Advancing the Science of a Read-across Framework for Evaluation of Data-poor Chemicals Incorporating Systematic and New Approach Methods

Abstract

The assessment of human health hazards posed by chemicals traditionally relies on toxicity studies in experimental animals. However, most chemicals currently in commerce do not meet the minimum data requirements for hazard identification and dose-response analysis in human health risk assessment. Previously, we introduced a read-across framework designed to address data gaps for screening-level assessment of chemicals with insufficient in vivo toxicity information (Wang et al., 2012). It relies on inference by analogy from suitably tested source analogues to a target chemical, based on structural, toxicokinetic, and toxicodynamic similarity. This approach has been used for dose-response assessment of data-poor chemicals relevant to the U.S. EPA’s Superfund program. We present herein, case studies of the application of this framework, highlighting specific examples of the use of biological similarity for chemical grouping and quantitative read-across. Based on practical knowledge and technological advances in the fields of read-across and predictive toxicology, we propose a revised framework. It includes important considerations for problem formulation, systematic review, target chemical analysis, analogue identification, analogue evaluation, and incorporation of new approach methods. This work emphasizes the integration of systematic methods and alternative toxicity testing data and tools in chemical risk assessment to inform regulatory decision-making.

2. INTRODUCTION

Read-across is a technique for filling knowledge gaps in toxicology information by analogy between well-studied and poorly studied chemicals. It is increasingly used to fill data gaps for both qualitative and quantitative chemical risk assessments. Read-across predictions often rely on conservative estimates to avoid underestimation of hazard. For example, read-across can be used to fulfill information requirements under European Union Registration, Evaluation, Authorisation and Restriction of Chemicals (REACH) regulations (EC, 2006) and the U.S. Environmental Protection Agency’s (EPA’s) High Production Volume Challenge Program (Bishop et al., 2012). Models and tools developed by the EPA to support read-across methodologies are available to assess chemical hazard under the Toxic Substances Control Act (TSCA) (https://www.epa.gov/tsca-screening-tools) and CompTox Chemicals Dashboard (https://comptox.epa.gov/dashboard). Additionally, workflows and technical guidelines for grouping chemicals based on category and analogue approaches have been developed to assist in read-across assessments (ECETOC, 2012; ECHA, 2017; OECD, 2014; Wang et al., 2012; Wu et al., 2010).

A critical mission for the U.S. EPA entails the evaluation of chemicals found at contaminated sites (e.g., Superfund sites), which commonly lack toxicity data and/or toxicity values and, therefore, are not considered in the quantitative interpretation of site- or media-specific hazards. To initially address this issue, a read-across framework was developed to identify and evaluate chemical analogues for quantitative read-across based on consideration of three main types of similarity: structural, metabolic, and toxicity-like (Wang et al., 2012). If sufficient similarity is demonstrated, a single “best” surrogate (i.e., source analogue) is identified, and its point of departure (POD) (e.g., benchmark dose or no-observed-adverse-effect-level) is adopted for screening-level quantitative assessment of the target chemical. Any known uncertainties and limitations of the adopted POD value are clearly documented. Although our work has focused on the assessment of environmental chemicals, we believe that the read-across framework and methods discussed herein are applicable to other use cases (e.g., occupational chemicals, food additives and pharmaceuticals).

The methodology outlined in the Wang et al. (2012) framework has significantly advanced the science of human health assessments at U.S. EPA. Its benefits have been demonstrated by its use to derive screening-level toxicity values for several data-poor chemicals in Provisional Peer-Reviewed Toxicity Value (PPRTV) assessments for the Superfund program (https://www.epa.gov/pprtv) since 2012. Through application of the methodology over several years, unforeseen challenges and circumstances have arisen, and a significant amount of practical information has been ascertained, specifically pertaining to the strengths and limitations of the original conceptualization of the framework. Additionally, there have been new developments in systematic literature search approaches, and computational tools and data streams that could be leveraged to improve existing read-across frameworks and to gain broader acceptance of read-across methodologies in regulatory decision-making (Patlewicz et al., 2018; Patlewicz et al., 2017). Altogether, an update of the read-across methodology initially presented in Wang et al., (2012) is imperative.

A set of case studies are presented herein to illustrate the application of the Wang et al., (2012) read-across framework for the U.S. EPA PPRTV program. These examples highlight general and case-specific learnings from the practical implementation of this methodology for quantitative read-across. We also propose a revised read-across framework that addresses key elements such as problem formulation; systematic review and target chemical profiling; expanded analogue identification; analogue evaluation by weight of evidence (WoE); and incorporation of data from new approach methods (NAMs).

3. CASE STUDIES

The case studies below were selected to demonstrate the practical application of a read-across framework for development of screening-level PPRTVs for noncancer repeat-dose toxicity via oral and inhalation exposure routes (Wang et al., 2012). Specifically, the case studies highlight the utility of metabolic and mechanistic considerations as a basis for the read-across justification. These examples also emphasize the role of expert knowledge in characterizing interdependencies between structural and toxicokinetic/toxicodynamic properties to build a similarity justification for read-across. Key learnings are also discussed in this section to support the changes and improvements in the revised read-across framework presented in Section 4.

2.1. Pentamethylphosphoramide (CASRN¹ 10159-46-3; DTXSID²60144035) and N,N,N’,N”-tetramethylphosphoramide (CASRN 16853-36-4; DTXSID80168576)

Hexamethylphosphoramide (HMPA) is the only candidate analogue that was identified for the target chemicals, pentamethylphosphoramide (PMPA) and N,N,N’,N”-tetramethylphosphoramide (TMPA) on the basis of structural similarity and existing oral toxicity values (Table 1) (U.S. EPA, 2020b; c). Because no candidate analogues with inhalation toxicity values were identified for the targets, the oral route was the focus for this read-cross evaluation. HMPA, PMPA and TMPA are phosphoric acid amine derivatives with common, general structural features: a phosphorous oxygen double bond, nitrogen-phosphorous bonds, and secondary or tertiary amides.

