The multi-dimensional embryonic zebrafish platform predicts flame retardant bioactivity

Abstract

Flame retardant chemicals (FRCs) commonly added to many consumer products present a human exposure burden associated with adverse health effects. Under pressure from consumers, FRC manufacturers have adopted some purportedly safer replacements for first-generation brominated diphenyl ethers (BDEs). In contrast, second and third-generation organophosphates and other alternative chemistries have limited bioactivity data available to estimate their hazard potential. In order to evaluate the toxicity of existing and potential replacement FRCs, we need efficient screening methods. We built a 61-FRC library in which we systemically assessed developmental toxicity and potential neurotoxicity effects in the embryonic zebrafish model. Data were compared to publicly available data generated in a battery of cell-based in vitro assays from ToxCast, Tox21, and other alternative models. Of the 61 FRCs, 19 of 45 that were tested in the ToxCast assays were bioactive in our zebrafish model. The zebrafish assays detected bioactivity for 10 of the 12 previously classified developmental neurotoxic FRCs. Developmental zebrafish were sufficiently sensitive at detecting FRC structure-bioactivity impacts that we were able to build a classification model using 13 physicochemical properties and 3 embryonic zebrafish assays that achieved a balanced accuracy of 91.7%. This work illustrates the power of a multi-dimensional in vivo platform to expand our ability to predict the hazard potential of new compounds based on structural relatedness, ultimately leading to reliable toxicity predictions based on chemical structure.

1. Introduction

The use of flame retardant chemicals (FRCs) has, until recently, been ubiquitous in the manufacture of furniture, electronics, building materials, textiles, and children’s clothing, since the mid-1970s. In 2005, several polybrominated diphenyl ethers (PBDEs) were voluntarily phased out over concerns about their environmental persistence, bioaccumulation, and adverse human health effects, including impaired neurodevelopment [1, 2]. Organohalogen and organophosphate flame retardant chemicals (OPFRCs) became the go-to alternatives to PBDEs, achieving high volume use in PVC plasticizers for which developmental and neurotoxicity data was already sparse. OPFRCs are now widely detected in indoor air, household dust, and wildlife samples [3, 4]. Studies over the years have shown that OPFRCs may be as hazardous as PBDEs, with a proclivity to disintegrate from their product matrix but persist in the environment, raising concerns of them being “regrettable substitutions” to their predecessors [5]. Structurally similar to organophosphate pesticides, it is plausible that OPFRCs could also potentiate neurotoxicity and developmental toxicity [6–9]. However, OPFRCs and several other alternative FRC classes are already in use despite the absence of a full profile of their environmental behavior and toxicological properties.

The hazard potential of 1^st generation and alternative FRCs contains persistent knowledge gaps. Comprehensively anchoring diverse FRC structures to a rich data set of vertebrate developmental outcomes could guide safer use of current, and selection of new, flame-retardant chemistries. Developmental zebrafish is an ideal model that provides high content data. The common use of zebrafish as a model is driven by a combination of physiological and molecular features- the genetic similarity amongst humans and zebrafish is ~70% [10], zebrafish and mammalian brain and heart share many anatomical and functional features along with a similar central nervous system [11, 12], zebrafish develop externally, are optically transparent, and most organs and systems are fully formed by 5 days post fertilization (dpf). These features in a small vertebrate model are ideal for evaluating a broad range of chemical-biological endpoints.

Concerns over the human and environmental safety of FRCs have resulted in a substantial body of in vitro data on FRC bioactivity [13–15]. However, pressing questions about tissue uptake, distribution, and metabolism has not been possible to address in vitro. Previously, we evaluated 44 halogenated and organophosphate flame retardants [16] for their ability to elicit adverse neurodevelopmental outcomes in zebrafish. To build upon this work, we procured a diverse library of 61 FRCs that had undergone quality assurance measures (45 from the US Environmental Protection Agency) and screened the library for developmental bioactivity [17] using different metrics such as a point of departure (POD), Lowest effect concentration (LEC) or benchmark dose (BMD) metric [18, 19]. The use of any of the metrics (POD, LEC, or BMD) was defined as the lowest concentration where a given response exceeded the background noise in the controls. The goal is to evaluate FRC bioactivity in one testing platform to identify hazard potential. We compared our data to the high throughput in vitro toxicology data for the 45 FRCs from US-EPA which included a subset of 12 FRCs previously identified as neurotoxicants in zebrafish [20], mouse embryonic stem cells [21], neural stem cells [21] and C. elegans [22]. Of the 12 classified potential developmental and neurotoxicant FRC, our zebrafish data using the LEC identified 91% correctly (10 out of the 11 bioactive FRC ). As the zebrafish data was highly informative, we built a classification model integrating 13 physicochemical properties and the 3 zebrafish assay results from the largest FRC library (Table 1) to date for vertebrate developmental bioactivity and neurotoxicity potential. By utilizing this multi-dimensional screening platform in conjunction with a battery of assays and alternative models, the field can rapidly prioritize chemicals for developmental and neurotoxicity testing in mammals.

