Gene Structure Prediction – Chasing the Horizon

Creating methods of automatic genome annotation will resolve bottlenecks in getting biological knowledge from volumes of eukaryotic genomic sequences generated by NGS technolgies. One of the stumbling blocks is the training of indispensible ab initio gene finders able i/ to identify genes missed by other methods, such as mapping of relevant protein sequences and available transcripts (produced by assembling of RNA-Seq reads), as well as ii/ to improve the results of mapping. Earlier we proposed and implemented iterative methods of unsupervised training (self-training) of eukaryotic gene finders^1,2  and demonstrated that accuracy of such algorithms is high for novel fungal genomes as well as for novel plant and animal genomes with compact size (below 500Mb).  
It was noticed that self-training meets difficulties if genome assembly is highly fragmented as well as if the average intron length is large (> 500nt), which is characteristic for large size genomes. Integration of transcript and protein data could provide a mechanism for stabilizing self-training, however, this integration is not trivial due to abundant expression of non-protein-coding RNA, noisy splicing as well as differences in exon-intron structures between genes coding for homologous proteins. We present here new algorithms for RNA-seq read mapping³ and for homologous protein mapping to genome.  Combination of these algorithms with self-training improved accuracy of parameter estimation and, subsequently, improved accuracy of gene prediction in eukaryotic genomes with complex features.  

1. Lomsadze et al.,  Nucl. Acids Res. 2005.
2. Ter-Hovhannisyan et al. Genome Res. 2008.
3. Burns et al. Nucl. Acids Res. 2013.