Genetic algorithms as a mechanism for modelling gene interactions using time-course data

Gene interaction models are weighted graphs derived from replicates of gene abundance time-course data, where each weighted edge is a probability of gene association. These interaction models are a tool to assist biological researchers in understanding gene relationships. Two new genetic algorithms, one fairly traditional and the other based on crossover with infrequent application of a chaotic mutation operator, are developed specifically to produce gene interaction models from sparse time-course abundance data. Both genetic algorithms evolve a new population from a current population of directed acyclic graphs, each representing a Bayesian model for possible gene interaction. The genetic algorithm fitness is the relative posterior probability that a Bayesian model fits the gene abundance replicates. These Bayesian likelihoods are computed using one of three analysis techniques: cotemporal, first order next state and second order next state. The weighted gene interaction models reflect the directed acyclic graphs and their likelihoods present in the final populations of numerous independent genetic algorithm executions. Using a simulated set of genes, these two genetic algorithms find the embedded signals and are consistent across analysis paradigms. Results from a set of biological gene abundance data, from Arabidopsis thaliana stimulated by the plant hormone auxin, are modeled.