(Q)SAR: A Tool for the Toxicologist

Abstract

A continuing education (CE) course at the 2014 American College of Toxicology annual meeting covered the topic of (Quantitative) Structure–Activity Relationships [(Q)SAR]. The (Q)SAR methodologies use predictive computer modeling based on predefined rules to describe the relationship between chemical structure and a chemical’s associated biological activity or statistical tools to find correlations between biologic activity and the molecular structure or properties of a compound. The (Q)SAR has applications in risk assessment, drug discovery, and regulatory decision making. Pressure within industry to reduce the cost of drug development and societal pressure for government regulatory agencies to produce more accurate and timely risk assessment of drugs and chemicals have necessitated the use of (Q)SAR. Producing a high-quality (Q)SAR model depends on many factors including the choice of statistical methods and descriptors, but first and foremost the quality of the data input into the model. Understanding how a (Q)SAR model is developed and applied is critical to the successful use of such a tool. The CE session covered the basic principles of (Q)SAR, practical applications of these computational models in toxicology, how regulatory agencies use and interpret (Q)SAR models, and potential pitfalls of using them.

Naomi Kruhlak from the Chemical Informatics Program with the US Food and Drug Administration discussed the basic principles of (Quantitative) Structure–Activity Relationships [(Q)SAR] including an introduction to the fundamentals of (Q)SAR model construction and application, model performance, chemical structure-based searching, and using (Q)SAR in drug development. The (Q)SAR uses computer-assisted learning to identify correlations between chemical structures and biological activity. It learns from actual laboratory testing data or clinical outcomes under the general assumption that similar molecules will exhibit similar physicochemical and biological properties. There are 2 basic types of models: quantitative and qualitative. Quantitative models, termed (Q)SARs, are most often statistically derived, whereas qualitative models, termed SARs, are frequently expert rule based. Collectively, they are referred to as “(Q)SAR.” Statistically derived models use algorithms such as partial least squares regression analysis , support vector machines, and k-nearest neighbors. These models can be built rapidly but vary in their interpretability. Examples of commercial software programs using statistically derived models include CASE Ultra, MC4PC, and Leadscope Model Applier. Expert rule-based models use human expert-derived correlations, which are often supported by mechanistic data and citations. These models are highly interpretable but are time consuming to build. Derek Nexus is an example of a commercial software program used in this method. ¹

Methods of validating a (Q)SAR model include noncross validation, which measures how well the model fits the data; cross-validation, which indicates the robustness of the model; y-scrambling, which provides a baseline performance of a random model; and external validation, which uses a new data set to test the model. In general, (Q)SAR models are capable of predicting most organic molecules under 1000 Da molecular weight. Polymers, inorganic molecules, and uncharacterized mixtures of organic molecules typically are unsuitable. In addition, most commercially available, global (Q)SAR models cannot differentiate stereochemical or geometric isomers. ²

Chemical structure-based searching can be used to supplement predictions obtained with a (Q)SAR model. There are 3 basic methods for searching a database using a chemical structure. These include exact structural match, substructural search, and whole molecule similarity. An exact match looks for a single molecular entity and is an unambiguous way to identify a compound, which may be stored under a multitude of chemical names or IDs. Substructural searching returns multiple hits containing a particular molecular feature, such as an epoxide. Whole molecule similarity considers all structural features in a queried compound and assigns a similarity index based on features in common with analogues.

(Quantitative) structure–activity relationships is used in drug development by both industry and regulatory agencies. Industry uses (Q)SAR in early toxicity screening to help identify and optimize lead compounds. In contrast, Food and Drug Administration (FDA) uses (Q)SAR models for later stage safety assessment when empirical data are limited or lacking. Overall benefits to using (Q)SAR models are that no synthesis of the test compound is required, results are rapid and inexpensive to obtain, and models can be tailored to meet the specific needs of the application with respect to predictive performance characteristics and the chemical space that they cover.

Next Samantha Gad, from GAD Consulting Services, presented on the practical applications of (Q)SAR in toxicology, focusing on areas in toxicology where (Q)SAR can be applied and a brief overview of some common (Q)SAR programs. In drug development, (Q)SAR can be used to not only predict the toxicological end points of a compound but also the potential metabolites and degradation products, and to predict absorption, distribution, metabolism, excretion, and toxicity characteristics, which can be helpful in optimizing lead candidate selection. Additionally, (Q)SAR can be used to predict the site of metabolism, and this feature is often used to predict different cytochrome P450-related metabolism end points. The (Q)SAR can also be used to address the risk assessment of compounds identified in leachable and extractable or migration studies where other data on the nature of the compound may be limiting. In the regulatory assessment of chemicals, (Q)SAR is becoming more important, and harmonized templates such as the (Q)SAR Model Reporting Format or the (Q)SAR Prediction Reporting Format are used when submitting data under Registration, Evaluation, Authorization, and Restriction of Chemicals (REACH). ³

The output of most (Q)SAR programs includes a list of surrogate compounds and this list offers an opportunity to supplement any missing data in the risk assessment of a compound with that of similar compounds. An example of using surrogate compounds in the development of a consumer product was presented. The steps involved identifying 3 surrogate compounds and any predictions of toxicity or alerts would be confirmed using a second (Q)SAR program. This is especially important in assessing mutagenicity, which would require running 2 complementary (Q)SAR programs if other existing data were absent.

