Clinical Observations

All rats were observed twice daily for signs of mortality or moribundity. Formal (out of cage) clinical observations were performed daily.

Body and Organ Weights

Animals were weighed during quarantine for randomization on the first day of dosing (study day 0) and on the day of necropsy (study day 5). A gross necropsy was performed on all rats that died spontaneously or were humanely euthanized due to moribund condition. During necropsy for all animals, the heart, liver, and kidneys were removed, and organ weights were recorded; bilateral organs were weighed separately.

Clinical Pathology

Animals were euthanized in random order by CO₂/O₂ (70%/30%) anesthesia 1 day after the final day of dosing. Blood samples were collected from each sex within a 1-hour window and were taken via vena cava or aorta. Blood was collected into tubes containing K₃ EDTA (tripotassium ethylenediaminetetraacetic acid) for hematology analysis and into tubes void of anticoagulant for serum chemistry and thyroid hormone measurements. The following hematology parameters were measured on an Advia^® 120 Hematology Analyzer (Siemens Medical Solutions USA, Inc., Malvern, PA): erythrocyte count, hemoglobin, hematocrit, mean cell volume, mean cell hemoglobin, mean cell hemoglobin concentration, white blood cell count and differential, reticulocyte count, platelet count, and nucleated erythrocyte count. Manual hematocrit was determined using a microcentrifuge and capillary reader. Blood smears were prepared, and qualitative evaluation of cellular morphology was performed per study protocol. The following clinical chemistry parameters were measured on a Roche cobas^® c501 Chemistry Analyzer (Roche Diagnostics, Indianapolis, IN): alanine aminotransferase (ALT), albumin, alkaline phosphatase (ALP), aspartate aminotransferase (AST), total bile acids, total bilirubin, direct bilirubin, cholesterol, creatine kinase, creatinine, glucose, sorbitol dehydrogenase (SDH), total protein, triglycerides, and urea nitrogen. Globulin, albumin/globulin (A/G) ratio, and indirect bilirubin were calculated based on direct measurements (e.g., indirect bilirubin = total bilirubin − direct bilirubin). Serum concentrations for thyroid stimulating hormone (TSH) and free thyroxine (fT4) were determined by immunoassay using commercially available immunoassay kits from EMD Millipore Corporation (Billerica, MA) for TSH and Biomatik Corporation (Kitchener, Ontario, Canada) for fT4. Serum concentrations of total thyroxine (total T4) and total triiodothyronine (total T3) were determined using a validated method as described elsewhere.¹³ Individual animal and summary clinical chemistry, hematology, and hormonal data are available in Appendix F.

Internal Dose Assessment

A screening level assessment of the internal dose was performed to determine whether the test chemical had bioaccumulative properties (i.e., if the half-life was >24 hours). Blood was collected from animals dedicated for internal dose assessment in the 4 and 37 mg/kg groups at 2 and 24 hours following the last dose administered on study day 4. At 2 hours postdose, blood was collected from the jugular vein of unanesthetized animals. At 24 hours postdose (study day 5), blood was collected from all study animals and dedicated internal dose assessment animals from the vena cava or abdominal aorta while animals were anesthetized with CO₂/O₂ (70%/30%). Blood was collected into tubes containing K₃ EDTA and kept on wet ice until plasma isolation, within 2 hours of collection. Samples were stored frozen (−85°C to −60°C) until analysis as described in Appendix A.

Sample Collection for Transcriptomics

Within 5 minutes of euthanasia, samples from the left liver lobe and right kidney were collected from all study animals for transcriptomic analysis. Half of the left liver lobe and half of the right kidney were processed for RNA isolation. Approximately 250 mg of each tissue was cut into small pieces (approximately 5 mm³) and placed into cryotubes containing RNAlater™. The tissue samples were stored at 2°C to 8°C overnight. The RNAlater™ was then removed and the samples were stored in a −85°C to −60°C freezer until processed for RNA isolation.

