Prediction of Biochemical Reactions Using Genetic Programming

To comprehend dynamic behaviors of biological systems, many models have been proposed. These models need literatures data to represent detailed and accurate dynamics. However, those data are sufficient only in few cases. To solve this problem, many techniques have been developed, including Genetic Algorithm (GA). Those methods require defining equations before predicting biological systems. To consider the case where even equation could not be obtained, we employ the Genetic Programming (GP) that was studied as a method to predict arbitrary equation from time course data without any knowledge of the equation. However, it is difficult for conventional GP to search the equations with high accuracy because our target biochemical reactions include not only variables but also many numerical parameters. In order to improve the accuracy of GP, we extended elite strategy to focus on numerical parameters. In addition, we added a penalty term to evaluation function to save the growth of the size of tree and consuming calculation time. By applying our improvements, we were able to predict biochemical reactions whose dimensions of variables were strictly the same as those of originals. The relative square error of predicted and given time-course data were decreased from 25.4% to 0.744%. Moreover, in experiments to validate the generalization ability of the predicted equations, we successfully decreased the relative square error of the predicted and given time-course data from 25.7% to 0.836%. The results of our numerical experiments indicate that our method succeeded in predicting approximation formulas without any definition of equations with reduced square error.