Finding optimal pairs of patterns

We consider the problem of finding the optimal pair of string patterns for discriminating between two sets of strings, i.e. finding the pair of patterns that is best with respect to some appropriate scoring function that gives higher scores to pattern pairs which occur more in the strings of one set, but less in the other. We present an O(N-2) time algorithm for finding the optimal pair of substring patterns, where N is the total length of the strings. The algorithm looks for all possible Boolean combination of the patterns, e.g. patterns of the form p Lambda -q, which indicates that the pattern pair is considered to match a given string s, if p occurs in s, AND (sic)q does NOT occur in s. The same algorithm can be applied to a variant of the problem where we axe given a single set of sequences along with a numeric attribute assigned to each sequence, and the problem is to find the optimal pattern pair whose occurrence in the sequences is correlated with this numeric attribute. An efficient implementation based on suffix arrays is presented, and the algorithm is applied to several nucleotide sequence datasets of moderate size, combined with microarray gene expression data, aiming to find regulatory elements that cooperate, complement, or compete with each other in enhancing and/or silencing certain genomic functions.