Acceleration of Covariance Models for Non-coding RNA Search

Stochastic context-free grammar (SCFG) based models for non-coding RNA (ncRNA) gene searches are much more powerful than regular grammar based models due to the ability to model intermolecular base pairing. The SCFG models (also known as covariance models) can be scored exactly using dynamic programming techniques. However, the computational resources needed to compute optimal scores using dynamic programming is too great for most applications. Pre-filtering of the database using regular grammar based models can lead to significant improvements in performance at little or no cost in terms of specificity or sensitivity. While pre-filtering is a major improvement, the algorithm is still way to slow. The use of an alternative search strategy for high scoring subsequences in the sequence database is explored in this paper. Rather than sequentially computing the best score at each database position and subsequence length as is done in the dynamic programming method, good suboptimal scores are found throughout the position and length search space and the search is expanded about these trial solutions.