanded dataset of toxicity values offer additional support for the original TTC values derived by Munro et al. (1996) reaffirming the utility of TTC as a promising tool in a risk-based prioritisation approach.

Highlights.

• Substances present in ToxValDB were assigned into their respective TTC categories

• Used ToxValDB toxicity values to derive new Cramer TTC values

• Evaluated whether the Cramer TTC values derived from the ToxValDB and Munro datasets were statistically equivalent

• Compared and contrasted the chemistry of the two datasets to rationalise any (dis)similarities in TTC values

• Study provides increased confidence in the existing TTC values based on the Munro dataset

Abbreviations

AChE inhibitors

acetylcholinesterase inhibitors

CDFs

cumulative distribution functions

EFSA

European Food Safety Authority

EPA

US Environmental Protection Agency

ECHA

European Chemicals Agency

FDA

US Food and Drug Administration

HTE

high-throughput exposure

K-S

Kolmogorov-Smirnov

LO(A)Els

lowest-observed (adverse) effect levels

NO(A)Els

no-observed (adverse) effect levels

OR

odds ratio

POD

point of departure

SMARTS

SMILES arbitrary target specification

SMILES

Simplified Molecular-Input Line-Entry System

TSCA

Toxic Substances Control Act

TTC

Threshold of Toxicological Concern

ToxValDB

Toxicity Values database

WHO

World Health Organisation