Quantitative Integration of Mode of Action Information in Dose–Response Modeling and POD Estimation for Nonmutagenic Carcinogens: A Case Study of TCDD

Abstract

Background:

Traditional dose–response assessment applies different low-dose extrapolation methods for cancer and noncancer effects and assumes that all carcinogens are mutagenic unless strong evidence suggests otherwise. Additionally, primarily focusing on one critical effect, dose–response modeling utilizes limited mode of action (MOA) data to inform low-dose risk.

Objective:

We aimed to build a dose–response modeling framework that continuously extends the curve into the low-dose region via a quantitative integration of MOA information and to estimate MOA-based points of departure (PODs) for nonmutagenic carcinogens.

Methods:

2,3,7,8-Tetrachlorodibenzo-p-dioxin (TCDD) was used as an example to demonstrate the new dose–response modeling framework. There were three major steps included: a) identifying and extracting key quantifiable events (KQEs), b) calculating essential doses that sequentially activate KQEs using the benchmark dose (BMD) methodology, and c) characterizing pathway dose–response relationship for MOA-based POD estimation.

Results:

We identified and extracted six KQEs and corresponding essential events composing the MOA of TCDD-induced liver tumors. With the essential doses estimated from the BMD method using various settings, three link functions were applied to model the pathway dose–response relationship. Given a toxicologically plausible definition of adversity, an MOA-based POD was derived from the pathway dose–response curve. The estimated MOA-based PODs were generally comparable with traditional PODs and can be further used to calculate reference doses (RfDs).

Conclusions:

The proposed framework quantitatively integrated mechanistic information in the modeling process and provided a promising strategy to harmonize cancer and noncancer dose–response assessment through pathway dose–response modeling. However, the framework can also be limited by data availability and the understanding of the underlying mechanism. https://doi.org/10.1289/EHP12677

Introduction

Chemical risk assessment, a vital tool to protect human and environmental health, has been widely used to evaluate the toxicity of substances by regulatory agencies.^1,2 Recently, in addition to typical toxicity data obtained from experimental animal tests, data from emerging methods^3–5 (e.g., high-throughput microarrays and computational toxicology tools) have been increasingly used to support chemical risk assessment. An important component in risk assessment is dose–response assessment,⁶ which primarily aims to identify a “safe” exposure level of a certain chemical for a target population typically based on two steps: point of departure (POD) derivation and low-dose extrapolation.⁶ Regardless of the methods to derive the POD [e.g., no observed adverse effect level (NOAEL), lowest observed adverse effect level (LOAEL), and benchmark dose/benchmark dose lower confidence limit (BMD/BMDL)⁷], the POD is mainly determined from the critical effect identified by risk assessors based on either toxicological or epidemiological studies,⁶ i.e., the most sensitive adverse effect. Conversely, the default low-dose extrapolation methods are different for cancer and noncancer effects,^8,9 which may not necessarily exhibit the mechanism of adverse effect adequately. With respect to noncarcinogenic effects, a threshold at low dose level is always assumed,¹⁰ and thus a nonlinear extrapolation method⁶ that generates a reference dose/concentration (RfD/RfC) by dividing the POD by multiple uncertainty factors to address variability and differences between the testing and target species/studies is applied. On the contrary, for carcinogenic effects, low dose extrapolation adopts a default linear approach to derive a cancer slope factor (CSF) for cancer risk estimation, under the assumption that all carcinogens are mutagens that pose a threat to health even at an extremely low concentration.^9,11 However, without considering the mode of action or mechanistic data of nonmutagenic carcinogens, the method may result in overly conservative estimates.

The latest Guidelines for Carcinogen Risk Assessment of the U.S. Environmental Protection Agency (EPA)⁹ suggest that mode of action (MOA) should be considered in selecting the method for low-dose extrapolation for cancer effect. In some cases, an RfD/RfC rather than a CSF can be developed if evidence suggests cancer results from a nonmutagenic MOA. Considering MOA in low-dose extrapolation is an important step forward in cancer risk assessment; however, adopting the refinement does not necessarily mean that the MOA information has been properly utilized in POD estimation. Notably, the POD is still derived from a single apical effect, and the two fundamental approaches of default low-dose extrapolation (i.e., linear vs. nonlinear) remain unchanged. Additionally, the RfD/RfC method abandons an important advantage of estimating a CSF, i.e., utilizing the CSF to calculate cancer risk.^10,12 Therefore, a new framework that utilizes and quantitatively integrates MOA information from various sources to avoid unnecessary dichotomized low-dose extrapolation is critically needed to improve dose–response assessment of nonmutagenic carcinogens.

The objective of this study is to develop a novel dose–response assessment framework that estimates an MOA-based POD by quantitatively integrating key events for nonmutagenic MOA of liver tumors into the modeling process using 2,3,7,8-tetrachlorodibenzo-p-dioxin (TCDD) as an example. Being one of the typical nonmutagenic carcinogens, extensive study results of this xenobiotic have been produced.¹³ Despite the restriction on the manufacture of TCDD since the late 1970s,¹⁴ human exposure to this compound has continued via the food chain or environmental media.^15,16 Primarily mediated by aryl hydrocarbon receptor (AhR) activation, TCDD exposure has been associated with tumor development (e.g., hepatocellular adenomas and carcinomas), demonstrated in many animal studies.¹⁷ Studies in the last four decades have provided relatively abundant toxicological data to propose a well-defined MOA of TCDD based on genomic, molecular, cellular, tissue, and organ essential effects. Besides, in light of sufficient epidemiological evidence, the International Agency for Research on Cancer¹⁸ determined TCDD as a human carcinogen. Therefore, TCDD can be used as an excellent example to demonstrate the proposed modeling methodology.

Methods

Overview

We proposed a new dose–response modeling framework to quantitatively integrate a series of dose-dependent biological events along a specific MOA over the entire dose continuum for toxicity evaluation so that a dichotomized low-dose extrapolation (i.e., linear vs. nonlinear) can be avoided. There were three key steps in the framework, which are graphically shown in Figure 1, ranging from information extraction to the construction of pathway dose–response relationship. In this study, the term “endpoint” referred to a particular biological change or alteration, and the term “event” was defined as a specific endpoint with a time stamp (so the same endpoints observed or measured at different time points were considered as different events).

Figure 1.

