Born out of computational chemistry and chemoinformatics, in silico methods for toxicology testing have brought new insight into several areas of toxicology, including new predictive tools and data-mining approaches to help make more effective use of large repositories of the results from in vitro and animal toxicology studies with xenobiotics. Substances that have been examined with these approaches include human pharmaceuticals, food ingredients, environmental agents and other chemicals to which there is significant human exposure. The use of in silico methods directed towards the evaluation of safety endpoints, as well as the deployment of chemoinformatics approaches in the analysis of genome responses after exposure to xenobiotics, has been supported by authorities through enacted legislation in the European Union--eg the Registration, Evaluation, Authorisation and Restriction of Chemicals (REACH) legislation--to help to reduce, refine and replace animal testing [1]. In addition, recommendations recently made by the US National Research Council in its report, Toxicity Testing for the 21st Century: A Vision and a Strategy, also lend strong support to the use of computer-based technologies for the assessment of toxicities, with an emphasis on those relevant to protecting public health [2]. The US Food and Drug Administration (FDA) recognises the need to develop and identify new product development and technical tools, for example, using computational, statistical, bioinformatics and genomics methods as part of its broad-based national strategy for transforming the way that FDA-regulated products are developed, evaluated and manufactured [3]. The FDA's Critical Path Initiative describes this goal in its landmark report, Innovation/Stagnation: Challenge and Opportunity on the Critical Path to New Medical Products [4]. Among other opportunities, the publication points out the potential to develop computer-based (in silico) methods as tools for the early detection of toxicity prior to significant human exposure to products that the agency regulates [4]. Consistent with the aforementioned initiatives, the development and evaluation of in silico methods and databases in toxicology have been of interest to the FDA Center for Drug Evaluation and Research (CDER), Office of Pharmaceutical Science (OPS) [5–11]. In this mini-review, in silico toxicology models and databases at FDA/CDER/OPS Science and Research Staff (SRS) will be described in their context as Critical Path Initiative toolkits.

One of the goals of the FDA Critical Path Initiative is to predict the toxicity of drugs early in their development, thus enabling the identification of safety signals prior to human exposure to, and significant investment in, new drug products. The in silico toxicology models that have been investigated at FDA/CDER/OPS primarily have been predictive tools. The principal approach has involved computational modelling of drug molecular structure under the (quantitative) structure-activity relationship ((Q)SAR) paradigm. The central axiom of (Q)SARs is that the activity of molecules is reflected in their structure, and this is quantifiable. Although structure-activity modelling is nothing new, there are specific examples of its success in early drug development [12]. Importantly, there are limitations--and many of these have been pointed out in recent in-depth reviews and commentaries [6, 13–15]. Without question, advancing computer technology facilitates high-end, and high-speed, analysis for dealing with the complexity of biological endpoints, using varying statistical algorithms to drive the many computational analyses. The FDA Critical Path Initiative of developing better evaluation tools and harnessing bioinformatics applies to innovation, science and technology to help to predict safety of drug products.

From the practical standpoint of industrial drug development, computational models like (Q)SARs and other approaches may support lead optimisation and identification of undesirable structural features in drugs, and evaluate metabolites and impurities (theoretical and known) in the discovery process. These attributes then may be removed by chemical modifications to reduce toxicity [16, 17]. The practical use of (Q)SAR models as instruments for regulators responsible for safety assessment of regulated products, however, is a much newer concept. Through applied regulatory research at the FDA OPS, insight into the application of (Q)SAR models, in terms of strengths and, importantly, limitations, has been learned over the past decade [6]. At this time, there is no specific guidance document available from the FDA on in silico models or computational toxicology for safety assessment. The use of (Q)SARs to identify structural alerts for predicting genetic toxicity is stated under a draft guidance regarding recommended approaches to the safety assessment of genotoxic and carcinogenic impurities in drug substances and products [18]. There are many questions that naturally could arise with respect to the applied use of (Q)SAR models, and this is being addressed at the FDA through regulatory research, committee review and, ultimately, public comment on the draft guidance. From this author's perspective, there are some reasonable questions that can be asked with respect to applying (Q)SAR models for predicting genetic toxicity for regulatory applications. These include:

The answers to these questions and many others could lead to criteria that would be important in establishing the validity of (Q)SAR models for regulatory use in the USA. There are clearly questions outstanding with respect to the use of computational modelling approaches to predicting drug-related toxicity from a regulatory science standpoint. A recent article from the pharmaceutical industry echoes the questions regarding the suitability of public data for predicting the chemical space held by proprietary drugs, and the widespread use of evaluating predictive performance based on weak cross-validation strategies, as opposed to the more rigorous method of external validation [19]. Cross-validation involves building an in silico toxicology model ((Q)SAR or similar), then removing a small part of the dataset (eg 10 per cent) and rebuilding the model to predict the toxicity endpoint being modelled (eg bacterial mutagenicity), using the test set that was removed. External validation is clearly more rigorous and representative of practical use of an in silico toxicology model. It involves testing the performance of the model using a set of test drugs that are foreign to the model and were never used in the construction of the in silico model. There are many such evaluations in the literature for genetic toxicity [20, 21].

Despite the limitations and need for improving research techniques in computational modelling, the goals and applied uses of in silico toxicology models and databases remain unchanged. It is through rigorous research and validation that a better process for understanding the in silico approach will be identified.

The FDA's Critical Path Initiative was launched in 2004 with the publication of a landmark report recognising that a) the decline in the number of new drugs approved in the USA is a public health concern; b) the rising costs of drug development is prohibitive; and c) there is a need to use scientific discoveries in biomedical science to modernise and develop safer and effective drugs [4]. The aforementioned research in drug database building and in silico computational modelling fits with the core mission of the FDA Critical Path Initiative. These methodologies are intended to predict drug-related toxicity earlier and to do this with emerging chemo- and bioinformatics approaches, combined with knowledge from over a hundred years of animal toxicology study data that the agency has received and reviewed. This work is expected to lead to better evaluation tools in predictive science, especially for preclinical and clinical drug safety areas. The newest approach, which we expect to be the most enabling and to provide breakthroughs in this area, is the use of clinical trial data that the agency has reviewed to construct predictive computational models directed at helping to support risk assessments in cardiac safety and, ultimately, to lead to safer and effective medicines for public health.