Exploring Current Read-across Applications and Needs Among Selected U.S. Federal Agencies

Abstract

Read-across is a well-established data gap-filling technique applied for regulatory purposes. In US Environmental Protection Agency’s New Chemicals Program under TSCA, read-across has been used extensively for decades, however the extent of application and acceptance of read-across among U.S. federal agencies is less clear. In an effort to build read-across capacity, raise awareness of the state of the science, and work towards a harmonization of read-across approaches across U.S. agencies, a new read-across workgroup was established under the Interagency Coordinating Committee on the Validation of Alternative Methods (ICCVAM). This is one of several ad hoc groups ICCVAM has convened to implement the ICCVAM Strategic Roadmap. In this article, we outline the charge and scope of the workgroup and summarize the current applications, tools used, and needs of the agencies represented on the workgroup for read-across. Of the agencies surveyed, the Environmental Protection Agency had the greatest experience in using read-across whereas other agencies indicated that they would benefit from gaining a perspective of the landscape of the tools and available guidance. Two practical case studies are also described to illustrate how the read-across approaches applied by two agencies vary on account of decision context.

1. Introduction

In recent years, “new approach methodologies” (NAMs) have been adopted as a broadly descriptive reference to any non-animal technology, methodology, approach, or combination of these that can provide information on chemical hazard and risk assessment. One of the most well-established NAMs is a data gap-filling technique known as read-across. Read-across is the process of using known information from one or more source substances to predict the same property for a (data-poor) target substance. The technique is applied within analog and category approaches. Briefly, an analog approach involves a target and a source substance whereas a category approach includes 2 or more source substances¹.

The field of read-across is extensively described in the peer reviewed literature (Ball et al., 2016; Patlewicz et al., 2018; van Leeuwen et al., 2009) as well as in technical guidance published by the Organisation of Economic Co-operation and Development (OECD) (OECD, 2014), the European Chemicals Agency (ECHA) (ECHA, 2008), and the European Centre for Ecotoxicology and Toxicology of Chemicals (ECETOC) (ECETOC, 2012). Read-across is also relied upon within internal company product stewardship processes (Wu et al., (2010) is an example). The approach gained significant attention in the run up to the European Union (EU) Registration Evaluation Authorisation and restriction of Chemicals (REACH) regulation (EC, 2006), as an efficient means to address endpoint-specific information requirements. The OECD/ECHA technical guidance (ECHA, 2008) aimed to clarify and re-define some of the common terms and concepts, demonstrate the linkages to (quantitative) structure-activity relationships ((Q)SARs), and provide structured workflows for the development and documentation of read-across justifications. A perspective on the evolution of the development and application of read-across has been described in more detail in Patlewicz et al. (2018). The technical workflows described in the OECD/ECHA technical guidance were additionally factored into the development of the OECD QSAR Toolbox (Dimitrov et al., 2016). The REACH regulation (EC, 2006) also provided impetus for other read-across tools to be developed e.g. ToxRead, Toxmatch, which are discussed later. A handful of publications summarize the state of the art of read-across practice with particular emphasis on the experiences gained under the REACH regulation (for example, Ball et al., (2016) which provides a good perspective including references within).

Although read-across is widely applied for regulatory purposes, many challenges remain for its use and acceptance. These have been discussed in much more detail elsewhere (Ball et al., 2016; Patlewicz et al., 2014; Schultz and Cronin, 2017) but include aspects such as the level of evidence and documentation required to support a read-across justification, the amount of residual uncertainty that can be tolerated, and whether the read-across approach will allow an accurate and credible assessment of the hazards to be made. Workshops (summarized in Ball et al., (2016); ECHA, (2016) OECD, (2008); Patlewicz et al., (2013a); and Stuard and Heinonen, (2018)) have helped to identify many of the issues and barriers to read-across acceptance. Frameworks have also been formulated to characterize the issues with read-across acceptance in a more consistent manner (examples include Blackburn and Stuard (2014); ECHA (2015; 2017a); Patlewicz et al., (2015); Schultz et al., (2015)).

Read-across can also form a component within integrated approaches to testing and assessment (IATA) (Tollefsen et al., 2014). IATA are structured approaches that integrate and weigh different types of data for the purposes of performing hazard identification, hazard characterization and/or safety assessment. They are also pertinent in the identification of data gaps and the associated test strategies to fill them. A number of IATA case studies underpinned by read-across have been developed in the OECD IATA Case Studies Project under the responsibility of the OECD’s Working Party for Hazard Assessment work programme in an effort to create a common understanding of using the approaches and extract lessons learned. This work of the OECD Case Studies Project is particularly focused on identifying major issues or additional technical guidance, such as revisions to the existing OECD grouping guidance (OECD, 2014) or other OECD adverse outcome pathway (AOP)-related guidance (OECD, 2016a; OECD, 2016b; OECD, 2017a). Findings and conclusions for IATA case studies are summarized in so-named ‘considerations’ documents (OECD, 2016c; OECD, 2017b; OECD, 2017c). The uncertainty assessment and reporting in read-across from a number of these OECD IATA case studies are also discussed (Schultz et al., 2019).

While the practice and application of read-across has been extensively described with the strong focus of EU needs under REACH (EC, 2006) and to a lesser extent the EU Cosmetic Regulation (EC, 2009), much less attention has been paid to the landscape of read-across uses, experiences, and needs among the various U.S. federal agencies. To date, the most well-known use cases are those within the Environmental Protection Agency (EPA) that have been already captured in the current OECD guidance (OECD, 2014), such as the experiences under the Premanufacture Notice (PMN) process or High Production Volume (HPV) program (see Worth and Patlewicz, (2007); van Leeuwen et al., (2009)) as well as the New Chemical Categories (NCC) document last updated in 2010 (https://www.epa.gov/reviewing-new-chemicals-under-toxic-substances-control-act-tsca/chemical-categories-used-review-new) which has been encoded in the OECD QSAR Toolbox as a profiling scheme.

Recently a new workgroup was established under the Interagency Coordinating Committee on the Validation of Alternative Methods (ICCVAM) for read-across (also known as the Read-Across Workgroup (RAWG)) which is one of several ad hoc groups convened to help actualize and implement the ICCVAM Strategic Roadmap (see https://ntp.niehs.nih.gov/go/natl-strategy) for establishing new approach methodologies. The RAWG includes representatives from eight federal agencies including the Consumer Product Safety Commission (CPSC), the Department of Defense (DoD), the National Library of Medicine (NLM); the Environmental Protection Agency (EPA); the National Institute of Environmental Health Sciences (NIEHS); the Food and Drug Administration (FDA); the Agency for Toxic Substances and Disease Registry (ATSDR), and the Department of Labor’s Occupational Safety and Health Administration (OSHA). International Cooperation on Alternative Test Methods (ICATM) partners such as the European Commission’s Joint Research Centre (EC-JRC), which operates the EU Reference Laboratory for Alternatives to Animal Testing (EURL-ECVAM) and the Japanese Center for the Validation of Alternative Methods (JaCVAM) are also members of the RAWG. ICATM partner activities in the area of read-across will be the subject of a forthcoming manuscript.

The RAWG’s scope and responsibility are described as follows:

1. To complement existing efforts, create a catalog of ongoing read-across experience and needs across the different agencies to highlight the different decision contexts of interest.

2. Create a catalog of existing read-across resources, including existing technical guidance and software tools, that will highlight the applicability of read-across tools to ensure appropriate context of use.

3. Identify high-quality data sources that may be used for evaluating and using read-across analyses for regulatory application e.g. repositories/databases such as ToxRefDB (https://epa.figshare.com/articles/Animal_Toxicity_Studies_Effects_and_Endpoints_Toxicity_Reference_Database_-_ToxRefDB_files_/6062545) that can be used in post-hoc validation of read-across evaluation using ‘known’ chemicals.