Table 1.

Candidate Analogue for Pentamethylphosphoramide and N,N,N’,N”-Tetramethylphosphoramide^a

Role	Targets		Analogue
Name	Pentamethylphosphoramide (PMPA)	N,N,N’,N”-tetramethylphosphoramide (TMPA)	Hexamethylphosphoramide (HMPA)
Structure
CASRN	10159-46-3	16853-36-4	680-31-9
DTXSID	60144035	80168576	6020694

^aAnalogue was identified by structural similarity search tools and availability of oral or inhalation toxicity values (U.S. EPA, 2020b; c).

HMPA is also a metabolic precursor of the target chemicals. PMPA and TMPA are primary intermediate metabolites of HMPA that result from sequential demethylation of the parent compound by cytochrome P450 (CYP450) (Figure 1). It is also notable that formaldehyde is produced from methylol intermediates generated after each demethylation step (Jones and Jackson, 1968). Both HMPA and formaldehyde target the upper respiratory tract (primarily the nasal cavity) and cause degenerative lesions in rats (Jones and Jackson, 1968; U.S. EPA, 2012). HMPA is a potent nasal toxicant via the oral and inhalation routes and the POD used in the derivation of oral toxicity values for this compound was based on nasal and tracheal lesions at doses ≥ 15 mg/kg-day (U.S. EPA, 2012). The adverse nasal effects in animals following HMPA exposure via the oral route are attributable to preferential deposition (Keller et al., 1997) and metabolism of the parent compound in nasal tissue (Dahl and Brezinski, 1985). Because HMPA, PMPA, and TMPA share a similar bioactivation pathway, the proposed mechanism of action for HMPA-induced nasal toxicity is plausible for PMPA and TMPA. As such, HMPA was considered a suitable analogue and the source chemical for deriving screening-level oral toxicity values for the target chemicals (U.S. EPA, 2020b; c). This case study provides an example of the use of a metabolic precursor of the target as a source analogue for read-across.

Figure 1.

Metabolism of HMPA and demethylated products (U.S. EPA, 2020b; c).

2.2. 4-Methyl-2-pentanol (CASRN 108-11-2; DTXSID2026781)

Five structural candidate analogues with existing oral or inhalation toxicity values were identified for the target, 4-methyl-2-pentanol or methyl isobutyl carbinol (MIBC), including its ketone derivative, methyl isobutyl ketone (MIBK) (Table 2) (U.S. EPA, 2017). MIBC and MIBK are metabolized to one another in a process that favors the formation of a major metabolite with comparable pharmacokinetics, 4-methyl-4-hydroxy-2-pentanone (HMP) (Figure 2) (Gingell et al., 2003; Granvil et al., 1994). The other candidate analogues are structurally related aliphatic C2-alcohol/ketones (2-propanol, 2-propanone, 2-butanone, and 2-hexanone) (Table 2). Similar to MIBC and MIBK, these analogues display rapid absorption, wide distribution in the body, bidirectional metabolism between C2-alcohol/ketone pairs (2-propanol/2-propane; 2-butanol/2-butanone; 2-hexanone/2-hexanol), and generally low acute toxicity in rodent lethality studies (LD50s range from 1,900 – 5,800 mg/kg and LC50s range from 8,000 – 100,000 mg/m³) (U.S. EPA, 2017).

Table 2.

Candidate Analogues for Methyl Isobutyl Carbinol^a

Role	Target	Analogues
Name	4-Methyl-2-Pentanol or methyl isobutyl carbinol (MIBC)	4-Methyl-2-Pentanone or methyl isobutyl ketone (MIBK)	2-Propanol	2-Propanone	2-Butanone	2-Hexanone
Structure
CASRN	108-11-2	108-10-1	67-63-0	67-64-1	78-93-3	591-78-6
DTXSID	2026781	5021889	7020762	8021482	3021516	0022068

^aSet of analogues identified by structural similarity search tools and availability of oral or inhalation toxicity values (U.S. EPA, 2017).

Figure 2.

Bidirectional Metabolism between methyl isobutyl carbinol (MIBC) and methyl isobutyl ketone (MIBK) (U.S. EPA, 2017).

Repeated-dose toxicity studies for the candidate analogues in animals exposed via the inhalation or oral route (and HMP, the shared metabolite between MIBC and MIBK) revealed common health effects. All analogues caused developmental/reproductive, neurological, liver, and kidney effects, but the relative potencies of such health effects varied across these chemicals (U.S. EPA, 2017). Limited information from a secondary source described an unpublished 6-week inhalation study of the target chemical in rats. It reported potential kidney-related effects: presence of ketone bodies in the urine in males and females at ≥ 825 mg/m³ and proteinuria and increased kidney weights in males at 3,700 mg/m³ (Blair et al., (1982) cited in (OECD, 2005). Notably, the most sensitive toxic effect for the candidate analogue 2-hexanone involves peripheral neuropathy associated with the production of its metabolite, 2,5-hexanedione (U.S. EPA, 2009). Formation of a similar dione product is not expected for the target chemical (Duguay and Plaa, 1995). Therefore, 2-hexanone was not considered a suitable analogue for MIBC.

