Predictive toxicology today: the transition from biological knowledge to practicable models

KEYWORDS:

• Toxicology
• alternative
• predictive method
• systemic toxicity
• high throughput
• ToxCast
• skin sensitization
• QSAR

1. Introduction

For a long time, animal toxicology has been the major source of information on chemical risk assessment. The study of the effects of chemicals on apical toxicity end points has provided evidence on potential toxic effects, as well as important clues on mechanisms of action (e.g. the patterns of affected target organs). The apical toxicity end points are the directly measured, whole-organism outcomes of exposure in in vivo tests, generally death, reproductive failure, developmental dysfunction, or cancer.

However, the need for totally or partially replacing the animal toxicity assays with shorter-term, animal-free toxicological methods has become more and more compelling in the recent decades. In the field of drug design, one primary need is the early recognition of potentially toxic molecules. In fact, attrition due to nonclinical safety represents a major issue for the productivity of pharmaceutical research and development [1]. In other fields, new regulatory policies (e.g. registration, evaluation, authorization and restriction of chemical substances and cosmetics) tend to drastically reduce the use of animals. As a consequence, there are considerable opportunities to accept ‘alternative’ approaches, both in silico (e.g. quantitative structure–activity relationships [QSAR]) and in vitro. This research is extremely active and has generated a large variety of approaches and proposals [2]. Particularly relevant are large-scale funded programs, like the European Union projects safety evaluation ultimately replacing animal testing and EU-ToxRisk [3] and the US Environmental Protection Agency ToxCast project [4].

Emerging trends in technologies include new in vitro screening assays, transcriptomics, stem cells, engineered microscale physiological systems, 3D organotypic culture models, and small model organisms (such as zebra fish and Caenorhabditis elegans). A great emphasis is being given to in-depth explorations of the mechanisms of toxicological action: the rationale is to identify the key events in the toxicity pathways and then devise new in vitro tests that could specifically measure those events. The ToxCast/Tox21 project [4] and the Adverse Outcome Pathway (AOP) concept and related activities [2,5] are examples of this new impetus. Another challenge is to develop tools for integrating multiple types of data from diverse experimental systems into unified risk assessment paradigms.

1.1. A central issue

A central issue is to what extent the proposed alternative approaches are able to provide correct predictions of the apical end points? The apical end points constitute the backbone of the present regulations, so novel, alternative approaches must prove to be able to replace satisfactorily the traditional assays or even to be superior. This is also a scientific issue. Since biology is an experimental science, any proposed predictive approach can be considered as scientifically sound only if it provides correct predictions of the toxicity of chemicals (recognized as such either in humans or in animal experimentation).

Given the vastity of the field, this Editorial will not attempt to review it in detail but will use and put into perspective a limited number of (recent and less recent) experiences with alternative approaches. The selection is obviously subjective. The aim is trying to derive take-home lessons of general relevance, with a special emphasis on the critical aspect of how to translate biological knowledge or hypotheses into operational, predictive models. This aspect is often overlooked, but it is of the outmost importance for creating a new predictive toxicology [6].

2. Case studies

2.1. Case study 1

A first case study is the research on skin sensitization. The skin sensitization ability of chemicals can be tested with established animal assays [7]. Since the chemical and biological pathways involved are relatively well characterized, the search for alternative methods is aimed at mimicking the main events of the hypothesized toxicological mechanism with in vitro assays. Several analyses have been published on how to best combine the in vitro assays within an optimized strategy. The overall predictivity of combinations of in vitro assays is quite good (80–90%) in respect to both animal and human skin sensitization results [7]. Remarkably, it has been shown that an in vitro assay (Direct Peptide Reactivity Assay [DPRA]) – mimicking the initiating event (i.e. reaction with skin proteins and haptenation) – alone is contributing with around 80% of predictivity. The combination of DPRA with other in vitro tests (e.g. hClat and KeratinoSens) that model intermediate key events only improves the predictivity of another 5–10% [8].

The above evidence suggests that the interaction of the chemical with the proteins (initiating event) is the rate-limiting step that dictates the overall sequence of events, and it is the basis for good predictive models. A complementary/alternative explanation is that mimicking intermediate key events with in vitro tests is experimentally more challenging than mimicking the initiating events.