Table 1. A table of the chemical structure for 61 FRCs and the 6 classification bins.

The classification bins are aryl phosphate ester (APE), polybrominated diphenyl ether (PBDE), other brominated, chlorinated phosphate ester (CPE), brominated phenol (BP), and others. Two mixtures (PentaDBE, and OctaBDE) and Burka metabolite structures are not listed. Four forms of Triisopropylated phenyl phosphate are listed together as they share the same chemical structure and CAS.

2. Methods and Materials

2.1. Chemicals

A comprehensive library of 61 flame retardants was procured from various sources: 42 from the US Environmental Protection Agency, 2 from the NIH National Toxicology Program (NTP) (Pentabromodiphenyl ether [CAS # - 32534–81-9] and Perfluorobutanesulfonate-K [375–73-5]) and 17 purchased from Sigma (Table S1A). FRCs from US EPA and NTP underwent third-party verification prior to shipment to Oregon State University. The data can be found at the US EPA chemistry dashboard (https://comptox.epa.gov/dashboard/chemical_lists/FLAMERETARD). The samples were provided in 100% Dimethyl sulfoxide (DMSO) at 20 mM (if possible) and stored in −80°C prior to testing. Samples purchased from Sigma had >98% purity and accompanying Certificate of Analysis. These were stored as recommended; all dried samples were suspended to maximum solubility (usually 20 mM) in DMSO.

2.2. Zebrafish Husbandry

Tropical 5D wild type zebrafish were housed at the Oregon State University Sinnhuber Aquatic Research Laboratory in a density of 1,000 fish in 100-gallon tanks. Each tank was kept at standard laboratory conditions of 28°C on a 14-h light/10-h dark photoperiod. Spawning funnels were placed into the tanks the night prior, and embryos were collected and staged [23]. The chorion was enzymatically removed using 2,277 U of pronase (90 μL of 25.3 U/μl; Roche, Indianapolis, In, USA) at 4 hpf using a custom automated dechorionator [24] to increase bioavailability.

2.3. Chemical Exposure

Six hour post fertilization (hpf) embryos were transferred into individual wells of a 96-well plate filled with 100 μL of embryo medium (EM) [24]. The FRCs were added to each well using the Hewlett-Packard D300 Digital Dispenser, an inkjet technology. Using a 20 mM stock solution, the D300 can precisely deliver the desired concentration directly into the experimental chamber. A lower energy mixing protocol was scripted for the HP D300 and implemented to automatically mix for 1 sec between all deliveries of at least 5 nL [17]. All wells were normalized to a final well concentration with 0.64% DMSO. For 54 of the FRCs, embryos were exposed to 10 concentrations (0.1 – 80 μM) with 32 animals per concentration. The other 7 (Tetrabromobisphenol A, 2,2–4,4-t-penta-,bromodipheyl ether, BDE-183, DE-71, Mono (2-chloropropyl) phosphate, and Tetrabromobisphenol A-2,3-dibromopropyl ether) were exposed to 6 concentrations up to 25 μM. The plates were sealed using Parafilm beneath the lid, to reduce evaporation and wrapped in aluminum foil to prevent photodegradation. The plates were placed on an orbital shaker at 235 rpm and 28°C for 16 hrs to aid in a homogenous test solution, thereafter in a static 28°C incubator.

After chemical exposure commenced, embryos were kept out of visible (VIS) light until tested for embryonic photomotor response (EPR) at 24 hpf [25]. EPR is a 1 minute behavioral assay of embryonic zebrafish tail flexions in response to visible (VIS) light. The embryos were kept in the dark until the initiation of the test. The test consisted of 3 phases: 30 s of darkness (infrared (IR) light, Background); a 1 s pulse of intense VIS light (13,000 lux), 9 s darkness (Excitation); the second pulse of VIS light, 10 s darkness (Refractory).

Following the EPR, the embryos are assessed for survival at 24 hpf, delays in developmental progression, altered spontaneous movement, and abnormal notochord development. At 120 hpf, the exposed larvae were subjected to a larval photomotor response assay (LPR) to assess flame retardant effect on locomotor behavior (putative neurological) using Viewpoint Zebrabox system (ViewPoint Life Sciences, Lyon, France). Larvae experience a total of 4 light cycles, each cycle consisting of 3 min of alternating VIS light (1000 lux) and dark (IR). The first cycle (6 minutes) was the acclimation period and removed from the analysis. The software tracked the total distance traveled every 40 ms and integrated these data in 6 s intervals over the 24 minute assay. The analysis was performed with the 3 sequential light cycles. Using a custom R script, wells with mortality or malformation were excluded from the analysis, and the total movement for each fish was calculated and the total motion between each treatment was compared to the control. Following the LPR assay, embryos were assessed for 18 developmental malformations including yolk sac edema (YSE) and pericardial edema (PE); body axis (AXIS), trunk length (TRUN), caudal fin (CFIN), pectoral fin (PFIN), pigmentation (PIG), and somite (SOMI) deformities; eye (EYE), snout (SNOU), jaw (JAW), and otolith (OTIC) malformations; gross brain development (BRAIN); notochord (NC) and circulatory (CIRC) deformities; swim bladder presence and inflation (SWIM); and touch-responses (TR). The presence or absence of each of the endpoints was recorded into a laboratory information management system called the Zebrafish Acquisition and Analysis Program (ZAAP) [26].

