Small peptides are presented by MHC molecules to T cells. A collaborative effort is proposed, involving the investigator and two senior investigators, to explore the potential of an innovative feature-space geometric approach for the design and implementation of MHC peptide binding models. The key concept formulated is the representation of amino acid sequences as vectors in a metric feature-space whose transformations make possible conceptual models of MHC peptide binding in terms of the amino acid sequence of the respective MHC alleles. Such models will allow the prediction of peptide binding not only for those HLA class I alleles for which there is sufficient data, but also for those other alleles for which there is no peptide binding data available yet. The investigator's preliminary work based on the use of this novel metric space approach has revealed probable critical positions in the HLA class I molecule - previously not identified - playing an important role in the selection of peptides for T-cell presentation; and clusters of HLA class I alleles with peptide-binding similarities not noticed before. Recent advances in neural networks by one of the senior investigators, a pioneer in this area and expert in its theory and its medical applications, make it possible to construct metric spaces heuristically. These advances involve data mining techniques that explore similarity measures heuristically with powerful non-linear transformation properties - referred to as universal approximators - that generate models with robust generalizations not otherwise possible. Preliminary collaboration with the other senior investigator, a recognized expert in MHC antigen processing and presentation, particularly in relation to HIV, has confirmed the prediction of peptide binding of HIV proteins. The main aim of this proposal is to model MHC peptide binding using the robust non-linear modeling properties of neural networks. The models designed and built will be validated experimentally by studying the interactions of specific peptides with various MHC alleles using cell lines that express a single HLA class I allele and in which the internal mechanism ofpeptide processing is blocked so that HLA molecules are stabilized by being exposed to extracellular peptides. This project relies on (1) a theoreticai framework to study amino acid sequences based on geometric and algebraic concepts, (2) the data-analysis power of neural networks, and (3) an experimental model to study allele-specific MHC peptide binding, already proved to be reliable and adequate to validate MHC-peptide binding predictions.
