High-Throughput Screening to Advance In Vitro Toxicology: Accomplishments, Challenges, and Future Directions

Abstract

Traditionally, chemical toxicity is determined by in vivo animal studies, which are low throughput, expensive, and sometimes fail to predict compound toxicity in humans. Due to the increasing number of chemicals in use and the high rate of drug candidate failure due to toxicity, it is imperative to develop in vitro, high-throughput screening methods to determine toxicity. The Tox21 program, a unique research consortium of federal public health agencies, was established to address and identify toxicity concerns in a high-throughput, concentration-responsive manner using a battery of in vitro assays. In this article, we review the advancements in high-throughput robotic screening methodology and informatics processes to enable the generation of toxicological data, and their impact on the field; further, we discuss the future of assessing environmental toxicity utilizing efficient and scalable methods that better represent the corresponding biological and toxicodynamic processes in humans.

INTRODUCTION

Pollution caused the death of 9 million people worldwide in 2015 (1). Due to the ever-growing amount of chemicals generated and released into the environment each day, there is a great need to determine the toxic profiles of these compounds quickly and efficiently. Currently, most toxicants are still evaluated using in vivo animal models; however, these tests are expensive, slow, low throughput, and may not accurately predict outcomes in humans. Recently, ethics have also become an increasing concern when performing these tests on animals (2, 3). With these shortcomings, it has become imperative to identify a novel solution to generating toxicological data.

In 2008, a US government collaboration, the Toxicology in the 21st Century (Tox21) program, was established to address the shortcomings with traditional toxicological testing. Tox21 is composed of multiple federal partners, including the Environmental Protection Agency (EPA), the National Toxicology Program (NTP), and the National Institutes of Health Chemical Genomics Center [NCGC, now a part of the National Center for Advancing Translational Sciences (NCATS)]. The Food and Drug Administration joined the effort in 2010 (4, 5). The specialties of each partner complement each other in the effort to transition from traditional in vivo animal experimentation to a target/pathway-specific, mechanism-based, high-throughput screening (HTS) approach (4). An overview of the Tox21 program, its partners, projects, and other information can be found online (https://tox21.gov).

Adopting advances in technology and data science, the Tox21 program has progressed through three phases. Phase I was essentially a proof-of-concept stage. The initial Tox21 compound collection, which included approximately 2,800 chemicals, was screened in over 75 cell-based and biochemical assays; a quantitative high-throughput screening (qHTS) approach was applied utilizing 1,536-well plates (4, 6, 7). These assays combined to form a comprehensive view of each compound with respect to cellular toxicity (DNA damage, apoptosis induction, and cytotoxicity), signaling pathways [antioxidant response element/nuclear factor erythroid 2–related factor 2 (ARE/Nrf2), cAMP response element binding (CREB), and hypoxia-inducible factor 1 alpha (HIF-1a)], inflammation modulation [nuclear factor kappa-light-chain-enhancer of activated B cells (NF-kB), tumor necrosis factor alpha (TNFa), and interleukin-8 (IL-8)], modulation of nuclear receptors [androgen receptor (AR) and estrogen receptor (ER)], and individual cellular targets [enzyme, human ether-a-go-go (hERG) channel inhibition, receptor binding, and protein-protein interaction disruption] (4, 6). These assays produced robust data from reproducible experiments using concentration-response curves, instead of a single concentration, which reduced the frequency of false positives and negatives (3), allowing the data to be used in computational modeling and hazard characterization, which are crucial steps in limiting animal model usage (2, 4, 6).

Phase II of the Tox21 program (production phase) began in 2010 and focused screening on stress response pathways and nuclear receptors incorporating phenotypic, target-specific, or mechanism-based assays. To date, more than 70 HTS assays have generated over 100 million data points (8). Many of these assays include a multiplex system wherein cytotoxicity can be measured alongside the assay readout to identify true positives; using this technique, cytotoxic compounds can be removed from active hits in antagonism assays (2, 6). To screen these assays, a new 10,000-compound (Tox21 10K) library was assembled through the combined effort of NTP, NCGC/NCATS, and the EPA. This unique library, representing the largest collection of environmental chemicals and related molecules assembled to date, was arrayed for screening in a novel 15-point concentration format, in triplicate, for each compound (4, 8); screening in this manner yielded comprehensive and robust bioactivity data for the constituent chemicals. After the completion of each assay, the data are made available for public consumption, thereby supporting transparency while encouraging further advancements in the field of toxicology.

Recently, Phase III of theTox21 program began by shifting the focus from hazard identification to more accurate representation of human health and disease utilizing physiologically relevant in vitro assays. The aims of this phase are to develop assays that are more predictive of human toxicity, alter existing methods to address current technical challenges, characterize previous in vivo toxicity studies and interpret the comparative in vitro data, establish an evaluation framework that confirms the performance of in vitro assays, and accurately extrapolate in vitro disposition data into in vivo models. Addressing these new goals requires closer collaboration among Tox21 partners, and therefore, multiple cross-partner projects were formed. Cross-partner projects require at least two Tox21 partners, have three-year terms, and are reviewed annually by leadership based upon progress and alignment with the goals of Tox21 (9). Phase III brings an exciting new aspect to the Tox21 program and will provide technologically advanced ways to assess toxicity rapidly and accurately.

