Implementation improvements to an rna pseudoknot prediction algorithm

The general problem of RNA pseudoknot prediction is computationally intractable. Most existing algorithms require the worst case CPU time O(N6) and RAM space O(N4) even for restricted pseudoknot categories. Such resource requirements make it infeasible to predict pseudoknots for RNA sequences of even a moderate length. This research work demonstrates two implementation techniques, memory mapping and parallel computation, to reduce resource usage in the algorithms by taking advantage of the ways the matrix is organized and computed. The techniques are applied to an automated and dynamic programming-based pseudoknot prediction algorithm we recently developed. Our experiments shows that savings in memory usage of approximately 97% and nearly 10-fold speedup (using 16 parallel processors) are achieved for the pseudoknot prediction tests conducted. Most existing RNA prediction algorithms are dynamic programmingbased and evolved from the CYK algorithm for stem-loop prediction; thus our techniques are general and applicable to these algorithms as well.