During the past few years there have been several reviews showing how in silico predictions of absorption, distribution, metabolism, elimination and toxicity (ADMET) processes can be used to help focus medicinal chemistry into more ideal areas of property space, minimizing the number of compounds needed to be synthesized to obtain the required biochemical and/or physicochemical profile [1–6]. There is no doubt that in silico models are increasingly being utilised. Their overall accuracy and the underlying understanding have increased over time [1–4]. The potential is clear, with an estimated size in the region of 1020–1024 molecules for the accessible virtual organic chemistry space. The use of computational models provides an approach to connect, use and extend existing experimental data, to assess and prioritise thousands of chemicals quickly. Furthermore Ekins et al. , highlight the prospect of in silico approaches being utilised for drug repurposing through an integration of knowledge from databases, networks and also in vitro screening while also raising the need for validated databases. However, although the potential is great at the same time we need to be fully aware of challenges and limitations faced by these tools.
Although different types of models [i.e. simple rules, 2D and/or 3D based quantitative structure–activity relationship (QSAR) or machine learning approaches] are available for different ADMET end points, many of these approaches tend to neglect direct structural information about the proteins involved in ADMET processes. In silico approaches based on the 3D structures of these proteins might therefore add extra benefits [4,8,9]. The increasing availability of 3D structures provides the means to explore the value of molecular understanding gained by taking into account the ligand and/or protein interactions. The report by Moroy et al.  focuses on recent in silico studies across a breadth of ADMET proteins and through a case study explores the value of this approach. The report also acknowledges the challenge caused by the promiscuous nature of some of the proteins involved in ADMET. Chen et al.  provides a detailed examination of these challenges in the case of P-glycoprotein (P-gp), which interacts with large numbers of structurally diverse compounds and has multiple binding sites. Although the high-resolution structures of P-gp are now available, the paper acknowledges that  ‘there are limited results for rationally translating this information into developing prediction models with satisfactory reliability’.
To paraphrase Machiavelli ‘Whoever wishes to foresee the future must consult the past’. For future prediction, it is important to learn as much we can from past and present facts and data. Therefore, it is not surprising that one of the first challenges for these in silico tools is that most of these techniques are ‘data hungry’. With the recent advances in high-throughput screening (HTS), chemical synthesis and biological screening, there is no shortage of publicly or commercially available databases that can be used as data sources for these models [5,The long term cost of inferior database quality: http://www.drugdiscoverytoday.com/view/22579/the-long-term-cost-of-inferior-database-quality/]. A recent article discusses some of these issues and has also evaluated the consequences of both random and systematic errors with chemical structure curations in well-known data sets .
Another issue is that none of these in silico models should be considered as the ‘finished article’ because with most other techniques the models are tuned as more information and/or knowledge becomes available to enhance prediction capability . Applicability domain is one of the main reasons for the QSAR and/or SAR model failure owing to the difference in chemical space of compounds that were used to develop and apply the models. In silico models might not perform well if a predicted chemical is beyond the chemical space where the models were developed [12–14]. Understanding the confidence of the prediction based on applicability domain is an area of increasing value to ensure that inappropriate application of models is minimised.
Acceptance of these tools outside the modelling community raises a different challenge of perception. Although some colleague’s embrace these methods others can be reluctant or sceptical. Some even fear (consciously or unconsciously) that these tools will challenge or even destroy their traditional tools and endanger their comfort zones . Although acceptance will come through showing the value of the application it can be enhanced by the transparency of the model. An awareness of the intended use and/or recipient of the data where non-experts are involved can help ensure that the use of complex modelling techniques does not consequently result in the benefits of the models being misunderstood. The goal of the in silico tools is not to produce a series of models to be used in place of laboratory tests, but rather to improve both the design and strategic use of test methods  and further integrate different methodologies to improve overall mechanistic understanding and predictivity [4–7] . Recent work has demonstrated that in silico methods will have a crucial role in bringing together new types of data for novel non-animal approaches to risk assessment [16,17].
In vitro and cell culture methods were started more than 100 years ago  and have been usefully deployed since the 1960s in the discovery of new chemicals. In comparison it should be acknowledged that in silico tools have been in existence for a relatively short time, and have only taken up a pace in the past 20 years. There are examples to show how some in silico approaches can be used to derive new drugs or even new uses for approved drugs [7,18]. Given the enormous speed of these in silico technologies a much quicker development can be expected as these tools become increasingly user-friendly and transparent. As more examples of successful applications are shown, and they are integrated with in vitro screening, it seems highly probable that in silico approaches will evolve rapidly. Recent advances in physiologically-based pharmacokinetic and/or pharmacodynamic modelling  that integrate these in silico chemistry tools to predict needed parameters to describe and give better understanding of the pharmacokinetics and dynamics of complex mixtures are steps in the right direction. There is also a need for these computational chemistry tools to align with other information sources (e.g. from systems biology, hazard, metabolites and exposure) to develop real or virtual models of tissues, organs and physiological processes that could be used for the risk assessments .
