In Silico Approaches In Carcinogenicity Hazard Assessment: Current Status and Future Needs

Abstract

Historically, identifying carcinogens has relied primarily on tumor studies in rodents, which require enormous resources in both money and time. In silico models have been developed for predicting rodent carcinogens but have not yet found general regulatory acceptance, in part due to the lack of a generally accepted protocol for performing such an assessment as well as limitations in predictive performance and scope. There remains a need for additional, improved in silico carcinogenicity models, especially ones that are more human-relevant, for use in research and regulatory decision-making. As part of an international effort to develop in silico toxicological protocols, a consortium of toxicologists, computational scientists, and regulatory scientists across several industries and governmental agencies evaluated the extent to which in silico models exist for each of the recently defined 10 key characteristics (KCs) of carcinogens. This position paper summarizes the current status of in silico tools for the assessment of each KC and identifies the data gaps that need to be addressed before a comprehensive in silico carcinogenicity protocol can be developed for regulatory use.

1. Introduction

Globally, cancer has generally been the second leading cause of death responsible for an estimated 9.6 million deaths in 2018 [1]. Carcinogenicity is a complex process that conceptually is often divided into three stages – initiation, promotion, and progression, the latter includes malignant conversion and metastasis (e.g., 2). Several public or private in silico models (e.g., CASE Ultra¹, Derek², Lazar³, the OECD QSAR Toolbox⁴, the Leadscope SAR Carcinogenicity Database⁵, the U.S. Food and Drug Administration (FDA) Center for Drug Evaluation and Research rodent carcinogenicity (Q)SAR models⁶, VEGA Hub⁷) have been developed to predict the rodent carcinogenicity potential of chemicals. However, despite recent improvement in performance, the overall concordance of these models is not sufficient to serve as a primary basis for drug regulatory safety assessments [7] and therefore, their use is limited to supporting drug safety assessments and to its application in a tiered testing framework for prioritization of environmental chemicals by chemical regulators. Thus, there remains a need for improved in silico carcinogenicity models to be developed, especially ones that are more directly applicable to predicting the risk for cancer in humans rather than in rodents, for use in research and regulatory decision-making.

A major obstacle in the development of in silico methods for some endpoints is the lack of sufficient reliable and robust training datasets while the main obstacle for their regulatory adoption is the lack of widely accepted procedures for performing and documenting such assessments. In response, an international inter-disciplinary consortium representing a number of stakeholders was organized to develop publicly available in silico protocols to support the hazard assessment of major toxicological endpoints [8]. To date, this effort has resulted in the publication of in silico protocols for genetic toxicology [9] and skin sensitization [10] while the development of similar protocols for acute lethality, endocrine activity, and skin/eye irritation and corrosion are in progress. Here, we report on a working group within the consortium that focused on developing an in silico protocol for the hazard assessment of carcinogens, to be based on the Smith et al. [11] publication that recognized 10 key characteristics (KCs) of carcinogens. The use of these KCs to identify carcinogens is challenging given the lack of relevant standardized assays for most KCs and thus a concomitant lack of KC relevant computational methods. However, a KC framework might facilitate the combination of different pieces of evidence to create a systematic approach for the evaluation and integration of mechanistic evidence with other types of evidence while allowing for hazard classifications that are scientifically defensible and appropriate for regulatory decision-making [12], as demonstrated recently by California under Proposition 65 [13] and by Jaechke et al. [14] in their respective analyses of the potential carcinogenicity of acetaminophen. Thus, we decided to summarize the current status of in silico tools for the hazard assessment of carcinogens and to identify data gaps that need to be addressed before a KC-based in silico carcinogenicity protocol could be developed.

2. Key Characteristics of Carcinogens

2.1. KC1: Is Electrophilic or Can Be Metabolically Activated

Electrophiles play a key role in all stages of carcinogenicity by forming covalent adducts with electron-rich nucleophiles, such as DNA, RNA, proteins, and lipids [15–18]. While seen most prominently causing DNA damage, electrophiles are also involved in post-translational protein modification, which can alter protein enzymatic activity, the transduction of signals within and between cells, and gene expression [19]. Nonreactive substances may be transformed to reactive electrophilic species via metabolic processes in vivo, with biotransformation facilitated by mammalian and bacterial enzymes depending on the species, substrate, and route of exposure [15,20,21]. The type of interaction and ultimate effect depends on the inherent characteristics of the chemical and its biological target [16,22,23]. Measurement of a chemical’s electrophilicity or its propensity to be metabolically activated to an electrophile is generally performed indirectly, most frequently via examination of protein or DNA adducts, measured in vitro or in vivo [24].

Ashby and Tennant [25] identified a set of generally reactive chemical features potentially capable of covalently binding to DNA; this list has been refined and updated to predict additional electrophilic actions as well as structural alerts for genotoxicity [16,18,26,27]. In chemico knowledge of how individual functionalities of a structure interact and affect its overall potential for electrophilicity is needed to understand reactions with important cellular targets [22,28–33]. However, relying on the identification of a structural electrophilic mechanisms alone for predicting activity is over-predictive, leading to false positives [34]. The computational arena is replete with models for predicting bacterial mutagenicity, which is an indirect marker for a structure being an electrophile or able to be transformed into one [15,35]. Such computational models have been shown to adequately predict a structure’s potential for mutagenicity (with or without metabolic activation) and their use has been accepted as a replacement for the bacterial reverse mutation assay for pharmaceutical impurities [36–38]. Stand-alone models for prediction of metabolism exist that consider human metabolism (e.g., 39–42), though the potential electrophilicity of the products must be analyzed in separate quantitative structure-activity relationship (QSAR) models [43–46]. Multiple tools exist to determine a substance’s electrophilicity, although they generally measure only indirect endpoints. The ability to link these different tools into an integrated hazard assessment strategy that better reflects the physiological situation remains a work in progress that needs to be addressed to improve on the in silico prediction of electrophilicity.