4. Identify case studies that demonstrate utility of read-across analyses in a regulatory setting and identify key data needs for regulatory acceptance.

5. Summarize best practices that are focused on the application and implementation of read-across within the different regulatory settings of interest.

6. Establish new collaborations that could be undertaken to address research gaps in assuring scientific confidence in read-across. These collaborations might address aspects such as how informatics approaches could be leveraged to transition from expert driven read-across assessments towards systematic, objective, and reproducible read-across predictions. This would entail investigating methods for assessing the performance of read-across predictions, combining different types of information, and quantifying the uncertainties.

7. Work with ICATM partners, and in conjunction with OECD, to coordinate efforts and ensure international harmonization on the use and application of read-across approaches.

This paper, a work product of the RAWG, aims to describe the progress made to date towards the first two of these objectives, which involve identification of read-across needs, resources, decision contexts and requirements, and application across the different U.S. federal agencies represented within the workgroup. Our aim with this article is to map the available guidance and tools as well as the current read-across use within the decision contexts of concern for the respective agencies, and to identify potential opportunities for regulatory application. This should help identify other supplementary guidance and tools that are in use but not necessarily broadly known, highlight existing capacity gaps, identify associated training requirements and needs not addressed by the existing guidance/resources and reveal other research opportunities. These will help to frame our implementation plan to meet the different needs of the RAWG members in coordination with ICATM partners and subsequently with OECD.

2. Cataloguing existing read-across resources

Read-across resources in the context of this article are intended to capture the available technical guidance and read-across software tools developed and/or applied under multi-national jurisdictions.

In the last decade or more, there has been a significant acceptance of computational tools for the evaluation of the toxicity profile of substances by both the regulated community and regulators. Application of tools varies depending on the decision context. Many of the tools applied are expert systems underpinned by (Q)SARs that provide de novo estimates of toxicity on the basis of chemical structure. Examples include Leadscope’s Model Applier (Leadscope Inc), Derek Nexus (Lhasa Ltd), TIMES (LMC, Bourgas), ToxGPS (MN-AM), and Vega (Mario Negri Institute), EPA’s OncoLogic™ tool for the evaluation of carcinogenicity (https://www.epa.gov/tsca-screening-tools/oncologictm-computer-system-evaluate-carcinogenic-potential-chemicals). Many of these expert systems play a valuable role in providing additional supporting information in the justification of a read-across assessment (ECETOC, 2012). For the purposes of this paper, however, we will only focus on software tools that have been specifically developed for read-across. While these include proprietary tools such as the ToxGPS read-across tool (launched in March 2018 by MN-AM) and REACHAcross™ (UL) (Luechtefeld et al., 2018), we only highlight in this paper publicly available tools that can be readily evaluated. In Patlewicz et al., (2017) a selection of these tools was described, namely the EPA’s Analog Identification Methodology (AIM), Toxmatch (Ideaconsult Ltd.), OECD QSAR Toolbox, AMBIT (Ideaconsult Ltd.), CBRA, ToxRead and CIIPro (Russo et al., 2017). The following section provides a brief description of each tool in the context of a generic read-across workflow and the reader is referred to the associated review (Patlewicz et al., 2017) which discusses these tools and their alignment in the read-across workflow in more detail.

2.1. Read-across workflow

There are a number of steps in the development of a category or analog approach. The exact number and names of these steps may vary depending on which technical guidance and publication is referenced (see ECHA (2008); ECETOC (2012); OECD (2014)). It is also important to note that this workflow only applies to target substances that are discrete organic chemicals. The workflow may vary when considering mixtures, polymers, inorganics or nanomaterials (see ECETOC (2012); Patlewicz et al., (2017)).

The seven key steps in the workflow are as follows:

1. Decision context

2. Data gap analysis

3. Overarching similarity rationale

4. Analog identification

5. Analog evaluation

6. Data gap filling

7. Uncertainty assessment

Decision Context

The first step in the process is defining the decision context as this will clarify the scope and purpose of the problem being considered and ensure fit-for-purpose application of read-across. The decision context might relate to a hazard level screening assessment, a risk assessment, or a prioritization scheme. Each of these decision contexts will dictate the extent to which uncertainty can be tolerated in the final read-across prediction and which resources (data, level of effort) need to be brought to bear.

Data gap analysis

Data gap analysis makes reference to assessing what data gaps exist for the target substance of interest and for which endpoints. Identifying existing gaps, will in turn dictate whether there are other data gap filling techniques that might be more applicable to use, e.g. a data gap for acute fish toxicity vs. skin sensitization vs. a prenatal developmental study. In this theoretical example, a QSAR approach could conceivably be the most practical and pragmatic solution for acute fish toxicity, a defined approach² making use of in chemico and in vitro assays might be appropriate for skin sensitization, whereas read-across could be a better tactic to meet the information requirements for a prenatal developmental endpoint.

Overarching similarity rationale

The overarching similarity rationale will inform how source analogs might be identified based on what is known about the target substance. For example, there may be some information about the target substance’s likely mechanism of action that will dictate what rationales might be most helpful to customize the search for source analogs. Other similarity rationales such as functional group reactivity, metabolism similarity etc. will also play a role.

Analog identification

Analog identification refers to the search for candidate source analogs. The search may be structured to consider specific aspects pertinent to the endpoint under study or alternatively may be an unsupervised search based on structure similarity.

Analog evaluation

In this step, the relevance and availability of toxicity information for the candidate source analogs identified is evaluated based on general and endpoint-specific considerations (discussed in Patlewicz et al., (2013b)).

Data gap filling

Data gap filling is the process of making the read-across prediction – this is either inferred by an expert or is derived algorithmically.

Uncertainty assessment

Finally, assessing the performance and characterizing the confidence associated with that prediction forms the basis of the uncertainty assessment step. This step considers whether the level of uncertainty is acceptable for the decision context or, if not, what further information might be needed.

2.2. Read-across tools

The available read-across tools address one or more of the workflow steps and vary in how they perform these steps. It is entirely reasonable to address the steps in a variety of ways so long as the approach is adequately documented and can be reproduced. The following brief summary of a selection of tools (in order of their release date) provides some context to their approach and which of the workflow steps they address (these are detailed in Table 1).

Table 1:

Selected freely available read-across tools - Adapted from Patlewicz et al. (2017)