Of the remining candidate analogues, MIBK was selected as the most suitable source analogue for MIBC for both oral and inhalation routes based primarily on the bidirectional metabolic relationship between these chemicals and similar in vivo pharmacokinetics. The POD used in the derivation of inhalation toxicity values for MIBK was adopted for screening-level inhalation assessment of the target chemical, MIBC (U.S. EPA, 2017). Since no oral toxicity values were available for the selected source analogue, screening-level oral assessment of MIBK was not attempted. The case study shows the use of a metabolite of the target as a source analogue for read-across.

2.3. 1-Bromo-2-chloroethane (CASRN 107-04-0; DTXSID4024775)

The search for structural analogues for the target, 1-bromo-2-chloroethane, yielded three candidates: 1,2-dibromoethane, 1,2-dichloroethane, and 1,2-dibromo-3-chloropropane (Table 3) (U.S. EPA, 2021c). The candidate analogues had oral or inhalation toxicity values and shared the following structural attributes: presence of two or three halogen substituents; halogens attached to adjacent carbon atoms (1,2-substitution patterns); and no more than one halogen per carbon atom.

Table 3.

Candidate Analogues for 1-Bromo-2-Chloroethane (CASRN 107-04-0)^a

Role	Target	Analogues
Name	1-Bromo-2-Chloroethane	1,2-Dibromoethane	1,2-Dichloroethane	1,2-Dibromo-3-Chloropropane (2R and 2S Isomers)
Structure
CASRN	107-04-0	106-93-4	107-06-2	96-12-8
DTXSID	4024775	3020415	6020438	3020413

^aSet of analogues identified by structural similarity search tools and availability of oral or inhalation toxicity values (U.S. EPA, 2021c).

Differences in the chemical reactivity of brominated versus chlorinated analogue chemicals is anticipated to influence the toxicity of these halogenated alkanes and was used as a basis for evaluating the suitability of the analogues for read-across. Experimental studies for the candidate analogues provide evidence of two major metabolic pathways for this group of chemicals: cytochrome P450 (CYP450) oxidation and direct glutathione (GSH) conjugation (U.S. EPA, 2021c). GSH conjugation has also been demonstrated for the target chemical in vivo and in vitro in rats (Jean and Reed, 1992; Marchand and Reed, 1989). Data for the candidate analogues indicate that GSH conjugation leads to the formation of reactive episulfonium metabolites that can bind to DNA and protein (reviewed by (Dekant and Vamvakas, 1993; Guengerich, 1994; U.S. EPA, 2021c)). Nonetheless, the rate of GSH conjugation is expected to be faster for 1-bromo-2-chloroethane and the brominated analogues due to the greater reactivity of the bromine leaving group. In this context 1,2-dibromoethane and 1,2-dibromo-3-chloropropane are considered to exhibit increased metabolic relevance, in comparison to 1,2-dichloroethane, which has a less reactive chlorine leaving group. This is supported by evidence of preferential GSH conjugation at the brominated versus chlorinated sites of 1-bromo-2-chloroethane and the analogue, 1,2-dibromo-3-chloropropane (Dekant and Vamvakas, 1993; Humphreys et al., 1991; Jean and Reed, 1992; Marchand and Reed, 1989).

Testicular toxicity is a sensitive effect for the analogues, 1,2-dibromoethane and 1,2-dibromo-3-chloropropane, in experimental animals via oral or inhalation exposure (U.S. EPA, 2021c). Additionally, occupational exposure to 1,2-dibromo-3-chloropropane has been associated with altered sperm parameters (decreased spermatogenesis and sperm count and altered sperm morphology) in humans (U.S. EPA, 2003; 2006). In particular, the testicular toxicity of 1,2-dibromo-3-chloropropane has been linked to the formation of reactive episulfonium metabolites via the GSH conjugation pathway (Omichinski et al., 1988; Søderlund et al., 1988). Indeed, reactive episulfonium metabolites induce testicular DNA damage, which results in degenerative testicular lesions (necrosis) and impaired spermatogenesis. The respiratory tract (primarily nasal lesions in rats and mice) was also a sensitive measure of toxicity following inhalation exposure to 1,2-dibromoethane and 1,2-dibromo-3-chloropropane, which may result in part from the release of HBr during metabolism of these chemicals (via CYP oxidation and GSH conjugation) (U.S. EPA, 2021c). HBr gas has been shown to be a nasal irritant in studies involving animals and humans (NRC, 2014). The absence of testicular and respiratory tract toxicity associated with exposure to the analogue, 1,2-dichloroethane, corroborates with the expected reduced chemical reactivity and slower rate of metabolic bioactivation for the chlorinated versus brominated chemicals (U.S. EPA, 2018).

Overall, the two brominated analogues (1,2-dibromoethane and 1,2-dibromo-3-chloropropane) were considered suitable analogues for the target, 1-bromo-2-chloroethane. Similarities occurred in chemical reactivity of the shared bromine leaving group and expected rate of potential metabolic activation (via GSH conjugation and CYP450 oxidation) associated with the primary toxicity effects of interest for this group of chemicals (testicular and respiratory tract effects). 1,2-dibromo-3-chloropropane was ultimately selected as the source analogue since it provided the most sensitive and health-protective POD for quantitative read-across via oral and inhalation routes (U.S. EPA, 2021c). This case study showcases the use of mechanistic information relating to chemical reactivity and metabolic activation as a primary basis for building a justification for read-across.

2.4. Key Learnings

The case studies were subject to an internal and external peer review process, and we believe that the read-across conclusions are well justified. However, through these and other read across applications, additional refinements and methodological advancements were considered to increase transparency and ensure continued defensibility in the science supporting regulatory decision making.