ADVANCING HIGH-THROUGHPUT METHODOLOGIES TO MEET BIG-DATA DEMANDS OF THE TWENTY-FIRST CENTURY

To meet the challenges of generating large amounts of in vitro data to support toxicology research, we have leveraged our HTS expertise, including assay development and computational modeling. This has been the area of expertise and focus of our team at NCATS for well over a decade; this review focuses on the NCATS experience and contributions to the Tox21 program. Alongside robotics, we have advanced the methodology of HTS to incorporate rapid and efficient generation of hundreds of thousands of concentration-response data sets per project (10) and the subsequent informatics pipelines. Further, we have leveraged the availability of large chemical libraries from NCATS, which include over 600,000 unique compounds, in a multiple-concentration format that can be screened in high-throughput in vitro assays (7). To begin assembling these large compound collections, NCATS assigns a unique identifier to each compound and manages the mother-daughter plate registration using the ActivityBase sample registration system (IDBS, United States). A compound management team uses a variety of automated instrumentation and barcoded racks and vessels that are coupled to a laboratory information management system (LIMS) developed in house. Compound management installed two SampleStores (Azenta Life Sciences, United Kingdom) for automated sample handling of up to 1.5 million solution tubes and 289,000 powder vials. The SampleStores automatically retrieve samples and deliver them in 96- or 24-well racks used by other instruments in the lab. The LIMS is also used to manage NCATS’s compound library inventory as well as streamline the sample request process from users through a web interface. A suite of Biomek NXP, FXP, and i7 liquid handlers (Beckman Coulter, United States) prepare compound plates in a 1,536-well plate format, which can be used for pintool or acoustic transfer. Labcyte Echo acoustic dispensers in Access workstations (Beckman Coulter, United States) prepare assay-ready plates used in cell-based and biochemical assays. This elaborate system allows the compound management team at NCATS to handle, store, and plate the compounds from large libraries, including the Tox21 10K compound collection. This shared library was formed collaboratively between NCATS, EPA, and NTP; originally, each partner contributed 3,764, 4,078, and 3,115 compounds, respectively, to the 10K library, which is described elsewhere in greater detail (8).

By 2012, there were 8,947 unique chemicals in the Tox21 10K library, each in 15 concentrations with a maximum concentration of 10–20 mM, dissolved in dimethyl sulfoxide (DMSO), and stored in 1,536- and 384-well plates. These plates were used not only for in vitro studies but also to perform analytical chemistry quality control (QC) analysis on each compound (8). Chemical quality and stability are important factors that affect the interpretation of data generated from high-throughput screens. Quality issues due to chemical degradation, insolubility, or volatility can lead to false positive or false negative findings. The entire Tox21 10K library was examined immediately following assembly and after the compound plates were exposed to storage at room temperature for 4 months. The major analytical techniques used to QC the compound library include liquid chromatography–mass spectrometry, gas chromatography–mass spectrometry, and nuclear magnetic resonance spectroscopy. Each chemical was assigned a QC grade based on purity, identity, and concentration. All QC results, including a PDF report for each individual compound with original analytical spectra, have been made publicly available online (https://tripod.nih.gov/tox/samples). More detailed analysis of the analytical QC results, including chemical quality and stability, will be published soon in a separate study.

NCATS performs high-throughput screens with key support from a robotics team. The Tox21 robot accesses workstations, including compound plate carousels, assay plate incubators, liquid handlers (BioRAPTR 2.0, Let’s Go Robotics, United States; Pintool station, Wako Automation, United States), a BlueWasher (BlueCatBio, United States), and readers [ViewLux and EnVision plate readers (Perkin Elmer, United States), Functional Drug Screening System (FDSS) 7000EX (Hamamatsu, Japan), and an Operetta CLS high-content imaging system (Perkin Elmer, United States)]. A high-precision robotic arm (Staubli, Switzerland) transfers plates to the surrounding instruments and workstations, which can screen thousands of compounds simultaneously (6) (Figure 1). The modular, robotic system can evolve with advancing lab technologies, as different readers and instruments can be added and removed as needed. The ViewLux and EnVision plate readers measure absorbance, luminescence, and fluorescence throughout the entire well of an assay plate, while the Operetta imager can visualize individual cells and quantify many cellular features in 1,536-well plates, such as in the green fluorescent protein–labeled LC3 assay (11). To conduct larger cellular imaging experiments, the Operetta has also been used to assess neurite outgrowth (12, 13), as well as to identify autophagy modulators (14, 15). The FDSS measures fluorescence in adherent or suspended cells in real time (kinetic measurements). For instance, it has been used to kinetically assess hERG modulation (16–18). Another important part of the robotic system is the PureTip (IonField Systems, United States), which is used to clean the slotted metal pins on the Pintool transfer head by triggering air ionization, resulting in a breakdown of chemical contaminants on the exterior and interior surfaces of the pins. Use of this cleaning technology greatly improved assay robustness and performance. This cleaning technology can reduce costs and waste associated with HTS.

Figure 1:

There are four components to the Toxicology in the 21^st Century (Tox21) robotic platform, including a central yellow arm surrounded by incubators, dispensers, and readers [such as ViewLux, EnVision, Functional Drug Screening System (FDSS), and Operetta], which are assembled together to perform high-throughput screens. The inset represents the front of the FDSS reader. Figure generated by Eric Wallgren using SolidWorks modeling software.