All raw data can be found in Table S1, and Table S2. Five summary LECs were computed as described in Truong et al [18] for any effect except mortality at 24 hpf (any.24h.MO24−), any effect at 24 hpf (any.24h.MO24+), any effect at 120 hpf except mortality (any.5d.MORT−), any effect at 120 hpf (any.5d.MORT+), any sublethal effect (any.MO−), and any effect throughout the experiment (any.MO+).

2.4. ToxPi

ToxCast data for 45 of the 61 chemicals were downloaded from the US EPA dashboard (https://comptox.epa.gov/dashboard; Table S1D) and used to visualize assay activity using the ToxPi GUI software [28, 29]. Assays missing all data or with zero-variance were removed from the ToxCast dataset. Slices were created by grouping assays for the 390 in vitro endpoints according to the technology platform (annotated by assay vendor summary names). In the overall model, each technology slice was given equal weight. To reflect potency (1,000,000 was inserted to indicate inactivity), assay data were scaled using the formula −log₁₀(x) +6, where x is the ToxCast Activity Concentration (AC₅₀) value. The length of the slice indicated the potency aggregated across all assays, where the negative-log scale resulted in longer slices (i.e. higher ToxPi scores) for chemicals that elicited more potent effects. The chemicals were clustered using Ward’s method, as implemented in the ToxPi GUI software.

2.5. Predictive model development

The chemical properties for the 61 FRCs were extracted from ChemSpider and EPA Episuite 4.1 (Table S1A). A total of 13 chemical properties were collected: log P, boiling point, density, enthalpy vaporization, flash point, index refraction, molar refractivity, molecular weight, number free rotating bonds, number H bond acceptor and donors, number rule 5 violations and the vapor pressure (Table S3D). A total of 6 classifications were built for this study. The first had only the 13 chemical properties, and another with the 13 chemical properties and the LEC from the 3 zebrafish assays (morphology, EPR, and LPR; Table S3A, S3E). An additional model was built with inputs of BMD values for the zebrafish morphology and the LPR data, rather than the LEC (Table S3I). A data matrix was created for all 61 FRCs and the 19 parameters (Table S3) in which FRCs that were not bioactive were assigned 10,000. Additional models: 1) all 19 parameters, 2) chemical properties + Morphology and the LPR AUC BMD endpoint, 3) chemical properties + Morphology and the LPR MOV BMD endpoint were constructed. Each FRC was clustered based on structure into 6 classes: aryl phosphate ester (APE), bromophenol (BP), chlorine phenol ester (CPE), other brominated (OtherBrominated) and polybrominated diether (PBDE). Due to limited commercial availability of some FRCs, structural class size varied: APE = 24, BP = 5, CPE=7, Other = 7, OtherBrominated = 7, and PBDE = 11. Using the CARET package in R [30], the complete dataset was split 80/20 into a training and test set, respectively. The training dataset was randomized with 10-fold cross-validation and repeated 3 times (i.e., 3X bootstrapping). Using the training dataset, the importance of each of the parameters was computed for each chemical class. This is a metric to assess its contribution to help build an accurate classification of an FRC based on the one parameter. Once a model was built from the training dataset, its ability to classify FRCs was evaluated using the test set. All model input and results can be found in Table S3. A glossary of commonly used terms is listed in Table S4.

2.6. Statistical Analysis

2.6.1. Morphology

Observations of morphology endpoints were recorded as binary (presence/absence) data.

Lowest effect Concentration (LEC)

Statistical significance was computed as described in Truong et al [18]. Briefly, significant differences between control and exposed fish were computed using a one-sided Fisher’s exact test, where adverse endpoints were tested to have a greater occurrence in exposed fish. The lowest effective concentration (LEC) is the lowest concentration where the exact test had a p-value less than 0.05.

Benchmark dose (BMD)

BMD value is the concentration that results in a 10% extra risk in response (10% benchmark response) compared to the negative control. It was calculated from the model that best-fits the concentration-response curve of each chemical-endpoint pair. First, the incidence data was transformed into a concentration-response curve by calculating the proportion of affected animals, P(d) at each concentration, d, which is the number of positively responding embryos/larvae (with hit = 1) divided by the total number at each concentration. Since the mortality at 24 and 120 hpf does not contribute to any value to the morphology endpoints, hit values for the morphology endpoints of embryos/larvae that died by 24 hpf or 120 hpf were assigned NA values. Therefore, the dead embryos/larvae were not included in the construction of concentration-response curves for the morphology endpoints. The concentration-response data (P(d) vs d) was then fitted – both with actual as well as log-transformed concentrations – using models recommended in EPA’s technical guidance document for benchmark dose calculations [31]. The models were logistic, logistic_bgr, probit, probit_bgr, log_logistic, log_probit, gamma, Weibull, quantal linear, and multistage 2. The model that best fits the curve was then used to determine the concentration (BMD) that corresponds to the 10% extra risk or benchmark response (BMR), defined as

$$BMR=0.1=\frac{P\left(BM{D}_{10}\right)-P\left(0\right)}{1-P(0)},$$

where P(BMD) is the proportion of affected animals at the BMD, and P(0) is the proportion in the negative control group.

