Optimizing genetic algorithm for motif discovery

Planted ( l , d ) -motif identification is an important and challenging problem in computational biology. In this paper, we present an original algorithm (GARPS) that combines Genetic Algorithm (GA) and Random Projection Strategy (RPS) to identify ( l , d ) -motifs. We start with RPS to find good starting positions by introducing position-weighted function and hash each of all l -mers in the input sequences onto the corresponding k -dimensional ( k < l ) subspace so as to deduce a set of candidate motifs. Then, we use the candidate motifs as the initial population of genetic algorithm to make series of iterations to refine motif candidates. Experimental results on simulated data show that GARPS performs better than the Projection algorithm and improves finding faint motifs. We also present experimental results on realistic biological data by identifying CRP binding sites in Escherichia Coli as well as PDR3 binding sites in the yeast Saccharomyces Cerevisiae. Furthermore, the implemented random projection algorithm RPS can be fine tuned for optimizing many heuristics methods starting from random seeds, such as simulated annealing, Gibbs sampling.