2.2. KC2: Is Genotoxic

The term genotoxicity relates to the ability of a chemical to damage DNA and/or chromosome structure or segregation; such changes might lead to permanent changes to the genome which can be transmitted to successive generations of cells, potentially increasing the risk of cancer [47,48]. Genotoxicity resulting in cancer is the best understood and most experimentally measured of any of the KCs. It is also the most common characteristic identified when analyzing known human carcinogens [49].

A comprehensive in silico protocol for the hazard assessment of genotoxic chemicals was published recently [9]. This protocol contains information on the available test methods to assess genotoxicity, the in silico methods that are being used to predict genotoxicity, and approaches for their integration with experimental data. Table 1 summarizes information on the in vitro and in vivo tests for genotoxicity generally used for making regulatory decisions, many of which are harmonized in Organisation for Economic Co-operation and Development (OECD) Test Guidelines.

Table 1.

Genotoxicity Test Methods, OECD Test Guidelines, and In Silico Methods^*

Test Method	OECD Test Guideline	Endpoint	In Silico Methods
In vitro Bacterial Reverse Mutation Assay	OECD 471 [50]	Detects DNA point mutations.	An extensive number of compounds tested. Primary endpoint for genetic toxicity QSAR models (rule-based and statistical-based) [37,51].
In vitro mammalian mutation tests	OECD 490 [52]	Detects gene mutations and chromosomal damage in endogenous thymidine kinase gene in the L5178Y TK+/− cell line (MLA) and the TK6 human lymphoblastoid cell line.	In silico MLA models are available, although limited expert rules exist.
	OECD 476 [53]	Detects mutations at HPRT and at the XPRT transgene.	There is not enough data on HPRT or XPRT mutations to support useful in silico model development. Currently used for read-across/weight of evidence.
In vitro micronucleus	OECD 487 [54]	Detects chromosome damage or other effects that lead to the formation of micronuclei in the cytoplasm of interphase cells.	Limited data published currently could be used for in silico derivation of structural alerts.
In vivo micronucleus	OECD 474 [55]	Detects chromosomal damage which results in the formation of micronuclei in erythrocytes in the bone marrow or peripheral blood cells of rats or mice	In silico in vivo micronucleus models available [56], although limited expert rules exist.
In vivo transgenic gene mutation assays	OECD 488 [57]	Measure mutagenic responses dependent upon in vivo metabolism, pharmacokinetics, DNA repair processes, and translesion DNA synthesis.	Limited data published, could be used for read-across/weight of evidence and derivation of expert alerts.
In vivo Pig-a assay	Validation documents only available [58,59]	Detects phenotypic mutation in an endogenous mammalian gene [60].
In vivo Comet assay	OECD 489 [61]	Measures DNA strand breaks and alkali-labile sites.

^*Modified from Hasselgren et al. [9].

HPRT = hypoxanthine-guanine phosphoribosyl transferase gene; MLA = mouse lymphoma assay; OECD = Organisation for Economic Co-operation and Development; XPRT = xanthine guanine phosphoribosyl transferase

New assays for detecting genotoxic chemicals, primarily in vitro, have been developed that greatly increase throughput and/or sensitivity. These include quantitative high throughput screens (HTS) that evaluate the up-regulation of specific DNA damage response elements using reporter genes, (e.g., GADD45A [62], γ-H2AX [63], ATAD5 [64], TP53 [65]); the detection of multiple DNA damage response elements via targeted transcriptomic platforms [66–70] or multiple reporter cell lines [71]; using next-generation sequencing technologies to assess somatic mutations [72] and high throughput flow cytometric systems that integrate multiple endpoints relevant to genotoxicity and toxicity [73,74]. Data generated by HTS have been used for developing computational methods while others have not yet generated a sufficient chemical training dataset.

2.3. KC3: Alters DNA Repair or Causes Genomic Instability

Genomic instability, a common feature of cancer, refers to the increased tendency of a cell to acquire heritable alterations. This is typically associated with deficiencies in DNA replication and maintenance [75–78]. This instability is considered an enabling characteristic that drives tumor development and facilitates other cancer hallmarks [79]. To preserve genomic integrity, cells rely on a DNA damage response system that maintains DNA replication fidelity and activates cell cycle checkpoints and/or the DNA repair machinery in response to damaged DNA [76,80–82].

In a cell, a decrease in DNA repair competency and genomic stability is characterized by increased levels of mutations and chromosomal instability. Assays used to evaluate the ability of a chemical to induce DNA damage [9] (Table 1) can also be used to assess a chemical’s ability to reduce DNA repair activity [83]. Typically, cells are treated with a known DNA damaging agent (e.g., ionizing radiation, ultraviolet light, a direct acting genotoxic chemical) in the absence and presence of a chemical being tested for its ability to reduce a specific DNA repair pathway and one or more genotoxic endpoints measured. Other than for developing potential drug candidates for cancer therapy (e.g., 84), in silico methods to predict the ability of chemicals to inhibit DNA repair systems have not yet been developed.

2.4. KC4: Induces Epigenetic Alterations

Epigenetics refers to stable changes in gene expression and chromatin organization that occur without intrinsic changes in the primary DNA sequence. These changes can be mitotically inherited over cell divisions [85]. The epigenetic code is characterized by plasticity, with large variability across different cell types, in the same cells at different developmental stages, and under the influence of environmental stimuli [86]. Deregulation of the epigenome has been shown to occur during carcinogenesis and some carcinogens may disrupt various epigenetic mechanisms, including changes in DNA methylation, microRNA expression, histone modifications, and chromatin structure (e.g., 49,85–91). These can lead to altered gene expression and, among other effects, to tumor promotion [86,92]. Additional information about the chemical-induced epigenetic changes that can occur are provided by Chappell et al. [93], Herceg et al. [86], and Wang and Yang [94].