The overview of the proposed dose–response modeling framework. Note: BMD, benchmark dose; BMDL, benchmark dose lower confidence limit; KQE, key quantifiable event. The y-axis “Duration” in Step 3 is the normalized exposure duration for individual events (between 0 and 1).

In Step 1, based on the components in a typical MOA framework, we first defined key quantifiable events (KQEs), the building blocks for MOA-based dose–response modeling in this study. Then, the KQEs characterized by various biological events were identified and extracted from peer-reviewed literature for MOA formation and dose–response data collection. Given the dose–response data for these identified KQEs at molecular, cellular, and organ levels, in Step 2, BMDs/BMDLs (i.e., daily administered doses in the unit of ng/kg/day) for individual events were estimated using the BMD methodology. Because the same endpoint can be observed or measured at multiple time points (e.g., pigmentation was observed at 31, 53, and 105 wk), we needed to determine the essential event of the endpoint, i.e., the event that was involved in the MOA pathway, based on the carcinogenesis process, data availability, and data suitability for BMD modeling (i.e., we usually chose the earliest event with clear and marked changes in response). An essential dose was then defined as a cumulative dose converted from the BMD or BMDL corresponding to a specific essential event. Step 3 integrated essential doses for all KQEs and corresponding temporal information to form a pathway dose–response relationship where “dose” was the essential dose on a log-scale and “response,” which reflects effect severity, was the normalized exposure duration of the various events. A beta regression with different link functions was applied to quantify the pathway dose–response relationship based on which an MOA-based POD was ultimately estimated.

Identifying and Extracting KQEs

In this study, we defined KQEs of TCDD as any quantifiable biological events that might be critical to TCDD-induced liver tumors (i.e., adenomas and carcinomas). The identified KQEs could include key events (KEs), associative events (AEs), and modulating factors (ModFs), as indicated in the typical MOA framework.¹⁹ Although KEs are biologically necessary and observable precursors contributing to a specific outcome,¹⁹ it is difficult or even impossible to directly measure them in some situations. Instead, as the indirect biomarkers corresponding to KEs, AEs play a key and reliable role in constructing the MOA.¹⁹ Conversely, ModFs modulate the time stamp of KEs’ occurrence and thus may alter the dose–response relationship in its shape and magnitude but cannot impact the necessity of KEs.^19,20 Herein, we considered these three components simultaneously when identifying and extracting KQEs. It is generally believed that the sustained AhR activation acts as the beginning of TCDD-induced liver tumors,^20,21 so this event was viewed as the first step toward hepatic carcinogenesis. As both dose and time accumulate, KQEs after the sustained AhR activation were assumed to follow sequentially biological alterations at molecular, cellular, and organ levels, which exhibited a concordance with the modified Hill criteria.⁹

To find and extract TCDD-related dose–response information to be utilized for building a pathway dose–response relationship, we performed a literature search in the PubMed database. We primarily focused on in vivo studies using female Sprague-Dawley rats as the testing animal because sufficient evidence suggests that TCDD-based MOA is generally sex-dependent and relatively abundant and adequate data are available for female Sprague-Dawley rats.¹³ The keywords that we used for identifying relevant studies were as follows: “2,3,7,8-tetrachlorodibenzo-p-dioxin,” “TCDD,” “female Sprague-Dawley rats,” “liver,” “hepatic,” “cancer,” “tumor,” and “neoplasm.” The search query string we utilized was as follows: (2,3,7,8 -tetrachlorodibenzo-p-dioxin OR TCDD) AND (female Sprague-Dawley rats) AND (liver OR hepatic) AND (cancer OR tumor OR neoplasm). The date when the search was last run was 15 March 2021. We focused on peer-reviewed articles written in English without restriction on publication date. Two reviewers independently worked on reviewing both the abstract and the full text of each article. Any articles related to TCDD ranging from specific receptor activation and gene expression to perceptible alterations in livers were included based on abstract screening. Then, the full texts of these papers were screened according to the following criteria: a) the eligible studies should explicitly report dose–response data that applied at least three dose levels (including the control group); b) any studies that were not able to show quantifiable events were excluded; and c) any quantifiable dose–response relationships with only initiation and/or partial hepatectomy designs were not taken into account due to potentially spurious results. After the two-step screening, we retrieved all KQEs and their corresponding dose–response data, both of which were used in developing a pathway dose–response relationship. Currently, despite limited evidence showing AhR relevance for a few quantifiable molecular, cellular, and organ responses occurring after the changes in the AhR-related gene expression, these events were still involved in TCDD-based MOA as part of uncertainty about the extant databases and the mechanistic pathways that have been partially understood.

Calculating BMDs/BMDLs of Individual KQEs Using the BMD Methodology

The dose–response data for individual KQEs typically fell into two categories, i.e., dichotomous and continuous data. Dichotomous data feature binary responses showing whether an adverse effect occurred or not (e.g., cancer incidence), while for continuous data, responses (e.g., body weight and index for proliferation) are measured on a continuous scale and typically expressed by mean and standard deviation. In this study, BMDs/BMDLs of KQEs were estimated using the Bayesian BMD modeling system (BBMD).²² We applied empirical dose–response models to estimate individual-model and model-averaged BMD and BMDL (which was named “BMA model” for short) as suggested by Shao and Shapiro.²² Although a physiologically based pharmacokinetic (PBPK) model can be taken into account to convert daily dose to body burden of TCDD, it is not necessary to do so due to the consistency of dose metric used in individual studies for key events, the linearity in dose metric conversion, and ultimate focus on oral administered doses for practical regulatory consideration.

To evaluate genomic changes, we utilized the transcriptomic responses (i.e., total RNA) to the AhR activation in rat hepatocytes (i.e., accession number GSE9838)²³ from the Gene Expression Omnibus database (GEO; https://www.ncbi.nlm.nih.gov/geo/).²⁴ Since modulation of gene expression was assumed to be an earlier KQE relative to altered focal cell growth, preneoplastic focal tissue changes, and liver tumors, it was rational to utilize the genomic data with short exposure duration, i.e., the dose–response relationship with three-consecutive-day exposure from the in vivo study (accession number GSE9838).²³ The BMD analyses using the gene expression data were conducted in BBMD following the procedure described in Ji et al.²⁵ including data preprocessing, dose–response trend screening, dose–response modeling, gene-level BMD estimation, etc. We chose 1% and 10% extra risk for dichotomous data and 1% and 10% relative change for continuous ones. According to the recommendation of the National Toxicology Program (NTP),²⁶ 0.1 and 1 standard deviation (SD) were selected as benchmark responses (BMRs) for genomic data BMD estimation.

