COMPARISON OF PROTEIN PREDICTION TOOLS

Comparison of Protein Prediction Tools



Comparison of Protein Prediction Tools

Introduction

Identifying the location of ligand binding sites on a protein is of fundamental importance for a range of applications including molecular docking, de novo drug design and structural identification and comparison of functional sites. Here, we describe a new method of ligand binding site prediction called Q-SiteFinder. It uses the interaction energy between the protein and a simple van der Waals probe to locate energetically favourable binding sites. Energetically favourable probe sites are clustered according to their spatial proximity and clusters are then ranked according to the sum of interaction energies for sites within each cluster.

There is at least one successful prediction in the top three predicted sites in 90% of proteins tested when using Q-SiteFinder. This success rate is higher than that of a commonly used pocket detection algorithm (Pocket-Finder) which uses geometric criteria. Additionally, Q-SiteFinder is twice as effective as Pocket-Finder in generating predicted sites that map accurately onto ligand coordinates. It also generates predicted sites with the lowest average volumes of the methods examined in this study. Unlike pocket detection, the volumes of the predicted sites appear to show relatively low dependence on protein volume and are similar in volume to the ligands they contain. Restricting the size of the pocket is important for reducing the search space required for docking and de novo drug design or site comparison. The method can be applied in structural genomics studies where protein binding sites remain uncharacterized since the 86% success rate for unbound proteins appears to be only slightly lower than that of ligand-bound proteins.

Anfinsen's experiment on the reversible denaturation of ribonuclease demonstrated that the tertiary structure of proteins in solution may be determined by the amino acid sequence, and extensive subsequent experiments have demonstrated thermodynamic reversibility for many small, single domain proteins and a number of more complex cases (for a recent survey and discussion see Dill). These results gave rise to the idea known today as the “thermodynamic hypothesis,” i.e., that the native form of a protein corresponds to global minimum on the conformational free energy surface. This, in turn, implies that a purely theoretical/computational approach to protein structure prediction from amino acid sequence should be feasible, at least in principle. In practice, the difficulty of determining accurate energy functions, the essentially rugged character of the free energy landscape, and the dimensionality of the configuration space of a polypeptide in solution, makes ab initio prediction of protein structure from sequence a very challenging problem. The field of ab initio structure prediction has spawned a large literature (for a recent review see Osguthorpe), and is featured prominently in the CASP series of blind prediction contests designed to objectively document progress in protein structure prediction by various methods.

To make ab initio protein structure prediction tractable, one requires a search engine that is rapid but not vulnerable to getting trapped in the rugged features of the configurational free energy ...