Structural similarity was used to compile a pool of candidate analogues for the target chemicals discussed previously and is routinely used as a primary basis for the analogue search in read-across assessments under the PPRTV program. Structural similarity searches are conducted using tools such as the U.S. EPA’s Distributed Structure-Searchable Toxicity database (DSSTox) (DSSTox, 2022)³ and National Library of Medicine’s ChemIDplus database (ChemIDplus, 2022). Structural similarity search tools employ a variety of quantitative methods (e.g., Tanimoto) and structural descriptors or fingerprints to identify and rank analogues based on similarity indices or scores. We have learned that similarity scores can vary widely across tools and can be disproportionately influenced by changes in, or absence of, single descriptors. The scores also fail to capture similarities or differences in general structural features, functional groups and chemical reactivity that may influence toxicokinetics and toxicodynamics and, thus, impact the expected toxicity. As a result, we determined that it was important to include information other than structure when identifying potential analogues. Because we rely on more than similarity scores, significant expert input is necessary to compile an adequate pool of analogues, recognizing any interdependencies between structural and biological (e.g., metabolism, toxicity, and mode of action [MOA]) properties as illustrated in the case studies.

As described in the Wang et al. (2012) workflow, only chemicals with toxicity values based on repeat-dose studies from authoritative sources are retained for further evaluation. Authoritative sources include, the U.S. EPA’s Integrated Risk Information System (IRIS) database (https://www.epa.gov/iris) and PPRTV electronic library (www.epa.gov/pprtv), the U.S. Department of Health and Human Services’ Agency for Toxic Substances and Disease Registry (ATSDR) Toxic Substance Portal (https://www.atsdr.cdc.gov/substances/index.asp) and California EPA’s Office of Environmental Health Hazard Assessment (OEHHA) chemical database (https://oehha.ca.gov/chemicals). The use of analogues from well-vetted toxicity databases and established protective toxicity values from several health agencies increases confidence in the read-across prediction. Also, limiting the sources to a few authorities facilitates the selection of a source analogue and its corresponding POD for quantitative read-across. Nonetheless, this requirement poses some challenges by restricting the pool of candidates and the evaluation of chemical categories as proposed by the Organisation for Economic Cooperation and Development (OECD) guidelines (OECD, 2014). However, we have found that this restriction does not excessively limit the read-across process.

The first two case studies emphasize the use of metabolic relationships, including metabolic precursors or metabolites of the target chemical as a basis for selecting source analogues for read-across. The third case study shows how mechanistic considerations relating to chemical reactivity and potential metabolic bioactivation can be used to inform and refine the selection of source analogues for read-across. Similarity comparisons for toxicokinetic and toxicodynamic properties relevant to read-across require experimental data on the target chemical and analogues, preferably from in vivo studies. However, this information may be sparse or altogether lacking for target chemicals with severe database deficiencies, necessitating reliance mostly on structural inference to support the similarity hypothesis for read-across.

To overcome these challenges, updates to the read-across framework are proposed (details are provided in Section 4). A comprehensive review of the available literature is included that applies principles of systematic review and profiles the target chemical using tools that integrate both chemistry and toxicity data. The information can be used to determine key attributes of the target chemical to refine the strategy for identifying and evaluating analogues for read-across. The process for identifying analogues has been expanded to encompass additional types of evidence (i.e., structural similarity, toxicokinetic and toxicodynamic properties, and membership in an existing chemical category) and to consider analogues that provide toxicokinetic and toxicodynamic information even if they lack toxicity values. Considerations for the evaluation of analogues via WoE and integration of NAM data and tools are discussed to build confidence and enhance expert judgement. Finally, a problem formulation step has been added to generalize the read-across process beyond PPRTVs.

4. REVISED READ-ACROSS FRAMEWORK

The revised framework is informed by practical experience and common read-across principles and methods previously outlined in the scientific literature (Patlewicz et al., 2015; Patlewicz et al., 2013; Patlewicz et al., 2018). It emphasizes biological similarity as a means for identifying suitable analogues, considers NAM data (e.g., in vitro bioactivity) and tools, and seeks to enhance transparency and scientific rigor in read-across assessments. Although the primary focus of the Wang et al. (2012) read-across approach has been on non-cancer quantitative assessment for PPRTV derivation, the revised framework is intended to be flexible and generalizable to other endpoints⁴ (e.g., qualitative hazard identification) and regulatory decision contexts. It can be easily adapted for both analogue and categorical approaches to read-across (OECD, 2017b). In an analogue approach, endpoints for the target chemical are estimated using test data from one or a few analogous chemicals, deemed to be sufficiently similar. In contrast, a categorical approach groups chemicals with similar toxicological properties or that follow a trend for a specific property based on structural attributes rather than matching pairs of target and analogue chemicals. The membership of the target chemicals in the category is determined, and the endpoint is estimated based on its membership in the category and relationship to other members of the category.

The five major steps in the revised framework are depicted in Figure 3 and discussed in more detail in the sections below: (1) Conduct a problem formulation stating the goals or risk-based decisions, methods of the assessment, and the acceptable level of confidence for the decision(s). (2) Systematically review the literature and use computational tools to identify information on the target chemical pertinent to the assessment. If the available information on the target chemical is inadequate to directly characterize the endpoint, attempt a read-across evaluation to fill the data gaps. (3) Initiate the read-across process by searching for candidate analogues on the basis of structural similarity, toxicokinetic, and toxicodynamic properties, or membership in an existing chemical category. Refine the pool of candidate analogues to select those that will be evaluated by WoE. (4) Assemble and weigh evidence for the target chemical and candidate analogues to evaluate suitability for read-across. (5) Characterize the endpoint by selecting the source analogue(s). If no suitable analogues are identified, then the problem formulation is reevaluated.

