Computational Analyses of Protein-Ligand Interactions

Protein-ligand interactions have a central role in all processes in living systems. A comprehensive understanding of protein interactions with small molecules is of great interest as it provides opportunities for understanding protein function and therapeutic intervention. The major aims of this thesis were to characterise protein-ligand interactions from databases of crystal structures and to apply molecular modelling techniques for accurate prediction of binding modes of molecular fragments in protein binding sites. The first aspect of the project was the analysis of hydrogen bond donors and acceptors in 187 protein-ligand complexes of resolution 2.5Å or better. The results showed that an extremely small fraction of them were not explicitly hydrogen bonded, with the hydrogen bond criterion of donor-acceptor distance ≤ 3.5 Å and H-bond angle of ≥ 90°. It was also noticed that a vast majority of such cases were explicable on the basis of weak interactions and weak donor/acceptor strength. The results were consistent with reported observations for buried protein regions. In a series of docking calculations, the fraction of lost hydrogen bonds was evaluated as a discriminator of good versus bad docking poses. Docking and scoring with a standard program, rDock, did not create incorrect poses with missing hydrogen bonds to an extent that would make lost hydrogen bonds a strong discriminator. The second aspect of the research is related to weak (CH-π and XH-π, X=N,O,S) interactions. In a survey of IsoStar, a database of protein-ligand interactions, subtle differences were noticed in geometric parameters of π interactions involving different types of ligand aromatic rings with strong and weak donor groups in binding sites. The results supported the hypothesis that energetically favourable interaction patterns are more frequent when there are electron-donating substituents attached to the aromatic ring. Finally, the applicability of a modelling technique, multiple copy simultaneous search, in terms of predicting energetically favourable poses of solvents and fragments in target binding sites, was explored in detail. Several factors such as re-scoring with a better scoring function, use of multiple receptor structures and good quality prediction of water binding sites led to a robust protocol for high quality predictions of fragment binding in test datasets.