Towards understanding protein-protein interactions: The AI approach

Abstract

We present the contribution to the eptiope prediction challenge of the DREAM5 experiment. The AI approach was used to model interactions between peptides and immunoglobuline protein. The protocol involved development of several independent feature sets for the description of peptides, including text-based representations with reduced alphabets, as well as features based on the chemical properties of peptides. Then several machine learning methods were applied to the resulting information systems, and their results were combined to form the final answer to the challenge. The results obtained in this way are close to the best achieved in the experiment. We show the relative merits of alternative feature sets and machine learning algorithms and present possible improvements to the methodology, both with respect to the feature construction and selection and with respect to the learning protocol. © 2011 Springer-Verlag Berlin Heidelberg.