Figure 3.

Revised read-across framework. Adapted from (Wang et al., 2012) and (Patlewicz et al., 2018)

3.1. Step 1: Problem formulation

The problem formulation step defines the problem to be addressed by an assessment, that may include read-across as data gap filling method (Patlewicz et al., 2018; U.S. EPA, 1992; 1998). In this step, the nature of the decision context, endpoint(s) of interest, and acceptable level of confidence in the methods and results are determined to the extent possible prior to initiating the assessment. The acceptable level of confidence is defined by the decision context and intended use of the assessment. These considerations will inform the necessary effort, information, and resources needed to reach the risk decision(s). They can also help to determine the appropriateness of read-across relative to other approaches for data gap filling in the absence of adequate chemical-specific information. If data generation is an option, simple inexpensive tests to determine physical-chemical attributes or fill metabolic and mechanistic data gaps may be appropriate. If available data are insufficient to identify an endpoint (e.g., POD) by read-across (e.g., no analogues with relevant endpoint data are identified), other available NAM data (e.g., from consensus structure-activity relationship models, high-throughput screening [HTS], or omics approaches) may be used to characterize the endpoint for certain decision contexts (e.g., screening and prioritization), in which case read-across may be used to inform conclusions about hazard, biological plausibility or dose-response analysis. Problem formulation is an iterative process that can be revised at any step of the revised framework as new insights and developments on the systematic review of the literature, target chemical profiling or analogue identification and evaluation arise.

A problem formulation for a read-across assessment should include, as appropriate:

1. A clear definition of the target chemical: names, formula, CASRN, DTXSID, or structural descriptor, as in SMILES⁵.

2. Endpoint(s): POD for a reference dose or cancer slope factor, acute median lethal dose, etc., for a route of exposure, including oral dose, dermal dose, and atmospheric concentration.

3. Possible data gap filling techniques: read-across, predictive model, other NAMs, etc.

4. Acceptable results: including defining the level of confidence required for the decision context.

5. Others: such as identification of susceptible populations or lifestages.

3.2. Step 2: Systematic review and target chemical profiling

Once the problem formulation has defined the desired information for the assessment, existing information on the target chemical is gathered by systematically searching and screening the literature and extracting the relevant information. Systematic review practices provide a consistent, transparent and reproducible approach for assembling information for evidence synthesis and evaluation (NTP, 2019; Suter et al., 2020; U.S. EPA, 2020a; 2021a; WHO, 2021). The methods for the literature search and inclusion/exclusion criteria for screening studies are identified and tailored to the objectives stated in the problem formulation. For PPRTVs, studies in humans and experimental mammalian models that are suitable for hazard identification and dose-response analysis are prioritized for extraction and synthesis. Experimental toxicokinetic and mechanistic studies (from in vivo and in vitro systems) and other data (e.g., physicochemical) that are potentially relevant to read-across are tracked as supplemental materials.

If the available information is deemed adequate for hazard and dose-response analysis according to the objectives outlined in the problem formulation, the assessment should proceed using chemical-specific information (see the decision box in Fig. 3). Otherwise, a read-across approach is pursued, and the information gathered during the literature search and screening is used as appropriate in subsequent analogue identification and evaluation steps.

In addition to the databases usually included in systematic reviews, computational tools such as the EPA’s CompTox Chemicals Dashboard (Williams et al., 2021) and OECD Quantitative Structure-Activity Relationship (QSAR) toolbox (OECD, 2022) can be used to profile the target chemical. These tools can help identify key structural features such as functional groups and structural alerts that may influence chemical reactivity, metabolism, or toxicity; experimental and predicted physicochemical and environmental fate properties relevant to bioavailability; information on existing chemical categories; in silico metabolism predictions (e.g., in vitro/in vivo rat metabolic simulators within the OCED QSAR Toolbox); in vitro high-throughput data and QSAR models for various toxicokinetic parameters (e.g., absorption, distribution, metabolism and excretion [ADME] module within the EPA’s CompTox Chemical Dashboard); toxicity data for non-cancer and cancer health effects from peer-reviewed studies, health assessment/regulatory databases and other sources of information not publicly available (i.e., grey literature); SAR/QSAR models for toxicity endpoints (e.g., acute toxicity and reproductive/developmental toxicity); and data from HTS or omics technologies that provide insights into potential mechanism of toxicity (e.g., in vitro bioactivity from EPA’s Toxicity Forecaster [ToxCast] and Tox21 assays). This information may be particularly useful for target chemicals with significant database deficiencies in formulating the search strategy for candidate analogues and facilitate similarity comparisons for the evaluation of analogues via WoE.

3.2.1. Adequate Information on target chemical?

This decision box in the revised framework determines whether a read-across evaluation should be pursued. As mentioned previously, if information identified in step 2 is sufficient to characterize the endpoint for the target chemical, the assessment proceeds using current U.S. EPA human health risk assessment guidance and practice. If not, proceed to step 3.

3.3. Step 3: Analogue identification

3.3.1. Compile initial pool of analogues

The first step in the read-across process is identifying relevant candidate analogues for the target chemical using structural similarity, toxicokinetic and toxicodynamic properties and/or category membership. The approach for the analogue search is guided by the results of systematic review and target chemical profiling (step 2). For instance, if the target chemical is known to be metabolized via a bioactivation or detoxification pathway associated with a specific adverse health outcome, then a search for candidate analogues that share the same pathway is prioritized. If the systematic review/target chemical profiling does not provide useful information, the analogue query would default to structural similarity.

An initial pool of candidate analogues is compiled based on one or more of the following approaches as appropriate.

