3METHODS

3.2.2.1. Database Search Term Development

The literature search focuses on the substance identifiers (name, synonyms, or trade names) with no date or language limits. Substance synonyms are identified by using synonyms in the EPA’s CompTox Chemicals Dashboard⁴ indicated as “valid” or “good”. The preferred chemical name, Chemical Abstracts Service Registry Number (CASRN), DSSTox substance identifier (DTXSID), and synonyms are used by EPA information specialists to develop search strategies tailored to each of the databases listed below.

3.2.2.2. Database Searches

The three databases listed below are queried for literature containing the chemical search terms, and all retrieved records are stored in the Health and Environmental Research Online (HERO) database⁵. Full details of the search strategy for each database are presented in the substance specific ETAP.

• PubMed (National Library of Medicine)⁶
• Web of Science (Clarivate)⁷
• ProQuest (Clarivate)⁸

After deduplication in HERO⁹, records are imported into SWIFT Review¹⁰ software (Howard et al. 2016) to identify those references most likely to be applicable to a human health risk assessment. In brief, SWIFT Review has pre-set literature search strategies (“filters”) developed by information specialists that can be applied to identify studies that are more likely to be useful for identifying human health content from those that likely do not (e.g., analytical chemistry methods). The filters function like a typical search strategy: studies are tagged if the search terms appear in title, abstract, keyword, or medical subject headings (MeSH) fields. The applied SWIFT Review filters focus on lines of evidence: human, animal models for human health, and in vitro studies. The details of the search strategies that underlie the filters are available online¹¹. Studies not retrieved using these filters are not considered further. Studies that include one or more of the search terms in the title, abstract, keyword, or MeSH fields are exported as a Research Information Systems (RIS) file for further screening in DistillerSR¹², as described below.

3.2.2.3. Other Resources Consulted

The literature search strategies described above are designed to be broad; however, as with any search strategy, studies may be missed for assorted reasons (e.g., specific substance is not mentioned in title, abstract, or keyword content; inability to capture “grey” literature not indexed in the databases listed above). Thus, in addition to the database searches, the sources below are used to identify studies that may have been missed. Records that appear to meet the PECO criteria are uploaded into DistillerSR, annotated with respect to source of the record, and screened using the methods described in Section 3.2.3. Other sources consulted include:

• Manual review of the reference list from final or publicly available draft assessments [e.g., EPA Integrated Risk Information System (IRIS), Agency for Toxic Substances and Disease Registry (ATSDR) Toxicological Profiles] or published journal review articles specifically focused on human health. Reviews may be identified from the database search or from ToxValDB.
• Manual review of the reference list of studies judged as PECO-relevant after full-text review.
• Electronic queries of European Chemicals Agency (ECHA) registration dossiers to identify data submitted by registrants¹³.
• Electronic queries of EPA ChemView database¹⁴ to identify unpublished studies, information submitted to EPA under Toxic Substances Control Act (TSCA) Section 4 (chemical testing results), Section 8(d) (health and safety studies), Section 8(e) (substantial risk of injury to health or the environment notices), and FYI (voluntary documents). Other databases accessible via ChemView include EPA’s High Production Volume (HPV) Challenge database and the Toxic Release Inventory (TRI) database.
• Electronic queries of NTP database of study results and research projects¹⁵.
• Electronic queries of the Organisation for Economic Cooperation and Development (OECD) Existing Chemicals Database and eChemPortal^16,17.
• Manual review of the list of references in ECOTOX database for the substance(s) of interest¹⁸.

3.2.2.4. Confidential Business Information

The methods described above are intended to identify evidence that is in the public domain, but additional existing information may not be publicly available. To avoid mislabeling substances as lacking repeated dose toxicity studies, searches of Confidential Business Information (CBI) databases may also be conducted to confirm data availability status. Although the results of CBI studies cannot be considered in many assessment products (including IRIS, PPRTV, ATSDR), confirmation of a true lack of data is an important consideration when determining whether to initiate new toxicological studies. In certain cases, CBI information may be utilized to determine whether an ETAP should be developed.