	AIM	Toxmatch	Ambit	OECD QSAR Toolbox	CBRA	ToxRead	CIIPro	GenRAa
Development timeline	Java based version is dated 2012. Initial development of web version was 2005.	First public version released in Dec 2006	Original AMBIT tool was developed in 2004-2005	Proof of concept released in 2008	Implementation of the Low et al., (2013) article	Implementation of Gini et al., (2014)	Implementation described in Russo et al., (2017)	Released August 2018, described in Helman et al. (2019)
Type of Tool	Standalone	Standalone	Web-based and standalone	Standalone and client/server	Standalone	Standalone	Web-based	Web-based via the EPA CompTox Chemicals Dashboard
Latest Version and any Updates	1.01 (Nov 2013) Static	1.07 (Jan 2009) Static	3.0.3 (2013-2015) Ongoing	4.3 (January 2019) Ongoing	0.75 (2013) First release	0.11 BETA (2014) Ongoing	First release	First release (Aug 2018)
Developed by	SRC, Inc.	IDEAconsult Ltd.	IDEAconsult Ltd.	Laboratory of Mathematical Chemistry, University Asen Zlatarov	Fourches Lab, North Carolina State University	Istituto di Ricerche Farmacologiche Mario Negri	Zhu Research Group, Rutgers University	NCCT, EPA
Available from	https://www.epa.gov/tsca-screening-tools/analog-identification-methodology-aim-tool	https://eurl-ecvam.jrc.ec.europa.eu/laboratories-research/predictive_toxicology/qsar_tools/toxmatch	http://cefic-lri.org/lri_toolbox/ambit/	www.qsartoolbox.org	http://www.fourches-laboratory.com/software	http://www.toxread.eu/	http://ciipro.rutgers.edu/	https://comptox.epa.gov/dashboard
Accepted Chemical Input	CAS, name, SMILES, structure drawing/import	CAS, name, SMILES, InChI	Name, SMILES, InChI	CAS, name, SMILES, structure drawing, structure data file (sdf)	Mol file, descriptors as txt	SMILES	PubChem CID, CAS, IUPAC, SMILES, InChI	Linked to the DSSTox inventory
Endpoint Coverage	N/A	Any; based on user input	IUCLIDb 5-supported endpoints (43 total)	Any, per available regulatory endpoints	Any; based on user input	Mutagenicity and bioconcentration factor	Any; based on user input	Repeat-dose toxicity endpoints covered in ToxRefDB v1.0
Analog Identification Approach	Fragment matching	Distance and correlation-based similarity indices based on descriptors or fingerprints	Substructure or similarity searching using structure, name, SMILES, InChI	Category definition followed by subcategorizations	Tanimoto distance using chemical and biological descriptors	VEGA similarity algorithm	Weighted estimated biological similarity	Jaccard similarity index based on different chemical fingerprints – Morgan, Torsion, Chemotype
Neighbor Selection	Automatic	Automatic	Manual	Automatic + manual filter	Automatic	Automatic	Automatic + manual filter	Automatic + manual filter
Data Source	Tool provides inventory index	User- or tool provided	User- or tool provided	User- or tool provided	User-provided	Tool-provided (EU ANTARES project)	User-provided; tool provides PubChem in vitro data	Tool-provided (ToxRef DB)
Quantitative vs Qualitative	N/A	Both	Qualitative (user-determined)	Both	Qualitative	Qualitative for mutagenicity, quantitative for bioconcentration factor	Qualitative	Quantitative binary predictions with measures of confidence of toxicity effects within the different study types
Visualization	None	Standard 2D plots, histograms and similarity matrix	None	Standard 2D Plots	Radial plot of neighbors	Interactive neighbor plot	Activity plot	Radial plot of neighborhoods, data matrix, data coverage
Output/Export	Output reports in the form of HTML, PDF, or XLSX	SDF or text files of data, image files of plots	Assessment report as DOCX or XLSX, data matrix as XLSX	IUCLID format, PDF and RTF files of prediction report, text files of data, image files of plots	NA	Image files of plots	NA	Output of predictions and associated data from source analogs as represented in the data matrix view as XLSX or CSV
Steps covered	Analog identification	All except uncertainty assessment	All except uncertainty assessment	All	All except uncertainty assessment	All except uncertainty assessment	All except uncertainty assessment	All

^aGenRA was not discussed in Patlewicz et al., 2017 given its release date

^bIUCLID: International Uniform Chemical Information Database; software for the administration of chemical substances data developed to fulfill EU information requirements under REACH

AIM – Analog Identification Methodology

The Analog Identification Methodology (AIM) tool was developed by SRC, Inc. for the EPA’s Office of Pollution Prevention and Toxics (OPPT) as a freely available standalone tool to help identify potential analogs for read-across. The current public version 1.01 was released in November 2013 (https://www.epa.gov/tsca-screening-tools/analog-identification-methodology-aim-tool). AIM addresses the analog identification step of the workflow. It uses a two-tiered system for identifying analogs based on a selection of chemical/structural fragments/atoms. The default approach selects analogs if all fragments/atoms and super fragments in the target substance are also contained in the source analogs proposed. If no analogs are identified that satisfy these criteria, a second-tier search is performed that relaxes the matching criteria used for large super fragments i.e. source analogs do not need to contain the large super fragment. The output is a candidate list of potential source analogs with some associated toxicity information. There is no ranking of the analogs and no structuring of the underlying toxicity information aside from providing hyperlinks to view publicly available information. Expert judgement is then used to select the most appropriate analog or analogs.

Toxmatch

Toxmatch is an open source application that encodes a variety of chemical similarity indices to group substances into categories to perform an objective endpoint specific read-across. It was developed by Ideaconsult Ltd. under the terms of an European Commission – Joint Research Centre contract (https://eurl-ecvam.jrc.ec.europa.eu/laboratories-research/predictive_toxicology/doc/Toxmatch_user_manual.pdf). Several case studies outlining its functionality were described by Gallegos Saliner et al., (2008) and Patlewicz et al., (2008). The current version of the software is 1.07, which was released in January 2009.

A dataset comprising structures, endpoint (activity) information and descriptors is either introduced into the software as a training set, or one of the pre-defined datasets can be selected for analysis. Toxmatch then calculates pairwise similarity measures for the dataset provided and creates a similarity matrix to visualize the structural diversity within the dataset. A range of different similarity indices can be computed, such as distance-like similarity indices or correlation-like similarity indices. The similarity information is then used to derive a prediction of the endpoint (activity) of interest. The way in which this prediction is calculated depends on the dataset in question and whether the endpoint is categorical or continuous in nature. If the endpoint is continuous in nature, the read-across prediction is effectively based on the weighted average of the activity values of the k nearest neighbors³, i.e. the activity of the most similar (closest) chemicals are averaged proportionately and used to estimate the activity of a chemical of interest. If the activity is categorical, the read-across prediction is based on a classification scheme. In this case, the source analogs are binned into groups and the similarity measure defines the likelihood that the target chemical falls into one or other groups. The procedure also relies on k nearest neighbors and classifies the target compound into the group where the majority of the k most similar compounds belong.

AMBIT

AMBIT is a cheminformatics software tool that was first developed by Ideaconsult Ltd. and sponsored by Cefic LRI in 2004. The functionality of the current version of AMBIT was developed with the specific read-across needs for REACH in mind (http://cefic-lri.org/lri_toolbox/ambit/). AMBIT enables both structure and data searches with IUCLID (the transaction infrastructure used for REACH), which in turn facilitates category development and read-across. Using AMBIT for read-across analysis creates a chemical assessment aimed at producing a report or document potentially suitable for regulatory submission in the EU. AMBIT identifies source analogs with ECHA provided non confidential registration data (https://echa.europa.eu/view-article/-/journal_content/title/echa-gives-out-registration-data-to-support-development-of-non-test-methods) and presents a data matrix view of the target and source analogs to facilitate an evaluation of their relevance based on the validity of their underlying experimental data. Data gap filling is an expert-driven step such that the read-across prediction from one or several source analogs is made on the basis of a manual inspection of the supporting experimental data.

OECD QSAR Toolbox

The OECD QSAR Toolbox is an application specifically intended for use by government agencies, the chemical industry, and other stakeholders for filling gaps in ecotoxicity, environmental fate, and toxicity data needed for assessing the hazards of chemicals. The Toolbox was developed through a series of phases, starting with a proof of concept which was released in 2008. The current version of the OECD QSAR Toolbox is v4.3. The Toolbox is developed and maintained by the Laboratory of Mathematical Chemistry, University Asen Zlatarov (Bulgaria), and its development is managed by OECD with funding from ECHA.

The OECD QSAR Toolbox incorporates information and tools from various sources into a workflow which actualizes the approach described in the OECD grouping guidance (OECD, 2014).

Dimitrov et al., (2016) have described the workflow and the major functionalities in detail; they can be summarized briefly as follows:

• A target chemical is introduced into the Toolbox, and the user “profiles” the chemical using one or more of the different schemes available. The profilers are structure-based, focusing on structural alerts for specific endpoints, empirical (e.g. structural similarity), or predefined schemes such as the New Chemical or High Production Volume categories.