Characterizing the Pathway Dose–Response Relationship for MOA-Based POD Estimation

One fundamental assumption for the pathway does-response relationship was that the KQEs along the pathway can be sequentially activated as dose cumulates and the essential dose was the cumulative dose (e.g., BMD/BMDL) that can activate the corresponding event. Therefore, prior to building the pathway dose–response relationship, all BMDs or BMDLs (unit: ng/kg/day) estimated in the previous step were converted to cumulative doses (unit: ng/kg) to achieve the relationship of a higher dose leading to a more severe response, based on the following equation: $\text{cumulative dose}=\text{BMD}\left(\text{or}\text{BMDL}\right)\times \text{days per week}\times \text{total exposure weeks}$. For the dose–response dataset of which exposure frequency was not explicitly reported, exposure frequency of once per day was assumed.

Besides using the essential dose level of each event (i.e., the cumulative dose that can activate the event) to form the x-axis, we also needed to carefully design the y-axis to form the dose–response data for the proposed pathway dose–response relationship. In this study, we defined the y-axis based on the time stamp of the identified events along the proposed pathway with two important considerations: a) the severity of the response should increase as the value on the y-axis increases; and b) the value range of y-axis should be appropriate for modeling. Therefore, we used the normalized exposure duration as the y-axis; that is, the time stamp of the identified event occurrence was divided by the time stamp of the apical event (i.e., the last event on the pathway). For example, the y-axis value of liver weight decrease observed at 13 wk was 13 divided by 105 (i.e., the number of weeks when liver tumor incidence was observed/measured) which was equal to 0.133. Consequently, this conversion method not only limited the range of the y-axis to [0, 1] but also quantitatively expressed the severity of the events. Additionally, because cumulative dose levels can range substantially (due to the exposure duration from 72 h to 105 wk), we used the log-transformed doses [i.e., ln(BMD) or ln(BMDL)] in dose–response modeling. Because the value of the dependent variable (i.e., the value of the y-axis) can be any number between 0 and 1 (not 0 or 1 like dichotomous data), a beta regression was applied to model the pathway dose–response data in software R version 4.0 with three link functions listed below.

$$\mathrm{\mu }\left(x_i\right)=\frac{e^{a+b\times x_i}}{\left(1+e^{a+b\times x_i}\right),}$$ (1)

$$\mathrm{\mu }\left(x_i\right)=pnorm\left(a+b\times x_i\right),$$ (2)

$$\mathrm{\mu }\left(x_i\right)=1-e^{-e^{a+b\times x_i}},$$ (3)

where Equations 1, 2, and 3 are the logistic, probit, and complementary log-log (cloglog) models, respectively, $x_i$ is the dose level on a log scale, $\mathrm{\mu }\left(x_i\right)$ is the mean of the response variables (i.e., the normalized exposure duration), and a and b are parameters to be estimated from the regression.

The concept of pathway dose–response modeling is graphically presented as Step 3 in Figure 1. Some alternative modeling settings (e.g., the selection of dose–response models, BMR values) in the step of BMD modeling for estimating the essential doses were compared to investigate their impact on the MOA-based POD, which are presented in the “Results.”

The fitted pathway dose–response curve basically described the relationship between cumulative doses and the severity of key events that occurred over the entire dose continuum. To calculate an MOA-based POD from the curve, we need to define an appropriate y-value reflecting the toxicologically plausible adversity. There were three possible approaches to define adversity and calculate the corresponding MOA-based POD of TCDD considered in this study. First, since short-term adverse outcomes did not necessarily persist in the long run, we felt that it was legitimate to select the first perceptibly irreversible endpoint as the critical event reflecting an adequately adverse effect of TCDD. The corresponding y-value of this event, thus, was viewed as a reasonable value used for MOA-based POD derivation. The second approach was to define adversity based on the precursor events of TCDD-induced liver tumors. In accordance with U.S. EPA,⁹ KEs were defined as precursor steps that are necessary components themselves or pivotal biomarkers corresponding to those components in the MOA, which suggested that we can choose the KEs, at least the commonly acceptable ones along a well-defined MOA as precursor events. The corresponding time point of the precursor event’s occurrence was then considered adverse. The third method was to focus on the events that occurred after the minimum duration of chronic exposure, which was defined as more than 3 months in Klaassen.²⁷ The minimum duration of chronic exposure was assumed to be the approximated adversity to show adequate untoward effects. Given the available durations in this study, we assumed a reasonable value for the y-variable, based on the third method, was 14 wk, which was slightly more than 3 months.

Given the fitted pathway dose–response curve and defined adversity, we calculated the dose level that caused the defined response level, i.e., the MOA-based POD. A bootstrap algorithm was employed to derive the confidence interval of the MOA-based POD; i.e., the essential doses of the identified KQEs along the pathway were resampled 1,000 times while keeping the sample size unchanged (i.e., the number of KQEs in each simulated pathway dose–response dataset stayed the same). The 5th and 95th percentile of the simulated MOA-based PODs represented the statistical lower and upper bound of the MOA-based POD.

Additionally, to address the question of how the MOA-based POD can be affected if some of the events were misclassified into an inappropriate category of KQEs, a sensitivity analysis was also conducted to examine the impact by simply using the earliest available event to form the pathway dose–response relationship.

Results

Identified KQEs of TCDD-Induced Liver Tumors and Data Availability

To develop an MOA-based dose–response modeling framework and illustrate this method using TCDD as a case due to extensive knowledge of the carcinogenicity of dioxins and abundant supporting data, we first identified KQEs, the essential components of the MOA regarding liver tumors. A total of 75 articles were retrieved following a literature search in PubMed. By screening the abstracts and full texts of these articles, eight studies were considered eligible to extract quantifiable events and corresponding dose–response data.^{13,23,28–33} Also taken together with the existing MOA information,^20,21 we ultimately identified six KQEs and a series of corresponding bioprocesses to establish the MOA of TCDD-induced liver tumors, as shown in Figure 2.

Figure 2.

