Advancing Computational Toxicology in the Big Data Era by Artificial Intelligence: Data-Driven and Mechanism-Driven Modeling for Chemical Toxicity

Abstract

In 2016, the Frank R. Lautenberg Chemical Safety for the 21st Century Act became the first US legislation to advance chemical safety evaluations by utilizing novel testing approaches that reduce the testing of vertebrate animals. Central to this mission is the advancement of computational toxicology and artificial intelligence approaches to implementing innovative testing methods. In the current big data era, the terms volume (amount of data), velocity (growth of data), and variety (the diversity of sources) have been used to characterize the currently available chemical, in vitro, and in vivo data for toxicity modeling purposes. Furthermore, as suggested by various scientists, the variability (internal consistency or lack thereof) of publicly available data pools, such as PubChem, also presents significant computational challenges. The development of novel artificial intelligence approaches based on public massive toxicity data is urgently needed to generate new predictive models for chemical toxicity evaluations and make the developed models applicable as alternatives for evaluating untested compounds. In this procedure, traditional approaches (e.g., QSAR) purely based on chemical structures have been replaced by newly designed data-driven and mechanism-driven modeling. The resulting models realize the concept of adverse outcome pathway (AOP), which can not only directly evaluate toxicity potentials of new compounds, but also illustrate relevant toxicity mechanisms. The recent advancement of computational toxicology in the big data era has paved the road to future toxicity testing, which will significantly impact on the public health.

INTRODUCTION

Traditional experimental testing procedures, both in vitro and in vivo, to identify compounds that can induce chemical toxicity are generally expensive and time-consuming.^1,2 Computational modeling is a promising alternative method for chemical toxicity evaluations. Existing computational models for risk assessment, such as quantitative structure–activity relationship (QSAR) models for various toxicity end points, can be used to quickly predict large numbers of new compounds in the risk assessment process and prioritize potential toxic compounds for experimental testing. However, critical issues of previous computational toxicology modeling studies, such as the small size of the data sets often being used in model development inducing coverage of a limited chemical space,³ activity cliffs,⁴ and overfitting,⁵ limit the applicability of existing models (e.g., QSAR models). The primary hypothesis of QSAR modeling (e.g., similar compounds will have similar activities) sometimes proves to be flawed and is the primary reason for activity cliffs.^6,7

Despite these limitations, regulatory acceptance of computational models remains an urgent demand in modern toxicology.^8,9 In 2016, the Frank R. Lautenberg Chemical Safety for the 21st Century Act (LCSA) became the first legislation since the Toxic Substances Control Act of 1976 (TSCA) to progress chemical risk assessment.^10,11 An essential component of the LCSA is a call for applicable computational approaches and associated predictive models for safety evaluation purposes.¹⁰ In the past decade, the development of new experimental protocols, especially high-throughput screening (HTS) assays, and the progress of combinatorial chemistry generated various biological data for millions of compounds.¹² Data sharing projects, such as PubChem,^13,14 have made chemical ”big data” publically available, which advanced modern toxicology studies into a big data era.^1,2,15,16 The available massive public data bring urgent requests for the development of innovative modeling approaches, driven by the recent progress of artificial intelligence, which can fulfill the current needs of chemical risk assessment.

On the basis of the Organization of Economic Co-operation and Development (OECD) guidance of QSAR model development for chemical toxicity, the predictions of computational models for new compounds need to be mechanistically explainable.¹⁷ However, the recently popular neural network approach to deal with big data typically performs as a “black box” algorithm,^18,19 which brings an uncertain future to computational toxicology.¹² Many HTS assays utilize human cells and tissues and have quantitative results that allow for mechanistic interpretation.¹² This data landscape enables researchers to create in silico models that incorporate the concept of the adverse outcome pathway (AOP)²⁰ with publically available big data, resulting in mechanism-driven modeling studies.^1,15,21 The resulting models of these studies can not only predict the toxicity of new compounds, but also illustrate toxicity mechanisms of importance in humans and animals, thereby filling the gap created by speculation about a possible lack of concordance between animal and human test data.²² The urgent need for advanced computational methods, availability of abundant HTS big data, and opportunity for incorporation of mechanistic analysis introduce new challenges and prospects to the modern computational toxicology area.