• In the next step of the workflow, available endpoint data are gathered from multiple sources that have been provided to the Toolbox. Some datasets are focused on specific endpoints whereas other sources are more encompassing in terms of the number of endpoints covered. The endpoint experimental data gathering step is critical for focusing on how the category should be defined for the subsequent data gap filling. The Toolbox is primarily intended to facilitate the endpoint specific data gap filling rather than developing a category of analogs to address more than one endpoint at the same time. After data is collected for the target, the user selects the endpoint of interest to focus the subsequent evaluation.

• The next step is to perform the data gap filling for the endpoint of interest, using one or more of the data gap filling approaches such as trend analysis, read-across or the use of QSARs.

• The last step in the workflow documents the prediction made. Prediction templates that follow a similar structure to the QSAR Prediction Reporting Format (QPRF) as reported in the ECHA guidance (ECHA, 2008) can be created, which document the logic and steps a user has made in deriving the prediction. Export files in IUCLID can also be generated, which is particularly relevant for industry users submitting registration dossiers to ECHA. Guided workflows and the possibility of including a template for read-across assessment framework (RAAF) elements (ECHA, 2015; 2017) into the reports are more recent features in the latest version of the Toolbox.

CBRA

Chemical Biological Read-Across (CBRA) (Low et al., 2013) investigated the feasibility of incorporating biological activity data in conjunction with chemical similarity to predict in vivo toxicity. The toxicity prediction is a similarity weighted average of the activities of nearest neighbors visualized as a radial plot. A software implementation of the approach was developed and is freely available from http://www.fourches-laboratory.com/software. This software explicitly addresses two steps in the previously described generic workflow, i.e., identification of analogs and data gap filling. The analogs and data used to make the prediction are provided by the user. Other examples where chemical and biological data have been combined together in a read-across application are discussed in Zhu et al. (2016).

ToxRead

ToxRead was originally developed by Gini et al., (2014) as a standalone Java tool to assess a single toxicity endpoint: Ames mutagenicity. The research was funded by two EU projects, CALIEDOS (http://www.life-caleidos.eu/) and PROSIL (http://www.life-prosil.eu/). The current version of the tool is v0.11 (http://www.toxread.eu/) and includes modules to make read-across predictions of both Ames mutagenicity and bioconcentration factor. A user inputs a structural identifier as a SMILES string and chooses a number of nearest neighbors and which endpoint to predict. The default is for three analogs which are presented in the form of a radial-like plot. In the center is the target chemical, surrounding it are source analogs that are structurally similar along with source analogs that present structural alerts for mutagenicity or bioconcentration. The read-across prediction is summarized based on the color of the target chemical, in the case of mutagenicity, a red color is indicative of a positive mutagenicity outcome whereas green indicates a non-mutagenic (negative) outcome. Double-clicking on the target chemical reveals a short report that summarizes the read-across prediction and provides a consensus prediction based on associated QSAR models. The tool explicitly addresses two of the steps in the workflow: selecting analogs based on a specific endpoint and making a read-across prediction (Benfenati et al., 2016; Manganelli and Benfenati, 2016).

CIIPro

CIIPro is a cheminformatics web portal freely available at http://ciipro.rutgers.edu/. CIIPro facilitates read-across predictions on the basis of chemical and/or biological similarity. The prediction output can be visualized by a similarity chart with associated similarity and confidence values. CIIPro is unique in that it takes advantage of the wide array of bioassay data publicly available from PubChem (https://pubchem.ncbi.nlm.nih.gov/). The underlying bioassay database is updated monthly by the CIIPro developers (Russo et al., 2017).

Generalized Read-Across (GenRA)

EPA’s National Center for Computational Toxicology (NCCT) has developed a read-across tool known as GenRA (Generalized Read-Across), which is incorporated as an add-in within the EPA CompTox Chemicals Dashboard (https://comptox.epa.gov/dashboard). This follows the category workflow (described in Helman et al., (2018); Helman et al., (2019)), although GenRA uses a data-driven approach to compute a read-across prediction using a similarity-weighted activity algorithm (Shah et al., 2016). The basis for GenRA mirrors that of CBRA’s approach, which identifies biological and chemical neighbors to make an inference of target substance activity. The current version of GenRA relies on chemical and/or biological similarity to make binary predictions of in vivo toxicity effects with quantitative measures of predictive performance. The chemical descriptors comprise different structural fingerprint approaches, e.g. Morgan fingerprints (Rogers and Hahn, 2010). The biological descriptors rely on high-throughput screening bioactivity data obtained from ToxCast (https://www.epa.gov/chemical-research/toxicity-forecasting) and Tox21 converted into fingerprint format (https://www.epa.gov/chemical-research/toxicology-testing-21st-century-tox21).

The underlying in vivo toxicity data is drawn from ToxRefDB v.1.0 (ftp://newftp.epa.gov/COMPTOX/High_Throughput_Screening_Data/Animal_Tox_Data/readme_toxrefdb_20141106.pdf). GenRA predicts up to 129 toxicity effects from 10 different repeated dose toxicity study types. Recent work has explored the impact that other similarity contexts, such as physicochemical similarity (Helman et al., 2018) have on read-across performance as well as extending the approach to quantitative endpoints (e.g. lowest observed adverse effect levels (LOAELs) ).

Read-across tool summary

AIM identifies analogs using chemical features/fingerprints, and as such only addresses this step in the workflow. Toxmatch facilitates endpoint-specific read-across using a range of different similarity metrics and user-provided data, and thus addresses the majority of the workflow steps. The OECD QSAR Toolbox follows the OECD/ECHA workflow for analog and category approaches to facilitate structure-based approaches to identify analogs for specific endpoints as well as providing the ability to sub-categorize on the basis of other similarity contexts (e.g. reactivity, physicochemical characteristics, and metabolism). AMBIT identifies analogs on the basis of associated REACH data, but the read-across prediction is derived manually by the end user. CBRA and CIIPro rely on chemical structure and biological activity data to make predictions. GenRA makes predictions on the basis of chemical and/or biological activity data using a user-guided workflow.

2.3. Existing technical guidance

In terms of technical guidance, the most established or best-known guidance is that developed for REACH implementation (EC, 2006), which is captured as part of the chemical information requirements in Chapter R6 ECHA guidance (ECHA, 2008). This guidance was identical to that from OECD since they were developed in collaboration. However, the OECD guidance was revised in 2014 to accommodate new advances in the field, including the potential application of AOPs.

There are also a number of publications in the literature that describe approaches and workflows for read-across. In Patlewicz et al., (2018), these were designated as workflows or frameworks for the development of read-across, versus frameworks for the assessment of read-across. Both of these types of frameworks can be used interchangeably to a large extent to develop and/or assess read-across, but we have chosen to differentiate them herein to capture the drivers of their initial development. For development of read-across, the most recognized regulatory frameworks are the workflows published in the OECD and ECHA guidance. In the literature, read-across development workflows proposed within companies or trade organizations include that published by Wu et al., (2010), ECETOC (2012) and Patlewicz et al., (2013b). Good read-across practices have since been summarized in Ball et al., (2016), and guidance on incorporating biological data in read-across was covered in Zhu et al., (2016). In terms of the assessment of read-across, there are several frameworks published; these include the ECHA Read-Across Assessment Framework (RAAF) (2015; 2017a), as well as frameworks from Patlewicz et al., (2015), Blackburn and Stuard (2014) and Schultz et al., (2015). These were compared and contrasted in Patlewicz et al., (2018), which observed consistency among these different frameworks. One other notable framework is by Wang et al., (2012). This framework, used as a component of screening-level quantitative risk assessments for substances of interest to the EPA’s Superfund program, is presently being updated to incorporate lessons learned from its application and recent progress in the science (Lizarraga et al., 2018). Table 2 summarizes the different frameworks.