Our established MOA of TCDD-induced liver tumors. KQE 1: Sustained AhR activation; KQE 2: Gene expression alterations; KQE 3: Enzyme and protein changes; KQE 4: Altered focal cell growth; KQE 5: Preneoplastic focal tissue changes; KQE 6: Liver tumor. Note: AhR, aryl hydrocarbon receptor; ARNT, aryl hydrocarbon receptor nuclear translocator; A4H, acetanilide-4-hydroxylase; EGF, epidermal growth factor; EROD, 7-ethoxyresorufin-O-deethylase; KQE, key quantifiable event; MOA, mode of action; PROD, 7-pentoxyresorufin-O-deethylase; TCDD, 2,3,7,8-tetrachlorodibenzo-p-dioxin.

1. We considered the first KQE to be sustained AhR activation. Few eligible in vivo dose–response studies directly representing AhR or aryl hydrocarbon receptor nuclear translocator (ARNT) protein changes in liver were available, because most studies (e.g., Pollenz et al.³⁴) did not evaluate at least three doses, which was required for BMD modeling. Therefore, the essential dose level of the AhR activation was not calculated.

2. The second KQE considered in this study was gene expression alterations. Since multiple xenobiotics, including TCDD, may affect expression of certain AhR-related genes,³⁵ we selected a set of AhR battery genes as the elements of KQE2, including Aldh3a1, Cyp1a1, Cyp1a2, Cyp1b1, and Nqo1, which are strongly associated with AhR activation^23,36 to reduce chemical-specific uncertainty.

3. KQE 3 was enzyme and protein induction. Sufficient and qualified data on the activities of 7-ethoxyresorufin-O-deethylase (EROD), 7-pentoxyresorufin-O-deethylase (PROD), and acetanilide-4-hydroxylase (A4H) were reported by NTP,¹³ and thus they were used to characterize TCDD-induced effects at molecular level in this study, considering their Cyp-associated characteristics. In addition, various induced cytokines [e.g., transforming growth factor alpha ($\text{TGF-}\mathrm{\alpha }$), interleukin 1 beta ($\text{IL-}1\mathrm{\beta }$)] and epidermal growth factor (EGF) acting as ModFs were potentially involved in TCDD-related responses and affected proliferation.²⁰ Given the availability of quantifiable dose–response data, we also leveraged altered properties of EGF receptor³³ (e.g., maximum binding capacity) as a component in KQE 3.

4. Accumulated changes at the molecular level that could further alter focal cell growth or homeostasis (e.g., apoptosis inhibition and/or cell proliferation) were considered as the fourth KQE. The endpoints including altered hepatic foci, hepatocellular bromodeoxyuridine (BrdU) labeling index, bile duct and/or oval cell proliferation, altered absolute and/or liver weight, and nodular hyperplasia were utilized to indicate cell proliferation.^13,29,30,32

5. Perceptibly preneoplastic focal tissue changes induced by the previous accumulative effects were considered as KQE 5, such as hypertrophy, pigmentation, steatosis, alterations of cell types and nuclei, toxic hepatopathy, inflammation, fibrosis, necrosis, etc.^13,28,29

6. In the final stage (i.e., KQE 6), hepatocellular adenoma and carcinoma would occur.^13,28,31 It was worth noting that other possible ModFs along the proposed MOA of TCDD, such as inhibited cell communication, altered estrogen, competitive AhR ligands, mitochondrial injury, mitoinhibition, etc., were not taken into account because of limited quantifiable data based on the the PubMed search.

Table 1 demonstrated detailed study design regarding dose level, exposure frequency, and exposure duration based on the eight eligible studies.

Table 1.

The summary of study design based on the eight eligible studies.

No.	Reference	Dose level (ng/kg/day)	Exposure frequency	Exposure duration
1	Goodman and Sauer28	0, 1, 10, and 100	Not reporteda	2 years
2	Harrill et al.29	0, 3, 22, 100, 300, and 1,000	4–5 days/week (total 19 doses)	4 weeks
3	Kociba et al.30	0, 1, 10, 100, and 1,000	5 days/week	13 weeks
4	Kociba et al.31	0, 1, 10, and 100	Not reporteda	2 years
5	Maronpot et al.32	0, 3.5, 10.7, 35.7, and 125	Once every 2 weeks	30 weeks
6	NTP13	0, 3, 10, 22, 46, and 100	5 days/week	14, 31, and 53 weeks
7	Sewall et al.33	0, 3.5, 10.7, 35.7, and 125	Once every 2 weeks	30 weeks
8	Silkworth et al.23	0, 300, and 3,000	Once per day	3 days

^a Exposure frequency was not explicitly reported in the corresponding article.

Essential Doses and the Exposure Duration of Events

Each adverse endpoint that was described in the previous section—together with the exposure duration—is listed in Table 2. The exposure duration was determined from corresponding essential events based on the carcinogenesis process (Figure 2), data availability (Table 1), and data suitability for BMD modeling (Figure S1–S8). To calculate the essential dose for these events using the BMD methodology, three strategies were employed as follows: a) using simple dose–response models to calculate BMD/BMDL (i.e., the linear model for continuous data and the logistic model for dichotomous data, both models have only two parameters); b) using complex dose–response models (i.e., the Hill model for continuous data and the dichotomous-Hill model for dichotomous data, both models have four parameters); and c) using Bayesian model averaged BMD estimation (i.e., BMA model). The estimated BMD or BMDL using the BMA method is listed in Table 2 as well. For more comprehensive modeling results, please refer to Table S1 in the supplemental material.

Table 2.

Benchmark doses (BMDs) and benchmark doses (lower confidence limit) (BMDLs) from model-averaged (BMA) BMD estimation.