2.2. Case study 2

Another case study is an alternative approach for predicting chemical carcinogenicity. Historically, the genotoxicity short-term tests have taken the pivotal role in the practice of prescreening of carcinogenicity: usually, in vitro tests (bacteria and/or mammalian cells based) are employed first, whereas in vivo assays (e.g. micronucleus) are used to confirm in vitro positives. However, there is evidence that this strategy is not sensitive enough to detect all genotoxic carcinogens and – obviously – cannot detect nongenotoxic carcinogens. To overcome the limitations of using only mutagenicity assays, a tiered strategy consisting of the in vitro Ames and Syrian hamster embryo (SHE) cell transformation assays, combined with structure–activity relationships, has been proposed. The Ames test identifies DNA-reactive carcinogens, that can also be identified through appropriate structural alerts (SAs). The chemicals negative in Tier 1 are supposed to be devoid of DNA reactivity and are studied for their ability to induce cancer through epigenetic/non-DNA-reactive mechanisms. Since the cell transformation assays are sensitive to both genotoxic and nongenotoxic carcinogens, SHE is used in Tier 2 together with SAs specific for nongenotoxic mechanisms. The tiered approach resulted able to identify both genotoxic and nongenotoxic carcinogens, with an estimated 90–95% sensitivity for rodent carcinogens. In addition, almost all International Agency for Research on Cancer human carcinogens (326/329, 99%) were correctly identified [8,9]. The above approach has been considered in a recent document of the Organization for the Economic Co-operation and Development as a prototype for building an integrated strategic testing system, that combines in vitro and in silico tools [10].

The take-home lesson from the two above case studies is that complex toxicological end points (skin sensitization and carcinogenicity) can be predicted with simplified approaches, that model the rate-limiting steps of the mechanisms of action.

2.3. Case study 3

Another important field of alternative methods is that of QSAR. This is a large area of approaches with different characteristics (from quantitative models using physical chemical parameters to more qualitative approaches based on the recognition of SAs). The underlying idea is that the properties of the molecules (including toxicity) derive from their chemical structure and that such structural features can be recognized and used to make predictions of activity [11]. One issue that has sometimes discouraged investigators from applying QSAR analysis has been the idea that a biological process is too complicated and involves too many steps to be successfully modeled. In spite of the expected difficulties, in the last 50 years thousands of successful QSAR analyses of biological activity have been generated. This includes, e.g., carcinogenicity induced by polycyclic aromatic hydrocarbons and by aromatic amines [12]. The QSAR equations obtained are concordant in showing that the difference between active and inactive compounds depends on a few factors/parameters that model the metabolic activation step, which results to be the rate-limiting step after which all the other, necessary steps (interaction with DNA or other targets, mutations, promotion, progression, etc.) follow in sequence [8,12]. Thus, the use of a limited number of parameters in the QSAR models permitted the prediction of a complex apical end point like carcinogenicity.

2.4. Case study 4

Another case study is an analysis of the ability of the in vitro high-throughPut (HTP) ToxCast assays to predict apical toxicological end points. ToxCast™ Phases I and II are testing a combined total of about 2000 chemicals, with around 900 assays [13]. The ToxCast HTPs are mostly related to phenomena like cell-to-cell interactions and signaling and not to, e.g., covalent interactions with DNA or proteins: in this sense, they largely code for putative intermediate events in toxicological pathways.

In a recent study, the ToxCast results were compared with those of a large series of established toxicological end point tests. The HTP assays did not correlate with in vitro and in vivo mutagenicity and with rodent carcinogenicity. The HTP assays showed relatively low correlations with acute, repeated dose, skin sensitization, and reproductive/developmental toxicity test results (average correlation coefficient = 0.36), with no clear specific patterns (i.e. clearly related to known toxicological pathways). Instead, the correlations with endocrine disruption end points were higher (average correlation coefficient = 0.50) and specifically related to estrogen/androgen receptors, thus pointing to mechanistically based relationships [8]. The limited predictive ability of ToxCast assays for the apical end points has been pointed out in a previous, independent study as well [14].

