Application of advanced in silico methods for predictive modeling and information integration

Abstract

Introduction: In silico predictive methods are well-known tools to the drug discovery process. In recent years, these tools have become of strategic interest to regulatory authorities to support risk-based approaches and to complement, and potentially strengthen evidence when considering product quality and safety of human pharmaceuticals.

Areas covered: This editorial reviews how chemically intelligent systems and computational models using structure-based assessments are important for providing predictive data on drug toxicity and safety liabilities considered at the FDA. The example of regulatory interest in application of in silico systems for mutagenicity predictions of drug impurities is discussed.

Expert opinion: The importance of information integration is emphasized toward the application of in silico predictive methods and enhancing data mining capabilities for safety signal detection. Modeling for cardiovascular drug safety based on human clinical trial data is one area of active testing of predictive technologies at the FDA. The FDA has taken appropriate steps in its strategies and initiatives aimed to enhance and support innovation for regulatory science and medical product development by developing and implementing the use of in silico predictive models and medical toxicity databases. This science priority area will ultimately help improve and protect public health.

Keywords::

• computational toxicology
• drug impurities
• drug safety
• genotoxicity
• in silico methods
• in silico toxicology
• QSAR
• SAR

1. Introduction

Computer technologies have no doubt made tremendous impact on efficiencies, generating accurate data for our society from telecommunications, and market demand forecasting, to protecting public safety with hurricane forecast models. Likewise, computer technologies have been an important strategy, investment and integral part of drug discovery. Computational strategies are implemented early on as a tool for creating innovation in the search, design and optimization of new drugs proposed to treat human diseases. Moreover, in silico methods help reduce drug attrition rates through use of pharmacovigilance computational data mining methods, medical bioinformatics approaches and predictive computer models for identifying potential drug toxicity. A major advantage of in silico methodologies is that they help speed the rate of production and screening of drug candidates based on analysis of calculated properties and prediction models for drug therapeutic targets and identification of safety liabilities all the while minimizing the need for expensive and time-consuming animal and in vitro assay laboratory work. What is gaining acceptance and recognition, although still at an early stage, is the use of in silico methods such as predictive modeling for toxicity and data mining geared toward supporting regulatory science and safety assessment needs. This editorial will consider how some drug discovery tools have turned into translational science tools to support regulatory work at the FDA including decision support to inform regulatory safety assessments.

2. Computational toxicology technologies

Computational toxicology technologies are multifaceted approaches, capable of chemical structure recognition, an array of property calculations, reasoning via human input rules or machine learning, and used for data transformation and management. Computational toxicology technologies are enabling for the identification of undesirable molecular structural features found to be associated with toxicity endpoints that are a safety liability. These technologies are typically used during lead identification and optimization, or selection of drug candidates, and even in later stages of development to inform safety assessment [1]. The use of private computational toxicology software programs by the pharmaceutical industry has been reported in the literature [2,3]. At the FDA, there is keen interest in applying computational toxicology technologies with in-house data for supporting a variety of regulatory science needs and initiatives [4]. Evaluation of several computational toxicology programs is in progress at FDA/CDER via agency-approved research collaboration agreements (RCAs) with not only software developers based in the US but also foreign companies [5,6]. Likewise, a recent review by FDA/CFSAN describes a long-term computational toxicology initiative to support regulatory safety assessments of certain food additives [7]. Thus in various regulatory programs at the FDA, there has been progress in not only exploring and evaluating computational toxicology technologies but also in harnessing agency-deidentified data to construct computational models to support regulatory decision-making on the safety of active pharmaceutical ingredient (API) and drug impurities in keeping with its mission to protect public health.

3. Predictive modeling and information integration

Many investigative drug discovery tools involving in silico methods are designed to optimize chemical design through modeling. Not only improved tools for predicting drug-related toxicities are needed by the pharmaceutical industry to decrease late-stage drug attrition, but reliable predictive tools are also needed by regulatory health authorities to support innovative regulatory science to complement or fill data gaps in support of safety evaluations. In silico predictive modeling offers an innovative opportunity for regulators and applied science investigators at the FDA to model drug-related toxicities in cases when evidence is equivocal or additional science-based information would complement or strengthen the evidence regarding a considered safety concern.