• Use structural similarity search databases (e.g., EPA’s CompTox Chemicals Dashboard and OECD QSAR Toolbox) to find candidate analogues, implementing a threshold cutoff value (e.g., > 50%) based on the number of analogues initially retrieved or any prior knowledge of the chemical category or group. Then narrow the field of candidates based on expert judgement, taking into account general structural features, functional groups, and chemical reactivity important for the toxicokinetic and toxicodynamic behavior of the target and analogues.

• Use toxicokinetic data to search for metabolic precursors, metabolites, reactive intermediates, and ultimate toxicants of the target, or chemicals with a similar detoxification/bioactivation pathway and consider them as candidate analogues. As previously suggested, when test data are limited, predictive software publicly and commercially available may be useful in retrieving information on metabolism to search for analogues (Boyce et al., 2022; Peach et al., 2012; Zakharov et al., 2012).

• Identify chemicals that share a similar toxicity pathway with the target chemical that is linked to a toxic effect or adverse outcome. This may be done by targeting the molecular initiating event or key events in an adverse outcome pathway (AOP) or MOA framework. In addition, HTS and omics approaches that cover a wide range of molecular and cellular responses can be leveraged to identify chemicals with a similar bioactivity or transcriptomic profile (See section 5 for more details).

• If the target chemical is part of an existing chemical category, consider other members of the category as candidate analogues, particularly those that are more similar in structure, toxicokinetics, and/or toxicity effect/pathway (i.e., nearest neighbor[s]). Examples of chemical categories include acid chlorides and alkoxysilanes as defined by the EPA New Chemicals Program (U.S. EPA, 2010). Understanding the basis for the chemical grouping and the boundaries of the existing category (e.g., carbon chain length, presence or absence of functional groups, or range of toxic potencies) is necessary to select analogues relevant to the read-across assessment.

3.3.2. Screen and refine analogues

The analogue pool is screened to ensure that the identified candidates can be of use. The essential attribute of a source analogue is available data for the endpoint or adverse health outcome. This is the information that is read across from the source analogue to the target chemical. For read-across in PPRTV applications, candidate analogues with available toxicity values from authoritative sources (e.g., EPA, ATSDR, CalEPA) are typically retained. In addition, candidate analogues without endpoint data that have toxicokinetic/toxicodynamic information may be included in the read-across particularly in the context of a categorical approach. Such chemicals can be used to make similarity comparisons across members of the category and to understand any interdependencies between structural and biological properties.

If a large number of candidate analogues have available data for the endpoint, expert judgment may be required to refine the list of candidates to advance to step 4 (analogue evaluation). Analogues identified by more than one approach described above can be prioritized. For example, analogues identified by metabolic or mechanistic similarity that also share a high structural similarity with the target are advanced for analogue evaluation. In addition, considerations for physicochemical properties relevant to bioavailability (e.g., molecular weight, water solubility, and octanol/water partition coefficient [log K_OW]), toxicokinetics beyond metabolism (e.g., rate of absorption, patterns of distribution, and rate and route of excretion), and toxicodynamics (e.g., expected toxicity and MOA) can be used to screen out analogues that exhibit notable differences from the target.

3.4. Step 4: Analogue Evaluation

To infer that an analogue is sufficiently similar to the target chemical, one must assemble the body of relevant evidence and weigh the evidence against the hypothesis of similarity of the target to the analogues previously identified. The appropriate method for weighing evidence depends on the available information. At one extreme, the only information that is available for comparison between the target chemical and an analogue is the chemical structure, so a narrative description of similarities of structure may be sufficient. At the other extreme, multiple types of evidence for analogy, including data from NAMs, may require a systematic method to clearly weigh the evidence (Suter and Lizarraga, 2022). Other considerations in choosing a WoE method may include the purpose and required level of confidence and the time and resources available for the assessment.

3.4.1. Assemble evidence for target and analogues

Information concerning the target chemical and analogues collected by the systematic review should be organized by type and should be determined to constitute evidence of toxicological similarity (Suter and Lizarraga, 2022). For read-across, we often use broadly-defined evidence types: structural attributes, physical-chemical attributes, toxicokinetics, and toxicodynamics (Table 4). These standard types provide a basis for aggregating information into sets of evidence that are sufficiently uniform to be evaluated in a common scale (Suter and Lizarraga, 2022). By presenting the types of evidence, we show how much of the potential range of evidence for toxicological similarity is available. Subtypes of evidence are also identified based on the available information (Table 4). For example, toxicokinetic evidence is divided into absorption, distribution, metabolism, and excretion. For target chemicals with limited or no experimental data from traditional toxicokinetic and toxicodynamic studies, evidence from NAMs (e.g., in silico model predictions and HTS or omics technologies) can be sought out to facilitate similarity comparisons between target and analogues (see section 5 for more details on the use of in vitro bioactivity and transcriptomics data to inform toxicodynamic similarity for read-across).

Table 4.

Types of Read-across Evidence for Evaluating Human Health Endpoints^a

Types of Evidence	Specific Types of Evidence
Structural attributes	Major structural features, moieties, functional groups, structural alerts and membership in an existing chemical category or group.
Physiochemical attributes	Experimental or model predictions related to bioavailability, bioaccumulation, or environmental fate (e.g., water solubility, log Kow, bioconcentration factor [BCF]).
Toxicokinetics — absorption	Studies in humans, animals, in vitro assays, or model predictions examining rates and extents of absorption relevant to the endpoint.
Toxicokinetics — distribution	Studies in humans, animals, in vitro assays, or model predictions examining patterns and extents of internal distribution relevant to the endpoint.
Toxicokinetics — metabolism	Studies in human, animals, in vitro assays, or models examining pathways, metabolites, rates, and extents of metabolism relevant to the endpoint.
Toxicokinetics— excretion	Studies in human, animals, in vitro assays, or models examining routes and rates of excretion relevant to the endpoint.
Toxicodynamics – toxicity	Studies in human, animals, in vitro assays, or model predictions examining target organs, health effects, and dose-response relationships for repeat-dose or acute exposure relevant to the endpoint.
Toxicodynamics — mechanisms	Studies in human, animals, in vitro assays, or model predictions related to the potential mechanisms of toxicity relevant to the endpoint.