All data generated by the robotic system are automatically saved locally to the control workstation, including assay data results and all logs generated by the connected instrumentation. Also, data are remotely sent to a shared database in the LIMS, which is cloud capable, allowing the easy and rapid sharing of data with NCATS collaborators. To perform all the assays mentioned above, this unique robotic system needs to be scheduled and maintained by a team of robotic specialists that can rapidly address technical issues. Scheduling occurs using custom software developed internally at NCATS and is another part of the LIMS. Overall, our in vitro toxicology robotics platform can store compound and assay plates, transfer plates between instruments and workstations, dispense liquids, and perform optical measurements in an automated fashion.

The main goal of our screening effort has been assessing the toxicological effects of drugs, industrial chemicals, and various environmental compounds (6). Any public health stakeholder, including academic, nongovernmental, private, or governmental organizations, may propose assays to be screened in the Tox21 program. An assay proposal must be written and submitted to the Tox21 Assay Evaluation and Screening working group for approval, which is based on toxicological relevance, alignment with Tox21 program goals, cross-partner project needs, and adequate assay performance (6, 19). To date, approximately 60 different assays have been incorporated into Tox21, which are listed in Table 1. The assays can be categorized as phenotypic, target, and pathway based. A full schematic of the Tox21 screening process is depicted in Figure 2. Once accepted, the NCATS Tox21 team optimizes and validates each proposed assay, using the Assay Guidance Manual (20) criteria, for multiple parameters (treatment duration, appropriate positive control usage, and assay signal window) and miniaturizes the assay into a 1,536-well format.

Table 1.

Validated assays for Tox21 screening

Target	Conduit	Pathway	Readout
5-Hydroxytryptamine receptor 2A	CHO-K1	GPCR/Ca2+ signaling	Fluorescence
Acetylcholinesterase	AChE-SH-SY5Y Biochemical	Stress response/neuronal signaling	Absorbance Fluorescence
Acetyl-(E)-2-(4-mercaptostyryl)-1,3,3-trimethyl-3H-indol-1-ium (acetyl-MSTI) binding	Biochemical	Electrophilic binding	Fluorescence
Aryl hydrocarbon receptor	AhR-HepG2	NR signaling	Luminescence
Activator protein 1	AP1-ME180	Stress response	Fluorescence
Androgen receptor (LBD)	AR-HEK293	NR signaling	Fluorescence
Androgen receptor (full)	AR-MDA	NR signaling	Luminescence
Antioxidant response element	Nrf2/ARE-HepG2	Stress response	Fluorescence Luminescence
Aromatase	Aromatase-MCF7	NR signaling	Luminescence
Autofluorescence	HEK293 and HepG2	Autofluorescence	Fluorescence
β2-Adrenergic receptor	CHO-K1	GPCR/cAMP signaling	Fluorescence
Constitutive androstane receptor	CAR-HepG2	NR signaling	Luminescence
cAMP response element–binding protein	CREB-HEK293	Stress response/neuronal signaling	Fluorescence
Caspase-3 and caspase-7	Caspase-3/7-HepG2	Stress response	Luminescence
CYP1A2	Enzyme	Cytochrome P450	Luminescence
CYP2C9	Enzyme	Cytochrome P450	Luminescence
CYP2C19	Enzyme	Cytochrome P450	Luminescence
CYP2D6	Enzyme	Cytochrome P450	Luminescence
CYP3A4	Enzyme	Cytochrome P450	Luminescence
Cytotoxicity and viability	HEK293	Stress response	Fluorescence Luminescence
	HepG2	Stress response	Fluorescence Luminescence
Dopamine receptor D2	DRD2-HEK293	GPCR/cAMP signaling	Fluorescence
DNA repair, Rad54/ku70 (−/−)	DT40	Stress response	Luminescence
DNA repair, Rev3 (−/−)	DT40	Stress response	Luminescence
DNA repair, wild type	DT40	Stress response	Luminescence
Ether-a-go-go-related gene potassium channel	hERG-U2OS	Cardiology signaling	Fluorescence
Estrogen receptor α (LBD)	ERα-HEK293	NR signaling	Fluorescence
Estrogen receptor α (full)	ER-MCF7	NR signaling	Luminescence
Estrogen receptor β	ERβ-HEK293	NR signaling	Fluorescence
Estrogen-related receptor α	ERR-HEK293	NR signaling	Luminescence
	PGC/ERR-HEK293	NR signaling	Luminescence
Endoplasmic reticulum stress response element	ESRE-Hela	Stress response	Fluorescence
Farnesoid X receptor	FXR-HEK293	NR signaling	Fluorescence
Glucocorticoid receptor	GR-Hela	NR signaling	Fluorescence
Gonadotropin-releasing hormone receptor	GnRHR-HEK293	GPCR/Ca2+ signaling	Fluorescence
H2A histone family member X	CHO	Stress response	Fluorescence Luminescence
HDAC I and HDAC II	HCT116	Epigenetics	Luminescence
Hypoxia-inducible factors	HRE-ME180	Stress response	Fluorescence
Heat shock element	HSE-HeLa	Stress response	Fluorescence
KISS1 receptor	KISS1-HEK293	GPCR/Ca2+ signaling	Fluorescence
Luciferase	Enzyme	Luciferase inhibition	Luminescence
Mitochondrial membrane potential	MMP-HepG2	Stress response	Fluorescence
Muscarinic acetylcholine receptor M1	M1-CHO-K1	GPCR/Ca2+ signaling	Fluorescence
Nonspecific cAMP signaling	HEK293	GPCR/cAMP signaling	Fluorescence
Nuclear factor κB	NF-κB-ME180	Stress response	Fluorescence
p53	p53-HCT-116	Stress response	Fluorescence
Peroxisome proliferator-activated receptor δ	PPARδ-HEK293	NR signaling	Fluorescence
Peroxisome proliferator-activated receptor γ	PPARγ-HEK293	NR signaling	Fluorescence
Progesterone receptor	PR-HEK293	NR signaling	Fluorescence
Pregnane X receptor	PXR-HepG2	NR signaling	Luminescence
Retinoic acid–related orphan receptor γ	RORγ-CHO	NR signaling	Luminescence
Retinoid X receptor	RXR-HEK293	NR signaling	Fluorescence
Retinol signaling pathway	RAR-C3H10T1/2	Developmental	Luminescence
Sonic Hedgehog pathway	ShhGli1–3T3	Developmental	Luminescence
Smad signaling pathway	SMAD-HEK293	Developmental	Fluorescence
Telomere length regulation protein ELG1	ELG1-HEK293	Stress response	Luminescence
Thyroid hormone receptor	TRE-GH3	NR signaling	Luminescence
Thyroid-stimulating hormone receptor	TSHR-HEK293	GPCR/cAMP signaling	Fluorescence
Thyrotropin-releasing hormone receptor	TRHR-HEK293	GPCR/Ca2+ signaling	Fluorescence
Vitamin D receptor	VDR-HEK293	NR signaling	Fluorescence