Changes in epigenetic patterns can be detected using standard molecular biology techniques and whole-genome approaches [24,95–98]. Alternatively, alterations in the protein levels and activity of epigenetic machinery can be measured. Appropriate in vivo models to study the impact of xenobiotics on the epigenome are lacking because of the complex nature of the epigenome among species, developmental stage, and across different tissues [86]. The OECD considers epigenetic mechanisms as crucial contributing information for an integrated approach to the testing and assessment of non-genotoxic carcinogens [97,99,100]. However, in silico methods for assessing this KC are limited. Epigenetic data are explicitly accounted for in the development of structural alerts for carcinogenicity [101,102], while computational approaches have been developed for identifying potential drugs and drug targets linked to mis-regulation of epigenetic mechanisms leading to disease [103–105]. A roadmap for investigating epigenome dysregulation and environmental origins of cancer includes strengthening causal relationships, finding consensus on an optimal approach for the analysis of population-based epigenome-wide studies, deciphering the complex exposure-to-phenotype pattern affected by early-life exposures, and understanding aging-associated epigenetic changes [89].

2.5. KC5: Induces Oxidative Stress

Oxidative stress is an imbalance between the generation and detoxification of reactive oxygen and/or nitrogen species (ROS/RNS). Sources of increased ROS formation include tissue inflammation, xenobiotic metabolism, interruption of mitochondrial oxidative phosphorylation, reduced turnover of oxidized cellular components, and an overwhelmed capacity of detoxifying enzymes [106]. Oxidative stress commonly occurs in neoplastic tissue and can be part of the tumor environment [107]. Major biomarkers and endpoints related to the formation of ROS and their effects include (1) cellular levels of oxidative stress/ROS, (2) oxidative changes to macromolecules (DNA, proteins, lipids), and (3) activation of the antioxidant response element (ARE) pathway and expression of downstream proteins and antioxidants [24]. Oxidative damage can alter the accuracy of DNA replication, integrity, and epigenetic features, inducing DNA replication stress and subsequent genome and epigenome instability [108]. The complexity of the carcinogenic process is demonstrated by the current understanding of the role that Nrf2 plays in oxidative damage and carcinogenesis (reviewed by He et al. [109] and Rojo de la Vega et al. [110]). This transcription factor is regarded as key to the cellular antioxidant response. However, Nrf2 activation in cancer cells promotes cancer progression and metastasis while conferring resistance to chemo- and radiotherapy. Thus, while activation in normal cells helps to prevent tumor initiation, prolonged or constitutive activation participates in tumor promotion and progression.

Several QSAR methods exist for the evaluation of the antioxidant properties of chemicals; however, no Model Reporting Format for one directly predicting oxidative stress, macromolecule oxidation, redox cycling, or ARE activation is listed in the Joint Research Centre QSAR database [35]. Published models are available that predict different aspects of oxidative stress, such as lipid peroxidation, redox-cycling, and ARE pathway activation [111–117]. Endpoints relevant to this KC, relevant assays, and the availability of corresponding in silico methods are detailed in Table 2. Some in silico models predict health outcomes based on ROS assays or oxidative stress as input parameters (e.g., insulin regulation [118], oxidative stress levels in workers [119], or phototoxicity [120]). Muller et al. [121] have developed a model for drug-induced liver injury, which, among other endpoints, uses oxidative stress as a descriptor. Most expert alert/rule-based systems recognize reactive structures like organic peroxides, hydroperoxides, or quinones, which are ROS or are known to release ROS. These alerts are then matched to endpoint damage like mutagenicity, micronucleus formation, or hepatotoxicity. While oxidative stress is not a direct predictor of cancer, it is considered a contributing factor in the prediction of carcinogenicity by in silico expert rule-based systems (e.g., Derek Nexus⁸).

Table 2:

List of endpoints and assays for the “Causes Oxidative Stress” (adapted from Smith et al.[24] with corresponding in silico models.

Endpoint	Assay	In Silico Methods
ROS formation	Redox-sensitive dyes [122–124] EPR [125] Genetically modified cells with GFP/YFP [124,126]	Expert rule-based systems that recognize for example, peroxides and quinones. Redox-cycling capability is dependent on the one-electron reduction potential of a compound. Quantum chemical models are available to predict the one-electron reduction potential from chemical structure [112,114].
Oxidative DNA damage	Comet assay modified with lesion-specific repair endonucleases [61,127,128] Bacterial reverse mutagenicity test with specific tester strains [50,129,130] GFP Reporter gene assays [71]	None located. DSB prediction [131] Some in silico models have training data for oxidation damage-sensitive tester strains (e.g., Leadscope Model Applier9, Lhasa Sarah Nexus10, MultiCASE11).
Oxidation of other macromolecules	Protein oxidation [123,132] Adductomics methods Lipid peroxidation: MDA, TBARS assay [123,132,133]	None located. None located. Models by Sierra et al. [115]
ARE/Nrf-2 activation	ARE/Nrf-2 transactivation assays [134–137] Expression and activity of ARE-related enzymes and molecules (e.g., CAT, SOD, GPx) [138,139]	Prediction of ARE response [111,113,117]. Prediction of Nrf-2/ARE activation based on toxicogenomics datasets [116]. Quantum model of Nrf-2/ARE activation None located.
Glutathione oxidation	Measurement of GSH level, GSH/GSSG ratio [140]	None located.