RNA Isolation, Library Creation, and Sequencing

RNA isolation was performed on tissue samples preserved in RNAlater™. Tissues were homogenized in QIAzol buffer (Qiagen Inc., Valencia, CA) using the TissueLyser II bead-beating system followed by RNA extraction using the Rneasy 96 QIAcube HT kits (Cat# 74171, Qiagen Inc., Valencia, CA) with a DNA digestion step. The concentration and purity of all isolated samples were determined from absorbency readings taken at 260 and 280 nm using a NanoDrop ND-8000 Spectrophotometer (NanoDrop Technologies, Wilmington, DE). The readings accurately determined the concentration of each sample while ensuring that an acceptable purity (A₂₆₀/A₂₈₀ ratio) between 1.80 and 2.20 was achieved. After quantification, RNA was stored at −70°C ± 10°C until further processing.

One microliter of each RNA sample (500–660 ng/µL) was hybridized with the S1500+ beta detector oligo pool mix (2 μL per sample) using the following thermocycler settings: 10 minutes at 70°C, followed by a gradual decrease to 45°C over 49 minutes, and ending with a 45°C hold for 1 minute. Hybridization was followed by nuclease digestion (24 μL nuclease mix addition followed by 90 minutes at 37°C), ligation (24 μL ligation mix addition followed by 60 minutes at 37°C), and heat denaturation (at 80°C for 15 minutes). Ten microliters of each ligation product were then transferred to a 96-well polymerase chain reaction (PCR) amplification microplate with 10 μL of PCR mix per well. Through 25 cycles of amplification, well-specific “barcoded” primer pairs were introduced to templates. Five microliters of the PCR amplification products from each well were then pooled into a single sequencing library. The TempO-Seq library was then processed with a PCR clean-up kit (Machery-Nagel, Mountain View, CA) prior to sequencing. Sequencing was performed using a 50-cycle single-end read flow cell on a HiSeq 2500 Sequencing System (Illumina, San Diego, CA). Processing of sequencing data was conducted using Illumina’s BCL2FASTQ software employing default parameter settings.

Sequence Data Processing

FASTQ files of TempO-Seq reads were aligned to the probe sequences from the target platform using Bowtie version 1.2.2¹⁴ with the following parameters: -v 3 -k 1 -m 1 --best --strata. This configuration allows up to three mismatches and reports the single best alignment. After alignment, the total sequenced reads, the percentage of reads aligning to the platform manifest, the alignment rate, and the percentage of expressed probes (≥5 reads per probe) were calculated for each sample.

Sequencing Quality Checks and Outlier Removal

Samples were flagged for values below the following thresholds: sequencing depth <300 K, total alignment rate <40%, unique alignment rate <30%, number of aligned reads <300 K, or percentage of probes with at least five reads <50%. Filtering on the percentage of expressed probes eliminates biased samples for which the sequenced reads only reflect a small portion of the measured transcriptome. In addition, FastQC was run on all samples to ensure adequate per base quality and per base N content, where N represents bases that could not be identified. This procedure resulted in one kidney 6:1 FTOH sample on one plate being flagged and removed (unique alignment rate of 1.86%).

Principal component (PCA), hierarchical cluster, and inter-replicate correlation analyses were performed. These analyses highlighted three additional outlier liver samples, which were removed before downstream analysis.

The processing of samples from the study of 6:1 FTOH was done in parallel with three other chemicals that were studied under a similar protocol, therefore allowing for a more powerful collective assessment of the data. Specifically, the samples from all four studies were distributed over twelve 96-well plates (i.e., one plate per chemical per tissue and four additional plates with overflow samples for three of the chemicals, with nine doses plus vehicle control). For kidney samples, average read depth per chemical varied across plates. Kidney samples on one of the overflow plates also clustered separately (in the PCA and hierarchical cluster analysis) from the other kidney samples for a given chemical. Therefore, kidney samples on that overflow plate were removed, resulting in one plate of data per chemical for the downstream analysis of kidney samples. The exclusion of these data had limited impact on the analysis as the samples from each dose group were randomly sorted into the overflow plates. The final sample counts that were used for benchmark dose (BMD) analysis of the transcriptomics data are shown in Table 1.

Table 1

Final Sample Counts for Benchmark Dose Analysis of the Transcriptomics Data.