BIG DATA IN CHEMICAL TOXICOLOGY

The term “big data” refers to data sets, structured or unstructured, that multiply quickly and are so large and multifaceted that they are impossible to treat using personal computers and traditional computational approaches.²³ Data sets with big data require advanced tools such as heterogeneous and cloud computing²⁴ that have capabilities beyond those of conventional data processing and handling techniques as well as dynamic data curation and sharing using algorithms such as those used to handle data streams.^25,26 These advanced techniques allow for rapid identification of target entities in these massive data sets in ways that manual data compilation and curation could never efficiently match, which has radical implications for the improvement of traditional computational toxicology modeling techniques like read-across.^15,16

Recent HTS programs and their associated data sharing efforts have revolutionized the landscape in many health fields, highlighted by the Big Data to Knowledge (BD2K) initiative by the National Institutes of Health (NIH), which emphasizes the usefulness of big data in biomedical research and critical need to capitalize on the amount of data available in the health field.^27,28 A significant HTS effort in toxicology is the United States Environmental Protection Agency (US EPA) research program called Toxicity Forecaster (ToxCast), which employed in vitro HTS tests and toxicogenomics techniques to quickly evaluate the toxicity of compounds and prioritize compounds for experimental testing.^29–31 Phase I of this project evaluated 300 unique compounds, mostly of agricultural interest (i.e., pesticides), using about 500 HTS assays.³⁰ Phase II evaluated an additional 767 compounds, including some failed pharmaceutical compounds, using about 700 HTS assays.³¹ Recently, the ToxCast initiative advanced to the Tox21 collaboration between the US EPA Office of Research and Development/National Center for Computational Toxicology (NCCT), NIH/National Institute of Environmental Health Sciences (NIEHS)/National Toxicology Program (NTP), and the NIH/National Chemical Genomics Center (NCGC), now part of the National Center for Advancing Translational Sciences (NCATS).^32–35 Phase I of Tox21 used 75 HTS assays, which were selected and refined from ToxCast assays, to screen an initial set of about 2800 compounds.³² Phase II began in 2010 to screen a more extensive set of approximately 10 000 environmental compounds.^32,34,35 As of 2018, the Tox21 program generated over 120 million data points for approximately 8500 chemicals.³³

Publicly available databases store much of the data obtained from the toxicology community, including data from HTS programs such as the ToxCast and Tox21 programs.^29–31,36 Table 1 describes a selection of significant sources representing publically available big data in the toxicology field. Among them, Aggregated Computational Toxicology Resource (ACToR),^37,38 Registration, Evaluation, Authorization, and Restriction of Chemicals (REACH),^16,39–42 RepDose,⁴³ Safety Evaluation Ultimately Replacing Animal Testing (SEURAT),⁴⁴ and Toxicology Data Network (ToxNET)⁴⁵ were specifically developed to share toxicity data. Chemical Effects in Biological Systems (CEBS),⁴⁶ ChEMBL,⁴⁷ Connectivity Map,^48,49 Comparative Toxicogenomics Database (CTD),⁵⁰ DrugMatrix,⁵¹ Gene Expression Omnibus (GEO),^52,53 and PubChem^13,14 share general biological data, including toxicity data. Most of these data portals are being updated frequently, and the total number of available data is increasing quickly with the above-mentioned HTS programs. In 2013, the total toxicity data pool contained over 70 million compounds and around 1 million assays.⁵⁴ Figure 1 shows the increase of the numbers of compound and bioassay records in PubChem since 2008.^54–64 From 2008 to 2018, the number of compounds in PubChem increased over three-fold from 25.6 million⁵⁴ to 96.5 million.⁵⁶ Similarly, the number of bioassay records increased from approximately 1500⁵⁴ to over 1 million.⁵⁶

Table 1.