Samantha reviewed a number of (Q)SAR programs highlighting their methodology (statistical vs expert rule based) and their end points, such as genotoxicity, carcinogenicity, hepatoxicity, and so on. Both commercially available (Q)SAR programs such as Derek, MultiCASE, and Leadscope were discussed as were programs which can be downloaded for free such as EPA Toxicity Estimation Software Tool (T.E.S.T.) and Toxtree. ⁴

In summary, when using (Q)SAR programs, the user should be aware that multiple programs are available and in general, a single (Q)SAR should not be used as a stand-alone method for risk assessment. As with all computer modeling, the user must take care in selecting the settings and interpreting the output. Finally, when submitting (Q)SAR data to a regulatory agency, always submit the full output from the program and be aware that the agency will assuredly run (Q)SAR to confirm the results.

Mark Powley from the Office of New Drugs of the FDA followed with a presentation entitled, “(Q)SAR and the Regulator,” which focused on the use of (Q)SAR in combination with expert knowledge to enhance risk assessment and regulatory decision making. There are a number of potential uses for (Q)SAR in the regulatory environment, such as providing information to support hypothesis testing, but currently the primary use of (Q)SAR at FDA/Center for Drug Evaluation and Research (CDER) is to evaluate the mutagenic potential of impurities as outlined by ICH M7 Guideline on Assessment and Control of DNA Reactive (Mutagenic) Impurities in Pharmaceuticals to Limit Potential Carcinogenic Risk. ⁵

Expert knowledge is any additional information that can supplement (Q)SAR analysis. Expert knowledge may include things such as the relevance of any structural alerts identified by (Q)SAR, the alert confidence, impact of mitigating factors, well-established structural alerts (known as a visual inspection), and empirical data from closely related analogs (known as a read across). Expert knowledge is typically supplied by someone with expertise in fields such as chemistry, genetic toxicology, computational toxicology, or others. It can maximize the confidence in the (Q)SAR prediction and can provide rationale for superseding a positive or negative (Q)SAR prediction, thereby improving sensitivity and specificity of the analysis. However, expert knowledge does have limitations. It is dependent on the knowledge and expertise of the individual and as such is subjective and cannot be validated or standardized. Furthermore, it is often not described in sufficient detail as some cases require extensive narrative to support the conclusions. Nonetheless, expert knowledge remains a valuable tool for the regulator in supplementing (Q)SAR analysis.

Since (Q)SAR is primarily used for impurity evaluations, Mark covered this application in more detail. When (Q)SAR is used in this arena, it should have adequate sensitivity to protect patient safety while at the same time balance specificity to prevent hindrance of drug development. Per ICH M7, an acceptable (Q)SAR evaluation will use predictions from 2 validated systems and will apply appropriate expert knowledge, if needed. While ICH M7 provides general recommendations on data needed to support impurity evaluations in drug development (eg, increasing information needed as development proceeds), Mark gave suggestions regarding desirable components of a (Q)SAR report. The reporting of (Q)SAR results in impurity evaluations should include details of the (Q)SAR system used, a summary of the individual (Q)SAR predictions, a final integrated prediction, and clearly describe any expert knowledge used in the analysis. Raw data in the form of outputs from (Q)SAR analyses should be appended to the report. The regulatory decision-making process is based on the (Q)SAR analysis but also considers risk–benefit specifically taking into account potential impurity exposures and the proposed clinical indication. ⁶

Finally, Nigel Greene from the Computational Sciences Center of Emphasis at Pfizer, Inc focused on the limitations of (Q)SAR in safety assessments. The discussion centered on 2 limitations of (Q)SAR: the problem of similarity and that of hazard versus exposure. The problem of similarity arises because most (Q)SAR approaches work on the assumption that similar chemicals will have similar properties and therefore similar biologic effects. But the definition of similar may vary depending on what is the biologic effect. ToxCast was an EPA-funded initiative that profiled over 2000 chemicals across 800 in vitro assays. ⁷ The goal of ToxCast was to create a method of prioritizing environmental chemicals for further evaluation. To help understand the relationships of the profiled chemicals with human health, some of the profiled chemicals were failed and marketed human pharmaceutical products. When the in vivo effects of some of the failed pharmaceutical products were compared, they showed widely different results. For example, TX6172 and TX6169 were 2 structurally similar PDE-4 inhibitors with similar properties and profiles across the in vitro assays. However, TX6162 caused severe inflammation of the heart and prostate whereas TX6169 did not. ⁸

The second limitation of (Q)SAR discussed was differentiating hazard versus exposure, especially in predicting organ toxicity. Using (Q)SAR to predict organ toxicity is problematic because the toxicity is often multifactorial or resulting from combinations of effects. For example, troglitazone was withdrawn from the market for liver failure. The in vivo toxicity profile showed that troglitazone was a potent inhibitor of the Bile Salt Export Pump transporter, caused mitochondrial dysfunction leading to cell death, and formed covalent adducts to cellular proteins. ^9,10 Further complicating the prediction of organ toxicity is dose does not always correlate well with the plasma exposure, and tissue-level exposure does not necessarily equal plasma exposure. In most safety assessments, exposure-driven toxicity, that is, where chemical accumulation in tissues and/or cells causes interactions with multiple biological processes and so simply overwhelms the cell resulting in cell death, predominates over mechanism-driven toxicity, that is, where the chemical in question interacts with a discrete mechanism or receptor that results in an adverse effect on the cell or tissue. However, current data sets often do not differentiate between the 2 and so the properties that correlate with toxicity may be misleading or inaccurate.

Despite these limitations, (Q)SAR can be a valuable tool in hazard identification. In the future, incorporation of exposure as a function of observed toxicity and differentiating toxicity driven solely by concentration from that of discrete, mechanism-based toxicity will be critical for enhancing the use of (Q)SAR.