^aSuter and Lizarraga (2022).

Information becomes evidence when it informs a hypothesis. For read-across, the hypothesis is that the target is sufficiently similar to one or more analogues to justify read-across from the analogues to the target. When organizing the evidence, it is important to identify how it informs that hypothesis, including identifying its logical implications (Suter and Lizarraga, 2022). For example, the implication of in vitro assays for the similarity hypothesis may be that the target and analogues exert similar estrogen receptor activity, which has been associated with reproductive effects in vivo (Lizarraga et al., 2019). In contrast, some properties (e.g., structural similarity scores) have vague or indirect implications for the endpoint or toxic effect and therefore would have low relevance to the hypothesis and could be misleading for similarity determinations. Because of the importance of avoiding irrelevant similarity, identifying the implications of evidence is imperative for a read-across toxicological assessment (Suter and Lizarraga, 2022).

3.4.2. Identify suitable analogues based on weight of evidence

Once evidence is assembled and organized, it is evaluated by WoE. Narrative WoE evaluates the similarity of chemicals by discussing the evidence. The narrative approach is highly flexible and can be convincing if the body of evidence is relatively small and points to an obvious conclusion. Systematic WoE methods are available that rely on considerations like Hill’s (1965) criteria (Owens et al., 2017; U.S. EPA, 2005) or inferential systems (EFSA, 2017; U.S. EPA, 2016) to evaluate the strength of the evidence and provide a judgement of the confidence in the conclusions reached. A systematic WoE approach for read-across has been previously proposed that assigns weights to each type of evidence of similarity with respect to its relevance, strength, and reliability or other properties of the evidence (Suter and Lizarraga, 2022). Several resources are also available that outline specific concepts for the evaluation of similarity and uncertainty in read-across assessments (Blackburn and Stuard, 2014; Schultz et al., 2015; Schultz et al., 2019). Systematic WoE methods provide consistency and transparency and can help clarify the relationship among multiple types of evidence for multiple analogues (Suter and Lizarraga, 2022) but the process is relatively labor intensive. The approach for WoE selected should reflect the goals of the read-across assessment and the level of confidence desired for the decision context in question.

The conclusions of a WoE analysis are obtained by evaluating the evidence for each analogue. For a categorical approach, evidence for the category rather than the individual analogues is assessed. The WoE results can be used to determine which analogues are deemed to be suitable and which is the most suitable or whether membership of the target in a chemical category can be established.

3.5. Step 5: Characterize the Endpoint

The last step of a read-across process is the characterization of the endpoint for the target chemical using information from one or more suitable analogues referred to as the source analogue(s). Suitability is determined by the WoE analysis and dependent upon the confidence required for the decision context to be supported. If a clear, most suitable analogue cannot be discerned from the WoE analysis in step 4, the closest structural analogue or most health protective analogue (e.g., most potent for the endpoint) can be chosen for the read-across. A trend analysis can also be implemented to estimate the endpoint for the target using test data from other members in a category by performing an internal QSAR model (OECD, 2014). If sufficient support for the selection of source analogues for read-across cannot be substantiated, then the scope and parameters of the assessment must be reexamined (Step 1: Problem formulation).

In the context of PPRTVs, the POD from the single best source analogue is adopted to derive screening-level non-cancer toxicity values for the target chemical. Uncertainty factors (UFs) are used for dose-response extrapolation to account for interspecies differences [animal-to-human], interindividual variability, subchronic-to-chronic exposures, lowest observed adverse effect level to no observed adverse effect level (LOAEL to NOAEL) extrapolation, and database uncertainties. As described in Wang et al., (2012) , the UF values used in the derivation of a toxicity value for the selected source analogue are not automatically applied to the value derivation for the target chemical, but rather should be considered using what is known about both the target and source analogue or the established chemical category. This may include any outstanding uncertainties identified in the WoE analysis in step 4. As an example, a default value of 10 is often applied for the database UF (UF_D) to account for the lack of adequate repeat-dose toxicity studies for the target chemical and the increased uncertainty inherent to the read-across process.

5. DISCUSSION

The revised read-across framework presented herein combines practical knowledge and developments in the fields of read-across and computational toxicology to advance read-across applications within EPA and increase transparency and scientific confidence in the evaluation of data-poor target chemicals. One important aspect of this framework is the incorporation of NAMs at key stages of the read-across process, including target chemical profiling and analogue identification and evaluation to support the selection of source analogues. NAMs can encompass a wide variety of technologies, methodologies and approaches that provide qualitative and quantitative information on chemical hazards to avoid or reduce animal testing (U.S. EPA, 2021b). The use of HTS and omics data to facilitate comparisons of toxicodynamic similarities between a target chemical and its analogues has garnered particular interest in the read-across context (Zhu et al., 2016). For example, transcriptomic studies (gene expression microarray, RNA-Seq) characterize gene expression changes associated with chemical exposures. HTS assays from the ToxCast and Tox21 databases profile in vitro bioactivity for environmentally relevant chemicals across a variety of cell-based and gene target endpoints (Richard et al., 2016). Similarities in gene expression or in vitro bioactivity patterns between a pair or within a group of chemicals may represent similar toxic effects. However, challenges with respect to the complexity and relevance of information derived from omics and HTS technologies must be addressed to increase acceptance and integration of NAMs into existing, read-across frameworks. Considerations for integrating and evalauting evidence from NAMs such as transcriptomics or in vitro bioactivity assays to inform WoE conclusions for read-across are outlined in this discussion.