The quality of the fit for each model was evaluated based on Chi-square test as Good-Fit, OK-Fit, or Bad-Fit: the fit was considered good if the p-value > 0.1 (Good-Fit); ok if the p-value ≤ 0.1 and R² > 0.7 (OK-Fit), or bad if p-value ≤ 0.1 and R² < 0.7 (Bad-Fit). A BMD value would not be calculated for a chemical-endpoint pair if all fits were bad fits. It is noted that only concentrations that had at least 8 fishes (from both plates combined) were included in the concentration-response curves. An attempt was made to fit a curve only if it had at least 3 such concentrations. Plates that did not have at least 8 negative control animals were not used.

95% confidence intervals for BMD values

The 95% confidence intervals for the BMD₁₀ values were calculated only for those curves with an adequate fit with adequate BMD values (BMD >= 1e-5 and < 1000). To calculate the confidence intervals, we used the likelihood-ratio method, as described in a previous study [32].

2.6.2. EPR

The recorded periods at the beginning and end of the experiment (immediately surrounding the initiation/termination of camera recording) were truncated to assure equivalence in recorded experimental period for all chemicals. The statistical analysis of activity considered only the Background (B), Excitatory (E), and Refractory (R) intervals. The overall pattern of activity within each B, E, or R interval was compared to that interval’s negative control activity using a combination of percent change (50% peak difference from control in either direction) and a Kolmogorov-Smirnov test (p< 0.01). The LEC is the lowest concentration where the KS test had a p-value less than 0.05.

2.6.3. LPR

Movement data from the 15 frames s⁻¹ capture was integrated into 6-sec bins. The analysis was conducted for only the first light-dark cycle (after the acclimation period). Control animals are typically 3 – 10 fold more active in the dark (IR) than in the VIS light.

Entropy Analysis

A statistical modeling method previously published using the entropy model [33] was applied. Briefly, the data were adapted to a uniform distribution with 95% and 5% quartiles to reduce bias caused by outliers. The differential entropy model allows us to measure the average surprisal of a continuous probability distribution. The area under the curve using the differential entropy values was computed for each exposure plate and averaged per chemical. The differential entropy for each treatment was statistically compared to the control using a 2 sample Kolmogorov-Smirnov test (K-S) test (p<0.01). The LPR at a given exposure concentration was considered valid only when statistical significance (K-S test, p < 0.01) was reached, and the relative AUC treatment: AUC control ratio was ≥ 1 or ≤ −0.3. Animals dead or malformed at the 120 hpf time point were excluded from the LPR data analysis. The LEC is the lowest concentration where the KS test had a p-value less than 0.05.

BMD

Using an established BMD model to assess the movement data, two endpoints were computed [19]. Briefly, a concentration-response curve was constructed for two metrics: MOV and AUC. The MOV endpoint represents the change in the movement at the light-to-dark transition time point. The AUC endpoint represents the change in the area under the movement versus time curve between the dark and light periods. Wells with dead animals were removed from the analysis and only live animals were used. For each metric, the concentration-response curve was constructed based on the number of animals that responded abnormally relative to the negative control for each chemical. Outliers were detected using a Tukey box-plot algorithm [34] and test plates with fewer than 50% of the controls responding as expected were removed. The statistical modeling used to calculate a BMD is described above in the morphology section of the statistical analysis method.

3. Results and Discussion

3.1. The goal for modeling toxicity and physicochemical properties of FRCs

Flame retardant chemicals have likely saved thousands of lives, buying people extra minutes or even just seconds to escape a fire. However, their ubiquity in everyday household items and indoor dust may constitute a chronic exposure hazard, particularly during development when the body is more sensitive to external stressors. Most reports associating developmental FRC exposure with toxicological outcomes have focused on one or a few FRCs. Comprehensive coverage of FRC structural diversity in toxicology studies is particularly lacking. We sought a practical approach to enable toxicologists to make better-informed decisions in the face of rapid innovations in FRC chemistry.