Several technical difficulties of the ToxCast assays (e.g. in vitro metabolic activation) are yet to be solved, and improvements are in progress [15]. Thus, at this stage, only the overall pattern of results – and not the details – is important. These seem to agree with indications from the skin sensitization analysis, in particular the evidence that intermediate key events are difficult to be mimicked, and/or their models (in vitro tests) give a relatively minor contribution to the final predictivity. It should be emphasized that the in vitro tests are conceived/used in isolation, whereas in vivo mechanisms are a continuum, with feedbacks. In such conditions, in vitro assays for intermediate events may be a poor representation of the reality. Another factor to be considered is that some systemic toxicity end points (e.g. repeated dose and reproductive toxicity) include disparate effects and are not mechanistically well defined as, e.g., carcinogenicity or skin sensitization; thus, it is more difficult to identify key events of general relevance.

3. Conclusions

In conclusion, the way to create a new and faster toxicology still seems long, but progress is apparent. Quantitative modeling of rate-limiting steps in skin sensitization and carcinogenicity predicts the majority of toxicants. Similarly, successful QSAR models exploit the quantification of only one or few rate-limiting steps. Most often, the rate-limiting steps are the initiating events of a toxicological pathway, while markers for intermediate events have a more limited correlation with most end points. It should be emphasized as well that the initiating events (e.g. DNA or protein reactivity) are – in principle – more prone to be modeled with QSAR methods. These successful stories from real-life examples seem in contrast with the current tendency to dissect more and more toxicity pathways based on the belief that the larger the number of key events considered, the higher the probability of creating efficient predictive models. As a matter of fact, the evidence discussed in this article agrees with the general experience of modeling: in different fields like ecology, systems biology, and macroeconomics, grossly simplified models capture important features of the behavior of incredibly complex interacting systems and permit successful predictions [6]

Since the types of toxicological pathways are very different in nature, it is not possible to apply a common simplified linear chain framework to their study. Empirical analysis of data is the only guide in the evolution of our understanding of how chemicals affect living systems. The formulation of biological hypotheses is aimed at identifying key events and devise ad hoc in vitro tests. However, the biological narration – even if a challenging step – is only a starting point. A further crucial passage is when the entire set of results is analyzed with rigorous data analysis methods, in order to (a) describe the type and nature of the pathway, (b) ascertain the presence of a solid correlation with the golden standards represented by the apical toxicological end points, and (c) build models that integrate the various types of evidence.

4. Expert opinion

The toxicology community has made significant progress developing assays and tools that will help achieve the predictive toxicology goals. Many efforts have been initiated toward developing new assays for toxicity testing and models to integrate large data sets into risk assessment frameworks. Emerging trends in technologies include new in vitro screening assays, transcriptomics, stem cells, engineered microscale physiological systems, 3D organotypic culture models, and small model organisms. Another large area of research is that of QSAR, that links the toxicological effects to chemical structures. Given the variety and diversity of the approaches and their work-in-progress character, it is clear that challenges and successes cannot be easily summarized. A common, important question is how in vitro data and in silico models can be used to understand and predict in vivo toxicity. This is a central issue because the apical toxicity end points constitute the requirements of the present regulations: novel approaches must prove to be a valid replacement of, or even to be superior to, traditional tests and have to generate reliable information for use in regulatory decisions. In addition, since biology is an experimental science, any proposed predictive approach can be considered as scientifically sound only if it provides correct predictions. A challenge is to develop tools for integrating multiple types of data from diverse experimental systems into unified risk assessment paradigms. In fact, the presently adopted in vivo regulatory assays are stand-alone systems, whereas it can be anticipated that an equivalent type of information will be provided by a range of different sources. This emphasizes the need of giving more attention to the techniques of quantitative data analysis and modeling. Fortunately, experience from different modeling fields (confirmed by a number of examples reported here) shows that successful predictive models usually are based on the quantification of only one or few rate-limiting steps. The challenge remains of the careful selection of these few crucial elements.

Declaration of interest

The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

Additional information

Funding

This paper was not funded.