Gene Finding with Semi-Supervised Training: Integration of RNA-Seq and Protein Data

We present a new algorithm, which i/ utilizes evidence derived from RNA-Seq and protein mapping to assist self-training of ab initio gene prediction algorithm, as well as ii/ generates gene prediction across the whole genome consistent with external evidence and nucleotide compositional patterns detected ab initio.

At the training stage the algorithm performs iterative estimation of parameters of hidden semi-Markov model. RNA-Seq reads mapping of splice junctions is notoriously noisy (low specificity) while protein mapping of closely related protein sequences is patchy (low sensitivity). Still it is possible to effectively combine both types of mapping with self-training procedure: to filter out RNA-Seq derived false positive introns while filling in gene structure elements missed by protein mapping. At the last step of iterations, elements of gene structure delineated with high scores by RNA-Seq read and protein mapping are integrated into the final gene prediction.

Algorithm is robust with respect to assembly fragmentation, genome size, and variation in volume of RNA-Seq read and protein data.  The algorithm is a new addition to the GeneMark line of gene finders
(topaz.gatech.edu/GeneMark).