Selected Publically Available Big Data Repositories

database	sizea	data type	web access
ACToR37,38	over 800 000 compounds and 500 000 assays	toxicity data (in vitro and in vivo)	https://actor.epa.gov/
CEBS46	over 11 000 compounds and 8000 studies	gene expression data	https://www.niehs.nih.gov/research/resources/databases/index.cfm
ChEMBL47	1.1 million bioassays, 1.8 million compounds, over 15 million activities	literature data for binding, function, and toxicity of drugs and drug-like compounds	https://www.ebi.ac.uk/chembl/
Connectivity Map48,49	about 1300 compounds and 7000 genes	gene expression data	https://portals.broadinstitute.org/cmap/
CTD50	over 14 000 compounds, 42 000 genes, 6000 diseases	relationships among compounds, genes, and diseases	https://ctdbase.org/
DrugMatrix51	about 600 drug molecules and 10 000 genes	gene expression data	https://ntp.niehs.nih.gov/results/toxfx/index.html
GEO52,53	over 4300 subdata sets	microarray, next-generation sequencing and other forms of high-throughput functional genomics data	https://www.ncbi.nlm.nih.gov/geo/
PubChem13,14,76	over 96 million compounds, 1 million bioassays, and 13 billion data points	toxicology, genomics, pharmacology, and literature data	https://pubchem.ncbi.nlm.nih.gov/
REACH16,39–42	21 405 unique substances with information from 89 905 dossiers	data submitted in european union chemical legislation	https://echa.europa.eu/information-on-chemicals/registered-substances/
RepDose43	364 chemicals investigated in 1017 studies, which resulted in 6002 specific effects	repeat-dose study data for dog, mouse, and rat	https://repdose.item.fraunhofer.de/
SEURAT44	over 5500 cosmetic-type compounds in the current COSMOS database web portal	animal toxicity data	http://www.seurat-1.eu/
ToxNET45	over 50 000 environmental compounds from 16 different resources	toxicity data (in vitro and in vivo)	https://toxnet.nlm.nih.gov/

^aOn the basis of live web counts or most recent literature articles as of October 2018; ACToR, Aggregated Computational Toxicology Resource; CTD, Comparative Toxicogenomics Database; CEBS, Chemical Effects in Biological Systems; GEO, Gene Expression Omnibus; REACH, Registration, Evaluation, Authorization, and Restriction of Chemicals; SEURAT, Safety Evaluation Ultimately Replacing Animal Testing; ToxNET, Toxicology Data Network.

Figure 1.

Increase of the number of (a) compound and (b) bioassay records in PubChem in the recent ten year period (from September 2008 to September 2018).

MODELING CHALLENGES CREATED BY BIG DATA: THE FOUR “V”S

The available big data for chemical toxicity brings new challenges to the future computational toxicology studies.^36,65 As the “big data” concept suggests, the volume of data is a critical characteristic. The nature of data that is relevant to various toxicity end points creates a large data volume.^1,15,65 These data stem from information obtained from compounds and original testing protocols including chemical information,^13,14 physicochemical properties, in vitro data,^16,36–42 in vivo data,^{16,37–45,47} and various–omics data^46–53 (Table 1). For example, the current PubChem bioassay database has around 240 million bioactivities as 30 gigabytes of Extensive Markup Language (XML) files. It is not feasible to apply traditional computational approaches or even Personal Computers (PCs) to deal with data with this kind of volume for modeling purposes. The recent progress of computer hardware, especially the application of Graphics Processing Unit (GPU),⁶⁷ makes it possible to deal with toxicity data with significant volume.

The progress of testing technology determines the velocity of big data. In the 1990s, combinatorial chemistry began to progress rapidly, creating large chemical libraries for screening in drug discovery.⁶⁸ The advancement of HTS protocols in the past decades makes the screening of these large chemical libraries (i.e., over one million compounds) feasible.^69,70 Automatic data analysis and the application of robots to replace humans in the testing procedures considerably lower the cost of testing a compound and rapidly grow the current big data sources.³² As a result, a substantial number of compounds have been tested against many assays. Table 2 shows the 20 compounds obtained from the Tox21 program with the most active responses in the PubChem bioassay database¹⁴ (accessed December 2018). For example, doxorubicin (CAS 25316–40–9), a drug that is used to treat cancer by killing cancer cells, showed 4452 active responses (Table 2). Vorinostat (CAS 149647–78–9), which is used to treat T-cell lymphoma that persists after treatment with other drugs, showed 4278 active responses (Table 2). Other well-characterized compounds, such as drugs and well-known pesticides, have similarly prolific response information available.

Table 2.

