Comparison of Artificial Neural Network, Genetic Programming, and Mechanistic Modeling of Complex Biological Processes

Abstract

Artificial neural network (ANN), genetic programming (GP), and mechanistic modeling approaches were compared to model cometabolic enzyme kinetics of simulated degradation of ammonia and trichlorethylene (TCE) in a quasi-steady-state bioreactor. ANN and GP approaches effectively modeled complex processes involving decreases and increases in metabolic activity, and made subsequent predictions of system behavior and responses to perturbation. The mechanistic model used fundamentally derived rate equations governing the cometabolic processes and applied them for modeling. All three approaches successfully modeled datasets containing known amounts of random error up to 5%, and the ANN approach was able to cope effectively with data gaps associated with unavailability of data. A valuable benefit of ANN and GP modeling was that the models did not require underlying descriptions or a priori knowledge of the physical processes governing the system. The essential requirement of ANN and GP modeling was sufficient numbers of observations for extracting correlations. The necessary datasets along with incorporated random errors were generated by Monte Carlo simulation. Results indicated that the models accurately predicted system responses and that they hold promise for further applications.