Computational Identification of Noncoding RNA Genes through Phylogenetic Shadowing

Although fairly accurate databases exist for protein-coding genes, little is known about another important class of genes known as noncoding RNA genes. These genes, which have been implicated in a wide variety of critical biochemical pathways including brain development [7] and viral defense [5], are not translated into polypeptides. Instead, their transcribed RNAs fold into stable, base-paired secondary and tertiary structures that confer catalytic ability. For the purposes of this paper, it is especially important to note that these secondary structures cause noncoding RNA genes to contain pseudo-palindromic sequences. Unfortunately, these pseudo-palindromic and other signals are not statistically sufficient for the computational identification of such genes [9]. Because they are difficult to detect even through biological techniques, it is important that accurate computational approaches be developed [10]. Although many heuristic and specialized methods have been suggested, comparative genomics approaches [8] have shown particular promise. However, even these approaches are primitive. For instance, current comparative genomics approaches are limited to two sequences, despite recent work showing the importance of using several related species [2]. Other problems include various heuristic approximations and poor scaling. Here, we briefly present a machine learning approach to genome-wide noncoding RNA gene prediction with multiple sequence alignments. The algorithm we describe scales linearly with respect to the length and number of genomes. More importantly, the approach is statistically sound, and allows the direct computation of probabilities through modular protein-coding, noncoding RNA, and intergenic sequence models.