Using Voronoï fingerprints to rescore Hex protein-protein docking models

Protein-protein docking procedures normally consist of two successive steps. Firstly, a search algorithm generates a large number of candidate conformations (this is the most CPU-intensive step), and a scoring function is then used to rank them in order to extract a near-native conformation. Thanks to modern GPUs, the Hex docking program can now generate and score billions of candidate conformations to produce a list of few hundred high quality candidate solutions in just few seconds. However, the Hex scoring function cannot normally identify a near-native conformation from this list. On the other hand, we have previously demonstrated that using Voronoi tessellations of protein interfaces using a relatively small number of surface properties for each Voronoi cell can be used to construct accurate scoring functions. However, such scoring functions are not yet sufficiently sensitive for large scale explorations of the interactome. Here, we introduce a hierarchical clustering of threedimensional docking predictions of the Docking Benchmark (3.0) to detect and and discriminate subsets of near-native complexes from decoys generated by Hex. This provides an a priori test of the ability of Voronoi fingerprints to score docking candidates based on their similarity to the interface fingerprints of known protein interfaces. Although this work is still in progress, we expect that this novel Voronoibased scoring scheme should give better docking predictions than using conventional shape-based or “knowledge-base” potentials. Nevertheless we had also applied a classical “knowledge-base” scoring function to rank and score the different shaped-based predictions.