Efficient Ant Colony Optimization Algorithms for Structure- and Ligand-Based Drug Design

Abstract

Within this work new algorithms based on Ant Colony Optimization (ACO) for the task of structure- and ligand-based drug design were developed. The ACO-algorithm PLANTS (Protein-Ligand ANT System) is capable of minimizing real-valued multi-dimensional objective functions. Therefore, all problems treated in this work were reduced to real-valued optimization problems, which were then tackled by the PLANTS approach.In the case of structure-based drug design, PLANTS was employed to treat the flexible protein-ligand docking problem. Two empirical objective functions aiming at the prediction of experimentally observed complex conformations were parameterized using a data set consisting of 31 protein-ligand complexes. The whole approach exhibited an excellent performance with respect to predicting experimentally observed complex conformations when docking back a ligand into its cocrystallized protein structure for publicly available benchmark data sets (CCDC / Astex). On average between 76% (213 complexes of the CCDC / Astex data set) and 87% (85 complexes of the Astex diverse data set consisting of drug-like ligands) of the complex conformations could be predicted correctly. A direct comparison with the protein-ligand docking software GOLD on the same test sets showed that the PLANTS approach was able to reach at least as high success rates as the ones observed for GOLD within shorter average search times. Further experiments were carried out for the task of docking a ligand into a non-native protein structure (cross-docking). In general, these experiments identified the limitations of the presented approach. However, the inclusion of additional degrees of freedom, like flexible protein side-chains and explicit, displaceable water molecules, significantly increased the predictive power of the approach.Besides the prediction of the complex conformation, the approach was also assessed with respect to the task of virtual screening. The inactive data sets were composed to match the property-distributions of the active ones to avoid one source of artificial enrichment. The performance of the approach regarding the discrimination between biologically active and inactive compounds was target-dependent and ranged from modest up to excellent enrichment factors.A coupled ACO-algorithm accounting for the parameterization of the objective function as well as the complex structure prediction was used to train a target-specific objective function. The target-specific objective function in general showed a better discrimination between active and inactive ligands, while no change in the enrichment factors at 1% and 2% of the ranked database was observed compared to the standard objective function.Furthermore, a graphics card accelerated version of the PLANTS approach was presented. This version is capable of exploiting the enormous floating point computation power of modern graphics cards for the structure transformation and objective function evaluation step. The graphics card accelerated version of the docking approach (Nvidia GeForce 8800 GTX graphics card) only needed approximately 20% of the computation time measured for the standard version of PLANTS on a test set consisting of 129 protein-ligand complexes, while reaching similar success rates.Finally, the PLANTS approach was used in the ligand-based part of the work for the problem of pairwise and multiple flexible ligand alignment. A similarity-based objective function, rewarding the alignment of identical pharmacophoric points in different ligands, was parameterized using 4 training data sets. The resulting parameterization showed a good predictive power when tested on the independent FlexS data set as well as for the concurrent alignment of multiple ligands. Additionally, ligand-based virtual screening experiments were carried out, which showed at least comparable results compared to the structure-based PLANTS approach at considerably shorter search times per ligand.