Selecting superior de novo transcriptome assemblies: Lessons learned by leveraging the best plant genome

Whereas de novo assemblies of RNA-Seq data are being published for a growing number of species across the tree of life, there are currently no broadly accepted methods for evaluating such assemblies. Here we present a detailed comparison of 99 transcriptome assemblies, generated with 6 de novo assemblers including CLC, Trinity, SOAP, Oases, ABySS and NextGENe. Controlled analyses of de novo assemblies for Arabidopsis thaliana and Oryza sativa transcriptomes provide new insights into the strengths and limitations of transcriptome assembly strategies. We find that the leading assemblers generate reassuringly accurate assemblies for the majority of transcripts. At the same time, we find a propensity for assemblers to fail to fully assemble highly expressed genes. Surprisingly, the instance of true chimeric assemblies is very low for all assemblers. Normalized libraries are reduced in highly abundant transcripts, but they also lack 1000s of low abundance transcripts. We conclude that the quality of de novo transcriptome assemblies is best assessed through consideration of a combination of metrics: 1) proportion of reads mapping to an assembly 2) recovery of conserved, widely expressed genes, 3) N50 length statistics, and 4) the total number of unigenes. We provide benchmark Illumina transcriptome data and introduce SCERNA, a broadly applicable modular protocol for de novo assembly improvement. Finally, our de novo assembly of the Arabidopsis leaf transcriptome revealed ~20 putative Arabidopsis genes lacking in the current annotation.