Prev. Article Next Article Table of Contents

Article

New and Original pK_a Prediction Method Using Grid Molecular Interaction Fields

Supporting Info

Francesca Milletti,^† Loriano Storchi,^‡ Gianluca Sforna,^‡ and Gabriele Cruciani*^†

Laboratory for Chemometrics and Cheminformatics, Department of Chemistry, Universit degli Studi di Perugia, via Elce di Sotto 10, 06123 Perugia, Italy, and Molecular Discovery Limited, 215 Marsh Road, Pinner, Middlesex, London HA5 5NE, United Kingdom

J. Chem. Inf. Model., 2007, 47 (6), pp 2172–2181

DOI: 10.1021/ci700018y

Publication Date (Web): October 2, 2007

†

Università degli Studi di Perugia.

‡

Molecular Discovery Limited.

Corresponding author phone: +390755855629; fax: +3907545646; e-mail: gabri@chemiome.chm.unipg.it.

Abstract

One of the most important physicochemical properties of a molecule is pK_a. It is known that two parameters imperative in ADME profiling, solubility, and lipophilicity are governed by pK_a, and receptor binding can be influenced by pK_a. Because most drugs are ionized in physiological conditions, pK_a is particularly relevant to medicinal chemistry. Despite the numerous advances in high-throughput measurements, in silico determination is still the fastest and cheapest way of obtaining pK_a. This paper presents a new original computational method for pK_a prediction of organic compounds. Descriptors were generated using the program GRID, and these descriptors are based on molecular interaction fields precomputed on a set of molecular fragments. The new method was developed, trained, and cross-validated by using a large and diverse data set of 24 617 pK_a values. This paper presents the results for a class of 421 acidic nitrogen compounds (RMSE = 0.41, r² = 0.97, q² = 0.87) and for a class of 947 six-membered N-heterocyclic bases (RMSE = 0.60, r² = 0.93, q² = 0.85). For external validation 28 novel compounds were selected that covered nine different ionizable groups, and 39 pK_a values could be experimentally determined by spectral gradient analysis (SGA). Comparison of experimental pK_a with calculated pK_a demonstrated that the predictive ability of the method is good (external set, r² = 0.85, RMSE = 0.90).