Twenty Tox21 Compounds with the Most Active Responses in PubChem Bioassays

chemical	CAS	number of active responses	number of inactive responses
Doxorubicin	25316-40-9	4452	119
Vorinostat	149647-78-9	4278	760
Paclitaxel	33069-62-4	3043	801
Colchicine	64-86-8	2043	1581
Etoposide	33419-42-0	1907	525
Fluorouracil	51-21-8	1804	1887
Acetazolamide	59-66-5	1794	2091
Sunitinib	341031-54-7	1702	138
Methotrexate	59-05-2	1687	1120
Lestaurtinib	111358-88-4	1414	35
Gefitinib	184475-35-2	1379	661
Diazepam	439-14-5	1320	628
Haloperidol	52-86-8	1309	1820
Bortezomib	179324-69-7	1276	212
Zidovudine	30516-87-1	1251	1964
Clozapine	5786-21-0	1204	1687
Efavirenz	154598-52-4	1184	537
Dasatinib	302962-49-8	1078	380
Mitomycin	50-07-7	1048	664
Nicotine	54-11-5	983	1283

Traditional modeling studies, usually using small in-house data sets for modeling purposes, often had a risk of overfitting and made poor predictions to new compounds.² The modeling community expected models to improve with more available data used for modeling purposes, thereby increasing knowledge about activity cliffs⁷¹ and decreasing the chance of overfitting.⁷² However, when using large data sets for modeling purposes, traditional machine learning approaches usually have flaws such as extended computational time and memory requirements that require adaptation of commonly used algorithms.⁷³ As a potential solution, deep learning with neural networks using GPUs might be more suitable for big data processing.⁷⁴

Additionally, the variety of big data brings new challenges to modeling procedures. Traditional modeling studies only deal with one object (i.e., a toxicity end point) using one type of attributes (i.e., chemical descriptors). However, existing big data repositories (e.g., PubChem) contain a diverse variety of information for compounds of interest, such as quantitative data obtained directly from assays and qualitative data as the screening read-out, which requires different data processing techniques. Integrating and curating data with high variety requires advanced artificial intelligent approaches.⁷⁵

Each source of big data contains a certain degree of data variability. PubChem, for example, contains data deposited by different sources including academia, pharmaceutical companies, government agencies, chemical vendors, screening centers, and journal publishers.^{13–15,70,76} The information obtained from each of these data sources is not consistent across assays or compounds, which creates an inherent variability that creates challenges in the following modeling procedure. For example, data generated from a Tox21 quantitative HTS (qHTS) assay to measure genotoxicity induced by small molecules in human embryonic kidney cells⁷⁷ exists in PubChem twice as (1) original data (AID 651632) and (2) conclusions by counting cytotoxicity (AID 720516). Under this condition, automatic data mining tools need to be able to distinguish this difference.

Furthermore, inconsistencies may also arise due to inaccurate chemical structures^16,78–80 and inherent experimental errors^1,80,81 resulting from data quality control (QC) issues of various experimental across sources. When aggregating data from multiple sources, it is common to encounter different representations to represent the same compounds (e.g., implicit versus explicit hydrogens and tautomeric forms). The quality of experimental data is also likely to vary across sources due to differences in protocols, compound purities, and other experimental errors. For example, Luechtefeld et al. reported that animal toxicity data obtained from various sources for the same compounds have consistency ranging from 70% to 90%, depending on the nature of testing protocols.^39,42 Therefore, the data curation of both chemical structures and experimental data is critical before using big data for the computational modeling procedure.^79,80,82,83 Because of the size of public data sets, automated curation workflows, such as those described in previous publications,^79,80,82,84 are necessary prior to modeling.

Another data variability issue is due to the complex, disorganized nature of public data and unbalanced distribution of HTS testing results (i.e., many more inactive results than actives). Although a wealth of data exists, there are many data gaps for compounds of interest since no compound has been tested against all assays, and many tests returned inconclusive results. For most existing data, a bias exists toward inactive responses due to the nature of HTS assays. For example, searching the recent PubChem database for 8367 Tox21 compounds yielded 812 assays that have at least 25 active results within these compounds including assays carried out by the Tox21 program and other sources (accessed in November 2018) (Figure 2). There are approximately 6.8 million data points in this bioprofile. However, the ratio of active versus inactive results is 1:11 (2% vs 22%) and the remaining data (76%) represent results from which no conclusion can be made(i.e., either “inconclusive” or “untested”). It is understandable that inactive results are much less informative than active results to determine chemical toxicity. In this situation, novel modeling approaches are needed to deal with missing data such as the method described by Zhang et al.⁶ and biased data,including cost-sensitive learning,⁸⁵ under-sampling,⁸⁵ and oversampling^86,87 algorithms.

Figure 2.

