Building better genome annotations across the tree of life

Recent technological advances in long read DNA sequencing accompanied by dramatic reduction in costs have made the production of genome assemblies financially achievable and computationally feasible, such that genome assembly no longer represents the major hurdle to evolutionary analysis for most non-model organisms. Now, the more difficult challenge is to properly annotate a draft genome assembly once it has been constructed. The primary challenge to annotations is how to select from the myriad gene prediction tools that are currently available, determine what kinds of data are necessary to generate high quality annotations, and evaluate the quality of the annotation. To determine which methods perform the best and determine whether the inclusion of RNA-seq data is necessary to obtain a high-quality annotation, we generated annotations with 10 different methods for 21 different species spanning vertebrates, plants, and insects. We found that the RNA-seq assembler Stringtie and the annotation transfer method TOGA were consistently top performers across a variety of metrics including BUSCO recovery, CDS length, and false positive rate, with the exception that TOGA performed less in plants with larger genomes. RNA-seq alignment rate was best with RNA-seq assemblers. HMM-based methods such as BRAKER, MAKER, and multi-genome AUGUSTUS mostly underperformed relative to Stringtie and TOGA. In general, inclusion of RNA-seq data will lead to substantial improvements to genome annotations, and there may be cases where complementarity among methods may motivate combining annotations from multiple sources.