3.2. Diverse FRC library anchored to diverse developmental impacts

Of the 61 FRCs tested in this study, 45 were also assessed as a part of the EPA ToxCast program and 34 halogenated and organophosphate flame retardants were assessed in this same platform [16]. Twelve of these FRCs were previously identified as potential teratogens and neurotoxicants in previous studies [21, 22] and underwent testing in a battery of in vitro neuronal assays, a zebrafish locomotor toxicity assay at the EPA, and assessment of impacts to three C. elegans morphology endpoints [21]. We anchored FRC physicochemical properties to multidimensional, whole animal readouts of their bioactivity. Of the 61 FRCs tested, 54 were assessed at 10 concentrations (up to 80 μM), 6 FRCs, due to chemical solubility limitations, were screened at 6 concentrations up to 25 μM, and one FRC up to 0.0064 μM (Figure 1). FRC-exposed embryos and larvae were evaluated for 2 behavioral endpoints - embryonic and larval photomotor response (EPR and LPR, respectively) and 22 morphology and mortality endpoints; LECs were modeled for these endpoints within Figure 1. The FRCs fell into 3 clusters: 1) not bioactive, yielding no significant incidences of an endpoint hit in any of the 3 assays; 2) morphometric/behavioral, associated with both morphological and hyper- or hypoactivity, and 3) behavioral, associated only with EPR or LPR effects.

Figure 1. Dot-heatmap representation of the developmental toxicity for 61 FRCs.

Dechorionated embryos were exposed to 61 FRCs to 4 dosing regimens from 6 to 120 hpf. The size of the dots represents potency, with the bigger circles representing a lower concentration; color dots represent morphological endpoints (red), while neurobehavioral endpoints are represented as yellow (hyperactivity) and blue (hypoactivity). The lowest effect concentration was computed for each endpoint in the morphology, 24h (24 hpf photomotor response), and 5d (120 hpf larval photomotor response) behavioral assays. For morphology, summary endpoints were created: any 24hr sublethal effect (any24h.MO24−), any 24 hr effect (any24h.MO24+), any 120 hpf sublethal effects (any.5d.MORT−), any 120 effects (any.5d.MORT+), any sublethal effects (any.MO−) and any effect throughout the exposure (any.MO+). The 24 hr behavioral assay detects motion in 3 intervals, Background, Excitatory, and Refractory, and a summary endpoint (all). At 5d, the behavioral assay measures distance traveled in the visible light phase and the dark (infrared) phase, and overall movement (All).

The hit concordance between the 34 halogenated and organophosphate flame retardants tested in Noyes et al (in the same lab) and this study was 100% for morphological endpoints, 93.9%, and 81.8% for EPR and LPR assays, respectively. When considering potency and direction (for the behavior assay), the concordance reduced to 79.6%, 42.4%, and 30.3%, respectively. Concordance for potency is defined as the log₁₀(ratio of the LELs) == 0 and for directionality for behavior, the LELs have to be in the same direction. The behavior concordance comes mainly from both studies identifying FRCs that do not cause aberrant behavior effects. By applying the additional criteria of directionality and potency, and the requirement of any difference in the ratio of the LEL, it explains the reduced concordance. Also, the use of LEL as the analysis input restricts the ability to classify disconcordance. With each study testing at different concentrations, it is difficult to use another metric. Considering all these factors, the use of hit concordance between the two studies demonstrates the confidence in the ability of the developmental zebrafish model to detect bioactivity.

Under the test conditions used, BDE-99 was the least bioactive, with no abnormal behavioral associations and only a modest incidence of teratogenesis. There were 16 FRCs associated exclusively with abnormal behavior and 45 FRCs associated with various levels of abnormal morphology. Consistent with a previous study [20], commonly used OPFRs such as TDCIPP, EHDP, and TPHP were associated with aberrant photomotor effects at 24 and 120 hpf. Several brominated flame retardants such as BDE-3, 15, and 153 were associated with hypo- or hyperactivity in these assays. Interestingly, BDE-47, a widely used brominated FR in the last decade and a known neurotoxicant [35], was not associated with aberrant photomotor behavior in our assays.

Pthalates have been used as plasticizers for decades and have been detected widely in human body fluids globally. Herein, exposure to bis(2-ethylhexyl) tetrabromopthalate was associated with strong photomotor hyperactivity, suggesting that this class of compounds impacts some aspect of neurodevelopment. Several FRCs were associated with mortality and abnormal morphology at 24 and 120 hpf. Triisobuytl phosphate, a solvent, plasticizer and anti-foaming agent, induced the highest incidence of mortality. Exposure to TPHP or TDCIPP was associated with abnormal embryo morphology, consistent with previous studies [9, 36, 37]. Overall, our results suggested that the OPFRCs were the most bioactive in developmental zebrafish.

