Quantum Mechanical/Molecular Mechanical Approaches in Drug Design

TELL TUTTLE

WestCHEM, Department of Pure and Applied Chemistry, University of Strathclyde, 295 Cathedral Street, Glasgow G1 1XL, UK

1.1 Introduction

Quantum mechanical/molecular mechanical (QM/MM) hybrid methods have their origins in the 1970s in the pioneering work of Warshell and Levitt. However, until the last two decades, the uptake of this methodology and its application to solving problems of biochemical interest was essentially non-existent. In the beginning of the 1990s the situation changed dramatically and since this time there has been an explosion in the number of publications that use QM/MM methods to study systems of biochemical interest. Several recent reviews in the literature have provided excellent coverage of the recent developments with a strong focus on the usefulness of QM/MM methods to studying enzyme reactions.

The dramatic uptake of QM/MM methodology has been driven by significant developments in the methodology accompanied by a rapid increase in computer power and decrease in computer cost. Moreover, along with the development in QM/MM methodology, the development of QM methods to deal with larger systems with ever increasing accuracy has allowed these high-level QM methods to be applied to biochemical systems. One can now find examples in the literature where enzyme-catalysed reactions have been modelled computationally to within chemical accuracy (ca. 1 kcal/mol) – a feat unimaginable only a few years previously. Despite these impressive strides in accuracy and efficiency, the application of QM/MM methods to the field of drug design has received far less attention until recently.

Computational drug discovery is a well-established discipline within the field of computational chemistry. Computational chemistry can be applied to the development of new drugs from both the perspective of the small molecule (drug-like compound) and the biological target (receptor). In cases where the drug target is unknown or a structure of the target receptor is unavailable, a quantitative–structure activity relationship (QSAR) can be developed by comparison of several ligands with varying biological responses (activities). One of the oldest and most popular 3D-QSAR methods is a Comparative Molecular Field Analysis (CoMFA), which uses an sp3-hybridized carbon with a charge of +1 to evaluate steric and electrostatic features of the compounds in the training set. By comparing the steric and electrostatic features of the compounds in the training set as a function of their activity a relationship between regions that benefit from having certain properties in relative molecular positions (e.g. decreased steric bulk next to a negatively charged region of the molecule) can be identified. While both ligand-based and receptor-based methods have seen significant successes throughout their application, in the current review the focus will be on receptor (structure) based methods.

Structure-based drug design (SBDD) stems from the lock-and-key principle originally proposed by Emil Fischer over 100 years ago. In the lock-and-key model the ligand and the receptor are both considered as rigid objects and the activity of a ligand results from the complementary matching of the steric and electrostatic features of the ligand (key) into the receptor's binding pocket (lock). This principle was subsequently revised in 1958 by Koshland with the proposal of the induced-fit model whereby both the ligand and the receptor are considered to behave dynamically and as such the 'lock' and the 'key' are both able to adapt to some extent to provide a better match where possible during the binding process. However, modelling the dynamic flexibility of a complete macromolecule (receptor) is still not computationally feasible when one wishes to screen a large database of potential drug-like molecules. As such, within modern docking methods a compromise between the lock-and-key method and the induced fit model is generally applied. These methods treat the ligand as a flexible molecule (or a series of different conformations of the ligand is used) and is docked into the rigid receptor, or semi rigid receptor (where flexibility of the side chains in the binding pocket is allowed). Irrespective of the approach used, the key requirement for SBDD is a reliable structure of the target macromolecule.

The SBDD approach usually involves multiple stages. These can be broadly classified as follows.

1. Generation of poses/conformations of the ligand in the active site – this process determines whether the ligand is able, and if so in which orientation, to physically fit into the active site.

2. Scoring and ranking the individual poses – in this process the static poses generated in the first step are given a score, usually based on electrostatic fit between the ligand and the residues that compose the binding site.

3. Refinement of the docked structures and their subsequent ranking – this third step is the most complicated and various different approaches exist to introduce ligand and binding site flexibility into the refinement of the structure (e.g. molecular dynamics simulations, random search methods, etc.).

Quantum mechanical/molecular mechanical (QM/MM) scoring refinement focuses on the third stage of the SBDD approach – the structural refinement and determination of the final ranking. This refinement process has several problems at a computational level. The methodology employed in the refinement needs to be efficient in order to deal with the large numbers of ligands and poses that need to be refined. Because of this the use of molecular mechanics (i.e. empirical force field based approaches) has been the method of choice. However, empirical methods are inherently limited in their accuracy to the quality/specificity of the parameterization. As a result, in repeated regular structures, such as proteins, which are combinations of only 20 amino acids, force field methods are remarkably successful (e.g. the CHARMM, AMBER, GROMOS, etc. force fields). However, in the case of ligands, which contain an infinite variety of chemical motifs, the parametrization of effective force fields is much more difficult. As such, docking and scoring programs generally rely on universal force fields for modelling both the ligand and the protein, which are much less successful, but offer a common level of accuracy between the protein and the ligand. In principle, the use of...