Dr. Koza applies GP with architecture-altering operations to determine if a given segment of protein is transmembrane. The amino acids sequence in alphabetical form represents the protein segments. A protein segment is usually transmembrane (embedded within cell membrane) if it consists of hydrophobic (water-hating) amino acids.
This problem has been attempted by numerous people using different techniques. One of which is implementation of algorithms that classify transmembrane protein segment based on biochemical properties of the amino acids. Another is the use of GP with pre-defined architecture of the classifying program to be evolved.
Dr. Koza goal is to find a classifying program using the architecture-altering operation. The starting condition for this classifying program is minimal. Coniditions such as number of automatically defined functions, or number of arguments for each funcion are initially unspecified. The evaluated classification by this program is compared with the known classification. The correlation between the two determines the fitness.
The best-of-run program produces an out-of-sample correlation of 0.9681, and in-sample error rate of 3%, and an out-of-sample error rate of 1.6%. The out-of-sample error rate is smaller than most previously attempted techniques, which has an error rate ranging from 2.5% to 2.8%.
Dr. Koza shows in this paper that genetic programming with architecture-altering operation and minimal pre-defined initial conditions has comparable performance as human generated solutions on non-trival problems.