Event	No.	Endpoints	Duration (wk)	$\text{BMR}=10\%$ or 1 SDa (BMD/BMDL, ng/kg/day)	$\text{BMR}=1\%$ or 0.1 SDa (BMD/BMDL, ng/kg/day)	Data source
KQE 1	NA	AhR or ARNT protein changes expected		NA		NA
KQE 2	1	Aldh3a1 b	0.429	698.38/469.04	71.35/47.39	Silkworth et al.23
	2	Cyp1a1 b	0.429	14.75/4.51	1.45/0.43	Silkworth et al.23
	3	Cyp1a2 b	0.429	34.88/9.49	3.72/0.86	Silkworth et al.23
	4	Cyp1b1 b	0.429	305.80/147.54	36.56/14.57	Silkworth et al.23
	5	Nqo1 b	0.429	556.98/50.45	57.58/5.29	Silkworth et al.23
KQE 3	6	Acetanilide-4-hydroxylase (A4H)	14	0.82/0.61	0.08/0.06	NTP13
	7	7-Ethoxyresorufin-O-deethylase (EROD)	14	$6.90\times {10}^{-3}/5.30\times {10}^{-3}$	$6.92\times {10}^{-4}/5.30\times {10}^{-4}$	NTP13
	8	7-Pentoxyresorufin-O-deethylase (PROD)	14	0.19/0.15	0.02/0.01	NTP13
	9	EGF receptor (maximum binding capacity)	30	2.21/0.87	0.21/0.08	Sewall et al.33
	10	EGF receptor (equilibrium dissociation constant)	30	13.73/1.16	4.95/0.10	Sewall et al.33
KQE 4	11	Altered hepatic foci volume (${\text{Foci}/\mathrm{cm}}^2$)	30	5.07/0.22	0.71/0.02	Maronpot et al.32
	12	Altered hepatic foci volume (${\text{Foci}/\mathrm{cm}}^3$)	30	5.30/0.20	0.89/0.02	Maronpot et al.32
	13	Bromodeoxyuridine (BrdU) labeling	31	4.59/1.26	0.48/0.13	NTP13
	14	Absolute liver weight	14	1.51/0.51	0.15/0.04	NTP13
	15	Relative liver weight	4	15.54/9.88	1.33/0.80	Harrill et al.29
	16	Bile duct hyperplasia	53	19.65/6.38	3.92/0.63	NTP13
	17	Oval cell hyperplasia	105	12.29/7.21	2.25/0.83	NTP13
	18	Nodular hyperplasia	105	42.40/29.75	14.45/7.43	NTP13
KQE 5	19	Hypertrophy	4	4.04/2.10	1.84/0.75	Harrill et al.29
	20	Pigmentation	31	2.26/0.74	0.43/0.07	NTP13
	21	Diffuse fatty change	53	42.64/18.41	16.98/2.28	NTP13
	22	Toxic hepatopathy	53	25.55/13.23	13.31/2.33	NTP13
	23	Mixed cell focus	53	5.78/2.46	0.64/0.23	NTP13
	24	Multinucleated hepatocyte	31	30.70/17.18	14.83/3.31	NTP13
	25	Eosinophilic focus 1	105	10.89/6.05	1.17/0.58	NTP13
	26	Eosinophilic focus 2	105	2.47/0.67	0.50/0.09	Goodman and Sauer28
	27	Portal fibrosis	105	53.70/42.23	26.48/14.91	NTP13
	28	Necrosis	105	36.91/20.91	4.12/2.03	NTP13
	29	Inflammation	4	35.47/15.51	3.56/1.47	Harrill et al.29
KQE 6	30	Hepatocellular adenoma 1	105	83.26/68.11	50.30/27.48	NTP13
	31	Hepatocellular adenoma 2	105	9.31/3.00	3.26/0.86	Goodman and Sauer28
	32	Hepatocellular carcinoma 1	105	65.88/29.50	14.56/2.96	Kociba et al.31
	33	Hepatocellular carcinoma 2	105	100.79/84.94	87.68/9.85	Goodman and Sauer28

Note: —, no data; AhR, aryl hydrocarbon receptor; ARNT, aryl hydrocarbon receptor nuclear translocator; BMD, benchmark does; BMDL, benchmark doses (lower confidence limit); BMR, benchmark response; EGF, epidermal growth factor; KQE, key quantifiable event; NA, not applicable; NTP, National Toxicology Program; SD, standard deviation.

^a 10% and 1% are BMRs for dichotomous and continuous dose–response data (genomic data are excluded), whereas 0.1 SD and 1 SD are BMRs for genomic data. Before and after the slash (i.e., “/”), BMDs and BMDLs are presented, respectively.

^b Distinct probes from GSE9838 microarray may correspond to the same gene. If so, that gene will have more than one essential dose derived from individual probes. Here, we average the essential doses corresponding to the same gene and use the mean value as cumulative BMDs or BMDLs.

For an endpoint with various time stamps, the basic rule that we proposed to determine the essential event was to select the earliest available event with three caveats. First, the exposure duration of the earliest event was biologically sufficient to cause the occurrence of that endpoint. For example, a several-day exposure may significantly alter gene expression but not result in cancer development. Second, to avoid unnecessary disturbance in support of the pathway-based dose–response modeling, the event with flipped responses at high dose levels (e.g., Fig. S3b1) or perceptible fluctuations (i.e., Fig. S5e1) was not considered as a proper essential event, even if the event could be observed at an early time point. Third, in some cases, the response was only observed at the highest dose level. To avoid uncertainty in dose–response modeling, the earliest event showing a response at more than one dose level served as the essential event.

Altered gene expressions were captured immediately after a 3-day exposure. Hinging on genes, the estimated BMDs and BMDLs varied greatly [e.g., $47.39\text{ng}/\mathrm{kg}/\text{day}$ (Aldh3a1) vs. $0.43\text{ng}/\mathrm{kg}/\text{day}$ (Cyp1a1)] (Table 2). Despite their early occurrence, the essential doses for these genes were not necessarily smaller than those for the following events in KQE 3 because the dose levels designed for the genes were relatively sparse and had a larger range [$0\sim \mathrm{9,000}\text{ng}/\mathrm{kg}$ (genes) vs. $0\sim \mathrm{7,000}\text{ng}/\mathrm{kg}$ (enzymes)] (Tables 1 and 2). The essential doses for the events in KQE 3 ranging from ${10}^{-2}$ to ${10}^2\text{ng}/\mathrm{kg}$ were generally smaller than the counterparts of KQE 4 (${10}^1$ to ${10}^4\text{ng}/\mathrm{kg}$) as expected, except for the altered hepatic foci volume ($\sim 0.3\text{ng}/\mathrm{kg}$) and increased absolute liver weight ($\sim 3\text{ng}/\mathrm{kg}$) which may serve as more sensitive biomarkers in KQE 4. The longer exposure duration and higher essential dose for each hyperplasia were consistent with the typical biological process. For KQE 5, the magnitude of essential doses fell in the range between ${10}^1$ and ${10}^4\text{ng}/\mathrm{kg}$, which was slightly higher than KQE 4, although some exceptions existed. Hypertrophy and pigmentation may be two earlier indicators in KQE 5. The apical events in KQE 6 indicating the most severe endpoints had the highest essential dose (i.e., $\sim \mathrm{14,500}\text{ng}/\mathrm{kg}$) because of the longest duration.