Data Normalization

The aligned read counts for attenuated probes were properly readjusted to calculate unattenuated equivalent counts using the attenuation factors provided in the platform manifest. To account for between-sample sequencing depth variation, unattenuated read counts were normalized at the probe level by applying reads per million normalization. A pseudo-read-count of 1.0 was added to each normalized expression value, and then the values were log2 transformed to complete the normalization. Principal component-based visualizations of the final expression data set used from modeling are available in Appendix C.

Statistical Analysis of Body Weights, Organ Weights, and Clinical Pathology

Two approaches were employed to assess the significance of pairwise comparisons between dosed and vehicle control groups in the analysis of continuous variables. Organ and body weight data, which have approximately normal distributions, were analyzed using the parametric multiple comparison procedures of Williams^15,16 and Dunnett.¹⁷ Clinical pathology data, which typically have skewed distributions, were analyzed using the nonparametric multiple comparison methods of Shirley¹⁸ and Dunn.¹⁹ The Jonckheere test²⁰ was used to assess the significance of dose-response trends and to determine whether a trend-sensitive test (Williams or Shirley test) was more appropriate for pairwise comparisons than a test that assumes no monotonic dose response (Dunnett or Dunn test). Trend-sensitive tests were used when the Jonckheere test was significant at p ≤ 0.01.

Prior to analysis, values identified by the outlier test of Dixon and Massey²¹ were examined by National Institute of Environmental Health Sciences (NIEHS) staff. Values from animals suspected of illness due to causes other than experimental exposure and values that the laboratory indicated as inadequate due to measurement problems were eliminated from the analysis.

A no-observed-effect level (NOEL) was identified as the highest dose not showing a significant (p ≤ 0.05) pairwise difference relative to the vehicle control group. A lowest-observed-effect level (LOEL) was identified as the lowest dose demonstrating a significant (p ≤ 0.05) pairwise difference relative to the vehicle control group. Throughout the results section for apical endpoints, interpretation of BMDs is made in relationship to NOEL and LOEL values for specific endpoints, as defined here, and are not meant to reflect an overall study NOEL or LOEL.

Benchmark Dose Analysis of Body Weights, Organ Weights, and Clinical Pathology

Clinical pathology, body weight, and organ weight endpoints that exhibited a significant trend and pairwise test were submitted in batch for automated BMD modeling analysis. For body weight, the BMD and benchmark dose lower confidence limit (BMD_L) were presented as not determined when there were no significant results. BMD modeling and analysis was conducted using a modification of Benchmark Dose Modeling Software (BMDS) version 2.7.0. Data sets were executed using the Python BMDS interface (https://pypi.python.org/pypi/bmds; version 0.11), which allows for batch processing of multiple data sets. Data for all endpoints submitted were continuous. A default benchmark response (BMR) of one standard deviation (relative to control) was used for all data sets. The following BMDS 2.7.0 models were used to model the means of the data sets:

• Linear
• Polynomial 2°, 3°, 4°, 5°, 6°, 7°, 8°
• Power
• Hill
• Exponential M2, M3, M4, M5

Multiple versions of the polynomial model were executed, from a polynomial of degree 2 to a polynomial of degree equal to the number of dose groups minus 1 (e.g., if a data set had five dose groups, a 2°, 3°, and 4° polynomial model would be executed). Models were initialized using BMDS 2.7.0 model defaults, including restricting the power parameter of the power model and n-parameter of the Hill model to >1 and the beta parameters of the polynomial model to positive or negative, depending on the mean response direction of the data set. For all models, either a constant or nonconstant variance model was selected as outlined in the EPA BMD technical guidance²² and was implemented in the BMDS 2.7.0 software.

After model execution, BMDs were selected using the model recommendation procedures generally described in the EPA BMD technical guidance²² and the automated decision logic described in Wignall et al.²³ and summarized in Appendix D, Table D-1. Models were placed into one of four possible bins, depending on the results and the bin recommendation logic:

1. Failure: model did not successfully complete
2. Nonviable model (NVM): model successfully completed but failed acceptability criteria
3. Not reportable (NR): model is identified and meets all acceptability criteria with the exception of the estimated BMD being below the lower limit of extrapolation (<1/3 the lowest nonzero dose tested); BMD reported as <1/3 the lowest nonzero dose tested and BMD_L is not reportable
4. Viable model: candidate for recommended model without warning

