COUPLED GENETIC ALGORITHM/KOHONEN NEURAL NETWORK (GANN) FOR PROJECTION OF THREE-DIMENSIONAL PROTEIN STRUCTURES ONTO THE PLANE

Abstract

An algorithm is presented for projecting — at the amino acid level — the three-dimensional crystal structure of a protein molecule onto a planar surface. The scheme is topologically consistent: if two amino acid residues are closely juxtaposed in three-dimensional space, they remain so upon projection. Through such projections, a single resulting picture captures the spatial relations amongst a protein molecule's amino acids. Operationally, a genetic algorithm is used to "evolve" a parameter set which serves as input for a self-organizing Kohonen neural network responsible for the projection itself. A fitness function characterizing the quality of the projections is defined and maximized via the genetic algorithm. The workings of both the genetic algorithm and neural network are discussed in detail. In this work, we seek to optimize projections resulting from the inherently "frustrated" task of collapsing a space-filling collection of amino acid residues onto a simpler surface. Ultimately, the chosen application is a testing ground for establishing the success of our coupled genetic algorithm/Kohonen neural network scheme which can easily be adapted for other uses.