PROTEIN HOMOLOGY FROM NEAREST NEIGHBOR NEURAL NETWORKS

Determination of homologies/sequence similarities between proteins provides insight into structure-function relationships and evolutionary processes. Current computer methods employ an initial alignment of protein amino acid sequences in order to recognize consensus regions and quantify similarities. Neural network methodologies have been used to study DNA and protein sequences to predict domains serving putative functions. We propose to improve on the self-organizing neural network (after Kohonen) to demonstrate structure-function relationships in proteins. Our method employs a novel means of identifying homologous relationships. For input the method uses 400 component vectors, the elements of which are normalized bipeptide nearest neighbor statistics. Initial evaluation of the method will focus on the G-protein coupling receptor super-family. These proteins exhibit hydropathy plots predicting seven transmembrane spanning regions and many diverse biochemical functions are transduced by its members. Our data set for this family presently contains 157 members, ranging from the chloride pump bacteriorhodopsin to neurotransmitter receptors in the human. The anticipated methodological improvements may eliminate ambiguities in the clusters produced by Kohonen-like methods and provide more useful indicators of homology or molecular evolutionary patterns.