1. Glesson, P. et al. (2011) In silico ADME models: a general assessment of their utility in drug discovery applications. Curr. Top. Med. Chem. 11, 358–381
2. Raunio, H. (2011) In silico toxicology – non-testing methods, frontiers in pharmacology, 2, 33
3. Dearden, J.C. (2007) In silico predictions of ADMET properties: how far we have come? Expert Opin. Drug Metabol. Toxicol. 3, 635–639
4. Moroy, G. (2012) Towards in silico structure-based ADMET prediction in drug discovery. Drug Discov. Today (in press) (DIO: 10.1016/j.drudis.2011.10.023)
5. Modi, S. et al. (2012) The value of in silico chemistry in the safety assessment of chemicals in the consumer goods and pharmaceutical industries. Drug Discov. Today 17, 135–142
6. Hartung, T and Hoffmann, S. (2009) Food for thought… on in silico methods in toxicology. ALTEX 3, 155–166
7. Ekins, S. et al. (2012) In silico repositioning of approved drugs for rare and neglected diseases. Drug Discov. Today 16, 298–310
8. Vedani, A. and Smiesko, M. (2009) In silico toxicology in drug discovery – concepts based on three-dimensional models. Altern. Lab. Anim. 37, 477–496
9. Chen, L. (2012) Computational models for predicting substrates or inhibitors of P-glycoprotein. Drug Discov. Today (in press) (DOI: 10.1016/j.drudis.2011.11.003)
10. Fourches, D. et al.(2010) Trust, but verify: on the importance of chemical structure curation in cheminformatics and QSAR modeling research. J. Chem. Inf. Model. 50, 1189–1204
11. Cronin, M.T. (2000) Computational methods for the prediction of drug toxicity. Curr. Opin. Drug Disc. Develop. 3, 292–297
12. Jaworska, J. et al. (2005) QSAR applicabilty domain estimation by projection of the training set descriptor space: a review. Altern. Lab Anim. 33, 445–459
13. Stanforth, R.W. et al.(2007) A measure of domain of applicability for QSAR modelling based on intelligent K-means clustering , QSAR & combinatorial. Science 26, 837–846
14. Dragos, H. et al. (2009) Predicting the predictability: a unified approach to the applicability domain problem of QSAR models. J. Chem. Inf. Comput. Sci. 49, 1762–1777
15. Veith, G.D. (2004) On the nature, evolution and future of QSAR in toxicology. SAR and QSAR Env. Res. 15, 323–330
16. Maxwell, G. and Mackay, C. (2008) Application of a systems biology approach to skin allergy risk assessment. Altern. Lab. Anim. 36, 521–556
17. Davies, M. et al. (2010) Determining epidermal disposition kinetics for use in an integrated non-animal approach to skin sensitisation risk assessment. Toxicol. Sci. 2, 308–318
18. Bisson, W.H. et al. (2007) Discovery of antiandrogen activity of nonsteroidal scaffolds of marketed drugs. Proc. Natl Acad. Sci. USA 104, 11927–11932
19. Zhao, P. (2011) Applications of physiologically based pharmacokinetic (PBPK) modeling and simulation during regulatory review. Clin. Pharmacol Ther. 89, 259–267
20. Abbott, A. (2010) Nature Germans cook up liver project. Nature 468, 879
Andrew White is a computational toxicologist in the Safety & Environmental Assurance Centre at Unilever PLC. Dr White has published and presented extensively on the applications of in vitro assay development, genomics and bioinformatics. His current focus is on the applications of computational toxicology for non-animal alternatives for risk assessment. He has represented Unilever on several external academic and industry research forums and currently represents Unilever on the industry forum for the European Bioinformatics Institute and as a scientific advisor for Cosmetics Europe on the Seurat 1 FP7 Research Initiative. Dr White received his PhD from the University of Newcastle upon Tyne and has spent the past 15 years in the consumer goods industry.
Sandeep Modi is working as Cheminformatician at Unilever and is involved in building several searchable databases and also develops integration approaches for risk assessments for various tox end points. He has published more than 70 papers in various field of chemistry. He is also representing Unilever on several external industry research projects. He received his PhD from the Tata institute of fundamental research (Mumbai, India). And before joining Unilever, he worked in computational chemistry group at GlaxoSmithKline where he was involved in building in silico models for various ADME end points.