Estimated MOA-Based PODs

Once the normalized exposure duration was calculated for each event with an estimated essential dose, the pathway dose–response relationship depending on both essential doses and durations was constructed. The fitted pathway dose–response curves using the three link functions (i.e., the logistic, probit, and cloglog models) are graphically shown in Figures 3–5 for various setting combinations. In general, all three models captured the sigmoidal pattern but were slightly off the points on the lower and upper ends which have been caused by the exposure dose levels used in different studies and the conversion from daily dose to cumulative dose. For example, experiments focusing on acute or short-term exposure tended to apply higher dose levels to obtain resounding results.

Figure 3.

The pathway dose–response relationships with the logistic model. (A1–A4) For individual events, essential doses from the simple dose–response model. (B1–B4) Essential doses from the complex dose–response model. (C1–C4) Essential doses from the model-averaged (BMA) model. The essential doses for panels A1, B1, and C1 are ${\text{BMD}}_1$ (i.e., BMDs with the 1% or 0.1 SD BMR). Those for panels A2, B2, and C2 are ${\text{BMDL}}_1$ (i.e., BMDLs with the 1% or 0.1 SD BMR). Those for panels A3, B3, and C3 are ${\text{BMD}}_{10}$ (i.e., BMDs with the 10% or 1 SD BMR). Those for panels A4, B4, and C4 are ${\text{BMDL}}_{10}$ (i.e., BMDLs with the 10% or 1 SD BMR). The black dots are essential events for different KQEs. The red solid curve between the two blue dotted curves is the fitting curve based on maximum likelihood estimation. The gray solid curves (or the gray area) represent the bootstrap estimation. The blue dotted curves show the 90% confidence interval. (C1) is exemplified to show MOA-based POD derivation. The green horizontal dotted lines demonstrate the potential adversity used to determine MOA-based PODs (e.g., 14, 31, or 53 wk), and the green vertical dotted lines show the estimated MOA-based doses on the log-scale. The corresponding results from the fitting curve are shown in red circles. Note: BMD, benchmark does; BMDL, benchmark doses (lower confidence limit); BMR, benchmark response; KQE, key quantifiable events; MOA, mode of action; POD, point of departure; SD, standard deviation.

Figure 5.

The pathway dose–response relationships with the complementary log-log (cloglog) model. (A1–A4) For individual events, essential doses from the simple dose–response model. (B1–B4) Essential doses from the complex dose–response model. (C1–C4) essential doses from the BMA model. The essential doses for panels A1, B1, and C1 are ${\text{BMD}}_1$. Those for panels A2, B2, and C2 are ${\text{BMDL}}_1$. Those for panels A3, B3, and C3 are ${\text{BMD}}_{10}$. Those for A4, B4, and C4 are ${\text{BMDL}}_{10}$. Other legends are the same as indicated in Figure 3. Note: BMD, benchmark does; BMDL, benchmark doses (lower confidence limit).

Figure 4.

The pathway dose–response relationships with the probit model. (A1–A4) For individual events, essential doses from the simple dose–response model. (B1–B4) Essential doses from the complex dose–response model. (C1–C4) Essential doses from the BMA model. The essential doses for panels A1, B1, and C1 are ${\text{BMD}}_1$. Those for panels A2, B2, and C2 are ${\text{BMDL}}_1$. Those for panels A3, B3, and C3 are ${\text{BMD}}_{10}$. Those for panels A4, B4, and C4 are ${\text{BMDL}}_{10}$. Other legends are the same as indicated in Figure 3. Note: BMD, benchmark does; BMDL, benchmark doses (lower confidence limit).

Although three potential definitions of adversity were discussed in the “Methods,” the definition based on precursor events might be the most plausible given various practical considerations. Using the first perceptibly irreversible endpoint to define adversity was not straightforward, because the reversibility of preneoplastic tissue changes for promoters (e.g., TCDD) has not been fully understood. Conversely, using the minimum requirement for being considered chronic to define adversity was relatively implausible because the uncertainty caused by the difference between long-term to short-term effects might be substantial. In other words, the MOA-based POD derived for the apical endpoint (i.e., liver tumor in this study) might be unreliable if the event used to define adversity had a weak causal link with the apical endpoint. Actually, from a quantitative perspective, a more conservative definition of adversity (e.g., based on precursor) may already include the less conservative definitions (e.g., based on the first irreversible endpoint), i.e., less conservative events can be prevented from happening. For the precursor-based definition, enzyme and protein changes (KQE 3), altered cellular growth (KQE 4), and changes in preneoplastic tissue (KQE 5) were commonly accepted precursor events.^20,21 Consequently, based on sustained activation of the receptor and the data availability, 31 and 53 wk were considered by the authors as two appropriate options to define adversity.

Given the advantage of the BMA model and the typical BMRs used in BMD analysis,^37–39 estimated MOA-based PODs using the essential doses estimated from two settings (i.e., ${\text{BMD}}_{10}$ and ${\text{BMD}}_1$ from the BMA method) and the three link functions with adversities defined at 31 and 53 wk are presented in Table 3. The estimated MOA-based PODs have been converted to daily dose on a regular scale by dividing cumulative estimates by corresponding exposure durations, i.e., the unit is ng/kg/day. Additionally, the 90th percentile interval of the MOA-based POD estimated from the bootstrap algorithm was presented together with the median value for each setting combination. Results for additional setting combinations are presented in Figure 6. For the two reasonable definitions of adversity for MOA-based POD calculation (i.e., 31 and 53 wk), the more conservative estimates at 31 wk were about 10-fold lower than the counterparts at 53 wk. The impacts on MOA-based POD estimates from additional factors were summarized below.

Table 3.

MOA-based points of departure (PODs) (median with 90% confidence interval) (ng/kg/day) based on the model-averaged (BMA) model.