Abbreviations: GPCR, G protein–coupled receptor; LBD, ligand-binding domain; NR, nuclear receptor; Tox21, Toxicology in the 21^st Century.

Figure 2:

Schematic of the Toxicology in the 21st Century (Tox21) screening process depicting the pathway through which an assay is brought forth, performed, and analyzed. Assays are nominated to an evaluation group that decides whether the assay is a good fit for Tox21. Once approved, an assay will go through development and optimization, validation, and eventually the full screen of the 10K compound library. Data are then analyzed and shared with partners and the public. Figure created with BioRender.com.

After acquiring optimal signal-to-background ratios (≥3), coefficients of variation (≤10%), and Z’ factors (≥0.5), an online validation is carried out using the robotic platform to ensure adequate assay performance and quality. Chemicals from the library of pharmacologically active compounds (1,280 compounds, Sigma-Aldrich, United States) are screened, along with 88 compounds from the Tox21 10K library. Once validated, the assays can be used to screen the Tox21 10K library. All of the data generated from each screen are publicly available through the PubChem database (http://pubchem.ncbi.nlm.nih.gov), enabling researchers to use these data sets for further studies, including computational modeling, after an internal evaluation by the Tox21 partners (2).

Before sharing the data with Tox21 partners and making them publicly available, the data go through a rigorous informatics pipeline. After a screen is completed, raw plate reads are first normalized relative to each respective positive control compound using the following equation:

$$\%\text{ Activity = }\left[\left({\text{V}}_{\text{compound}}{\text{ - V}}_{\text{DMSO}}\right)/\left({\text{V}}_{\text{pos}}{\text{ - V}}_{\text{DMSO}}\right)\right]\times 100,$$

where ${\text{V}}_{\text{compound}}$ denotes compound well values, ${\text{V}}_{\text{DMSO}}$ denotes median values of the DMSO-only wells, and ${\text{V}}_{\text{pos}}$ denotes median values of the positive control wells. Each plate is then corrected using compound-free control plates (i.e., DMSO-only plates) at the beginning and end of the compound plate stack to remove background patterns and subtle abnormalities, such as tip effects or blotting from cell dispensing (21). Corrected plate data are then pivoted to form concentration-response series, which are subsequently fit to a four-parameter Hill equation (22) yielding concentrations of half-maximal activity (AC₅₀) and maximal responses (efficacy) (23). Based on the efficacy, quality of fit, and number of data points observed above background activity, concentration-response curves are categorized into classes 1–4; problematic curves are automatically placed into class 5 and are manually inspected to correct the curve class, if necessary. Furthermore, each curve class is combined with an efficacy cutoff yielding a numerical curve rank; more potent and highly efficacious curves are ranked higher and should be viewed as having a higher compound activity. The curve fitting results from the replicate assay runs are then assessed for activity reproducibility and identified as active match, inactive match, mismatch, or inconclusive based on the compatibility of each curve class and rank. The final activity outcome of a compound is determined based on its reproducibility and multichannel activity (e.g., multiple readouts within one well) (24). Once the data parsing and assessment at NCATS are complete, the concentration-response data, curve fitting results, raw plate data, and assay conditions are shared with theTox21 partners and then released to the public domain such as PubChem and the NCATS Tox21 Data Browser (https://tripod.nih.gov/tox21/pubdata/).