If only one model was in the viable model bin, it was selected as the best-fitting model. If the viable bin had more than one model, consistent with EPA guidance,²² either the model with the lowest Akaike information criterion (AIC) or lowest BMD_L was selected. If the range of BMD_L values was sufficiently close (less than threefold difference), the AIC value was used; otherwise, the BMD_L value was used. If no model was recommended, no BMD was presented in the results. Details on the analysis criteria and decision tree are provided in Table D-1 and Figure D-1, respectively. To avoid effects of model extrapolation, BMD values derived from viable models that were threefold lower than the lowest nonzero dose tested were reported as <1/3 the lowest nonzero dose tested, and corresponding BMD_L values were not reported. Finally, all modeling results from apical data yielding a BMD were reviewed by a subject matter expert to determine the validity of the modeling results and potency estimates.

Benchmark Dose Analysis of Transcriptomics Data

The BMD analysis of the transcriptomic data was performed in accordance with the National Toxicology Program (NTP) best practices for genomic dose-response modeling as reviewed by an independent panel of experts in October 2017. These recommendations are described in the 2018 publication, National Toxicology Program Approach to Genomic Dose Response Modeling.²⁴

Dose-response analyses of normalized gene expression data were performed using BMDExpress 2.30.0507 BETA (https://github.com/auerbachs/BMDExpress-2/releases). A trend test (the Williams trend test^15,16 p ≤ 0.05, 10,000 permutations) and fold change filter (1.5-fold change up or down relative to the vehicle control group for probe sets) were applied to the data set to remove probe sets demonstrating no response to chemical exposure from subsequent analysis. These filter criteria were empirically determined with the goal of balancing false discovery with reproducibility. The criteria are consistent with the MicroArray Quality Control recommendations to combine the nominal p value threshold with a fold change filter to maximize replicability of transcriptomic findings across labs.²⁵ The following dose-response models were fit to the probe sets that passed the trend test and fold change filter:

• Hill
• Power
• Linear
• Polynomial 2°
• Exponential M2, M3, M4, M5

All gene expression data analyzed in BMDExpress were log2 transformed, and thus nearly all probes (also known as detection oligos or DO) were assumed to exhibit constant variance across the doses. For this reason and for efficiency purposes, each model was run assuming constant variance. Lacking any broadly applicable guidance regarding the level of change in gene expression considered biologically significant, a BMR of one standard deviation (relative to the fit at control) was used in this study. This approach enables standardization of the BMR between apical endpoints and transcriptomic endpoints and provides a standard for use across multiple chemicals tested in this rapid screening paradigm. The expression direction (upregulated or downregulated) for each probe was determined by a trend test intrinsic to the model executables (provided by EPA) contained in BMDExpress.

To identify the best-fit model for each fitted probe, the AIC values for each fitted model were compared and the model with the lowest AIC was selected. The best model for each probe was used to calculate the BMD, BMD_L, and BMD upper confidence limit (BMD_U). The specific parameter settings, selected from the BMDExpress software when performing probe-level BMD analysis, were as follows: maximum iterations – 250, confidence level – 0.95, BMR factor – 1 (the multiplier of the standard deviation that defined the BMD), restrict power – no restriction, and constant variance – selected. The specific model selection setting in the BMDExpress software when performing probe set-level BMD analysis was as follows: best poly model test – lowest AIC, flag Hill model with “k” parameters – <1/3 the lowest nonzero dose tested, and best model selection with flagged Hill model – include flagged Hill model. The inclusion of the flagged models is a deviation from EPA BMD analysis guidance.²² The justification for this deviation relates to subsequent use of the data in which the probe BMD values are grouped into gene sets from which a median BMD is derived. If the probes were removed from the analysis or forced to another model, the probe might not be counted in the gene set analysis and could lead to loss of “active” gene sets. Importantly, most of the probes that produce flagged Hill models show highly potent responses and should therefore be counted in the analysis.