Link function	31 wk		53 wk
	${\text{BMD}}_{10}$ a	${\text{BMD}}_1$ b	${\text{BMD}}_{10}$	${\text{BMD}}_1$
Logistic	1.25 (0.414, 3.904)	0.179 (0.0556, 0.613)	11.6 (3.23, 612)	1.86 (0.509, 44.8)
Probit	1.13 (0.316, 4.71)	0.142 (0.0421, 0.638)	11.8 (3.10, 2,944)	2.04 (0.472, 55.9)
Cloglog	1.43 (0.531, 4.27)	0.184 (0.0503, 0.731)	10.9 (4.42, 286)	2.25 (0.595, 18.0)

Note: BMD, benchmark does; BMDL, benchmark doses (lower confidence limit); BMR, benchmark response; Cloglog, complementary log-log; SD, standard deviation.

^a ${\text{BMD}}_{10}$: BMDs with the 10% or 1 SD BMR.

^b ${\text{BMD}}_1$: BMDs with the 1% or 0.1 SD BMR.

Figure 6.

The distribution of MOA-based PODs for various setting combinations. Panels A, D, and G are the boxplots based on the logistic model with the 14th, 31st, and 53rd week specified as the adversity, respectively. Panels B, E, and H are the boxplots based on the probit model with the 14th, 31st, and 53rd week specified as the adversity. Panels C, F, and I are the boxplots based on the complementary log-log (cloglog) model with the 14th, 31st, and 53rd week as the adversity. Gray, yellow, blue, and purple boxes, which are indicated as the number of 1, 2, 3, and 4, use ${\text{BMD}}_{10}$ (i.e., BMDs with the 10% or 1 SD benchmark response BMR), ${\text{BMDL}}_{10}$ (i.e., BMDLs with the 10% or 1 SD BMR), ${\text{BMD}}_1$ (i.e., BMDs with the 1% or 0.1 SD BMR), and ${\text{BMDL}}_1$ (i.e., BMDLs with the 1% or 0.1 SD BMR) as essential doses, respectively. The diamonds for different boxes are the mean values. Some mean values are not able to be detected because they are beyond the y-axis limit that we set up. For panels G, H, and I, boxes belonging to complex and model-averaged (BMA) dose–response models are zoomed in above. Note: BMD, benchmark does; BMDL, benchmark doses (lower confidence limit); BMR, benchmark response; MOA, mode of action; POD, point of departure; SD, standard deviation.

The link function (logistic vs. probit vs. cloglog).

We considered three link functions, i.e., the logistic, probit, and cloglog models. Because these three models are two-parameter models with a sigmoidal shape, the impact on the MOA-based POD estimation was very limited.

The simple vs. complex vs. BMA model.

The main differences between these three different methods that were applied to estimate essential doses were the uncertainty in modeling fitting and BMD estimation mainly reflected by the BMDL estimates. So, generally, the BMA method by considering model uncertainty provided the lowest MOA-based POD and was followed by the situation of using the complex model.

BMDs vs. BMDLs.

The MOA-based PODs estimated using BMDs as the essential doses were typically four to six times higher than the counterparts using BMDLs, and the difference was generally higher when the BMR was smaller, indicating that the uncertainty was larger in the lower dose region. Although using BMDLs can make the estimate of MOA-based POD more conservative, such a protective POD can be achieved through other ways (e.g., using 2.5% or even 1% instead of 5% lower bound of the MOA-based POD estimate). BMD is a more reliable estimate (with less uncertainty) than BMDL from a statistical perspective, so using BMDs as the essential doses may be a more appropriate choice for MOA-based POD estimation.

For the sensitivity analysis, as the estimates listed in Table S2 show, newly estimated MOA-based PODs were close to the corresponding values listed in Table 3, especially those with the adversity of 31 wk.

The Comparison of MOA-Based POD with Traditional POD

A comparison was performed between MOA-based POD in Table 3 and the traditional POD (ng/kg) derived from a single endpoint as shown in Table 2. When the 31-wk adversity definition and higher BMR (i.e., $\text{BMR}=10\%$ or 1 SD) were applied, the MOA-based POD was about $1.2\text{ng}/\mathrm{kg}/\text{day}$ with a lower bound of $0.4\text{ng}/\mathrm{kg}/\text{day}$. Conversely, the estimated BMD of the endpoints that occurred around 31 wk (which can be considered as single-endpoint-based POD) ranged from 2.21 to $30.70\text{ng}/\mathrm{kg}/\text{day}$ with corresponding BMDL estimates in the range of $0.20\sim 17.18\text{ng}/\mathrm{kg}/\text{day}$. So, the MOA-based POD was comparable with the range of the traditional POD, and a similar pattern has been observed if using the lower BMR. If we focused on the 53-wk time point to define the adversity, then the MOA-based POD was again compatible to the corresponding traditional POD (i.e., $11\text{ng}/\mathrm{kg}/\text{day}$ vs. $5.78\sim 42.64\text{ng}/\mathrm{kg}/\text{day}$).

Discussion

In this study, we quantitatively integrated MOA information in dose–response modeling and estimated MOA-based PODs for TCDD. The strategy included three main steps: a) identifying and extracting KQEs, b) calculating essential doses that sequentially activated KQEs using the Bayesian BMD methodology, and c) characterizing pathway dose–response relationship for MOA-based POD estimation. Overall, six KQEs and 33 essential events were identified. The MOA of TCDD proposed in this study required quality assessment. Budinsky et al.²⁰ has provided relatively complete and elaborate estimations with regard to whether the established MOA conformed to the modified Hill criteria and was of human relevance. Due to the similarity of our established MOA and Budinsky et al.’s,²⁰ we believe that the MOA framework developed in this study satisfied the criteria of causality and relevance. It is worth pointing out that developing an MOA based on available literature is an important step but not the strength/focus of the proposed framework or something the framework aims to improve. The most significant improvement is that instead of choosing the linear or threshold approach to perform low-dose extrapolation, the framework aimed at building a dose–response model that can continuously extend the curve/relationship into the low-dose region. The modeling framework proposed here was demonstrated using a nonmutagenic carcinogen, but it can be further generalized to noncarcinogens by integrating their MOA information and constructing corresponding pathway dose–response relationships. This means that the apical event for a substance may not necessarily be cancer effects. An earlier effect may serve as the apical event as long as the corresponding MOA characterized by KQEs can be established. In addition, unlike the current method focusing on one critical effect or apical event, the new framework incorporated a series of biological events in dose–response modeling, which may better characterize mechanistic uncertainty that has not been adequately addressed by the traditional single-endpoint-based POD estimation method. Consequently, the MOA-based POD may be a more biologically plausible starting point for deriving a human reference dose.