ARE = antioxidant responsive element; CAT = catalase; DSB = double-strand break; EPR = electron paramagnetic resonance; GFP = green fluorescent protein; GPx = glutathione peroxidase; GSH = glutathione; GSSG = oxidized glutathione; MDA = malondialdehyde; Nrf-2 = nuclear factor erythroid 2–related factor 2; qHTS = quantitative high throughput screening; ROS = reactive oxygen species; SOD = superoxide dismutase; TBARS = thiobarbituric acid reactive substances; YFP = yellow fluorescent protein

2.6. KC6: Induces Chronic Inflammation

Inflammation is an adaptive response of the body to a loss of tissue homeostasis triggered by internal or external stimuli, including non-infectious and infectious agents [141,142]. Chronic inflammation results in an imbalance of tissue homeostasis due to prolonged release of inflammation mediators and persistent recruitment and activation of inflammatory cells [87,143]. This can predispose the host to various disease conditions including cancer [144–148]. The perpetuated inflammatory conditions can affect all stages of carcinogenesis [145,146,149] by increasing the risk of cancer, promoting tumor progression, and supporting metastatic spread [79,150]. Poor sensitivity and complex aspects related to exposures and interindividual variability make it difficult to identify diagnostic biomarkers of chronic inflammation that lead to cancer [151,152]. Furthermore, inflammation that elicits cancer and cancer-induced inflammation are not easily distinguished [153].

To our knowledge, computational toxicology has not been applied to predict either tissue inflammation or inflammatory signaling as detected by biomarkers or in vitro phenotypic profiling. Phenotypic in vitro assays such as those used in the ToxCast and Tox21 programs detect the ability of chemicals to up- or down-regulate key pathways associated with chronic inflammation [154]. These assays provide reasonable training sets, although the short-term nature of these assays poses a limitation for the evaluation of the prolonged inflammatory responses [24]. Specifically, NF-κB transcription factors have been the subject of several computational studies focusing on their inhibition; however, further investigation is needed to assess their predictive accuracy for chronic inflammation [155]. The mechanisms underlying chronic inflammation are complex and there is a limited understanding of the factors that induce this state or promote its prolongation in a spatiotemporal framework [156]. A high throughput in vitro ROS assay coupled with structure-activity relationship analysis of the known carcinogens might be a good starting point for in silico predictive methods for supporting the development of an inflammation-mediated adverse outcome pathway (AOP). Datasets including inflammatory responses based on biomarkers or measurements of tissue inflammation in humans may also offer a possible means for developing in silico models for chronic inflammation. However, the use of biomarkers is hampered by the difficulty in establishing a causal relationship between inflammatory biomarkers and cancer development [24].

2.7. KC7: Is Immunosuppressive

Immunosuppression is a reduction in the capacity of the immune system to respond effectively to foreign antigens, including antigens on tumor cells, and is associated in humans with an increased risk of certain tumors. Mechanisms have been proposed to explain the inter-relationship between the immune system and cancer [157,158]: these include tumor immunoediting [159], oncogenic viruses (e.g., HIV [160]), chronic inflammation [161], and chronic B cell stimulation [162]. Neoplasm types associated with drug-induced immunosuppression include a 20-fold increase in lymphoma, non-melanoma skin cancer, and Kaposi sarcoma seen in renal transplant patients [163]. Immunosuppression may not directly transform normal cells into potential tumor cells but rather potentiate neoplasia by facilitating the survival and proliferation of pre-neoplastic cells [157,163]. Furthermore, immune suppression of the adaptive immune system can lead to activation of the innate immune system to cause inflammation [164], another KC of carcinogens.

Experimental methods for assessing immunosuppression that focus on the ability of chemicals to interfere with various immune pathways are described by the FDA [165]; this draft document states that the standard 2-year carcinogenicity studies are not specifically designed to detect carcinogenicity caused by drug-induced decreases in tumor surveillance particularly when the increased tumor risk is caused by recrudescence of latent viral oncogenes, infectious agents, or chronic inflammatory states. However, as reviewed by Alden et al. [166], the p53+/− and the AKR mouse strains harbor a lymphoma virus and viral oncogene, respectively, and both develop lymphomas upon chronic treatment with immunosuppressive drugs suggesting these strains might be useful for assessing the cancer risk from immunosuppressive drugs. The diversity and mechanisms of action (MOA) for immunosuppressive reagents make both measurement and in silico predictions of this endpoint challenging.

2.8. KC8: Modulates Receptor-Mediated Effects

Receptor-mediated effects can involve alterations in receptor expression and activation or altered levels of receptor ligands or cofactors [167]. Receptors are broadly divided into two categories: intracellular receptors (e.g., nuclear receptors) and cell-surface receptors (e.g., ligand-gated ion channels, G-protein coupled receptors (GPCRs), receptor tyrosine kinases (RTKs)) [168]. Many human carcinogens display receptor-mediated activity [49,167,169]. While many receptors can play a role in carcinogenesis (e.g., activation of PI3/AKT signaling through GPCRs and RTKs [170]), the focus here is primarily on nuclear receptors as these [171] and their co-regulators [172] are critically involved in normal physiological processes and therefore, serve as important targets in carcinogenesis. Hormone-dependent cancers such as estrogen-dependent breast [173] and endometrial [174] cancers and androgen-dependent prostate cancer [175] represent the most well-defined examples of nuclear receptor involvement in cancer.

