Symbolic regression via genetic programming in the optimization of a controlled release pharmaceutical formulation

Symbolic regression via genetic programming (GP) was used in the optimization of a pharmaceutical zero-order release matrix tablet, and its predictive performance was compared to that of artificial neural network (ANN) models. Two types of GP algorithms were employed: 1) standard GP, where a single population is used with a restricted or an extended function set, and 2) multi-population (island model) GP, where a finite number of populations is adopted. The amounts of four polymers, namely PEG4000, PVP K30, HPMC K100 and HPMC E50LV were selected as independent variables, while the percentage of nimodipine released in 2 and 8 h (Y2h, and Y8h), respectively, and the time at which 90% of the drug was dissolved (t90%), were selected as responses. Optimal models were selected by minimization of the Euclidian distance between predicted and optimum release parameters. It was found that the prediction ability of GP on an external validation set was higher compared to that of the ANNs, with the multi population and standard GP combined with an extended function set, showing slightly better predictive performance. Similarity factor (f2) values confirmed GP's increased prediction performance for multi-population GP (f2 = 85.52) and standard GP using an extended function set (f2 = 84.47).