(445b) Applications of Genetic Algorithms for Variable Selection in Pharmaceutical Development

Genetic algorithms GA have been demonstrated to be powerful and efficient tools in the optimization of variable selection for empirical model generation [1]. In pharmaceutical process development the problem of variable selection is encountered in two distinct applications. The application of PAT technology generally requires the development of a quantitative relationship between analyte concentration and a multi-wavelength spectral response. The determination of which wavelengths should be selected for an optimal model is a challenging and time consuming task. Another area in which variable selection is a key optimization task is the development of surface response curves which requires careful selection of the factor terms that should be included to optimize model accuracy and predictive power. This presentation discusses detailed applications of genetic algorithms in the generation of partial least square models for PAT applications and surface response models for process risk assessment evaluation. Generally the solution from a genetic algorithm is a family of genomes of similar fitness from which the appropriate model is subsequently selected. This presentation will also address strategies for optimal subject selection from the final population representing the GA solution. The results will be compared to models obtained via careful statistical analysis of the factor-response relationships.

[1] Leardi, R. Journal of Chemometrics 15, 559-569 2001