Many chemicals that have endocrine disrupting potential (e.g., phthalates, bisphenol A, polychlorinated biphenyls) display carcinogenic potential as well [176–178]. The molecular initiating event associated with receptor-based effects is binding of ligand to receptor. However, receptor binding alone is insufficient to determine downstream effects as the nature of such interactions (i.e., agonism versus antagonism) determines subsequent gene and protein expression alterations. An example of an AOP leading from estrogen receptor (ER) activation to breast cancer is presented in Morgan et al. [179].

There are numerous nuclear receptor assay methods in the literature. Methods that directly or indirectly measure receptor modulation are reviewed by Smith et al. [24]. However, these assays do not directly indicate carcinogenic potential and virtually all of the in vitro assays lack metabolizing systems which significantly limits their sensitivity and relevance. Open source computational models for ER [180] and androgen receptor (AR) [181] modulation that integrate results from multiple high throughput in vitro assays are bridging the gap between in vitro and in silico methods. Computational models that predict ER activity include read-across, molecular docking, and virtual screening tools are available for in silico hazard identification [182–187]. However, generating the necessary carcinogenicity data can be challenging due to knowledge gaps regarding the tissue-specificity of many receptor-mediated effects as well as uncertainty about which receptor-mediated effects contribute to carcinogenesis. Empirical data are required to identify the receptor-mediated events and the corresponding tissues and cancer types for which receptor activity is the critical initiating event in carcinogenesis, rather than being a downstream indicator of disease progression. Research efforts should be directed toward identifying the inflection point between normal receptor function and inappropriate receptor activity that triggers disease. Of interest is the observation that the conclusion of a recent evaluation of non-genotoxic agrochemical rodent carcinogens was that the most common MOA involved nuclear receptor activation [188]. Another concern is that some receptors (e.g., aryl hydrocarbon receptor (AhR), peroxisome proliferator-activated receptor (PPARα)) have a critical role in rodent liver carcinogenesis through non-genotoxic mechanisms, but with questionable relevance to humans [189]. These and other species- and sex-specific tumors highlight the need to better understand the specificity of receptor involvement in human carcinogenesis.

2.9. KC9: Causes Immortalization

Escaping cellular senescence and becoming immortal is a prerequisite step in carcinogenesis [190]. However, while immortalization per se is not sufficient for tumorigenesis, it may allow for the accumulation of the additional genetic changes needed for malignancy [191,192]. Senescence is regulated by telomere shortening and some carcinogens have been shown to activate the enzyme telomerase, thus preventing their shortening and/or increasing their length [193]. Several other mechanisms involving altered signaling pathways (e.g., p53/p21 and p16/Rb pathways) can influence senescence and immortalization [191,192]. Genotoxicants are thought to make cells immortal through direct inactivation (via point mutations and deletions) of the effector pathways (e.g., direct inactivation of the tumor suppressors p53 and p16 [194], while epigenetic mechanisms have been linked to immortalization induced by some non-genotoxic carcinogens [191].

Table 3 provides a list of endpoints and assays, adapted from Smith et al. [24], as well as corresponding in silico methods. Data for the Syrian hamster embryo (SHE) in vitro cell transformation assays are collected in the OECD QSAR toolbox¹² [185]. However, the molecular mechanism governing the processes in cell transformation are not fully understood and therefore are not used for making regulatory decisions [195–197]. No other in silico methods currently exist for predicting the ability of chemicals to induce immortalization.

Table 3.

List of endpoints and assays with corresponding in silico models for the “Causes immortalization” KC.

Endpoint	Assay	Type of assay	In Silico
Changes of in vitro transformation activity	Cell transformation assays	Syrian hamster embryo (SHE) assay [198–201]	QSAR models for SHE cell transformation from public literature [5,6,202,203] and Instem13.
Alterations of cellular senescence markers	Changes in β-galactosidase, p16, p21, and p53 protein levels	Non-human and human in vitro assays and in vivo assays and biomarkers [204]	-
Alterations of telomere length and/or telomerase activity	Telomerase activity assay	Non-human and human in vitro assay and in vivo assays and biomarkers [205]	-
	Telomere length by real-time PCR	In vivo assays and biomarkers [206–209]	-
Alterations in stem cell genes	Expression of Myc, Oct3/4, Klf4, Sox-2	Non-human and human in vitro assay [210,211] and in vivo assays and biomarkers [210,212]	-

SHE = Syrian hamster embryo

2.10. KC10: Alters Cell Proliferation, Cell Death, or Nutrient Supply

Tumor size is determined by cell proliferation, cell loss via apoptosis or necrosis, and the vascular supply. Cancer cells often exhibit unusual cellular energetics, relying on glycolysis for energy even under aerobic conditions [213] and converting glucose to lactate. Alterations in cellular replication and/or cell-cycle control linked to carcinogenesis have been described as predisposing to unrepaired DNA damage leading to cancer-causing mutations in replicating cells, sustained replication, and the ability of a transformed cell to escape normal cell-cycle control. A component common to these is evasion of apoptosis or other terminal programming [214]. Factors involved in cell proliferation and cell death include inflammatory signals [215–217], apoptosis and autophagy [218] and angiogenesis [49]. Dysregulation of gap junction intercellular communication (GJIC) can cause loss of contact inhibition, leading to abnormal cell growth [87]. GJIC is dysregulated or absent in a majority of human cancer cells [219–221]. GJIC is also involved in metastasis as reduced connexin expression may promote physical detachment of the cells from their substrate [222,223]. Nongenotoxic mechanisms may be responsible for GJIC inhibition [224,225], including inhibition by tumor promoters [226–230].

