Improving Maker Gene Annotations in Grasses through the Use of GC Specific Hidden Markov Models

The applicability of whole genome sequencing depends on accurate gene prediction and annotation. Hidden markov models (HMMs) are used in the structural annotation process to algorithmically predict genes and are dependent on training-sets of genes to identify gene structure features.  Typically, HMMs are trained on a random selection of genes that should represent the total variation in gene structure features found in a given genome. One example of this variance is GC content, which differs across plant species but is bimodal in many grass genomes.  This bimodality presents a challenge in accurately predicting gene structure using HMMs.  We hypothesized that genes with extreme GC contents would not be well predicted with HMMs trained on a random selection of genes.  To address this, we developed training gene sets with both high and low GC content from the genomes of sorghum (Sorghum bicolor) and Brachypodium (Brachypodium distachyon) for the ab initio prediction programs SNAP and AUGUSTUS.  The HMMs developed through this work were employed by the MAKER annotation pipeline to predict new gene models for these species. The gene predictions identified by high, low and random GC HMMs were then combined through MAKER to create new annotations.  Using this approach, gene prediction accuracy was improved for both species.  This work has enhanced the structural annotations of both the sorghum and Brachypodium genomes and has created a novel approach to GC-specific HMM development for use with the MAKER annotation pipeline.