One current area for in silico predictive models to support regulatory decision-making is in the assessment of the genotoxic potential of drug impurities identified in drug products under review [8,9]. Drug impurities carry no benefit to patients and so specific steps need to be taken to control exposure to genotoxic impurities in drug product proposed for clinical trial in order to protect patient safety. ICH Quality Guidelines Q3A and Q3B address aspects of controlling the different classes of drug impurities potentially found in drug product and drug substance. Current FDA draft guidance on how to evaluate the safety of genotoxic and carcinogenic drug impurities in drug products and substances under clinical development as investigational new drugs and for marketing applications as new drugs provides recommendations on safety qualification and identification of structural alerts [8]. In the FDA draft guidance, the use of in silico predictive models using structure–activity relationships (SAR) and quantitative structure–activity relationship (QSAR) approaches is recommended for identification of structural alerts and to assess potential genotoxicity using these systems if a structural alert is present below the qualification threshold. At a recent public Drug Information Agency/FDA workshop on SAR and QSAR approaches for assessing genotoxic impurities in pharmaceuticals, consensus was reached that application of in silico SARs and QSARs are useful for early genotoxicity assessment of drug impurities [10]. The timing of the workshop was important for input on ICH M7 guidance on mutagenic impurities. ICH M7 might address the nature of in silico structure-based assessments such as SARs and QSARs generated from computational toxicology software programs to help identify potentially mutagenic drug impurities in drug products under development. Therefore, it is important to understand what constitutes an acceptable SAR or QSAR assessment for mutagenicity and what would be an appropriate validation test for a model used to provide predictive data on the mutagenic potential of drug impurities. ICH M7 has opportunity to address this and other questions such as what are the preferred predictive performance characteristics (sensitivity/specificity/concordance/negative–positive predictivity) of an in silico SAR/QSAR model for mutagenicity, how many different SAR or QSAR model predictions are needed, how might experimental evidence from genetic toxicity testing of API aid or not in the assessment of potential mutagenicity of associated drug impurities and how might modulation of mutagenic activity (e.g., substructural mitigating factors, metabolic activation) be incorporated into an in silico prediction.

It is well recognized that there are many different computational models and platforms for predicting genotoxicity. Many of these are commercially available [5,6], others are at no cost [11] and some systems are proprietary in-house [1,3]. Due to data-sharing initiatives and widespread availability of public data sources on mutagenicity of chemicals and drugs, it is also of interest to assess integration of public and proprietary sources data for predictive modeling and validation. Clearly, each in silico prediction program has its own advantages and disadvantages; however, there are relatively few published investigations on external validation of the predictive performance of these methods on a comparative basis. A recent assessment comparing predictive performance for Ames bacterial mutagenicity between four well-known computational toxicology software found that when comparing statistical QSAR software programs to human expert rule software programs, the human expert rule software had higher sensitivity based on evaluation of 7000 chemicals foreign to the model training set [12]. In addition, the same study found that none of the commercial software performed well with an in-house pharmaceutical proprietary data set, but did predict well with public structures. Although there are many factors that go into model construction, nature of the data, algorithms as well as descriptor approaches, there are many promising commercial software that have not been tested as such. Still the study provided an informative and objective assessment. Another possibility is that commercially available off-the-shelf software packages may not, without significant modification/training, adequately predict a particular chemical ‘space’ and thus many groups within foods, cosmetics and pharmaceuticals build ‘in-house’ models not to avoid using commercial tools, but rather simply to produce models that can predict their particular chemical structures under development or regulatory purview.

4. Expert opinion

Over its 100 plus years of history, the FDA has collected a vast amount of high-quality data that many would consider the richest health and medical data sets on the planet. This brings to light the importance of arriving at viable informatics approaches for harnessing this information into useful, translational knowledge for use in data mining, modeling and information integration and management. There is tremendous opportunity to tap into this information for building predictive models to support risk-based regulatory science, address certain drug safety evaluations and facilitate a more cost-effective drug discovery and development process. The FDA has made important progress in this area over the past several years and importantly has been responsible with its data compilations for the protection of public health and for aiding the basic science required for food, drug and cosmetic regulations. The FDA/CDER Office of Pharmaceutical Science (OPS) has recently developed new predictive clinical computational models to support internal cardiovascular safety evaluations on QT interval effects in humans. These models are based on clinical trial data of drugs tested in thorough QT studies and employ predictive technologies available to FDA via an agency-approved RCA with the Prous Institute for Biomedical Research (Barcelona, Spain) [6]. In addition, CDER/OPS has examined several predictive technologies for assessing potential mutagenic and carcinogenic drug impurities and continues to do so through similar agency-approved RCAs with technology developers and ensuring protection of proprietary data [13]. This work is consistent with FDA's published strategies and initiatives in regulatory science and ethically supports avoiding, when possible, unnecessary delays and experiments, especially in vivo experiments. Two major reports recently issued by FDA to the public support the use of computational in silico modeling as part of the agency's plan for advancing regulatory science [14] and strategic priorities for 2011 – 2015 [15]. These reports have as an implementation strategy to modernizing toxicology and enhancing product safety evaluation, the use and development of computational methods and in silico modeling. This science priority area will support improvements in medical product development and encourage application of innovative approaches for science-based decision-making. In doing so, the FDA is committed to excellence in protecting public health by ensuring the safety and efficacy of innovative new medical products.

Declaration of interest

L Valerio declares no conflict of interest and has received no payment in the preparation of this manuscript.

Acknowledgement

The content of this publication does not necessarily reflect the views or policies of the Department of Health and Human Services, nor does mention of trade names, commercial products, or organizations imply endorsement by the US Government. This paper reflects the current thinking and experience of the authors.