Methods for assessing cell proliferation, cell death/apoptosis, cell viability, and angiogenesis in vivo or in vitro as well as methods for assessing cellular bioenergetics and metabolism are presented in Table 4. Other than for potential drug candidates, global in silico methods to assess the ability of chemicals to induce cell death, cell proliferation, and alteration of nutrient supply have not been developed. However, local QSARs (i.e., those dedicated to a specific group or class of chemicals) have been developed to: (1) facilitate the discovery of novel inhibitors of enzymes and signaling molecules associated with cellular processes, whose aberration might lead to cancer (e.g., 231–237); (2) predict the antiproliferative activity of specific classes of candidate anticancer agents [238,239]; and (3) predict the ability of structurally-related compounds such as perfluorinated fatty acids, dialkyl phthalates, polycyclic aromatic hydrocarbons, and polychlorinated biphenyls, to alter GJIC function [102,240–242].

Table 4:

List of endpoints and assays for the KC “Alters Cell Proliferation, Cell Death, or Nutrient Supply”. For these endpoints, corresponding in silico models have not been identified.

Endpoint	In Vitro and In Vivo
Cell proliferation	DNA content: 5-bromo-2’-deoxyuridine (BrdU) nucleoside labelling in S phase cells by (ELISA), flow cytometry, immunofluorescence, and high content imaging [243]. In situ measurements of proliferating proteins (i.e., proliferating cell nuclear antigen (PCNA), Ki67, and Phospho-histone H3) by western blot (WB), immune fluorescence (IF), immunohistochemistry (IHC), flow cytometry, and ELISA [243].
Cell viability	Dye staining: DNA intercalating dye PI -propidium iodide as an indirect measure of the health of the population and number of cells in flow cytometry applications [244]. Metabolic assays: reduction of substrates (formazan dyes) by mitochondrial enzymes (e.g., MTT) mainly used in vitro. Cell death assays to assess the proportion of dead cells, hence necrosis, including: Loss of membrane integrity or indirect measure of cell viability; live/dead counting (e.g., intracellular and secreted markers such as High Mobility Group B1, HMGB1; lactate dehydrogenase, LDH; interleukin, IL-1β); T-cell killing assays (release of pro-inflammatory cytokines IL-1β and IL-18, by flow cytometry or live cell imaging) [245].
Apoptosis	Annexin V assay, to measure membrane asymmetry or the irregular distribution of different lipid species on either side of the plasma membrane; Caspase-3 cleavage and activity; Chromatin condensation; TUNEL Assay which, measures DNA strand breaks and DNA fragmentation; Cytochrome c release assay to measure translocation from the mitochondria to the cytosol triggered by apoptotic stimuli. All of these methods can be assessed using flow cytometry, immunofluorescence, ELISA, immunoblotting, or in situ IHC applications [243].
Nutrient supply, as metabolism	ATP measurement assays: to determine the number of viable, metabolically active cells in a sample, by colorimetric, fluorometric, or luminescent readout [246].
Cell oxygen consumption & bioenergetics	O2 consumption and mitochondrial bioenergetics to identify redox state, mainly in in vitro systems, including EPR oximetry; the Clark electrode; the MitoXpress fluorescent assay; the Seahorse XF24 Extracellular Flux Analyzer [247,248]. Alterations in signaling transduction pathways (e.g., Wnt/β-catenin, Notch, Raf, TGFβ, IGF, PI3K/Akt/mTOR, NF-kB, Hedgehog, ERα-/ERβ-/AR); Expression of senescence markers (i.e., expression of pH-dependent β-galactosidase activity) [249–251].
Angiogenesis	Angiogenesis markers (e.g., VEGF) to investigate tumor development either in in vitro or in vivo systems. Gene expression profiles of short-term in vitro models [68,154] and in short-term in vivo models [252].
Gap junction intercellular communication (GJIC)	Metabolic cooperation, electrocoupling, dye injection, scrap loading/dye transfer, fluorescence recovery after photobleaching, electric cell-substrate impedance sensing, and expression of gap junction genes (e.g., connexin 32) can be used to measure GJIC [102,224,253,254]. In vitro [87] and in vivo [255–257] models are available.

Akt = “Ak” refers to the AKR mouse strain while the t stands for thymoma, ATP = adenosine triphosphate, AR = androgen receptor, ELISA = enzyme-linked immunosorbent assay, EPR = electron paramagnetic resonance, ER = estrogen receptor, IGF – insulin growth factor, mTOR = mechanistic target of rapamycin, NF-kB = nuclear factor kappa-light-chain-enhancer of activated B cells, Raf = Rapidly Accelerated Fibrosarcoma, TGF = transforming growth factor, TUNEL = terminal deoxynucleotidyl transferase dUTP nick end labeling, VEGF = Vascular endothelial growth factor, Wnt = a blending of Wingless and Int-1.

3. Discussion and Conclusion

Anticipating the carcinogenic potential of small molecules remains a major challenge that poses a serious threat to the population when unintentionally hazardous drugs or impurities reach the market [258]. Furthermore, humans are exposed to thousands of chemicals that have inadequate data on which to predict their potential for toxicological effects, including cancer [259]. The KC construct provides a pragmatic platform for collecting data and organizing information regarding carcinogens and it has recently been used in combination with read-across principles to support a recommendation to include chemicals in the NTP Report on Carcinogens that did not have animal cancer data [260]. Such structured information may help to support the development of in silico models possibly using AOP concepts. Here, we indicate where experimental methods and in silico models currently exist for each KC to support the hazard assessment for carcinogenicity and, importantly, where gaps exist. The “is genotoxic” KC is the only characteristic where global in silico methods for predicting carcinogenicity hazard have been developed that are being used for making regulatory decisions [37,38]. In silico predictions of electrophilicity exist also but are not accepted for making regulatory decisions alone due to low specificity. Mechanistic challenges as well as the absence of relevant methods for measuring certain KCs and/or the lack of robust experimental data, along with their complex causal relationship to carcinogenesis, make the other nine KCs less tractable for the development and application of in silico prediction models. This is especially true for non-genotoxic carcinogens where achieving a dose that exceeds homeostasis and exposure duration are important elements. Many of the KCs lack specificity for carcinogenicity by being involved in non-carcinogenicity-related disease processes as well. A critically important requirement for their regulatory use is a specific effect related directly to tumor development. However, as proposed by Madia et al [261], integration of relevant KC data derived from other toxicological studies into carcinogen hazard assessments is useful as it will maximize exploitation of existing data while minimizing the need to develop additional resources to evaluate non-genotoxic mechanisms.

