Software → DOx

• Introduction
• The Search Module
• The Scoring Function
• Validation on Standard data sets
• Validation on therapeutically relevant targets
• References

Introduction

Protein-ligand docking methods attempt to identify optimal positions and orientations of a ligand or small molecule with respect to a given protein receptor or enzyme.  Docking programs typically are composed of two main components:

1. A search component that explores position and orientation of the ligand with respect to the protein.
2. A scoring component which evaluates each generated molecular configuration.

The Search Module

Search modules come in two main types, exhaustive and stochastic.
Exhaustive or systematic search methods move and rotate the ligand into every possible position and orientation within the search space using a given “granularity” of search. The success of such programs is often limited by efficiency considerations due to the complexity and scale associated with large proteins and receptors. Virtual high-throughput screening can therefore be hampered by large processing times. Such methods also encounter problems when presented with flexible molecules due to the exponential increase in the search space size.
A variety of stochastic and systematic search techniques are used by docking programs. For example, AutoDock [8] and GOLD [1] use variations of the Genetic Algorithm (GA) [4] method for this purpose. DOCK [1] uses an incremental construction and random conformation search-based method to search for optimal poses. FlexX [6, 11] incrementally constructs the molecule to sample the conformation space and iteratively places it within the active site. Earlier versions of AutoDock used the simulated annealing [5] method to perform the search. All such methods attempt to achieve the correct balance between the efficiency and accuracy of the search.
In principle, a docking search is ideally suited for implementation within a GA as the search space can be represented using 6 or more real numbers which describe the position and orientation of the ligand with respect to the protein. In principle, a GA should be able identify the optimal configuration for a given optimisation problem. This is however limited by the design of the evaluation system and the parameters used during the running of the GA. Efficiency concerns normally do not allow for large execution times and small population sizes and limited execution times can force the algorithm to settle within local minima. Modern GAs usually operate by performing large variations (and therefore large configuration changes) at the early stages of execution, performing the final optimizations by incrementally changing the mutation rate, cross-over rate and the level of elitism. The use of an optimal GA configuration for the problem space is very important and was a main consideration for the design of the DOx search component.
DOx uses a GA-based search method with a gradient-based optimization module. AutoDock [8] also uses a similar implementation of a hybrid GA by incorporating Lamarckian rules to the operation of the algorithm.  DOx also uses a novel chromosome design which fragments the translational and rotational coordinates of a new configuration into several values of varying magnitude to allow the GA to perform its search using different step-sizes.

The Scoring Module

The scoring module is used to evaluate the favourability of a generated molecular configuration.  A variety of scoring functions have been developed over the past decade [13]. Several recent studies have also evaluated many collections of these scoring functions for accuracy. These studies [13, 14] have indicated the effectiveness of the XScore, DrugScore, PLP and G-Score scoring functions.  The PLP and XScore functions have been implemented in DOx.
However, many of the scoring functions available have been developed, tested and evaluated against distinct classes of proteins and may therefore return different results for generalized cases. The best scoring function to use for a particular class of target protein can be difficult to predict. Therefore, a more recent approach to the construction of the scoring module involves the use of consensus scoring. Consensus scoring simply involves the use of two or more scoring functions for the prediction of the binding affinity. The construction of the final score could be done in many ways. The simplest would involve the normalisation of two scores (e.g. PLP and XScore) and using the largest or smallest. A more refined approach could involve the scoring of the ligand against a collection of scoring functions and constructing the final score by the parameterized addition of the different scoring function scores. The values of each parameter could be dependent on the class of the protein being used.

References