Bioprofile of 8367 Tox21 compounds represented by data from 812 PubChem assays. Active results (1) were represented by red; inactive results (−1) were represented by blue; and inconclusives or untested results (0) were represented by gray.

DATA-DRIVEN COMPUTATIONAL TOXICOLOGY MODELING

Despite the challenges posed by big data in computational toxicological modeling, the advancement of data-driven technology and development of computational tools to overcome the challenges of the four “V”s create new opportunities for existing model improvement and novel model development (Figure 3). For example, some early developed data mining tools, such as Chem2BioRDF⁸⁸ and HTS Navigator,⁸⁹ can link various public data sources to target compounds. Additionally, some data sharing portals have relevant data mining tools, such as rpubchem⁹⁰ and ToxCast pipeline (“tcpl”),⁹¹ which exist to simplify and optimize the parsing and collection of data from PubChem⁹⁰ and the ToxCast database,⁹¹ respectively. The recently developed online tools REACHacross^2,16 and the Chemical In Vitro–In Vivo Profiling portal (CIIPro)⁹² provide new methods to extract data from the REACH and PubChem databases automatically. Additionally, online tools such as Chembench⁹³ and the Chemistry at Harvard Macromolecular Mechanics web-user interface (CHARM-Ming)⁹⁴ are available to streamline the development and distribution of curated toxicity data and QSAR models.

Figure 3.

General workflow for construction of data-driven and mechanism-driven models for chemical toxicity.

In 2014, the National Center for Advancing Translational Sciences (NCATS) launched a challenge for the development of computational models of nuclear receptor and stress response pathways using HTS data generated by the Tox21 program.⁹⁵ This challenge inspired the creation of around 400 models using a variety of different modeling techniques based on 8043 compounds, which have been tested against 12 HTS assays.⁹⁵ All the models were used to predict an external test set of 296 new compounds, which were experimentally tested using the same protocols but kept aside until all models were submitted to NCATS. The top model, which has the best prediction accuracy of these new compounds, employed a neural network approach.⁹⁶ The high performance of the resulting ensemble model demonstrates the potential of neural networks for the improvement of big data models. In 2017, another similar initiative of big toxicity data modeling was organized by the National Toxicology Program Interagency Center for the Evaluation of Alternative Toxicological Methods (NICEATM) and the Interagency Coordinating Committee on the Validation of Alternative Methods (ICCVAM).⁹⁷ Thirty-two international research groups submitted models using deep learning, classic QSAR, and cluster-based methods. These models were developed based on 8994 compound training set tested in rodent studies for acute toxicity. Consensus predictions were made for a large set of 40 000 compounds of environmental interest as a resource for the toxicology community.

In recent studies, neural networks, as a popular artificial intelligent approach, showed advantages to deal with large data sets. In 2017, Xu et al. reported three neural network models developed to predict acute oral toxicity end points based on a training set of 8080 compounds.⁹⁸ All three models (i.e., a regression model for LD₅₀ values, a multiclassification model for US EPA hazard categories, and a multitask model to simultaneously predict both of these end points) simultaneously outperformed previously reported models for these end points. Wen et al. also reported a deep learning model developed to predict interactions between drugs and their biological targets based on 15 524 drug–target pairs obtained from the DrugBank database.⁹⁹ This model employed a pretraining feature extraction step to predict whether specific drug-target pairs will interact and overall outperformed classic QSAR approaches. The high performance of these models demonstrates not only the advantages of using neural networks to model large data sets, but also to advance feature selections. On the other hand, there is a study that showed that neural network models are not better than traditional models using machine learning approaches.¹⁰⁰ Currently, there is still no universal criterion to select modeling approaches for big data sets.

Read-across was initially introduced as a technique to fill toxicity data gaps by making predictions based on similar compounds.¹⁰¹ Traditionally, read-across relied only on chemical similarity calculations. In some recent studies, extra parameters, such as physicochemical properties, biological interactions (e.g., metabolism potential), including QSAR predictions, were also used to identify similar compounds.^102,103 However, in the big data era, the inclusion of comprehensive toxicity data in the read-across study may improve model quality.^{15,16,104,105} Luechtefeld et al. recently described a method for automating the read-across process using read-across structure–activity relationship (RASAR) models.¹⁰⁵ The RASAR models developed in this study were based on the REACH database consisting of over 10 000 unique compounds with various toxicity end points, making use of hazard classifications rather than the raw data from in vivo tests. According to this study, due to the integration of large collections of toxicity data and the application associated with new modeling approaches, such as automatic feature selection, the resulted RASAR models showed better average sensitivity than the reproducibility of experimental animal tests, which indicated potential abilities to overcome experimental errors and data inconsistencies by computational modeling approaches.