A DECADE OF ACHIEVEMENTS

With the use of robotics,NCATS has screened thousands of compounds for multiple end points in a relatively short amount of time. Since its inception, our Tox21 team has screened more than 70 in vitro assays, including many stress response pathway and nuclear receptor assays (Figure 3) encompassing 15 concentrations of each compound, in three independent runs, which generated over 100 million data points (https://pubchem.ncbi.nlm.nih.gov/#query=tox21&tab=bioassay) (25). The data generated from each screen have passed through stringent QC and generated high reproducibility rates. The pregnane X receptor (PXR) and constitutive androstane receptor (CAR) are two nuclear receptors that have been studied and screened against the Tox21 10K compound library. PXR modulation may be responsible for the metabolism of nearly 50% of marketed drugs through its regulation of drug-metabolizing enzymes and transporters, while also playing a role in energy homeostasis, immune response, and cancer (26–29). In a recent study, the NCATS Tox21 team identified 11 potentially novel or not-well-characterized PXR activators within the 10K library by utilizing a HTS followed by confirmation studies (27, 30). Whereas PXR has a promiscuous ligand-binding pocket, CAR has a much smaller, and therefore more selective, binding pocket. However, CAR still has an important impact on human health by modulating energy metabolism, tumor progression, and cancer therapy outside of its traditional role as a transcription factor. One unique factor of CAR is the constitutive activity when no ligand is present in immortalized cells, making it difficult to fully study its activity. However, the NCATS team overcame this issue by incorporating the known CAR antagonist PK11195 into each assay well, translocating CAR back into the cytoplasm. Due to the increasing number of roles CAR plays in the human body, the Tox21 10K compound library was screened for novel CAR activators (31). From this screening, four compounds—neticonazole, diphenamid, phenothrin, and rimcazole—displayed CAR agonist activity with rimcazole, demonstrating potential CAR selectivity.

Figure 3:

Assay configuration. This chart depicts the percentage of each type of assay performed in the Toxicology in the 21st Century (Tox21) program at the National Center for Advancing Translational Sciences. G protein–coupled receptor (15%), nuclear receptor (33%), stress response (30%), cytochrome P450 (8%), developmental (5%), and other (8%) assays encompass the types of pathways explored by screening the Tox21 library. Other assays include neuronal signaling, electrophilic binding, autofluorescence, cardiology signaling, epigenetics, and luciferase inhibition.

As previously stated, the goal of Tox21 has expanded in recent years to include collaborative efforts between its partners. One of these cross-partner projects involved the identification of acetylcholinesterase (AChE) inhibitors. Acetylcholinesterase is a cholinergic enzyme primarily found in neuromuscular junctions and cholinergic brain synapses (32, 33), where it is responsible for the termination of neurotransmission by hydrolyzing acetylcholine (ACh) into acetate and choline. After this transformation, the presynaptic nerve takes up choline and combines it with acetyl-CoA to produce ACh through the action of choline-acetyltransferase (34). Inhibition of AChE can induce ACh accumulation in the synaptic space, which stimulates nicotinic and muscarinic receptors, disrupting neurotransmission (35) that can cause toxicity.

There are still many compounds that have yet to be identified as AChE inhibitors, especially among those in the Tox21 10K chemical library. To measure the activity of AChE, the Tox21 cross-partner project team developed homogenous enzyme- and cell-based assays employing both fluorescence and absorbance readouts (36, 37). One limitation of AChE assays is that some chemicals become AChE inhibitors only after bioactivation (e.g., organophosphorus pesticides) and do not display AChE inhibitory effects in their parent form. To address this issue, human or rat liver microsomes (HLMs and RLMs, respectively) were incorporated into another HTS, developed in a 1,536-well format to account for metabolism (37, 38). To validate this assay, a group of organophosphorus pesticides, containing both parental compounds and their active metabolites, were screened for AChE inhibition activity. The assay utilized recombinant human AChE protein with HLMs or RLMs. Many parental organophosphorus pesticides only showed the inhibitory effects after metabolism was introduced into the reaction, indicating the need for bioactivation. Together, these data demonstrated the ability to profile AChE inhibitors in vitro using metabolic activation.

Once the AChE assays (with or without microsome addition) were screened against the Tox21 10K compound collection, a group of AChE inhibitors were evaluated in SH-SY5Y cells and human neural stem cells (39) using a monolayer or spheroid culture conditions. Some compounds displayed greater AChE inhibition potency in the monolayer culture, while other compounds were more potent in the spheroid culture.To investigate which cytochrome P450 (CYP) enzymes were involved in bioactivation, both CYP activity (P450-Glo assay) and CYP gene expression (quantitative reverse transcription polymerase chain reaction assay) were measured. For neural stem cells, the expression level of CYP3A4 and CYP2D6 was higher in spheroids than in the monolayer culture. Molecular docking was also employed to study the binding mode of these AChE inhibitors. The docking results predicted that all newly identified AChE inhibitors would bind to the active site of AChE, confirming the inhibition observed in vitro. From this study, a group of novel AChE inhibitors were identified, mostly pharmaceutical compounds. To further assess the toxicological relevance of these compounds, their AChE IC50 values were compared to human plasma concentrations (C_max) found in the literature. Several compounds had IC₅₀ values within tenfold of C_max, indicating that these compounds could be physiologically relevant; therefore, these AChE inhibitors should be prioritized for further in-depth study (40).

