Optimization of Gene Prediction via More Accurate Phylogenetic Substitution Models

Determining the beginning and end positions of each exon in each protein coding gene within a genome can be difficult because the DNA patterns that signal a gene’s presence have multiple weakly related alternate forms and the DNA fragments that comprise a gene are generally small in comparison to the size of the genome. In response to this challenge, automated gene predictors were created to generate putative gene structures. N SCAN identifies gene structures in a target DNA sequence and can use conservation patterns learned from alignments between a target and one or more informant DNA sequences. N SCAN uses a Bayesian network, generated from a phylogenetic tree, to probabilistically relate the target sequence to the aligned sequence(s). Phylogenetic substitution models are used to estimate substitution likelihood along the branches of the tree. Although N SCAN’s predictive accuracy is already a benchmark for de novo HMM based gene predictors, optimizing its use of substitution models will allow for improved conservation pattern estimates leading to even better accuracy. Selecting optimal substitution models requires avoiding overfitting as more detailed models require more free parameters; unfortunately, the number of parameters is limited by the number of known genes available for parameter estimation (training). In order to optimize substitution model selection, we tested eight models on the entire genome including General, Reversible, HKY, Jukes-Cantor, and Kimura. In addition to testing models on the entire genome, genome feature based model selection strategies were investigated by assessing the ability of each model to accurately reflex the unique conservation patterns present in each genome region. Context dependency was examined using zeroth, first, and second order models. All models were tested on the human and D. melanogaster genomes. Analysis of the data suggests that the nucleotide equilibrium frequency assumption (denoted as πi) is the strongest predictor of a model’s accuracy, followed by reversibility and transition/transversion inequality. Furthermore, second order models are shown to give an average of 0.6% improvement over first order models, which give an 18% improvement over zeroth order models. Finally, by limiting parameter usage by the number of training examples available for each feature, genome feature based model selection better estimates substitution likelihood leading to a significant improvement in N SCAN’s gene annotation accuracy