3.4.1.1. Chemical Purity

Substances evaluated in an ETAP are typically procured from a commercial source, synthesized, or obtained from a reliable third party. The purity of the chemical substance is typically provided by the commercial source and also evaluated independently using the most appropriate analytical method [e.g., liquid chromatography-mass spectrometry (LCMS), gas chromatography-mass spectrometry (GCMS), Nuclear Magnetic Resonance (NMR)]. Quantitative structure activity relationship (QSAR) models may be used to identify the probable physical form, acidity, and analytical (Lowe et al. 2021; Mansouri et al. 2018; Mansouri et al. 2019)method. For most studies, the purity of single chemical test article of 95% or greater is acceptable. The purity of a single chemical test article less than 95% may be acceptable but will be documented accordingly. For mixtures, technical grade chemicals, and formulations, the relative purity and composition should reflect, as close as possible, the relevant human exposure context.

3.4.1.2. Vehicle Selection and Stability

For oral gavage studies²², a set of dosing vehicles are evaluated for chemical solubility and stability. The vehicles may include 1:1:8 Kolliphor:ethanol:deionized water, deionized water with ≤2% Tween® 80, corn oil, deionized water, as well as other options depending on physicochemical properties of the substance. The solubility is assessed visually and/or through analytical measurements. If an aqueous vehicle is used, the pH of the solution with test chemical should be determined, as too low or high of a pH can adversely affect the animal. Chemical stability in the dosing vehicle will also be assessed over a seven-day period.

3.4.1.3. Dose Identification

The approach used to select the dose range for the study will depend on a number of factors that may be specific to the substance of interest. For the ETAP study design, a minimum of eight dose levels plus a vehicle control will be evaluated. The dose range will be based on the highest dose with the dose levels decreasing at half-log₁₀ intervals except for the lowest dose, which will be a full log₁₀ lower than the second lowest dose. Given the intended application of ETAP and its pre-screening criteria, neither in vivo repeated dose toxicity data nor suitable human evidence will be available. Selection of the highest dose will depend on a number of factors that may be specific to the chemical of interest. If existing acute toxicity data are available for the substance of interest, the selection of the highest dose may consider the doses from such studies. If no acute toxicity data are available, in silico approaches (e.g., QSAR modeling) or pilot tolerability/dose range finding studies with limited numbers of animals may be used to inform selection of the highest dose.

3.4.5.1. Sequence Alignment

Raw sequencing reads (FASTQ files) are aligned to known probe sequences listed in the TempO-Seq probe manifest to compute a matrix of read counts for each probe in each sample. Initial quality checks are performed post-alignment to identify samples with insufficient sequencing depth or input RNA to yield reliable results. Each FASTQ file is aligned to the TempO-Seq probe manifest using HISAT2 (Kim et al. 2015; Kim et al. 2019). The alignment results are imported directly into SAMtools (Li et al. 2009) to compute probe-level counts for each individual FASTQ file. Samples are examined for additional quality statistics and those not meeting minimum quality standards are removed from the analysis. Samples that do not pass quality checks may be subjected to reprocessing for RNA isolation and RNA-seq. Quality metrics include (Harrill et al. 2021a):

• Sequencing depth (i.e., total number of mapped reads). Samples with < 10% of target depth are removed from further analysis.
• Fraction of uniquely mapped reads. Samples with < 50% of reads uniquely mapped to known probes are removed from further analysis.
• Probe coverage (i.e., total probes with at least 5 reads). Samples with < 1,200 covered probes are removed from further analysis.
• Signal distribution (i.e., the minimum number of probes that capture 80% of total mapped reads in the sample). No cutoff is applied, but this metric is considered when evaluating potential outlier samples (see below).