It is also important to note that the MOA-based POD should be interpreted and used differently than a traditional POD. The traditional POD is typically calculated from one critical endpoint (i.e., the most sensitive adverse effect or its precursor) and then extrapolated to a human reference dose by dividing by a number of uncertainty factors, including one for the animal to human extrapolation, one for human variability, one for database completeness, one for short-term to long-term extrapolation, and a decreasingly used one for LOAEL to NOAEL extrapolation.⁸ The use of one or more uncertainty factors depends on the properties of the selected critical endpoints. If a BMD/BMDL were derived from a chronic effect using animal data, then the first three uncertainty factors mentioned above are typically used.⁸ Conversely, we believe that the pathway dose–response relationship consisting of a number of temporally and biologically connected events from early changes to the apical effect better characterizes the dose-dependent biological changes from a mechanistic perspective. Therefore, when extrapolating the MOA-based POD to human RfD, the key factors to consider should be whether the MOA is human-relevant and human variability, as the concerns about chronic effects and dataset completeness have been well considered in the temporal pathway dose–response relationship (see Step 3 in Figure 1). A critical advantage of the proposed method over the traditional one is that the uncertainty of the tumor formation mechanism is quantified in the modeling process by integrating various KQEs even if they are not confirmed to be on the pathway. This is probably the main reason that the MOA-based POD was within the range determined by multiple traditional single-endpoint-based POD. From a conservative perspective, as presented in Table 3, the MOA-based POD of TCDD based on the adversity definition of 31 wk and ${\text{BMD}}_1\text{-based}$ essential doses was about $0.05\text{ng}/\mathrm{kg}/\text{day}$, leading to an RfD of $5\times {10}^{-4}\text{ng}/\mathrm{kg}/\text{day}$ if a typical $10\times 10$ uncertainty factor was applied (considering the animal to human extrapolation and human variability). This estimated MOA-based RfD was very close to the oral RfD of $7\times {10}^{-4}\text{ng}/\mathrm{kg}/\text{day}$ proposed in the EPA’s Integrated Risk Information System (IRIS) program, which was derived from epidemiological studies,⁴⁰ and updated tolerable weekly intake of $3\times {10}^{-4}\text{ng}/\mathrm{kg}/\text{day}$ reported by the European Food Safety Authority (EFSA),⁴¹ whereas it was slightly smaller than the chronic minimum risk level of $1\times {10}^{-3}\text{ng}/\mathrm{kg}/\text{day}$ estimated by the Agency for Toxic Substances and Disease Registry (ATSDR),¹⁷ given that the uncertainty factor of three was assigned to the animal to human extrapolation by ATSDR.

As demonstrated in “Results,” the proposed MOA-based POD can be affected by a few factors, especially those in the three steps of the strategy. Some key events can be observed or measured at multiple time points and were selected to form the pathway dose–response relationship primarily based on their occurrence time point with additional considerations as described in “Results.” In the case study of TCDD, there are a few events that were not selected at their earliest appearance. The results of the sensitivity analysis suggested that the proposed dose–response modeling framework was relatively robust regarding the identification and placing key events.

The MOA-based PODs can be affected by various setting combinations generally involving the settings to derive essential doses (i.e., BMD) of selected events (such as dose–response models, BMRs, and using BMD or BMDL), the choice of link functions, and the definition of adversity to calculate the MOA-based POD. The settings of BMD modeling for estimating essential doses have a major impact on the final MOA-based POD through influencing the estimated essential doses but not through changing the shape of the pathway dose–response relationship. Changing certain settings in the BMD modeling for essential dose estimation can shift the entire curve; for example, using BMDL instead of BMD or setting $\text{BMR}=1\%$ instead of 10% can move the entire pathway dose–response relationship to the left on x-axis. Therefore, the settings to derive the essential doses are an important way to incorporate risk assessors’ judgment on the severity of the apical effect. The three link functions had limited the impact on the MOA-based POD estimation primarily because they were not very structurally different. More complicated link functions can be used and may have more significant impact on the MOA-based POD estimation, but model uncertainty should be properly addressed. Then, the definition of adversity on the pathway dose–response curve is a critical factor in determining the MOA-based POD. Just like the BMR definition in BMD modeling, the definition of adversity should be handled carefully case by case.

The modeling strategy developed in this study is more aligned with the probabilistic framework proposed by Chiu and Slob⁴² in which a target human dose was proposed. The adversity chosen to calculate the MOA-based POD also reflects the magnitude of the adverse effect, and the resulting POD distribution can be used to derive the probabilistic target human dose by integrating additional distributional adjustment factors. Actually, the proposed modeling framework can be used to quantify animal-to-human uncertainty and human variability by integrating the appropriate essential doses derived from epidemiological studies in the pathway dose–response relationship. It is expected that the additional human data will introduce extra variance in the MOA-based POD calculation (resulting in a smaller lower bounder), which reflects that inter- and intraspecies uncertainty and variability that are considered and quantified. This application will be presented and discussed in another paper.

This study was limited by the availability of quantitative data on early and late biological events. First, there were few eligible in vivo dose–response data (i.e., at least three dose levels) on the sustained AhR activation characterized with AhR and ARNT protein changes in liver, so we hardly knew about the essential doses corresponding to this event. Although altered gene expression may also serve as an indication of sustained AhR activation,²¹ the range of dose levels designed for this event was usually large to ensure the sufficient alterations to be detected. Therefore, the essential dose of altered gene expression was not necessarily lower than those of the following events, which would cause great uncertainty in the lower end of the pathway dose–response relationship. Second, the adversity heavily relied on data retrieved and was mainly determined by the NTP’s study design (i.e., 14, 31, and 53 wk)¹³ in this study because the NTP report was the main source of essential events in KQE 3–5. In general, the proposed framework requires more data than the traditional approach that relies on one critical endpoint; however, the proposed framework is also relatively robust because no data to support one or two key events (like the TCDD case presented in the current study, there were no data available for the first KQE) will not invalidate the approach. Lack of data will impact both the proposed and traditional approaches through enlarged uncertainty.

Overall, the proposed modeling framework made a major step forward in dose–response assessment by quantitatively integrating mechanistic information in the modeling process and providing a potential strategy to harmonize cancer and noncancer dose–response assessment through pathway dose–response modeling. However, the modeling strategy was also limited by the availability of appropriate data and by the understanding of the apical effect’s MOA.