PREDICTIONS AND INTERPRETATIONS OF NEW TOXICANTS BY MECHANISM-DRIVEN MODELING

In 2007, the National Research Council (NRC) laid out the framework for risk assessment using modern techniques, urgently calling for mechanistic rather than empirical interpretation.¹² Traditional QSAR models often perform as “black boxes”, which provide toxicity predictions without clear mechanistic interpretations.¹⁹ Although the neural network modeling showed certain advantages when dealing with big toxicity data, the resulted neural network models still cannot resolve the above challenge. In vitro assays often investigate mechanistically relevant information on toxicants. However, one or few assays cannot represent the complexity of whole organisms, and the results obtained from in vitro and in vivo tests always have obscure relationships, making in vitro–in vivo extrapolation (IVIVE) challenging.¹⁰⁶

Mechanism-driven modeling, initiated and advanced by the concept of the adverse outcome pathway (AOP), allows for mechanistic extrapolation of toxicity evaluations of new compounds, filling a critical need of applying alternatives for regulatory toxicology studies.²⁰ An AOP starts with a molecular feature (e.g., a chemical fragment), which indicates potential interactions with biomolecules such as receptors. This molecular initiating event (MIE) that triggers a cascade of measurable key events at the cellular level, and lead to tissue, organ, and eventually in vivo organism level adverse outcomes²⁰ (Figure 3). The identification and organization of MIEs and key events in a pathway that leads to an adverse outcome define the associated toxicity mechanisms of interest for risk assessments. Mechanism-based assays outcomes can be used within this pathway to systematically assess whether a compound is likely to induce the target adverse outcome.¹⁰⁷ Currently, AOPs are being developed for various types of toxicities such as acute inhalation toxicity,¹⁰⁸ neurotoxicity,^109–111 skin sensitization,^112–114 estrogen receptor bindings,^115,116 forestomach tumors not induced by genotoxic events,¹¹⁷ and drug-induced cholestatic liver injury.¹¹⁸

One of the major research goals of toxicology HTS programs, such as ToxCast and Tox21, is to perform mechanism-driven computational toxicology modeling, which presents a practical way to increase the quality of IVIVE by employing comprehensive testing batteries consisting of associated in vitro assays related to animal toxicity end points.¹ These programs generate a large amount of mechanistic data that paves the way for AOP modeling that are more interpretable than traditional computational toxicology studies, which are always questioned as “black boxes.” Carefully considering the biological relevance of the experimental data selected, including associated biological pathway information, for modeling (i.e., HTS assay measurements and readouts) the target animal toxicity end point of interest is to integrate the AOP concept into the development of computational models with both high prediction accuracy and a meaningful biological interpretations. Therefore, the current critical features of resulting AOP models are (1) biological relevance of data to target toxicity; (2) computational approaches to identify/organize mechanistic assays; and (3) both predictive and interpretable pathway models.

The current literature also documents the results from profiling of ToxCast and Tox21 assay data using computational clustering techniques to elucidate previously unknown compound–receptor interactions, pathway perturbations, and toxicity mechanisms.^119–123 One such profiling effort was described by Sipes et al. in 2013.¹¹⁹ This study clustered 976 compounds from 330 ToxCast Phase I and II bioassays based on chemical structure and bioassay responses, which led to the identification of possible modes of action of compounds. For example, a pharmaceutical compound Anthralin (CAS 1143–38–0) that has a therapeutic use for the treatment of psoriasis with unknown mechanism of action was identified to show active responses in the same assays as a known inhibitor of inflammation, tannic acid. This connection gives insight into the possible mode of action of Anthralin and demonstrates the value of data generated from the ToxCast program in identifying previously unknown mechanistic information for target compounds.