Table 2:

Selected read-across development and assessment frameworks

Framework	Regulatory Technical guidance	Literature guidance
Development of read-across	• ECHA REACH guidance (ECHA, 2008) • OECD grouping guidance (OECD, 2014)	• Wu et al.,(2010) • ECETOC guidance (2012) • Patlewicz et al., (2013a) • Ball et al., (2016) • Zhu et al., (2016) • Wang et al., (2012)
Assessment of read-across	• ECHA RAAF (ECHA, 2015; ECHA, 2017a)	• Blackburn and Stuard, (2014) • Patlewicz et al., (2015) • Schultz et al., (2015)

3. Existing tools, resources, or guidance applied by U.S. agencies

Surveying the ICCVAM RAWG members found that the extent of familiarity and application of tools and resources and guidance varies across the different U.S. agencies. In some cases, there is a preference towards using QSAR approaches versus read-across, and the tools used vary between using the OECD QSAR Toolbox or other tools as well as considering other novel data streams (e.g. ToxCast data). This stems in part from a lack of familiarity with the landscape of resources currently available. Indeed, of the agencies profiled, EPA OPPT had the greatest awareness of the OECD grouping guidance having contributed to the initial 2007 version and its more recent revision in 2014. The other agencies surveyed responded that they would benefit from some basic scene-setting to gain a perspective of the landscape of the tools and available guidance. To that end, the review by Patlewicz et al., (2018) formed a helpful introduction to harmonize the baseline awareness among the RAWG members and is a recommended resource to provide additional background information. Here we summarize the current familiarity and use of the different resources applied by the different agencies.

3.1. Agency for Toxic Substances and Disease Registry

Although the Agency for Toxic Substances and Disease Registry (ATSDR) is not a regulatory agency, it uses read-across and QSAR approaches to assist risk assessors at Superfund sites and in emergency response situations and public health disasters. Several of the established read-across frameworks listed in Table 2 are routinely applied depending on the decision context. The computational toxicology laboratory at ATSDR uses several state-of-the-art software read-across and (Q)SAR programs, both freely and commercially available. These include OECD QSAR Toolbox (Dimitrov et al., 2016), ToxRead (Gini et al., 2014), TOxicity Prediction by Komputer Assisted Technology (TOPKAT) (BIOVIA https://www.3dsbiovia.com/products/collaborative-science/biovia-discovery-studio/qsar-admet-and-predictive-toxicology.html), Leadscope (Leadscope, Inc. https://www.leadscope.com), SimulationsPlus (SimulationsPlus, Inc. https://www.simulations-plus.com/software/admetpredictor/toxicity/), and CaseTox (MultiCASE, Inc. http://www.multicase.com/). To further advance the application of these approaches to protect public health, ATSDR collaborates with software vendors including Health Designs, Inc. (Mumtaz et al., 1995; Ruiz et al., 2012; Ruiz et al., 2011) and more recently Leadscope as part of its In Silico Protocol effort, a vendor collaboration funded by a National Institutes of Health (NIH) grant that comprises a consortium of more than 55 partners from industry, regulatory bodies and academia (Myatt et al., 2018).

3.2. Department of Defense

The Department of Defense (DoD) is not a regulatory agency and has no statutory requirements mandating the collection and use of toxicity data. However, DoD generates and uses toxicity data with the goal of protecting human health and the environment from harmful chemical exposures relevant to DoD activities. Branches of DoD covered in this article include the Army, Navy, and Air Force.

DoD branches vary in their use and awareness of read-across resources. The Army is developing a read-across tool that relies on chemical and bioactivity information to make toxicity predictions for substances of interest (Burgoon, personal communication). Other branches of the DoD rely on the OECD QSAR Toolbox to undertake read-across assessments for specific endpoints of interest.

3.3. EPA Office of Pollution Prevention and Toxics

EPA’s OPPT regulates new and existing chemicals under the authority of the Toxic Substances Control Act (TSCA) (EPA, 2008). TSCA, originally passed in 1976, requires a company to submit a pre-manufacture notice to EPA OPPT prior to either manufacturing or importing a new chemical or a Significant New Use Notice (SNUN) before initiating a new use of an existing chemical if it is subject to a Significant New Use Rule (SNUR). Although toxicity testing is not required, the submission must include any existing test data. The 2016 Frank R. Lautenberg Chemical Safety for the 21^st Century Act (called the Lautenberg Chemical Safety Act (LSCA)) amended TSCA to introduce mandatory deadlines and other statutory requirements (https://www.epa.gov/assessing-and-managing-chemicals-under-tsca/frank-r-lautenberg-chemical-safety-21st-century-act). This changed EPA OPPT’s authority to regulate both new and existing chemical substances in the U.S. The LCSA also requires that OPPT promote the development and implementation of alternative test methods and strategies to reduce, refine, or replace vertebrate animal testing under TSCA. Such methods and strategies include computational approaches, read-across, and in vitro testing (EPA, 2017b). An EPA Strategic Plan for meeting this requirement was recently published (EPA, 2018).

In addition to evaluating new chemicals prior to their entry into U.S. commerce, the LSCA requires that EPA OPPT evaluate risk to existing chemicals. The recently updated TSCA Inventory includes 86,228 chemicals of which 40,655 have been determined to be “active” in US commerce in the past 10 years (https://www.epa.gov/newsreleases/epa-releases-first-major-update-chemicals-list-40-years). Note: The definitions of what constitutes ‘active’ vs ‘inactive’ are described in more detail (https://www.epa.gov/tsca-inventory/tsca-inventory-notification-active-inactive-rule). Briefly, per an EPA rule, industry registrants were required to report chemicals manufactured (including imported) or processed in the U.S. over a 10-year period. Substances were reported or exempt are referred to as ‘active’ in US commerce. The amended TSCA requires EPA OPPT to prioritize existing chemicals to determine which ones require risk evaluation (i.e., “high-priority”) or may be considered of low priority such that risk evaluation is not warranted at this time. This prioritization requirement may involve the use of read-across to determine whether there are any data gaps. OPPT can request new information to fill those gaps if deemed necessary for prioritizing or conducting a risk evaluation (EPA, 2017a; EPA, 2017b).

OPPT experience with read-across approaches is relatively mature due to its experience with tools they have commissioned, such as AIM, ChemACE (a structure clustering tool), and Oncologic (a cancer screening expert system). They also use the New Chemical Categories document. These tools are described in more detail elsewhere (see https://www.epa.gov/tsca-screening-tools/using-predictive-methods-assess-hazard-under-tsca#models). OPPT also contributed to the OECD grouping guidance (OECD, 2014), and many of the insights developed from the EPA High Production Volume categories were incorporated in this guidance.

3.4. EPA Office of Research and Development

EPA has two research centers engaged in the area of read-across. The National Center for Environmental Assessment (NCEA) applies read-across routinely within its Superfund Health Risk Technical Support Center (STSC). Applications include development of screening level Provisional Peer Reviewed Toxicity Values (PPRTVs) and providing technical assistance through the STSC Hotline to identify suitable analogs for chemicals detected at contaminated sites. PPRTVs are toxicity values primarily derived for use in EPA’s Superfund Program that are derived from a review of the relevant scientific literature using EPA methods, sources of data and guidance for value derivation (see the case study in Section 6.1 for more details). The read-across approach relies on an expert-driven framework introduced by Wang et al., (2012) which is presently being updated to expand the scope and decision context of read-across applications (Lizarraga et al., 2018). NCEA incorporates information from several publicly available databases and resources into read-across assessments to aid in the identification and evaluation of analogs. These resources include the National Library of Medicine’s ChemIDplus database, EPA’s DSSTox database (which has been superseded by the US EPA CompTox Chemicals Dashboard), ChemACE, and the OECD QSAR Toolbox.