Data derived from omics and HTS studies are diverse and complex. Therefore, these data should to the extent possible be interpreted in terms of their implications for the endpoint or adverse health outcome. Interpretation may be guided by systematic frameworks such as AOP, MOA, and key characteristics of toxicants. The interpretation of the information can then be used to organize, synthesize, and integrate mechanistic data, including data from NAMs, to augment WoE conclusions in a read-across assessment. AOP and MOA frameworks help identify a sequence of key events, linking molecular and cellular responses driven by exposure to a toxic chemical to an adverse health effect. Examples include the AOP on the activation of Cyp2E1 leading to liver cancer (Webster et al., 2021) or the AOP on thyroperoxidase inhibition leading to adverse neurodevelopmental outcomes (Crofton et al., 2019). Similarly, the Key Characteristics approach describes properties of toxicants that encompass different types of mechanisms, which are relevant for hazard characterization and risk assessment. For example, the following twelve key characteristics of hepatotoxicants have been identified: 1) Is reactive and/or is metabolized (bioactivated) to reactive moieties; 2) Causes death (apoptosis and/or necrosis) of liver cells; 3) Affects liver cell proliferation and/or tissue regeneration; 4) Disrupts transport function; 5) Induces oxidative stress; 6) Triggers immune-mediated responses in liver; 7) Causes mitochondrial dysfunction; 8) Activates stress signaling pathways; 9) Causes cholestasis; 10) Disrupts cellular cytoskeleton; 11) Causes liver fibrosis; and 12) Disrupts liver metabolism, including of lipids and proteins (Rusyn et al., 2021).

Molecular or cellular changes identified by omics or HTS studies corresponding to key events or mechanisms associated with a particular MOA or AOP are highly relevant and can be applied to evaluate analogues for read-across by facilitating similarity comparisons and establishing biological plausibility. Indeed, the incorporation of AOPs has been proposed for the development of Integrated Approaches to Testing and Assessment (IATA) under OCED guidance (OECD, 2017a). IATA case studies are available demonstrating the use of AOPs to assemble and interpret NAM data to enhance read-across efforts (Van der Stel et al., 2021). An important limitation for integrating MOA/AOP-based approaches into read-across assessments is the availability of well-characterized or established mechanisms of toxicity relevant to risk assessment. Additionally, many environmental chemicals interact with numerous gene targets and signaling pathways, exhibiting multiple mechanisms of toxicity that complicate the characterization of plausible MOAs or AOPs and their relative contribution to the expected adverse health outcome.

Statistical methods linking patterns of gene expression or in vitro bioactivity to a specific toxicity or adverse health effect can also provide relevant information for read-across assessments. Many genomic studies have focused on identifying gene expression signatures or biomarkers for predicting qualitative and quantitative endpoints related to carcinogenicity (Dean et al., 2017; Fielden et al., 2007; Thomas et al., 2009; Thomas et al., 2007; Thomas et al., 2013). Likewise, associations between in vitro ToxCast/Tox21 assays and in vivo non-cancer endpoints such as reproductive and developmental effects (Kleinstreuer et al., 2011; Martin et al., 2011; Sipes et al., 2011) and endocrine disrupting activity (Judson et al., 2015; Kleinstreuer et al., 2017) have been characterized. Comparison of genomic signatures and bioactivity patterns anchored to phenotypic outcomes among a group of chemicals can be used to assess toxicodynamic similarity, helping to elucidate potential mechanisms of toxicity in support of read-across justifications (Lizarraga et al., 2019; Nakagawa et al., 2022; Nakagawa et al., 2020; Nakagawa et al., 2021; Vrijenhoek et al., 2022)

Another strategy for applying omics or HTS data to inform read-across involves chemical grouping on the basis of similarity of global gene expression or in vitro bioactivity. If the survey of genes or pathways is comprehensive such is the case of genome-wide microarray studies, chemicals with a similar gene expression or bioactivity fingerprint may be expected to behave biologically or toxicologically similar (Low et al., 2013; Shah et al., 2016). Even in the absence of an established relationship to an adverse effect or toxicity, such information can be particularly useful when initially grouping chemicals and identifying candidate analogues for evaluation via read-across (Sipes et al., 2013).

6. CONCLUSION

The case studies depicted in this manuscript outline considerations for the use of metabolic and mechanistic similarity as a basis for evaluating chemicals for screening-level quantitative assessment via read-across. The lessons from these case studies and from the practical knowledge gained through the implementation of read-across for evaluation of data-poor chemicals of interest to the EPA’s Superfund program were applied to update our previous read-across framework (Wang et al., 2012). Key areas of improvement include problem formulation to expand the scope and decision context of potential read-across applications; systematic review and target chemical profiling to identify relevant information on the target and inform the analogue identification and evaluation strategy; an expanded analogue identification approach based on chemical and biological similarities; analogue evaluation by WoE; and considerations for incorporating evidence from NAMs. Future case studies are planned to illustrate the applicability of the revised framework in advancing read-across efforts within the EPA to continue addressing data gaps for environmental chemicals with little to no toxicity data.

Highlights.

• Case studies are presented applying a previous read-across methodology

• A revised framework is proposed incorporating experience and scientific advances

• The framework expands the scope and decision context of read-across applications