The ToxCast initiative and Tox21 program have also inspired the creation of mechanistic models for developmental toxicity,^124,125 estrogen receptor activity,^126–128 and acute oral toxicity¹²⁹ that incorporate mechanistically relevant HTS assay data by computational models into pathways leading to adverse outcomes. For example, Browne et al. developed a computational model that incorporates HTS data from 18 ToxCast assays that comprise an adverse outcome pathway leading to endocrine disruption.¹²⁷ The authors of this study also demonstrated a generalizable performance-based validation procedure to evaluate the robustness of a computational AOP model for regulatory use. To be considered as a viable alternative for regulatory evaluation purposes, computational models must perform equivalently or better than the existing approved protocols. This computational endocrine disruption AOP model was validated by predicting a set of in vitro reference chemicals identified by ICCVAM and in vivo reference chemicals curated through literature review.⁷⁵ The computational model pre-dictions of the 42 in vivo reference chemicals with at least two independent concordant results (active or inactive) from guideline-like uterotrophic studies showed 84% accuracy. A comparison of the computational predictions with the results from in vivo protocols identified false negative, which showed activity in multiple independent in vivo tests but inactivity in the 18 ToxCast assays, potentially due to its volatility.

The strong potential for computational approaches to support risk assessment grounded in mechanistic interpretation has inspired the creation of other mechanism-based studies. These studies, similar to AOP models developed by using mechanistic ToxCast and Tox21 assays, can predict chemical toxicity by interpretable results. However, instead of manually selected assays to be integrated into pathways, these studies relied on computational approaches to prioritize useful biological data, which are suitable for big data modeling. For example, Virtual AOP (vAOP) models were developed by using the currently available big data for hepatotoxicity.¹³⁰ In this study, an automatic profiling tool that can extract bioassay data from PubChem was used to identify assays relevant to hepatotoxicity and oxidative stress.^92,130 Data from several PubChem assays can be combined to predict hepatotoxicity for compounds with specific structural alerts. The resulting vAOP models provided insight into possible new mechanisms leading to hepatotoxicity.¹³⁰ The identified vAOP contained two chemical fragments as MIEs and four PubChem assays (AIDs 686978, 743067, 743140, and 743202), which are all relevant to oxidative stress, such that if a new compound contains one of these MIEs and has an active response in at least one of the four assays, it is predicted to be hepatotoxic by inducing oxidative stress.

Luechtefeld et al. reported a procedure for the recursive importance-based elimination of chemical and biological features that are irrelevant to target toxicity.¹³¹ The application of this technique involves ranking features based on relative importance to identify the assays and chemical fragments that contribute the most critical information to a resulting model of an in vivo toxicity end point. In this study, they identified in vitro assays and chemical fragments of mechanistic significance to skin sensitization and then used this information to train models that incorporate dose–response data, which showed an advantage when compared to models trained without these data. The success of this modeling process was validated by better predictivity and mechanism interpretations.

Computational techniques and models based on chemical structures, such as structural alerts, also could advance traditional QSAR studies by predicting toxicity mechanisms for large data sets.^107,132 For example, recently, there have been reports of modeling studies to predict MIEs relevant to hepatic steatosis.^133,134 Mellor et al. evaluated binding interactions of 12 713 compounds in the ChEMBL database with nuclear receptor structure files in Protein Data Bank to develop a basic alert-based workflow to identify compounds that may bind to nuclear receptors and induce hepatic steatosis.¹³³ Another study by Gadaleta et al. involved the development of QSAR models to predict compound activity in ToxCast assays that are relevant to MIEs that lead to hepatic steatosis.¹³⁴

OTHER AREAS OF COMPUTATIONAL TOXICOLOGY IN THE BIG DATA ERA

A critical aspect of mechanism-based toxicity evaluation is the incorporation of toxicokinetic information on compounds of interest to predict dose-dependent Effects of compounds. For example, Bhhatarai et al. recently reported a modeling study of acute toxicity, which incorporated simulations of absorption and metabolism into the modeling process.¹²⁹ Strope et al. also reported a modeling study that resulted in an ionization constant (pK_a) model for a set of 32 413 chemicals.¹³⁵ The applicability of this model was evaluated by using the pK_a predictions to estimate distribution ratio into tissues for 22 compounds with steady-state volume of distribution data. Modeling studies that incorporate toxicokinetic information are becoming applicable with tools such as the High-Throughput Toxicokinetics (“httk”) package in R that was designed to make use of the data from programs such as ToxCast and Tox21.¹³⁶