EPA’s NCCT is engaged in read-across approaches from the perspective of developing new approaches to systematically and objectively evaluate the performance and quantify the uncertainties associated with read-across predictions. NCCT has developed an approach called GenRA (Shah et al., 2016) and have actualized this approach into a graphical user interface that is implemented into the US EPA CompTox Chemicals Dashboard (Helman et al., 2019; Williams et al., 2017).

3.5. National Library of Medicine (NLM)

The National Library of Medicine (NLM) within the National Institutes of Health (NIH) provides access to citations of read-across publications included in PubMed, and often free full text access via PubMed Central. Also, NLM includes read-across information for some substances in its TOXNET suite of databases, e.g., the Hazardous Substances Data Bank and International Estimates for Risk databases, and TOXNET’s ChemIDplus can be used for identifying structurally similar substances. Furthermore, NLM provides information about read-across in its “ToxTutor” online self-paced tutorial covering key principles of toxicology and related topics ( https://toxtutor.nlm.nih.gov/ ).

4. Decision contexts and needs for read-across applications by U.S. agencies and partners

The read-across use cases to date have largely been dominated by new chemicals assessment by EPA and classification and labelling and risk assessment needs under REACH; the decision contexts among the different U.S. Agencies are quite varied in scope. These decision contexts will have an impact on the resources that need to be brought to bear as well as the level of uncertainty that could be tolerated. Table 3 summarizes the decision contexts for each agency.

Table 3:

Agency-specific decision contexts and desired activities for read-across

Agency	Decision context	Desired activities
ASTDR	• Supporting emergency response • Filling data gaps for chemicals of interest • Hazard assessments of chemicals found at waste sites	• Guidance for the use of read-across for emergency response and for chemicals found at waste sites • Read across application to chemical mixtures • Identifying best practices for the use of read-across for specific contexts
CPSC		• Training and guidance on application of read-across • Identifying best practices for the use of read-across for specific contexts
DoD	• Screening for occupational safety • Screening for environmental safety • Army: Read-across and weight of evidence (WoE) approaches for emergency response, product registration and exposure limits for use by internal clients	• Training and guidance on application of read-across • Identifying best practices for the use of read-across for specific contexts.
EPA NCEA	• Support STSC-related activities in site-specific screening and prioritization and quantitative risk assesment of data-poor chemicals at contaminated sites	• Integration of software tools and new approach methodologies data to augment expert-judgement and increase confidence in read-across assessments • Guidance on systematic approaches for WoE to evaluate similarity and uncertaintity
EPA OPPT	• Prioritization (existing chemicals) • Risk evaluation (new and existing chemicals)	• Refinement to the New Chemicals Categories (NCC) • List of acceptable new approach methodologies
FDA Center for Devices and Radiological Health	• Risk assessment of medical devices	• Guidance for the use of read-across within a toxicological risk assessment of medical devices • Qualify6 read-across as a tool to use for device evaluation and support regulatory decision making using the Center’s Medical Device Development Tools (MDDT) program (FDA, 2017)
FDA Center for Food Safety and Nutrition	• May use QSAR and read-across to assist in hazard identification and prioritization	• Guidance for the use of read-across for contaminants, dietary ingredients within dietary supplements, food contact substances and cosmetic ingredients
ICCVAM	• Support U.S. agency needs and decision contexts	• Facilitate collaboration and cooperation across Federal Agencies to establish scientific confidence in read-across approaches for different types of decision contexts

C&L = Classification and Labelling, FHSA = Federal Hazardous Substances Act; NOAEL = No Observed Adverse Effect Level; PPPA = Poison Prevention Packaging Act; PPRTV = Provisional Peer Reviewed Toxicity Values; STSC = Superfund Health Risk Technical Support Center; WoE = Weight of Evidence

⁶Qualify has a specific meaning in FDA terminology. Briefly, qualification of a biomarker is a determination that within the stated context of use, the biomarker can be relied upon to have a specific interpretation and application in drug development and regulatory review (www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM628118.pdf).

5. Ad hoc survey of tools by selected US agencies

Workgroup members were also surveyed on the frequency of use of different tools within their respective agencies and the specific purposes for which they are used. This was intended to provide a perspective of how different tools are applied in practice (e.g., application of the OECD QSAR Toolbox for identifying analogs only vs. deriving predictions).

Feedback was provided from workgroup members representing four agencies, including the Department of Defense (DOD), the Food and Drug Administration (FDA), the Environmental Protection Agency’s National Center for Environmental Assessment (EPA NCEA), and the Agency for Toxic Substances and Disease Registry (ATSDR). Each of the workgroup members indicated that their agencies’ decision contexts are primarily hazard assessments. Furthermore, each agency applies the workflow process a few times a year or monthly. Table 4 summarizes the types of tools used by the agencies and the steps of the workflow in which they are applied.

Table 4.

Tool coverage of workflow steps and use by various U.S. agencies

Agency	Tools Used for Workflow Step 4: Analog identification	Tools Used for Workflow Step 5: Analog evaluation	Tools Used for Workflow Step 6: Data gap filling
DOD	OECD QSAR Toolbox	No tools identified; expert judgement used	OECD QSAR Toolbox
FDA	In house tools; Commercial tools
EPA OPPT	AIM ChemIDplus OECD QSAR Toolbox	No tools identified; expert judgement used	OECD QSAR Toolbox ECOSAR
EPA NCEA	OECD QSAR Toolbox; ChemACE;; ChemIDplus; DSSTox	No tools identified; expert judgement used
ATSDR	OECD QSAR Toolbox	In-house tools; Commercial tools; Other freely available tools; e.g. ToxRead

6. Case studies

Of the workgroup members, two agencies identified specific case studies to illustrate how read-across is applied in practice.

6.1. EPA evaluation of n-heptanal

Many chemicals of interest to the EPA Superfund program have limited or no toxicity data for the derivation of human health reference values (RfV) via traditional risk assessment practices. This poses a significant challenge to field toxicologists and risk assessors that rely on these values for cleanup and remediation decisions at contaminated sites. The Wang et al., (2012) read-across framework, discussed earlier in Section 2.3, was introduced to address toxicity data gaps and assist in screening-level quantitative risk assessment of Superfund chemicals. The method uses three primary types of analogs/similarity contexts (structure, toxicokinetics, and toxicodynamics) to build a read-across justification for the target chemical. Typically, analogs are identified by conducting a structural similarity search using publicly available databases such as ChemID-plus, DSSTox, ChemACE and the OECD QSAR Toolbox. The list of potential analogs is cross-referenced with regulatory health assessment repositories to select analogs with existent RfVs. Sources include the EPA’s Integrated Risk Information System, its PPRTV electronic library (https://hhpprtv.ornl.gov), ATSDR reports, and the California EPA database. Information on functional moieties, reactivity, physicochemical properties, toxicokinetics and toxicodynamics is compiled for the target and analogs and analyzed via a weight-of-evidence approach to select a single best source analog. Finally, the point-of-departure (POD) for the source analog is adopted for screening-level quantitative risk assessment. A case study of the application of the Wang et al., (2012) methodology for the derivation of PPRTVs for n-heptanal is presented below, discussing important lessons and ongoing challenges.

In the assessment of n-heptanal (CASRN 111-71-7), the derivation of subchronic and chronic PPRTVs using chemical-specific information was precluded due to the lack of adequate in vivo data for hazard identification and dose-response analysis (EPA, 2017c). Instead, three analogs with published reference values for the inhalation route (acetaldehyde, propionaldehyde, and glutaraldehyde) were identified via structural similarity comparisons. Although the similarity scores for the analogs were relatively low (10−36% for OECD, 33−69% for DSSTox) and appeared to be biased by the carbon chain-length descriptor (C7 for the target and C2-C5 for the analogs), all four chemicals shared structural and physicochemical properties consistent with other saturated aliphatic aldehydes (see Table A-1 of PPRTV assessment in U.S. EPA (2017c)).

