Invited Talk: Ab Initio Gene Finding Engines: What Is Under the Hood

摘要

I will revisit the statistical and computational foundations of ab initio gene finding algorithms that best fit current challenges in analysis of genomic data. With the number of new sequenced genomes rapidly growing, there is a need to generate high quality gene annotations in less time. In recent gene prediction competitions, the organizers described in great details the sets of experimentally confirmed eukaryotic genes that the contest participants were supposed to use for training statistical models, the key parts of ab initio gene finding algorithms. However, the gene prediction algorithm developed in our lab is only one of its kind that does not require a training set at all. It is using an unsupervised training approach and exhibits the same or better level of accuracy of gene identification as the algorithm trained on a sufficiently large training set. With more than 600 eukaryotic genome sequencing projects registered, as of February 2007, the self-learning gene finders become important tools able to accelerate extraction of biological information from newly sequenced eukaryotic genomes.