Toxicogenomics uses techniques such as proteomics, metabolomics, and genetic sequencing to study the toxicity of compounds and provides insight into how cells exposed to a toxic chemical express genes, proteins, and metabolites, which yields critical information for elucidating and understanding toxicity pathways.^17,137 The toxicogenomics data complement the results of in vitro assays that focus on interactions with and activation of nuclear receptors and specific cellular stress responses.¹³⁸ In the current big data era, the toxicogenomics data landscape continues to grow. For example, in 2008, the CTD contained 116 067 compound–gene interactions.⁶⁶ By 2016, this number increased more than 10-fold to 1 379 105 compound–gene interactions.⁵⁰ A collaboration among Agilent, Inc., Brown University, Georgetown University, the Hamner Institute, the Johns Hopkins Center for Alternatives to Animal Testing (CAAT), and the US EPA led the Human Toxome project that also aims to generate omics data, with an end goal of developing a process to map and evaluate the specific molecular mechanisms that underlie AOPs.¹³⁹ As toxicogenomics technologies continue to progress and molecular mechanisms become well-understood, it will become feasible to assess differences in chemical toxicity pathways that may arise due to the genetic variation that is inherent among individuals.

CONCLUSIONS

Computational modeling is a promising alternative method to replace, reduce, and refine traditional animal models for chemical toxicity evaluations, especially in the current big data era. As big data repositories continue to grow at rapid velocity and new techniques to deal with big data sets are being developed, computational models become applicable to large chemical space analysis, diverse biological data optimization, and complex mechanism studies. This innovative movement will allow not only for the predictions of new compounds, but also for toxicity mechanism illustrations of potential toxicants. This currently growing big toxicity data landscape and the advancements in modeling technique developments to handle the wealth of toxicity information available together create a new direction that is optimal for the integration of computational models into the mechanism-based chemical risk assessments, which are urgently required by regulatory agencies.

ABBREVIATIONS

ACToR

Aggregated Computational Toxicology Resource

AID

BioAssay Identifier

AOP

Adverse Outcome Pathway

BD2K

Big Data to Knowledge

CAAT

Center for Alternatives to Animal Testing

CAS

Chemical Abstracts Service

CEBS

Chemical Effects in Biological Systems

CHARMMing

Chemistry at Harvard Macromolecular Mechanics

CIIPro

Chemical In Vitro–In Vivo Profiling

CTD

Comparative Toxicogenomics Database

GPU

Graphics Processing Unit

GPU

Graphics Processing Unit

HTS

High-Throughput Screening

ICCVAM

Interagency Coordinating Committee on the Validation of Alternative Methods

IVIVE

In Vitro–In Vivo Extrapolation

LCSA

Frank R. Lautenberg Chemical Safety for the 21st Century Act

LD₅₀

Median Lethal Dose

MIE

Molecular Initiating Event

NCATS

National Center for Advancing Translational Sciences

NCCT

National Center for Computational Toxicology

NCGC

NIH Chemical Genomics Center

NICEATM

National Toxicology Program Interagency Center for the Evaluation of Alternative Toxicological Methods

NIEHS

National Institute of Environmental Health Sciences

NIH

National Institutes of Health

NRC

National Research Council

NTP

National Toxicology Program

OECD

Organization for Economic Cooperation and Development

PC

Personal Computer

pK_a

Ionization Constant

QC

Quality Control

QSAR

Quantitative Structure–Activity Relationship

RASAR

Read-Across Structure–Activity Relationship

REACH

Registration, Evaluation, Authorization and Restriction of Chemicals

SEURAT

Safety Evaluation Ultimately Replacing Animal Testing

TGx

Toxicogenomics

Tox21

Toxicity Testing in the 21st Century

ToxCast

Toxicity Forecaster

ToxNET

Toxicology Data Network

TSCA

Toxic Substances Control Act of 1976

US EPA

United States Environmental Protection Agency

vAOP

Virtual Adverse Outcome Pathway

XML

Extensive Markup Language

Biographies

Heather L. Ciallella currently is a Ph.D. student in CCIB under the mentorship of Dr. Hao Zhu, where her research focuses on the applications of deep learning algorithms to chemical toxicity predictions.

Hao Zhu is an Associate Professor in the Chemistry Department and Center for Computational and Integrative Biology (CCIB) at Rutgers, The State University of New Jersey in Camden. He received his Ph.D. in Computational Chemistry from Case Western Reserve University in 2002. Dr. Zhu has authored or coauthored over 60 peer-reviewed publications and book chapters in the applications of cheminformatics to chemical toxicity assessments, computer-aided drug discovery, and rational nanomaterial design.