In another cross-partner project, a microsome-derived metabolic activation assay for measuring p53 activation has been optimized into a cell-based 1,536-well plate format (41). Because p53 is a tumor suppressor and induced in cells subjected to various types of cellular stress, especially DNA damage, the p53-bla assay can identify potential genotoxicants in a simple and rapid manner (42). Due to the assay’s extensive quality and high reproducibility, the p53-bla assay (43) was utilized as an example of a cell-based qHTS method that could be retrofitted with metabolic capability to represent a more physiologically relevant experiment. The Tox21 10K compound library was screened employing the p53-bla cell line under three conditions: without microsomes, with HLMs, or with RLMs. Several organophosphorus pesticides were identified as p53 inducers when in the presence of microsomes. These compounds were cherry-picked and retested using the same assay conditions; however, due to product availability, different lots of HLMs and RLMs were utilized in the experiments. There were large lot-to-lot variations when utilizing HLMs in the cell-based assay and a poor (10%) confirmation rate for the identification of p53 inducers. In contrast, the confirmation rate was around 90% when cells were cotreated with RLMs. This newly developed method identified several novel and known p53 inducers in a physiologically relevant cell system, verifying the assay’s capability of evaluating compounds that need metabolic activation to induce a biological response. However, it is well known that the metabolites produced by RLMs might not be the same as those produced in humans (44). In conclusion, this cross-partner project demonstrated the importance of metabolism on p53 inducers and species differences between RLMs and HLMs.

Another important cross-partner project featured an in vitro skin model and used it to assess skin hazards and irritants as an alternative to animal models. The adverse outcome pathway of skin sensitization is composed of four key elements: covalent modification of self-proteins, keratinocyte activation, presentation of new antigens by dendritic cells, and an inflammatory response of lymphocytes (45, 46). The direct peptide reactivity assay (DPRA) involves measuring the molecular initiating event of the skin sensitization pathway, which involves chemicals that covalently bind with an endogenous protein. The Organization for Economic Co-operation and Development published a test guideline for the DPRA, but it requires high concentrations of chemicals and peptides that can interfere with ultraviolet detection, leading to false results. Furthermore, the original high-performance liquid chromatography–mass spectrometry DPRA protocol (47) is labor intensive and time consuming; to address this issue, a RapidFire solid-phase extraction system coupled with tandem mass spectrometry (MS/MS) has been applied that conducts high-speed, solid-phase extraction to quantify peptides and other analytes. Recently, NCATS generated a 384-well high-throughput modified DPRA protocol utilizing a RapidFire solid-phase extraction system coupled with MS/MS method, which decreased the amount of peptide usage (46). This newly generated platform identified sensitizers (e.g., cinnamic aldehyde and ethylene glycol dimethacrylate) in a reliable and sensitive method, while drastically improving throughput.

Keeping abreast with emerging technologies is an important goal of the NCATS Tox21 team. NCATS recently applied bioprinting (3D biofabrication), a developing technology, to skin toxicity testing. Identifying skin irritancy, reversible local skin tissue damage, is one of the regulatory requirements for evaluation of safety when industrial and consumer products are being considered for release to the marketplace. The Draize test is a commonly used method that directly applies a test chemical to the skin of a rabbit (48). However, recent mandates in Europe, for example, prohibit the use of animals when evaluating cosmetics (49, 50). In response, reconstructed human epidermis (RhE) models have become important tools; yet there are some disadvantages, including being low throughput and lacking physiological complexity. Although 2D cellular models are high throughput, they also lack physiological relevance and have limited predictive value. However, a combination and modification of these models can establish an improved platform to screen potential skin irritants. The Tox21 team at NCATS first implemented a cytotoxicity assay with monolayer keratinocytes and screened a group of about 500 topically applied compounds. Active hits were then further evaluated using biofabricated 3D skin tissue models (50). These models, RhE and full-thickness skin, are printed in a controlled layer formation to create a reproducible and physiologically complex mimic of human skin. The use of these bioprinted models enabled accurate and relevant high-throughput assays to efficiently screen environmental chemicals for skin irritation and/or sensitization potential.