3.3. FRC profile in high throughput in vitro assays

We tested 61 FRCs, 45 of which were previously evaluated for bioactivity in 390 in vitro endpoints as part of the US EPA ToxCast/Tox21 Program. Because the ToxCast/Tox21 in vitro endpoints were numerous and often not directly analogous to vertebrate endpoints, we aggregated the in vitro data at the level of technology platform (annotated by assay vendor summary names) and collapsed in vitro endpoints with strong direct correlations to each other into a generic endpoint. ToxPi was employed to rank each FRC based on its bioactivity across all the assay vendors and this study zebrafish data. As illustrated in Figure 2, almost half (23 FRCs; Figure S1D) were bioactive in a ToxCast in vitro assay. The top 5 bioactive FRCs, which exhibited a high slice score in 3 vendor assays (Attagene, ATG; BSK, BioSeek; and Tox21, assays in the Tox21 consortium toolbox), were TDCIPP, TCBPA, TPHP, TPP, and chlorpyrifos. These outcomes, based on in vitro assessments, were consistent with the data seen in our embryonic assessments (gray slice in the toxpi), however, the ones that were the least bioactive in the ToxPi (e.g., Trimethyl phosphate, 2-Bromopropanoic acid) caused aberrant swimming behavior, which would not be identified in the ToxCast in vitro assays. The vendor slice that was most potent was BioSeek, with 14 out of the 45 FRCs. BioSeek consists of 84 assays that model complex human disease and tissue biology of primary human cells under various stimulatory conditions [38]. The 8 cell types are stimulated with a variety of important biological effectors by measuring cytotoxicity and proliferation on a total of 87 BioMAP endpoints. The BioMAP profiles of FRCs that ToxPi-clustered together in Figure 2, had similar BioMAP profiles when compared to the BioMAP reference database of >3000 compounds. An advantage of having a large reference database is that it provides an opportunity to identify other compounds that have the same BioMAP profile and can induce the same biological effects. Of the remaining 31 FRCs, ToxPi scored 23 inactive in the in vitro assays, but were either teratogenic or induced aberrant behavior. In vitro assays offer biological insight when compared to a large reference library, but also have significant blind spots that the zebrafish model reveals.

Figure 2. Clustering of 45 FRCs assessed in ToxCast assays.

From the 61 FRCs, 45 were previously assessed in 8 vendor groups that are part of the ToxCast or Tox21 program and also in zebrafish. The vendor’s data are visualized as a slice of the pie (OT; Odyssey Thera NVS: Novascreen; CEETox; BSK: BioSeek; ATG: Attagene; APR: Apredica; ACEA, Tox21 assays in the Tox21 consortium toolbox, ZF: this study zebrafish).

3.4. Comparing Point of Departure (POD) for FRCs and Tox21 assays to detect developmental and neurotoxicity

Since the phase-out of penta-BDEs in 2005, OPFRC replacements and have been detected in the environment at levels comparable to polybrominated diphenyl ethers (PBDEs) [39]. The National Toxicology Program (NTP) evaluated 8 OPFRCs for teratogenicity and neurotoxicity due to their structural similarity to organophosphorus insecticides: TPHP, IPP, EHDP, BPDP, TMPP, IDDP, TDCIPP, TCEP. The NTP study also included 4 brominated flame retardants: BDE-47, BDE-99, BDE-71, and TBBPA. The NTP concluded that more comprehensive studies are needed to define the hazard potentials of replacement OPFRs, except for TCEP [21]. The assays reported within this study used point-of-departure (POD) values, which denote the level at which a chemical-induced response exceeded the background noise. The calculation for the PODs was based on a range of in vitro and in vivo assessments, including developmental toxicity assays using mouse embryonic stem cells, C. elegans and zebrafish and neurotoxicity assays using human and rodent neuronal cells. We used a similar approach for teratogenicity, EPR and LPR endpoints. For teratogenicity, we calculated the lowest effective concentration (LEC) from the concentration-response data and a benchmark dose (BMD) [19] for each of the 12 FRCs (8 OPFRCs and 4 brominated FRCs) reported on by NTP. For the 120 hr LPR assay, a LEC was calculated in addition to the BMD for the photomotor response to the changing of the light. Two endpoints were calculated for the LPR using BMD for two metrics: the area under the curve, and the response transitioning from the light to the dark phase [19]. Overall, 11 of the 12 FRCs were identified as hits across all 3 zebrafish assays (regardless of analysis model). Our whole animal assays showed that BMD 120 LPR metric was the lowest for eight of the FRCs: BDE-99, BPDP, IDDP, IPP, TBBPA, TCEP, TDCIPP, and TMPP (Figure 3A). The zebrafish behavior assay was not the most sensitive for four FRCs: BDE-47, DE-71, EHDP and TPHP, where C elegans larval development was the most sensitive assay. TCEP was the only FRC that was a negative amongst all the NTP in vitro and in vivo assessments. The use of the BMD computation method for the LPR assay identified TCEP as being bioactive, while the other assays (teratogenicity and EPR) were not identified as hits. Two FRCs, BDE-71 and-99 were not common to both zebrafish studies, and BDE-47 bioactivity was discordant between the NTP studies and the Tanguay zebrafish endpoints (Figure 3B). There are instances where the zebrafish assays showed both the highest and lowest sensitivities for the same compound. But the key point is that in no such instance was bioactivity ever missed; that is, as a biosensor of chemical activity, the developmental zebrafish was the most sensitive all around. In our whole animal assays, the aryl phosphate FRCs were the most bioactive, followed by brominated and chlorinated.

Figure 3. Comparison of flame retardant toxicity to other high throughput assays.

