Can We Estimate the Accuracy of ADMET Predictions ? Igor

Teaser: Is there a distinct relationship between the accuracy of prediction and molecular similarity? Can we estimate the accuracy of property prediction for new compounds? This manuscript contains 17 pages including an abstract, 1 table and 3 figures. 2 This article reviews recent developments in methods to access the accuracy of prediction and applicability domain of ADMET models and methods to predict physico-chemical properties of compounds in the early stages of drug development. The methods are classified into two main groups, namely, methods based on the analysis of similarity of molecules and methods based on the analysis of calculated properties. Using the example of octanol-water distribution coefficients we exemplify consistency of estimated and calculated accuracy of the ALOGPS program (http://www.vcclab.org) to predict in house and publicly available datasets. The importance of the methods for improvement of the quality of the high-throughput screening and hits triage, and in particular to avoid improper filtering of compounds standing far from the investigated chemical space is discussed. 3 Each year an increasing number of computational methods devoted to the development of predictive ADMET models is published. Despite the fact that their importance for the drug discovery process is well recognized [1], the available methods are not yet sufficiently reliable and are limited in their application [2]. For example, recent reviews [3,4] indicate that as many as 50 articles devoted to methodological developments to predict lipophilicity and aqueous solubility are projected to be published in 2005. This is about a 5-fold increase compared to 1995. However, the prediction accuracy for proprietary datasets remains disappointingly low [5-8]. One can describe such relative levels of failure in terms of the applicability domain (AD) of the models. In the " ontology " classification of the model failure, one can distinguish at least two major problems: experimental design and diversity of compounds. The experimental design problems can result from different end-points of the models, [9] i.e. agreement of protocols used in the development of the models, data consistency and quality and model applicability. The second reason for model failure is the difference in chemical space of compounds that were used to develop and apply the models. This problem can also be attributed to experimental design problems: in predictive models, both training and test set compounds have to be from the same chemical space [10-14]. However, there are at least two principal reasons making such a situation unlikely. First, because …