Chemicals that disrupt mitochondrial function have an important toxic impact on human health. Due to the importance of mitochondrial toxicity, a Tox21 cross-partner project was initiated to measure mitochondrial membrane potential (MMP) in HepG2 cells. Mitochondria, which are membrane-bound organelles, generate most of the energy found in cells in the form of adenosine triphosphate (ATP) via the respiratory chain (51). Under normal physiological conditions, the MMP generated by a proton pump is needed for the respiratory chain in order to synthesize ATP, making it a key parameter in assessing mitochondrial health and compound-induced mitochondrial toxicity (52). Mitochondria also have an important role in metabolism, cell signaling, and apoptosis (53). To identify mitochondrial toxicants, NCATS screened the Tox21 10K chemical library with a homogenous cell-based assay using the fluorescent MMP indicator m-MPI (54). M-MPI is a water-soluble, membrane-permeable fluorescent dye that detects cellular MMP changes (55) caused by various compounds, including uncouplers, electron transport inhibitors, and apoptotic agents. MMP inhibitors identified from the Tox21 10K compound collection were clustered based on structural similarity to identify the structural features associated with decreases in MMP (56). Diverse compound groups from the cluster analysis inhibited MMP, including cardiac glycosides, chlorinated organic insecticides, flavonoids, quinolone-based dyes, parabens, thiazolidinedione-based drugs, and triarylmethane dyes; some of these groups were previously reported to impair mitochondrial functions, thereby confirming our assay. Potent compounds were further studied through a tiered approach to prioritize in-depth mechanistic evaluation (57). Mechanistic studies included the MMP assay in rat primary hepatocytes and human neural stem cells, measurement of reactive oxygen species, upregulation of p53, Nrf2/ARE modulation, oxygen consumption, and ATP status in Caenorhabditis elegans. This approach identified the known disruptors of mitochondrial functions, as well as several undercharacterized potential mitochondrial toxicants. To further evaluate selected mitochondrial toxicants in a specific biological context, human cardiomyocytes (AC16) were chosen for a global proteomic analysis following treatment with the potent uncoupler of oxidative phosphorylation in mitochondria, trifluoromethoxy carbonylcyanide phenylhydrazone (FCCP), and four undercharacterized mitochondrial toxicants (dinoterb, picoxystrobin, pinacyanol, and triclocarban) (58). The profiles suggested that the compounds dysregulated mitochondrial proteins as well as proteins involved in lipid metabolism, stress response, and cytoskeleton protein changes. Therefore, dysregulation of these proteins leads to changes in various metabolic processes. Through this battery of assays, thousands of chemicals were screened for mitochondrial toxicity, elucidating their potential toxicological mechanisms involved in disruption of MMP and prioritizing the screened compounds for further in vivo testing.

The Tox21 partners also assessed how toxicants impact cultured cell lines using medium-throughput transcriptomic technologies that incorporate multiple concentration-response experiments (59, 60). Adapting the NCATS algorithm used for analyzing qHTS data (24), we developed a point-of-departure (POD) algorithm to analyze concentration-response transcriptomic data. The POD algorithm identifies genes that respond to a treatment in a concentration-dependent trend and calculates the threshold concentration at which the significant response occurs. In this method, a gene must show a significant response at one or more concentrations by Student’s t-test, and a three-point trend with a significant slope must be present. Additionally, the direction of change needs to be agreed, that is, an increase from the control must be matched with a positive slope, or a decrease from the control must be matched with a negative slope. If these conditions are met, data are iteratively fitted to the Hill equation, and the POD is designated as the concentration at which the line exceeds three standard deviations from the controls. Use of the Hill equation provides consistency between experiments and prevents overfitting, which may result from fitting to multiple or insufficiently constrained equations (61).

The POD concept is widely used in toxicology as a measure of the lowest toxicant concentration that causes a statistically significant response and is analogous to the lowest observed effect level (62). This POD algorithm not only increases the stringency for identifying treatment-responsive genes over a t-test at a single concentration but also allows separation of biological pathways that respond to chemical binding to different targets at low versus high concentrations. The POD algorithm has recently been submitted for publication and will be made publicly available on a website to encourage its use.

IMPACT OF SCREENING DATA

Throughout the past decade, the data generated by the Tox21 program have impacted the fields of in vitro toxicology and computational modeling. In 2015, a data challenge was proposed to further the quality of predictive models of nuclear receptor and stress response pathway assays (63). Researchers worldwide were asked to use a panel of screened nuclear receptor and stress response pathway assays as a training set for identifying biochemical and cellular pathways based on chemical structure data. Nearly 400 submissions, from 18 different countries, were acquired, and the winning models achieved greater than 80% accuracy, with several reaching over 90%. This important data challenge was able to prioritize novel chemicals for further testing with respect to human health concerns. This challenge had important implications since it featured a combination of cell-based assays with computational modeling approaches. Tox21 and computational labs inside NCATS demonstrated that using in vitro data to generate a toxicity computational model results in a more accurate prediction of human toxicity than in vivo animal studies (64).

In 2010, the Deepwater Horizon oil platform spilled a massive amount of oil into the Gulf of Mexico; consequently, large volumes (>1 million gallons) of oil dispersants were scattered into the Gulf to remediate the spill. These dispersants had little toxicological data available, posing public health concerns. Some dispersants degrade into nonylphenol, which is a known endocrine disrupter; therefore, there was a need for rapid toxicological data on the dispersants. NCATS screened eight commercial oil dispersants with in vitro high-throughput assays focusing on the ARs and ERs, alongside a variety of other biological pathways (65). While none of the dispersants modulated the activity of the AR, two displayed a weak ER signal in one assay. Although eight oil dispersants were screened, the dispersant used for the Deepwater Horizon oil spill, Corexit 9500, did not demonstrate ER activity. This study demonstrated how the Tox21 can respond to data gaps associated with an unprecedented environmental disaster.