To perform Gene Ontology (GO; annotation accession date: 07/15/2020) gene set analysis, only GO terms with ≥10 and ≤250 annotated genes measured on the gene expression platform were considered. Before sorting genes into the GO terms, the best-fit model for each probe was subjected to a filtering process to remove those probes (1) with a BMD greater than the highest dose tested, (2) that mapped to more than one gene, (3) that had a global goodness-of-fit p value ≤0.1, and (4) with a BMD_U/BMD_L ratio >40. GO terms that were at least 5% populated and contained three genes that passed the criteria mentioned above were considered “active” (i.e., responsive to chemical exposure). For this report, GO terms populated with identical sets of differentially expressed genes were filtered to limit redundancy in reporting based on the following selection criteria: (1) highest percentage populated and (2) most specific/highest GO level. Redundant GO terms failing to differentiate on the basis of these criteria were retained and reported. A complete list of “active” GO terms can be found in Appendix F. To avoid effects of model extrapolation, GO terms exhibiting BMD values below the lower limit of extrapolation (<1/3 the lowest nonzero dose tested) were reported as <1/3 the lowest nonzero dose tested and corresponding BMD_L and BMD_U values were not reported.

To perform Individual Gene Analysis, the best-fit model for each probe was subjected to a filtering process to remove those probes (1) with a BMD greater than the highest dose tested, (2) that mapped to more than one gene, (3) that had a global goodness-of-fit p value ≤0.1, or (4) with a BMD_U/BMD_L ratio >40. For genes that had more than one probe represented on the platform and passed this filtering process, a median BMD was used to estimate the BMD, BMD_L, and BMD_U values. To ensure only genes with a robust response were assessed for potency, genes with probes that had a median fold change <|2| were removed prior to reporting. A complete list of genes and their corresponding metrics can be found in Appendix F. To avoid effects of model extrapolation, genes exhibiting BMD values below the lower limit of extrapolation (<1/3 the lowest nonzero dose tested) were reported as <1/3 the lowest nonzero dose tested and corresponding BMD_L and BMD_U values were not reported.

A summary of the BMDExpress gene expression analysis pipeline used in this study is shown in Figure D-2.

Empirical False Discovery Rate Determination for Genomic Dose-response Modeling

The genomic dose-response analysis pipeline is a complex multistep process with multiple modeling steps and parameter variables. Because of this complexity, traditional statistical models for determining false discovery rates for the genes and pathways are not straightforward to apply. To overcome this issue, an empirical false discovery rate was determined on the basis of the totality of the analysis pipeline. This was done through the evaluation of synthetic null data sets derived from vehicle control data from four short-term repeat dose toxicogenomic studies including 6:1 FTOH (each with 10 vehicle control samples). The other toxicogenomic studies, which are reported in separate NIEHS reports, are of perfluorohexanesulfonamide,²⁶ 1,1,2,2-tetrahydroperfluoro-1-dodecanol,²⁷ and 2,3-benzofluorene.²⁸ Samples from all four studies were processed as a group and subjected to sequencing at the same time and were visually inspected to ensure there was no batch effect between the different studies.

To create synthetic null data for a given group (tissue per sex combination), up to 40 vehicle control samples from the original studies (10 replicates × 4 chemicals) were used to generate the data sets, with outliers excluded from the analysis. Each computationally generated sample was created by mixing two randomly selected vehicle control samples via a weighted average approach through which weights were obtained from random uniform (0,1) distribution. A total of 55 samples (10 vehicle control samples + 45 dosed samples [9 doses × 5 replicates]) were computationally generated per data set and assigned doses spaced by approximately half-log. A total of 20 data sets were generated per group (i.e., 20 data sets each for female kidney, male kidney, female liver, and male liver) and analyzed using both the individual gene-level and GO biological process (gene set) analysis pipeline employed to analyze the data from each study. The median empirical false discovery rates across the 20 null sets in each group for gene-level analysis across each group were 0.037%, 0.037%, 0%, and 0% (female kidney, male kidney, female liver, and male liver, respectively). The median empirical false discovery rate for each of the 20 null data sets in each group using the GO biological process (gene set) level analysis was 0%. Details of the empirical false discovery rate analysis are available in Appendix C. The associated bm2 analysis file that is the basis of the empirical false discovery rate can be found in Appendix F.

Data Accessibility

Primary and analyzed data used in this study are available to the public at https://doi.org/10.22427/NIEHS-DATA-NIEHS-07.²⁹