3.4.5.2. Sample Normalization

Prior to performing downstream gene expression across samples, probe counts for each sample are normalized to adjust for differences in sequencing depth. For each exposure regimen in each sex and tissue, raw probe counts for all samples (including matched controls) are normalized within each sample as follows:

• All probes with a mean read count < 5 are removed, as these probes lack sufficient signal for reliable analysis.
• Each remaining probe is normalized to Counts Per Million (CPM), which is probe count * 1,000,000 / sum of all remaining probe counts in sample.
• CPM values are transformed to log₂ scale with added pseudo-count of 1 to prevent taking log of zero counts and ensuring a positive value for dose response modeling.

To identify potential outlier samples or batch effects, a principal component analysis (PCA) is performed on subsets of samples corresponding to: 1) all samples corresponding to same substance, tissue, and sex, including matched vehicle controls (“chemical exposure PCA”); and 2) all matched vehicle controls corresponding to the same tissue and sex (“vehicle PCA”). Samples not meeting the sequencing quality metrics (e.g., < 50% of uniquely aligned reads) are excluded prior to PCA analysis. Outlier samples are identified based on the following considerations:

• Individual samples separated from all remaining samples on either principal component #1 (PC1) or principal component #2 (PC2) by >2x the span of all other samples on the corresponding PC are considered strong outliers and removed from further analysis.
• Individual samples separated by <2x the range of all other samples are considered moderate outliers, and additional exclusion criteria are considered:

  • Vehicle samples that appear as moderate outliers on both a chemical exposure PCA and vehicle PCA are excluded unless multiple controls from the same group appeared as outliers.
  • Moderate outlier samples with lower quality than corresponding tissue samples by one or more sequencing quality metrics (e.g., percentage of uniquely mapped reads) are excluded.
  • Samples that appear as moderate outliers in both PC1 and PC2 with a relatively large Euclidean distance from all other remaining samples are excluded.
  • Moderate outlier samples that are especially distant from corresponding replicates or similar doses are excluded.

When multiple outlier samples are present on the same PCA, they are only removed if each outlier sample corresponds to a different dose group, as these are unlikely to represent any reproducible dose-dependent effect. A minimum number of two samples that pass quality control and outlier detection is required for each dose level; individual dose levels not meeting this criterion will be excluded from subsequent analysis. A minimum number of three vehicle control samples that pass quality control and outlier detection is required to proceed with dose response modeling for a given tissue, sex, and exposure regimen.

3.4.5.3. Dose Response Analysis

Once the sequencing data are aligned and normalized, and all low quality and outlier samples removed, dose response modeling is performed. Each data set consists of the series of remaining replicates for all concentrations of a single chemical and matched vehicle controls in the same sex and tissue. The dose response modeling is performed independently on each probe and for each data set using the peer-reviewed BMDExpress software version 2.3 (Phillips et al. 2019; Yang et al. 2007). The dose response analysis procedures are consistent with the NTP Approach to Genomic Dose Response Modeling (NTP 2018), but have been adapted for the specific gene expression platform used in this method (EPA 2024):

• Normalized Log₂(CPM) with added pseudo-count of 1 is used as input.
• For each data set (specific combination of exposure, sex, and tissue), the analysis of variance (ANOVA) pre-modeling test is used to confirm that at least one probe has significant response with a false discovery rate (FDR) < 0.05. If no probes have a significant response, the particular sex and tissue combination is determined to be inactive for the chemical and dose range tested.
• Pre-filtering of probes suitable for dose response modeling is performed using a William’s trend test (p < 0.05) and a mean absolute fold-change relative to vehicle controls of 1.5x or greater in at least one dose.
• Model fitting and BMD determination are performed on each probe passing the pre-filtering criteria:

