James Bradford 

Supervisor: Dr David Westhead

School: Biochemistry and Molecular Biology, University Of Leeds

We are currently using Support Vector Machines (SVMs) to predict binding sites on protein surfaces.  There are two key features to our method:

1.        We attempt to classify surface patches of contiguous surface residues as either part of or outside the interface.  The attributes that describe each surface patch are based on hydrophobicity, residue propensity, conservation, electrostatics and shape.

2.        We separate our data set into four complex types: homodimers, enzyme-inhibitors, heterodimers-transient or heterodimers-permanent.  We find that training an SVM on each of these sets rather than on all complexes together increases prediction success. 

Binding site prediction compliments our work on the docking problem.  Our docking algorithm uses methods from graph theory to find maximal regions of surface complementarity between two protein molecules.  Knowledge of at least one binding site would significantly reduce the search space resulting in a decrease in number of possible solutions generated.