There is widespread industry and regulatory interest in developing a framework for the prediction of human carcinogenicity using a combination of available experimental in vitro and in vivo data and inputs from in silico models. Such frameworks, termed “In Silico Toxicology Protocols,” [8] have been successfully developed for genetic toxicology [9] and skin sensitization [10] following a pattern that is generally applicable to any toxicological endpoint. For carcinogenicity, data would be integrated across the 10 KCs to give an overall hazard assessment. Each KC would be evaluated using available experimental data and the results of statistical models, expert alerts, and read-across evaluation to make an assessment in terms of hazard and the associated reliability of hazard determination. Once all 10 KCs are evaluated to the extent possible, the data are combined into an overall assessment of carcinogenicity and an associated confidence level (low, medium, high) assigned.

Integration of the 10 KCs into an overall hazard assessment is complex and not simply additive. The majority of KCs are not truly independent but rather interact with each other in a complex network that differs depending on the species, sex, individual, tissue, and cancer type, as well as with other factors related to environment and personal life-styles. For example, although electrophilicity is a common attribute of genotoxicants, electrophilic targets also include cellular constituents other than DNA that might result in altered homeostasis, increased genomic instability, and/or altered gene expression.

The contributions of the different KCs to carcinogenesis are not equal. Rather, the 10 KCs are meant to comprehensively account for the different mechanisms involved in tumor initiation, promotion, and progression, regardless of prevalence. Krewski et al. [49] analyzed the key characteristics of 86 agents known to cause cancer in humans for which mechanistic information was available. The most prevalent key characteristic was “is genotoxic” with 85 of the 86 known human carcinogens demonstrating that activity, followed by 47 agents with “alters cell proliferation, cell death, or nutrient supply” and 40 agents with “induces oxidative stress”. The 86 human carcinogens were associated with one to nine of the 10 KCs, with an average of four KCs. However, it is likely that the agents associated with very few KCs and especially a single KC (e.g., thiotepa with KC2 only) reflect the lack of relevant studies rather than a unique profile. Illustratively, we used these data published by Krewski and colleagues [49] to generate Figure 1 This figure provides a model that places the 10 KCs in perspective with the three stages of carcinogenesis and with each other. The KC boxes are color-coded based on the percentage of the 86 human carcinogens with data supporting the involvement of that KC [49]; these percentages might not reflect reality but are currently the best available. The arrows indicate likely interactions at the molecular and cellular levels and effects. Given the lack of experimental data supporting some of these interactions, we consider the relationships to be biologically plausible but not necessarily demonstrated experimentally. However, this figure aids in demonstrating the complexity of the interactions between the various KCs and the 3 stages of carcinogenesis. We also hope that by indicating where likely intersections occur, it will prompt the development of useful AOPs perhaps based on those intersections. In this figure, interactions between the various KCs are not specified but rather from each KC to its primary molecular or cellular level events/effects and from there to one or more downstream events and/or to one or more of the three stages of carcinogenesis.

Figure 1. Interactions among the 10 KCs and with the Three Stages of Carcinogenesis.

KC1 = is electrophilic, KC2 = is genotoxic, KC3 = alters DNA repair or causes genomic instability, KC4 = induces epigenetic alterations, KC5 = induces oxidative stress, KC6 = induces chronic inflammation, KC7 = is immunosuppressive, KC8 = modulates receptor-mediated effects, KC9 = causes immortalization, KC10 = alters cell proliferation, cell death, or nutrient supply. Supporting information for the different arrows can be found in the text or are considered by the working group to be biologically plausible. KCs are color coded according to the association between KCs and 86 chemicals carcinogenic to humans (International Agency for Research on Cancer Group 1 Agents) as determined by Krewski et al. [49].

Related to Figure 1 is Table 5, which summarizes the in silico methods currently available for each KC and the extent to which these methods are currently used for making regulatory cancer-related decisions. The table also summarizes the likely relationships among the different KCs in terms of the initial (i.e., first level) common molecular or cellular level event/effect each KC affects.

Table 5:

Relationships between the different KCs, their contribution to one or more stages of carcinogenesis, and availability of global in silico methods and whether used for making regulatory decisions related to cancer.

Key Characteristic (KC)	First level, common molecular or cellular level event/effect among KCs	In Silico Methods
		Availability	Used for making cancer-related regulatory decisions
KC1: Is Electrophilic or Can Be Metabolically Activated	KC2, KC3, KC8	Well-established for adduct formation, but otherwise limited	Yes, but limited to DNA reactivity
KC2: Is Genotoxic	KC1	Well-established	Yes
KC3: Alters DNA Repair Or Causes Genomic Instability	KC1	None	No
KC4: Induces Epigenetic Alterations		Limited	No
KC5: Induces Oxidative Stress	KC6, KC7	Not for cancer but for assay specific endpoints	No
KC6: Induces Chronic Inflammation	KC5, KC7	Not for cancer but could be developed for assay specific endpoints	No
KC7: Is Immunosuppressive	KC5, KC6	None	No
KC8: Modulates Receptor-Mediated Effects	KC1	Yes, for multiple nuclear receptors1	No
KC9: Causes Immortalization		Only for in vitro Syrian hamster embryo cell transformation	No
KC10: Alters cell proliferation, cell death, or nutrient supply		None	No

First level refers to arrows from two or more KCs that point to the same first molecular or cellular level event/effect in Figure 1. As noted in Figure 1, KCs 1–8 contribute to all three stages of carcinogenesis (initiation, promotion, progression), while KCs 9 and 10 contribute to promotion and progression.