  • The following dose response models are used in the analysis – linear, second degree polynomial, power, Hill, second degree exponential, third degree exponential, fourth degree exponential, and fifth degree exponential.
  • Models are run assuming a constant variance.
  • For the power model, power is restricted >=1.
  • The model with the lowest Akaike information criterion (AIC) is selected as the best-fit model except in cases where the “k” parameter for the Hill model is less than one-third the lowest dose. In cases for which the “k” parameter for the Hill model is out of bounds, the Hill model is excluded from the final selection (Rowlands et al. 2013; Thomas et al. 2013b).
  • The Benchmark Response (BMR) is set to 1.349 * standard deviation of replicate vehicle control samples (Thomas et al. 2007). Based on EPA guidance, a BMR of 1 standard deviation for continuous data approximates a 10% increase in risk for normally distributed effects when the direction of the effects is known (EPA 2012). However, for most gene expression changes, the direction is not known a priori. To provide an equivalent 10% increase in risk, a BMR of 1.349 * standard deviation is required (Thomas et al. 2007).
  • The BMD, BMDL, and BMD upper confidence bound (BMDU) are calculated for each probe.
  • Only probes meeting the following BMD modeling criteria are included in the next step for gene set summarization (EPA 2024; NTP 2018):

    • BMD < highest dose used in the study
    • Model fit p-value > 0.1
    • BMD/BMDL < 20

3.4.5.4. Gene Set Summarization

BMD results for each exposure/sex/tissue are aggregated into Gene Ontology (GO)²³ biological process classes to identify BMD values. The gene set summarization process is performed as follows:

• Probes are mapped to associated genes. For genes with multiple probes, the BMD/BMDL values from valid probes are averaged. Probes mapping to multiple genes are excluded.
• Genes with conflicting probes are flagged for further review. Using the default setting in BMDExpress, conflicting probes are defined as those with a correlation cut-off of < 0.5 across doses.
• The BMD values for the individual genes are aggregated into GO biological process classes using the current annotations available in BMDExpress.
• GO classes containing fewer than 3 genes with valid BMDs meeting the above criteria are removed from the analysis.
• The BMD and BMDL for each GO class are calculated as the respective medians of corresponding values from the associated genes.

3.4.5.5. POD Identification

The GO biological process class with the lowest median BMD value is identified separately for each tissue examined in each sex. For each tissue and sex, if the lowest median BMD corresponds to more than one GO biological process class, the GO class with the lowest median BMDL value is selected. If there is more than one GO biological process class identified with identical median BMD and BMDL values for the same tissue and sex, then the following criteria are used in the order provided to select the most informative GO class:

• Only the GO classes with the highest number of dose-responsive genes are retained.
• Only the GO classes with the highest percent coverage of the gene set are retained.
• Only the GO classes with the highest (most specific) GO level are retained.

If multiple GO classes remain after applying the selection criteria above, then the additional remaining GO classes will be reported in a footnote.

If the lowest median BMD value identified across all tissues and each sex is more than 3-fold below the lowest positive dose, a ‘no value’ ETAP is declared, and the dose range tested is reported. A follow-up study with an extended dose range may be considered. If the lowest median BMD value identified is less than 3-fold below the lowest positive dose or within the tested dose range, the ETAP is considered valid. The GO biological process class with the lowest median BMD value across all tissues and each sex is then identified. If the lowest median BMD is identical in more than one tissue or sex, the GO biological process class with the lowest median BMDL value is selected. If there is more than one tissue or sex with identical median BMD and BMDL values, then the same criteria described above are used to select the most informative GO biological process class. The median BMDL associated with the identified GO biological process class is selected as the transcriptomic POD. The transcriptomic POD is defined as the dose at which there were no coordinated transcriptional changes that would indicate a potential toxicity of concern. The coordinated transcriptional changes used to identify the POD do not necessarily discriminate between specific hazards, adverse or adaptive effects, nor are they used to infer a mechanism or mode of action. If no tissue in either sex passes the pre-modeling filter nor produces at least one valid GO class, a ‘no value’ ETAP is declared and the dose range tested is reported. A follow-up study with an extended dose range may be considered.