The reactive aldehyde moiety common in this group of chemicals is associated with nasal lesions, the expected inhalation toxicity endpoint. The presence of an additional aldehyde moiety in glutaraldehyde was expected to increase its airway reactivity and toxicity in comparison to n-heptanal and the related monoaldehyde analogs, acetaldehyde and propionaldehyde (see Table A-3 of PPRTV assessment in U.S. EPA (2017c)). Therefore, glutaraldehyde was excluded as a potential source analog for n-heptanal.

Of the remaining analogs, both acetaldehyde and propionaldehyde had similar acute toxicity targets (central nervous system and respiratory tract toxicity) and LD50 values to n-heptanal (see Table A-3 of PPRTV assessment in U.S. EPA (2017c). A slight reduction in acute toxicity potency was noted with increased carbon chain length, which is consistent with previous reports examining short-chain aldehydes (C1-C4) (Bombick and Doolittle, 1995; Koerker et al., 1976; Skog, 1950). This effect of carbon chain length was also evident for the physicochemical properties of n-heptanal and the monoaldehyde analogs (see Table A-1 of the PPRTV assessment in U.S. EPA (2017c)). In addition, acetaldehyde and propionaldehyde had close similarity with regards to PODs and critical effects for subchronic and chronic inhalation toxicity (see Table A-3 of PPRTV assessment in U.S. EPA (2017c)).

In terms of toxicokinetic properties, metabolism and excretion pathways for n-heptanal, acetaldehyde and propionaldehyde are similar to those of other saturated aliphatic aldehydes, which are oxidized to carboxylic acid by aldehyde dehydrogenase (ALDH) and the β-oxidation pathway (see Table A-2 of PPRTV assessment in U.S. EPA (2017c)). Metabolism is expected to have a detoxifying role in the elimination of these aldehydes, and comparison of ALDH2 activity in human liver samples revealed commonalities in metabolic rates for n-heptanal, acetaldehyde, and propionaldehyde (Wang et al., 2002).

Altogether, acetaldehyde and propionaldehyde were considered suitable analogs for n-heptanal based on the shared reactive aldehyde moiety, metabolism pathway similarities leading to detoxification, and comparable LD50 values. However, taking into account the effect of carbon chain length on the acute toxicity and physicochemical properties of these aldehydes, propionaldehyde was ultimately selected as the closest analog to n-heptanal. In addition, the POD value for propionaldehyde is derived from a principal study of longer exposure duration (7 weeks) than that of acetaldehyde (4 weeks), increasing confidence in the principal study as a source for the derivation of screening subchronic and chronic inhalation RfVs.

The Wang et al. (2012) methodology has been implemented to derive screening-level PPRTV assessments for data-poor chemicals in the Superfund program. However, many obstacles remain to be addressed for chemicals with severe database deficiencies. One of the major limitations is the requirement for analogs to have existent human health RfVs for consideration and inclusion in the read-across analysis, which restricts the evaluation of chemical categories as proposed by the OECD and ECHA frameworks for chemical grouping and read-across (ECHA, 2017a; OECD, 2014). Another challenge has been the expansion of the search strategy used to identify potential analogs. Currently, this is done mostly by structural chemical grouping. This relies on similarity metrics such as Tanimoto scores and requires significant expert evaluation to capture similarities and differences in key functional moieties, chemical reactivity, and physicochemical properties important for toxicokinetics and toxicodynamics, as in the n-heptanal example. Additional computational tools and software need to be integrated to augment expert judgement and provide different means of grouping chemicals on the basis of both chemical and biological properties. Likewise, the inclusion of data derived from new approach methodologies such as in vitro toxicokinetic studies, transcriptomics or bioactivity from high-throughput or high-content screens could assist in filling data gaps for Absorption, Distribution, Metabolism and Excretion (ADME) and toxicity/mechanistic plausibility, which would increase confidence in read-across predictions. Finally, more systematic or quantitative approaches are required for the evaluation and integration of evidence when assessing similarity and uncertainty in read-across justifications.

6.2. FDA toxicological risk assessment for a medical device

A toxicological risk assessment is a necessary part of chemical characterization and biocompatibility studies to understand or predict the human response to materials in medical devices. Thus, the U.S. Food and Drug Administration (FDA) recommends that applicants derive and justify in their premarket submission a POD such as no-observed-adverse-effect level (NOAEL). This enables evaluation by agency reviewers of the safety of materials of manufacture and/or leachable chemicals and degradation products released from a finished medical device (FDA, 2017). The POD may be based on data generated by the applicant, or from available toxicity literature and other publicly available information. If data are not available to derive a NOAEL, the concept of threshold of toxicological concern (Munro et al., 2008) can be used to assess systemic toxicity.

Applicants have tried to use read-across as a less burdensome approach for finding a surrogate (source analog) with available POD data to help estimate NOAELs for extractables/leachables released from a finished device. However, FDA has not provided any guidance to date that would assist stakeholders on how to adequately use read-across in medical device premarket applications. FDA-recognized standards specific to medical devices such as ISO 10993 Part 17 are also not currently helpful to inform a toxicological risk assessment in this regard. This case study performed using the OECD QSAR Toolbox illustrates some practical challenges encountered by FDA reviewers when assessing an industry-submitted read-across prediction. For reasons of confidentiality, the precise target substance could not be disclosed in this case study, but a summary of the steps that the industry applicant followed are described in brief.

An assessment of target substance X required repeat-dose toxicity information that was lacking. Substance X was characterized by its 2D chemical structure and 1D SMILES notation; no CAS registry information was available. The SMILES code was used to search the OECD QSAR Toolbox for any available information for substance X as well as to identify candidate source analogs.

Substance X was ‘profiled’ using several rulebases within the Toolbox that identify whether a substance is a member or potential member of an existing regulatory category such as the OECD HPV categories or EPA New Chemical Categories. However, the substance was not identified as a member of any existing regulatory chemical category. Accordingly, it was simply identified as a discrete substance. It is worth noting that the FDA has not established any material(s) of manufacture/chemical categories applicable to medical devices.

The applicant then used two other profilers to categorize the substance, the OncoLogic Primary Classification for carcinogenicity and the Aquatic toxicity classification by ECOSAR profilers. No rationale was provided to substantiate these selections. A category was first defined using the OncoLogic Primary Classification and a separate category was then defined using the ECOSAR profiler. The pool of potential analogs from the OncoLogic profiler was larger and hence was examined first for available repeat-dose toxicity data. No subcategorization was performed on this defined category.

No Observed Effect Level (NOEL) values from repeat-dose studies conducted in rats were found for over 30 chemicals in the pool of potential analogs. This subset was carried forward into the data gap-filling step of the workflow. A trend analysis was performed using log Kow as the correlating parameter to derive a simple linear regression to predict the NOEL value for the target substance. The predicted log Kow of the target substance X (as determined by the applicant and verified by agency reviewers) exceeded the range of log Kow values for the source analogs with NOEL data. The applicant also presented information that showed that the structural similarity of the target substance relative to the source analogs ranged from 10-20%.

Given the low structural similarity of the target substance relative to the source analogs and the fact that the log Kow exceeded the range of the source analogs, the applicant chose to use the candidate analogs identified using the ECOSAR profiler. However, the predicted log Kow of the target still exceeded the range of log Kow values for the source analogs identified. Comparing the source analogs identified by the two profiler approaches, revealed that the most similar source analog was common to both profiling approaches but was still only 10-20% similar to the target substance. Given the lack of confidence in the estimated rat NOEL values from the two profiling approaches attempted, the preferred source analog was rejected.