The lowest effect concentration (LEC), point of departure (POD), and benchmark dose (BMD) for 12 FRC that were previously assessed in in vitro models, zebrafish and C. elegans. (A) The LEC and POD across the 11 assays are plotted for the 12 FRCs (–log₁₀ (value/1000,000)) with the higher the value, the more potent. (B) The 11 assays are broken up into 3 screening outcomes (developmental toxicity, developmental neurotoxicity, and acute neurotoxicity) and the LEC/POD expressed in μM is listed per FRC class. Dash (−) represents no toxicity observed at the tested concentrations. NT denotes FRCs not tested in certain assays. For the zebrafish behavior assay, the values in the bracket represent the phase the LEC/POD was calculated.

Among the 12 FRCs tested, zebrafish teratogenicity measured by BMD or LEC was the most sensitive for DBE-71, TBBPA, BPDP, TMPP, TDCIPP (5/12, 41.67%) (Figure 3A). Previously, a US EPA zebrafish lab (Padilla ZF) evaluated ten of the FRCs for teratogenicity and found that IDDP and TCEP were not toxic, and TPHP was the most toxic, concordant with our study (Figure 3B). While there were assay differences between ours and Padilla ZF with respect to exposure duration (5 vs. 6 days), static vs. daily renewal, and sample size (n=32 vs. 4), we were still able to identify the extremes – the most and least toxic FRCs. Only for BDE-47 were the results completely discordant; Padilla ZF identified a teratogenicity POD of 28.2 μM, while our assay did not identify BDE-47 as a teratogenic hit.

To assess developmental and acute neurotoxicity, the NTP utilized human stem cell neuroprogenitor, rat neurite outgrowth, human stem cell neurite outgrowth, and rat neural network activity in vitro assays to assess 10 FRCs (Figure 3B) and found that nine were hits. Our 24 hpf EPR and 120 hpf LPR (in combination) also identified these nine as hits. The one FRC that was disconcordant between our assays (no activity) and other in vitro models was EHDP. In another study employing a 3–5 dpf zebrafish behavioral assay [36], both TCEP (negative in all the studies discussed within Figure 3) and EHDP showed hypo- or hyperactivity. The limited discord observed for some OPFRC bioactivity was likely not due to variation inherent in the model, but to differences in laboratory practices.

For five of the nine OPFRCs (55.56%), the zebrafish model was more sensitive than the in vitro models. Among the zebrafish behavior assays, the 120 hpf LPR was the most sensitive toward FRC bioactivity, identifying seven of the nine FRCs that were bioactive among the four neuro-specific in vitro assays reported in Behl et al. 2015 [21]. BDE-47 was identified as a hit in our LPR assay, consistent with previous studies where BDE-47 disrupted neuronal growth and behavior during zebrafish development [40]. This may indicate that BDE-47, in the concentration range tested, specifically targeted neurodevelopment.

Previously, we demonstrated that BMD can be calculated from zebrafish teratogenicity and that it was more precise than LEC [19]. The use of BMD is often used by regulatory agencies as an estimate of the point of departure. Of our nine FRC hits in teratogenicity, we found five were within the 95% confidence estimation of BMD (Table S1; IPP, BDPD, TMPP, EHDP, TDCIPP). The LEC analysis missed BDE-71 and TPHP as hits, while BMD detected them. For TBBPA, the LEC was more sensitive than BMD (0.5 vs 5 μM). The difference in the POD could be explained by calculation of the method where, for BMD, the fit of the dose-response curve dictates the POD, while LEC calculations are agnostic of the concentration-response curve. However, when both LEC and BMD analyses were applied, eleven of the twelve FRCs (91.67%) were concordant with the C. elegans larval model, mouse stem cell differentiation, and Padilla ZF model data.

The tri-phenyl OPFRCs were the most bioactive class. A recent study from the NTP [36] identified tri-phenyl OPFRCs as cardiotoxic in zebrafish embryos. TPHP, a widely used OPFRC, and plasticizer, displayed the lowest LECs across the studies and was associated with strong photomotor hypoactivity and a moderate incidence of mortality and abnormal morphology in our assays. Other studies have associated TPHP-induced cardiotoxicity and hepatotoxicity through the disruption of nuclear receptors and metabolic pathways [41, 42], and neurotoxicity through disruption of neurotransmitters and alteration of transcript levels of neurodevelopmental genes [43]. Such information emphasizes the need for more thorough study of the developmental impacts of TPHP and similar OPFRCs. Among the tri-phenyl OPFRCs, IPP was identified as bioactive in all ten assays, and BPDP and TMPP were positive hits in at least eight of the ten assays. Exposure to these compounds has elsewhere been associated with abnormal morphology and behavior during development [37, 44]; however, further studies are needed to elucidate their mechanisms of toxicity.

3.5. A classification model of FRC hazard potential

Flame retardant chemicals are intended to prevent ignition and slow the combustion rate, which is clearly a human health benefit. Rather than discount the value of this technology in light of its potential exposure hazards, a more productive strategy might be to understand structure-bioactivity relationships in a way that guides chemists to those of lowest hazard potential. We demonstrate a rapid method for achieving this goal. By using impacts on multiple dimensions of zebrafish development, we have begun to predict which hazard class FRCs fall into.