3.6.1.1. Intraspecies Variability Uncertainty Factor (UF_H)

The intraspecies UF_H is applied to account for variation in susceptibility within the human population (interindividual variability) and the possibility (given a lack of relevant data) that the database available is not representative of the exposure/dose response relationship in the subgroups of the human population that are most sensitive to the health hazards of the substance being assessed. As the reference dose is defined to be applicable to “susceptible subgroups,” this UF is used to account for uncertainty in that regard. The adjustment of the intraspecies UF_H from 10 should be considered only if data are sufficiently representative of the exposure/dose response data for the most susceptible human population(s) (e.g., early and late lifestages). The UF_H may be presumed to entail aspects of both toxicokinetic (TK) and toxicodynamic (TD), thus providing an opportunity to integrate traditional and/or NAM-based information that might support reduction in the UF or quantitative application of a data-derived extrapolation factor (DDEF) for human TK (DDEF_HK) and/or human TD (DDEF_HD) for the UF_H (EPA 2014).

For transcriptomic PODs identified in the ETAP, a UF_H of 10 is applied. However, if information is available that informs intraspecies variability or unique sensitivities or susceptibilities of relevance to human populations (e.g., toxicokinetic and/or toxicodynamic variation[s] in human populations), then expert judgment may be used to consider the weight of the evidence to support application of a DDEF_HK and/or DDEF_HD in place of the standard UF_H of 10. However, should human intraspecies information be identified during the evidence mapping phase, consideration should be given to transitioning such a substance to another assessment product line outside of ETAP.

3.6.1.2. Animal-to-Human Interspecies Uncertainty Factor (UF_A)

The interspecies UF_A is applied to account for the extrapolation of laboratory animal data to humans, and it generally is presumed to include cross-species TK and TD uncertainties. With chemical-specific data that informs cross-species scaling of TK (e.g., clearance or plasma T_1/2), the TK half of the UF_A may be reduced from a 3 (i.e., 10^0.5) to a 1 through the development and application of a dosimetric adjustment factor (DAF) that accounts, in general, for differences in TK between animals and humans. In the absence of chemical-specific TK data, a DAF may be applied to a transcriptomic POD obtained from in vivo animal oral exposure study designs using standard EPA guidance and practice, such as BW^3/4 allometric scaling (EPA 2011a). This results in the derivation of a POD human equivalent dose (POD_HED, such as a transcriptomic BMDL_HED).

The UF_A is intended to also account for differences in TD-related species sensitivity between the laboratory animals used for testing and humans. Seldom are there chemical-specific data available to inform TD differences between species, and one-half the standard 10-fold interspecies UF_A (i.e., 10^0.5) is assumed to account for such differences. Unless data support the conclusion that the laboratory test species is more or equally as susceptible to a chemical substance as are humans, and in the absence of any other specific TK or TD data, a UF_A of 3 (in conjunction with calculation of a POD_HED) is applied for the ETAP.

3.6.1.3. Subchronic-to-Chronic Duration Uncertainty Factor (UF_S)

EPA defines a chronic duration as repeated exposure by the oral, dermal, or inhalation route for more than approximately 10% of the life span in humans, corresponding to more than approximately 90 days to 2 years in typically used laboratory animal species (EPA 2002, 2011a). Subchronic duration is defined as repeated exposure by the oral, dermal, or inhalation route for more than 30 days, up to approximately 10% of the life span in humans (more than 30 days up to approximately 90 days in traditional laboratory animal species) (EPA 2002, 2011a). In traditional risk assessment practice, if no chronic duration study is available, information from a subchronic study may be used to support the derivation of an RfD with the application of a UF_S of 10 to the subchronic POD.

Duration extrapolation in the context of an ETAP is informed by multiple previous studies that have demonstrated dose-concordance between traditional apical effect-based PODs derived from longer-term (i.e., chronic) duration studies and gene set-based transcriptomic PODs derived from shorter-term studies (EPA 2024). The concordance was robust across species, sexes, routes or modes of exposure, and technological platforms. In the analysis performed to inform the choices and parameters used in the transcriptomic dose response modeling process, the error in the concordance of the 5-day transcriptomic BMDs with the apical effect BMDs from chronic rodent bioassays was approximately equivalent to the combined inter-study variability associated with the 5-day transcriptomic study and the chronic rodent bioassay (EPA 2024). This demonstrates that the observed differences between the 5-day transcriptomic and chronic apical BMDs are largely driven by inter-study variability in the BMDs, rather than systematic differences. As a result, when using 5-day transcriptomic PODs for non-cancer health effect domains in the ETAP, a UF_S of 1 is applied for considerations of duration in the derivation of a TRV.

