How Much Should We Rely on Computer Annotation for Identifying Genes in New Genomes?

With the advent of large numbers of low quality “draft” genomes the research communities in non-model organisms rely heavily on computationally annotated genes. The huge variation in the quality of these draft genomes and the predicted gene lists affects research and development. The draft Actinidia (Kiwifruit) genome was a big step forward in our understanding of this important horticultural crop. Typical of draft genomes, a large proportion of the sequence was not assigned to chromosomes, and the gene list missed key published genes. To address this we developed an international consortium firstly to improve the genome construction and secondly to manually annotate the whole genome.

Manual annotation was performed using the WebApollo software and employed tracks of computationally predicted genes, EST sequences and mRNA-seq data from a variety of different tissues. In total 93 researchers from around the world produced evidence-based gene models for over 37000 genes. Our subsequent analysis showed that the majority of computational gene predictions in kiwifruit do not align to the manually edited genes. In particular, the computationally generated genes did not accurately predict intron-exon structures or translation start and stop regions. Non canonical splice sites were particularly poorly annotated. From this exercise we now have a second generation genome to start to understand the biological processes associated with this plant and a method for the scientific community to improve the accuracy and utility of draft genomes.