A predictive model of only the chemical properties was developed using randomForest (RF) with two trees for robust evaluations of descriptor importance. A total of 13 chemical descriptors was used in this model. Before building the model, the dataset was divided 80/20 into a training and test set, respectively. The train set consisted of 49 FRCs, while the test set consisted of 12 FRCs. The model underwent 10-fold cross-validations and 3X bootstrapping. When assessing a classification model, it is important to determine if it can accurately predict whether an FRC belongs in a certain class (sensitivity) and not (specificity), but the overall performance is assessed by using a balance accuracy metric. The model we have built has an overall balance accuracy of 83.3% with a p-value of 0.019. Using only the chemical property model, the balanced accuracy was better for the Other, and Other-brominated classes (100%). The second highest were CPE, and APE FRCs with 95.45%, and 91.67% balance accuracy, respectively, where the density and log P were the most important variables. For PBDE, density, number of H bond acceptors, index of refraction, number rule 5 violation, and number of free rotating bonds were equally important variables (~55–60 importance score; Table S3F).

When a model was built with the 16 descriptors (13 chemical properties and the LEC from our 3 zebrafish assays) as described above, the overall balance accuracy was 91.67% with a p-value of 0.003. The model was best at classifying APE, Other, Other-brominated, CPE, and PBDE’s with balance accuracies of 91.67, 100, 100, 100, and 100%, respectively (Figure 4A). From the 61 FRCs, a total of 24 were classified as aryl phosphate ester, 7 as Other-brominated, and 11 for PBDE (Table S1). The classification model was able to predict with 91.67% balance accuracy for the APE class, which is the bin the OPFRCs are a part of. The top 4 parameters of this class (all with > 60% importance) were density, number of H bond acceptors, index of refraction, and zebrafish aberrant morphology data calculated using LEC (Table S3E). However, for the Other-brominated and PBDEs, the zebrafish data had an importance of < 30%. The drivers for these classes were the number of H bond acceptors and the number of free rotating bonds. For all 3 of these classes, the sensitivity with the zebrafish data and 13 chemical properties was 100%. The addition of the zebrafish LEC dramatically improved the overall ability to classify FRCs and the individual classes.

Figure 4. Classification modeling performance.

We used RandomForest to build a classification model for 6 FRC classes: Other, OtherBrominated, PolyBrominated Diphenyl Ether (PBDE), aryl phosphate ester (APE), brominated phenol (BP), and chlorinated phosphate ester (CPE). The balanced accuracy (BA) for each of the FRC classes, and their sensitivity, specificity, and positive predictive value for the models (A) 16 parameters of 13 chemicals with the LEC for all 3 zebrafish assays and (B) 14 parameters with the BMD for the zebrafish morphology only. Both models yield a BA of 91.67%.

As the BMD is a more commonly used POD than LEC, a classification model was built to evaluate whether the balance accuracy would differ. A random forest model was built using the 13 chemical descriptors, the BMD zebrafish morphology value, and two LPR BMD endpoint measurements (16 parameters total). Using both LPR BMD endpoints did not improve the balanced accuracy (BA) of the model (75%), however when considering only the AUC computing method for the LPR data, the BA was similar to a model with the chemical properties only (83.33%). The best classification model using the BMD was when only the BMD morphology data was considered (in addition to the 13 chemical descriptors). The overall balance accuracy was 91.67% with a p-value of 0.0032 (Figure 4B). The APE, BP, Other, and other brominated classes had 100% sensitivity and balance accuracy. The BMD had the highest importance in BP (~35, Table S3K). A potential reason the mild improvement over the 13 chemical property model was observed for the LEC (morphology and LPR) and BMD (morphology only) models could be due to the use of only the zebrafish gross endpoints, which seems to be the least sensitive measured endpoint. The LEC classification model had the lowest BA for the BP class, while for the BMD morphology only model, the lowest BA was for PBDE. Overall, the use of either BMD morphology or LEC for building predictive models to classify FRCs has a good balance accuracy (91.667%), with each model having its strength in predicting certain classes of FRCs.

Our model highlights a comprehensive effort to bin FRC structures according to their mode of action to discover how it depends on physicochemical properties, with the ultimate goal of predicting the toxicity from structural relatedness. Demand for FRCs will increase as industrial and consumer product flammability regulations become more stringent. This presents an important opportunity for toxicology to develop chemical structure-bioactivity relationships that guide FRC chemists toward structures with the lowest hazard liability. Developing predictive classification models with a balanced accuracy >80% will be an important consideration in the decision-making process for the use of FRCs.

Highlights.

• Identified the bioactivity of 61 flame retardant chemicals (FRCs)

• The zebrafish assay detects bioactivity of 12 of the 10 known developmental neurotoxic FRCs.

• Classification model using physicochemical properties and 3 zebrafish assays yields a balanced accuracy of 91.7%