3.6.1.4. Lowest Observed Adverse Effect Level (LOAEL)-to-No Observed Adverse Effect Level (NOAEL) Uncertainty Factor (UF_L)

The current EPA approach for dose response assessment prioritizes the application of BMD modeling to identify potential PODs for effects. However, in traditional human health risk assessment practice, when dose response data are not amenable to BMD modeling, point estimates such as LOAELs and NOAELs are identified as potential PODs. A LOAEL is defined as the lowest exposure level at which there are statistically and/or biologically significant increases in frequency or severity of adverse effects between an exposed population and a corresponding control group. A NOAEL is the highest dose level tested at which the specified adverse effect is not produced. Generally, a LOAEL-to-NOAEL uncertainty factor (UF_L) is applied to derive a non-cancer reference value using an apical effect LOAEL if a NOAEL is unavailable. This UF_L is employed to estimate an exposure level below the LOAEL expected to be in the range of a NOAEL. Importantly, the underlying biology leading to and/or resultant of cell, tissue, or organ/system level toxicity invariably involves changes in gene expression. Selecting the gene set with the lowest BMD and BMDL is not necessarily associating transcriptional events with a specific adverse event per se, rather, it is thought to be a dose that approximates a NOAEL.

The gene set summarization of the gene expression changes is described in Section 3.4.5.4 and is suggested as the minimum unit of transcriptional activity to be used in the identification of a POD. That is, BMDLs for single genes are not recommended for POD identification; rather, only those groupings of genes that constitute a GO biological process class in accordance with the criteria outlined in 3.4.5.4 are considered for potential POD (e.g., GO biological process-based BMDL) identification. When GO biological process-based BMDL values are successfully identified for one or more classes using methods consistent with the ETAP, a UF_L of 1 is applied.

3.6.1.5. Database Uncertainty Factor (UF_D)

In traditional human health risk assessment, the UF_D is intended to account for the potential for deriving an under-protective RfD as a result of an incomplete characterization of the substance’s toxicity via the oral exposure route. In addition to identifying data gaps in toxicity information, review of existent data may also suggest that a lower reference value might result if additional data are available. Consequently, in deciding to apply this factor to account for deficiencies in the available data set and in identifying its magnitude, the assessor should consider both the data lacking and the data available for health outcome domains, tissues, or organ systems, as well as life stages. In the context of the ETAP, previous studies have demonstrated that GO biological process-based transcriptomic BMD values following 5 days of exposure are in agreement with BMD values for histopathological effects in two-year chronic rodent bioassays (EPA 2024). Responses in other health effect domains, such as developmental, reproductive, endocrine, neurotoxicity, or immunotoxicity, may not necessarily be accounted for in 5-day in vivo transcriptomic studies. Therefore, a UF_D of 10 should be applied to account for data gaps in the derivation of a TRV for an ETAP.

3.6.1.6. Derivation of the Transcriptomic Reference Value

Using the BMDL_HED from Section 3.5, the standard calculation of the TRV is summarized based on the following equation; however, the exact calculation may vary in unusual circumstances based on the considerations discussed above:

The TRV is defined as an estimate of a daily oral dose to the human population that is likely to be without appreciable risk of adverse non-cancer health effects over a lifetime. The TRV is derived from a transcriptomic POD with uncertainty factors applied to reflect limitations of the data used. While a TRV is expressly presented as a chronic value in an ETAP, it may also be applicable across other exposure durations of interest including short-term and subchronic. This approach has been previously used by EPA in certain risk assessment applications (e.g., PPRTV assessments) wherein a chronic non-cancer reference value has been adopted as a conservative estimate for a subchronic non-cancer reference value when data quality and/or lack of duration relevant hazard and dose response data preclude direct derivation.