The applicant then investigated the feasibility of defining a category using “functional groups”. No analogs were identified that contained all of the functional groups in the target substance, and those with partial similarity lacked repeat-dose toxicity data. Therefore, no read-across could be attempted. Finally, the applicant sought to identify an in vivo metabolite of the target substance using the rat liver S9 metabolism simulator within the Toolbox. This simulator mimics the expected metabolic transformations expected to be observed in Ames mutagenicity studies. Of over 30 metabolites predicted, one had available repeat-dose toxicity data, and its NOEL value in OECD QSAR Toolbox was subsequently used as a POD for margin of safety calculations on target substance X.

Several lessons were learned from this case study. The lack of scientific justification provided to support the profiling selections made the assessment by agency reviewers difficult, as there was no tractable hypothesis presented to support or reject the analogs identified. From the different starting points chosen in terms of defining the category within the Toolbox, there was also a lack of consistency in the predictions made since the resulting number of source analogs differed. As such, it was difficult to evaluate the commonality in source analogs identified by one or other profiling scheme. Moreover, the chemical inventories and the availability of repeat-dose toxicity information appeared to be of limited relevance to the chemistries of interest to the FDA. This meant that the breadth of relevant analogs was limited. Characterizing the uncertainty of the POD predictions being derived was also identified as an issue, since quantitative adjustment(s) by use of additional uncertainty factors is not currently viewed as an approach to address uncertainties in the identification of a POD from source analogue(s) as a surrogate for a data-poor target chemical.

7. Discussion

ICCVAM convened a read-across workgroup as part of a strategy to actualize the use of new approach methodologies, as described in its Strategic Roadmap (ICCVAM 2018). The RAWG agreed to a number of objectives to chart out its own scope and direction. The initial objectives of the RAWG have focused on understanding what read-across approaches US agencies require, routinely apply, or are familiar with, depending on the different regulatory responsibilities and decision contexts, and what tools are being routinely used or are under development for read-across applications.

Among the federal agencies surveyed, use of read-across is by far most firmly established within the EPA. However, the types of approaches used and the decision contexts vary substantially even within this single agency. For new chemicals, existing software tools commissioned by OPPT have served EPA well in terms of screening for potential hazards of limited numbers of substances within a tight timeline. That said, opportunities remain to embrace the latest state of the science in read-across to refine how some of these new chemical assessments might be performed. Noteworthy are advances in the types of read-across tools that have been developed in the last decade, such as the OECD QSAR Toolbox, that could play a complementary role to the legacy tools in current use. Following the amended TSCA Regulation, the anticipated challenges of prioritizing large numbers of substances will necessitate other approaches such as systematic data-driven read-across (e.g. GenRA (Helman et al., 2019), REACHAcross (https://www.ulreachacross.com/about.html). These tools exploit chemical and/or biological similarity and can derive predictions for many thousands of substances in a reproducible manner with quantitative measures of uncertainty. On the other hand, NCEA relies upon an expert-driven framework to develop PPRTVs that can then be taken up in screening-level quantitative risk assessments which typically entail a much smaller number of substances. The level of uncertainty in a read-across prediction in the context of a quantitative risk assessment is generally much lower than can be tolerated in the context of prioritizing thousands of substances. Opportunities to consider new approaches and make aspects of the currently used framework more objective have been noted in the discussion of the EPA case study and have been proposed in Lizarraga et al., (2018).

For other agencies, there is a reasonable level of familiarity with the OECD guidance and notably the OECD QSAR Toolbox as a tool to facilitate endpoint read-across. Challenges exist in the use of the Toolbox in terms of understanding the robustness of a prediction if the starting profiler selected is different or is not well justified, as was evidenced in the FDA case study. Further challenges are presented by the evolution of software versions; when additional data is incorporated into the tool, the read-across predictions may change over time. Detailed documentation around version control will help address this issue.

Several agencies are interested in read-across for mixtures. Unfortunately, this is a major gap in the read-across field, with only minimal guidance available in the OECD grouping guidance and a considerations document offered by ECHA (ECHA, 2017b). Concawe, a division of the European Petroleum Refiners Association has a project on read-across approaches for Chemical Substances of Unknown or Variable Composition, Complex Reaction Products and Biological Materials (UVCBs) (See https://www.concawe.eu/cat-app/). This is an area where biological information as the basis for read-across may become an extremely valuable tool, particularly when similarity can be assessed based on in vitro bioactivity profiles relevant to the endpoint of concern, e.g. as defined by mapping to AOPs.

A common concern raised by several agencies that is exemplified in the case studies is how to adequately characterize the scientific confidence of a read-across prediction. Guidance is needed on how to accomplish this in a reproducible manner and the best practices or principles that could be followed to achieve this. Scientific confidence considerations have been described in several published assessment frameworks, including the RAAF (ECHA 2015, 2017b) and in the scientific literature (Blackburn and Stuard, 2014; Patlewicz et al., 2015; Schultz et al., 2015). However, while these identify sources of uncertainty, they lack practical guidance of the strategies to reduce uncertainty and how to determine whether the residual uncertainty is acceptable. The difficulty lies in determining what level of residual uncertainty can be tolerated, which in turn depends on the decision context. Another factor affecting the confidence that can be placed in a read-across prediction is the quality, completeness in reporting and variability of the underlying in vivo data that is frequently used to generate predictions. Without a clear characterization of the variability of reference data, the performance of any alternative approach such as a read-across prediction cannot be reasonably assessed. Data-driven approaches to algorithmically codify read-across are an obvious starting point to transition away from current expert-driven approaches. A proposal for a harmonized hybrid workflow to begin such a transition was recently articulated in Patlewicz et al., (2018).

Other recurrent themes from the agencies concerned included access to sufficient high quality and curated toxicity data and the relevance of this information for the types of chemistries of interest, as well as ongoing efforts to generate additional data to support read-across approaches. It therefore behooves the agencies to compile and characterize their chemical landscape, which would facilitate an evaluation of the applicability of existing current tools and specific needs for additional information. Recent work in this area has focused on development of open-source models (Mansouri et al. 2018) to predict physicochemical properties to supplement structural features in read-across approaches (Helman et al., 2018), and prospective testing of chemical categories such as Per- and Polyfluoroalkyl Substances (PFAS) (Patlewicz et al., 2019). If the domain of applicability in terms of chemical coverage is outside of that of the inventories of interest for agencies, strategies to conduct targeted in vitro testing to chart the landscape and anchor this to the existing legacy in vivo data may well be a critical undertaking to realize the benefits and promise of read-across in the future.

8. Future directions

A number of areas were identified in scoping the current needs of the respective agencies (e.g. building capacity in read-across tools, best practices for different decision contexts, guidance on approaches to evaluating the similarity of analogs). Next steps are to prioritize these and focus on one area to tackle in more detail taking advantage of complementary activities within OECD or other ICCVAM workgroups. The RAWG is currently considering potential applications for the consensus read-across model for acute oral systemic toxicity developed for a project initiated by the ICCVAM Acute Toxicity Workgroup. The ATWG project leveraged a large animal dataset and the expertise of the international QSAR modeling community to build models for multiple endpoints of regulatory concern defined by ICCVAM agencies (Kleinstreuer et al., 2018 ). The submitted models were then evaluated and combined into a consensus read-across modeling suite that demonstrates equivalent performance to the reproducibility of the animal studies. The consensus model approach could support members of the RAWG to use and interpret read-across predictions for chemicals of interest to their respective Agencies. This type of case study should help build confidence in how data-driven read-across methods are developed, and how predictions can be generated and interpreted in the appropriate context. ICCVAM agencies could also contribute significantly to improving utility and acceptance of read-across approaches through systematic consideration of confidence in QSAR outputs with relevant examples, and the strategic development of additional in vitro data, anchored to existing in vivo data, to increase relevant applicability domains.

Highlights.

• The ICCVAM read-across workgroup (RAWG) scope and charge is presented

• Read-across applications and needs across selected U.S. agencies are summarized

• Two case studies illustrate how read-across differs between 2 of the agencies