The 99% prevalence of the “Is Genotoxic” KC among the known human carcinogens supports a regulatory decision tree approach for carcinogenicity hazard assessment that stops at in vitro or rodent genotoxicity if positive, with the other KCs evaluated only if the genotoxicity results are negative or inconclusive [18]. However, the more universal approach presented here would, once the data gaps have been filled, potentially provide for a better mechanistic understanding of carcinogenicity hazard as well as human relevance. Such data would also be useful for risk assessment, especially in situations where multiple chemicals or mixtures are being considered. A firmer grasp of the relationships between chemical structure, relative potency, and mechanisms of carcinogenicity would help to facilitate rational molecular design strategies during chemical development and strengthen risk-based decision-making when identifying lead compounds [262,263]. There is also tremendous value in having effective in silico carcinogenicity models which can provide rapid and inexpensive assessments to support the development of new drugs and treatments.

As noted earlier, global and regulatory accepted in silico models for the prediction of 9 of the 10 KCs do not currently exist. This is due in part because the development of relevant, sensitive, and specific assays depends on a better understanding of the biological processes involved in most of the KCs as well as because experimental data for these specific KCs are generally not required by regulatory agencies. Therefore, robust datasets on which to develop useful in silico models are generally not available. Chiu et al. [154] determined that the high throughput screening (HTS) assays conducted within the U.S. Environmental Protection Agency’s ToxCast program included assays that were relevant to all but three KCs (3, 7, and 9). However, the extent of coverage within the seven KCs with HTS assays varied widely. Subsequently, Iyer et al. [264] identified and mapped 137 ToxCast cancer pathway-related assays to 5 of the 10 KCs, concluding that the biological coverage of the ToxCast assays related to cancer pathways was currently limited. These two publications indicate where the development of additional HTS assays would be useful in terms of the KCs and also where chemical rich datasets already exist for the development of in silico models. Currently, a major limitation of these HTS assays is the general lack of metabolic competency, meaning that test results are limited to the parent compound and any known metabolites that also might have been tested.

The development of approaches integrating in silico predictions with in vitro and/or in vivo experimental data could serve as a more mechanistic, hypothesis-driven and tailored carcinogenicity assessment strategy in the future. Furthermore, the concept of a predictive computational model built upon pathways of toxicity (i.e., a quantitative AOP (qAOP)) is gaining interest due to its ability to address the question of how much perturbation in any of the upstream key events and under what conditions will adverse outcomes likely occur [265]. There is an increased interest in global genomics methods that evaluate changes in transcriptomics and/or proteomics in tissues and cells when exposed to carcinogens. The goal is to identify responding pathways and the point at which cellular homeostasis is exceeded (i.e., the tipping point), potentially leading to an increased risk for tumor development [266,267]. A network perspective on cancer encourages the application of computational modeling approaches [268], which aid the integration of multiple assays’ results to generate either qualitative or quantitative outcomes. Methodologies for computational analysis can vary widely depending on the question being posed and the available experimental data, ranging from highly abstracted models using correlative regression to networks and logic modeling techniques and highly specific models using differential equations [268]. It is important to ensure that the developed computational models are fit for their intended purpose, as determined by problem formulation of the question to be addressed or the regulatory use under consideration [269].

Currently, there are no integrated approaches to testing and assessment (IATAs) for the assessment of carcinogenicity. However, an OECD working group has addressed the possibility of an IATA for the detection of non-genotoxic carcinogens [97,99,270,271]. In this context, the ‘Hallmarks of Cancer’ [79] and key characteristics defined by Smith et al. [11] were further analyzed with the goal to elaborate grouping format/overarching KCs to address specifically non-genotoxic carcinogens. Such IATAs could also serve as a framework to integrate the various existing in silico methods for the evaluation of absorption, distribution, metabolism, excretion (ADME) and toxicokinetics [272–275] into carcinogenicity hazard assessment. The assays useful for evaluating the ability of chemicals to induce a specific KC can also be integrated (and weighted) with historical pathology data (e.g., pre-neoplastic lesions [276–278]) and/or human molecular epidemiological data. These may provide evidence that reinforces biological plausibility [279] and therefore establish more meaningful relationships across biology, exposures, susceptibility, and disease that significantly contribute to an understanding of the underlying pathways of carcinogenicity [85]. Although it remains a significant challenge to integrate all of the modelled and observational data into a single in silico carcinogenicity protocol, developing such a protocol would contribute to a more complete mechanistic understanding of cancer and its development, significantly advancing other areas of research and medicine.

This position paper summarizes the current availability of in silico methods for each of the 10 KCs of carcinogenicity as well as indicating where experimental methods need to be implemented and robust databases generated to support the development of relevant in silico methods for underrepresented KCs. Given our increased scientific understanding of processes related to carcinogenicity and other diseases that occur in response to some or many of the same key characteristics, as well as the continued development of better assays and computational methods, we can expect incremental advances in our ability to predict the carcinogenic hazard of compounds.

Highlights.

• Summarizes the 10 key characteristics (KCs) of carcinogens

• Assesses how current in silico methods address each of the KCs

• Indicates where experimental methods need to be implemented and robust databases generated

• Highlights interactions among the KCs with the different stages of carcinogenesis