Alongside projects conducted amongTox21 partners, our team also collaborates with many labs both external and internal to NCATS. Recently, the world has been impacted by the coronavirus disease (COVID-19) pandemic caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). In cooperation with other laboratories at NCATS, a biological activity-based modeling approach was used to discover potential drugs that target SARS-CoV-2; these could serve as future candidates for treating COVID-19 (14). Also, potential COVID-19 drug candidates, currently in treatment trials, were evaluated for their effects against various targets and signaling pathways based on the existing Tox21 screening data sets due to several cellular signaling pathways getting recruited upon viral infection (66). Other examples of collaborations include screening botanicals and dietary supplements (67), prioritizing chemicals for genotoxic characterization (68, 69), identifying chemicals/drugs that inhibit fetal CYP3A7 activity (70, 71), characterizing methylene blue as a promising therapy against flaviviruses such as the zika virus (72), and revealing AP-1 and autophagy pathways as potential drug therapy targets against COVID-19 (73). One major benefit of Tox21 is the publicly accessible data. Many laboratories worldwide have analyzed Tox21 data, including assessing multiple species for acute toxicity of compounds (74) and generating a high-throughput proteomics sample preparation platform (75). These are just a few examples demonstrating the impact of scientific collaboration by the NCATS Tox21 team.

CURRENT CHALLENGES AND OPPORTUNITIES FOR HIGH-THROUGHPUT IN VITRO TOXICOLOGY

Although progress has been made within the Tox21 program to use in vitro assays for toxicological testing, there are some challenges that need to be addressed to improve the physiological relevance of Tox21 data. A significant limitation for interpreting Tox21 data is neglecting the pharmacokinetics (absorption, distribution, metabolism, and excretion) of a compound, which can have a major impact on overall toxicity. Not incorporating these data can lead to false positives or negatives. Metabolism is one pharmacokinetic parameter currently being investigated for incorporation into existing Tox21 assays. Although it is difficult to incorporate metabolism into HTS, a small number of assays (e.g., AChE inhibition and p53 induction) have demonstrated that metabolic systems may be compatible with screening. Aside from these two assays, most of the projects to date in the Tox21 program have not considered the impact of metabolism. The aim for the next phase of this program is to enhance screening capabilities by addressing this issue and including metabolism into additional assays, as well as developing more cell models with physiologically relevant systems (76, 77). Identifying the influences of metabolism along with other pharmacokinetic factors (absorption, distribution, and excretion) is important to determining the full physiological context of a toxic compound (78) and, therefore, needs to be incorporated into HTS data.

In addition to pharmacokinetic considerations, accurate ways to understand the true concentrations of chemicals being tested in vitro are also important. The Tox21 10K chemical library contains compounds with a wide range of physicochemical properties (8), including molecular weight, hydrophobicity (logP), functional groups, and reactivity. These properties can lead to chemicals adsorbing to multiwell plates, protein binding, insolubility, and instability in media, as well as changes in cell membrane permeability. Models that can account for these factors (79) may better inform researchers on the true extracellular and intracellular concentrations of compounds being screened in various in vitro toxicological assays.

Another broad challenge is data interpretation. While we have shared approximately 100 million data points of screening data with the community, it is important to know how to interpret these data. Mining through as much data as our screening effort puts out and understanding the caveats associated with the assay technologies can be very challenging; this is why a data analysis pipeline has been optimized to account for multichannel readouts and assay artifacts and been set forth by the NCATS computational experts (24). Alongside understanding data, the informatics modeling itself can be improved upon by incorporating the combination of in vivo and in vitro data to predict human toxicity more accurately.

FUTURE PERSPECTIVE

Addressing the issues in the preceding section is just the beginning of what the future holds for our HTS efforts; evolving with novel technology and updated scientific advancements is one of our main goals. One of the ways in which future screening efforts can help advance toxicological testing is to find new methods to reduce the use of animal models. Previously, we described the use of a high-throughput DPRA assay that quantifies chemical-peptide conjugation (46). In the future, we believe this improved method could replace one of the current dermal testing models and are hoping to implement this to screen additional compounds. Physiological cellular models are also currently being investigated and optimized with the goal of reducing in vivo animal models. Once an experiment can accurately include metabolism, it can predict the physiological outcome to a much higher degree and therefore be able to prioritize chemicals for further in-depth study. Since these experiments will be cell based, it will be quick and efficient to screen compounds in multiple assays and generate a thorough characterization of the mechanism of action and toxicity for each chemical. For example, after optimization, there will be 18 different assays identifying modulation of the ER (80), which can then further be studied to determine endocrine disruption. Once these methods are adopted, outdated in vivo assays can be replaced. Stem cells have also recently been used to understand person-to-person variability, as well as genetic components of disease states (81). Utilizing these physiologically relevant cells in Tox21 assays would vastly improve the accuracy of predicting human toxicity.

Generating a comprehensive view of the toxic outcome of each compound will take a thorough knowledge of processing the Tox21 data. Due to the intricate nature of analyzing each compound through multiple assay end points, the next step is to teach a workshop involving understanding these data sets and how to accurately interpret them. After streamlining the NCATS data to link animal and human studies into a prediction algorithm, this knowledge could then be used in a regulatory sense. These approaches would subsequently lessen the large amount of in vivo animal studies that are necessary for toxicity profiling. Tox21 has come a long way since its inception; however, there are always further advancements and improvements to be made within the scientific community, and, to this end, we will continue to engage and expand upon HTS methods with the end goal of characterizing hazardous chemicals in a quick and accurate manner.