The experimental determination of peptide binding affinity to Major Histocompatibility Complex (MHC) molecules yields information which is often important to our understanding of the mechanics of autoimmunity, a widely occurring phenomenon giving rise to extensive pathological changes in many disease conditions. However, the most common biochemical technique by which this information is obtained (Competitive binding assaying) is a time consuming and costly procedure.
Consequently, many groups have been using computational methods to try and predict which peptides are likely to bind to MHC molecules, and so replace the experimental approach with the cheaper and faster computational alternative. However, whilst some of the systems produced over the last decade have attained reasonably high levels of accuracy, they have all suffered from debilitating limitations which have hampered their widespread use in one way or another. Most commonly, the systems are not only restricted to predicting peptide binding in the context of a single MHC allele, but they also require large volumes of experimentally determined peptide binding data for use during their calibration, or `training'.
Presented here is a system which focuses on predicting peptides which are likely to bind to members of HLA-DR, a large and commonly occurring subset of (class II) MHC molecules which are expressed in humans. The system is based upon a unique three dimensional structural modelling technique which (a). is able to produce peptide binding predictions with comparable accuracy to the current `state-of-the-art', yet (b). only requires a fraction of the training data, and (c). when trained, is not specific to any single allele, but applicable to any member of the HLA-DR family.