Optimizing de novo assembly of short-read RNA-seq data for phylogenomics

Abstract

Background: RNA-seq has shown huge potential for phylogenomic inferences in non-model organisms. However, error, incompleteness, and redundant assembled transcripts for each gene in de novo assembly of short reads cause noise in analyses and a large amount of missing data in the aligned matrix. To address these problems, we compare de novo assemblies of paired end 90 bp RNA-seq reads using Oases, Trinity, Trans-ABySS and SOAPdenovo-Trans to transcripts from genome annotation of the model plant Ricinus communis. By doing so we evaluate strategies for optimizing total gene coverage and minimizing assembly chimeras and redundancy.

Results: We found that the frequency and structure of chimeras vary dramatically among different software packages. The differences were largely due to the number of trans-self chimeras that contain repeats in the opposite direction. More than half of the total chimeras in Oases and Trinity were trans-self chimeras. Within each package, we found a trade-off between maximizing reference coverage and minimizing redundancy and chimera rate. In order to reduce redundancy, we investigated three methods: 1) using cap3 and CD-HIT-EST to combine highly similar transcripts, 2) only retaining the transcript with the highest read coverage, or removing the transcript with the lowest read coverage for each subcomponent in Trinity, and 3) filtering Oases single k-mer assemblies by number of transcripts per locus and relative transcript length, and then finding the transcript with the highest read coverage. We then utilized results from blastx against model protein sequences to effectively remove trans chimeras. After optimization, seven assembly strategies among all four packages successfully assembled 42.9-47.1% of reference genes to more than 200 bp, with a chimera rate of 0.92-2.21%, and on average 1.8-3.1 transcripts per reference gene assembled.

Conclusions: With rapidly improving sequencing and assembly tools, our study provides a framework to benchmark and optimize performance before choosing tools or parameter combinations for analyzing short-read RNA-seq data. Our study demonstrates that choice of assembly package, k-mer sizes, post-assembly redundancy-reduction and chimera cleanup, and strand-specific RNA-seq library preparation and assembly dramatically improves gene coverage by non-redundant and non-chimeric transcripts that are optimized for downstream phylogenomic analyses.

Figure 1

Chimera compositions among assembled transcripts before post-processing. Oases MN: Oases-M merging single k-mer assemblies of 21, 31, 41, 51 and 61; MW: Oases-M merging single k-mer assemblies of 19–71, with increment of 2; Trans-ABySS MK: Trans-ABySS merging single k-mer assemblies of 21, 31, 41, 51 and 61.

Figure 2

Parameters for choosing the representative transcript for each locus in Oases. (A) Only considering transcripts that are longer than 0.3 of the longest transcript in the same locus; (B) only considering transcripts that are longer than 0.85 of the longest transcript in the same locus. A data point is plotted only when there are five or more loci of the same size in each data set.

Figure 3

Overall comparison among assembly strategies. (A) Number of transcripts in each category; and (B) percent reference coverage, redundancy and chimera rate among assembly strategies. Cap3: redundancy reduction using cap3; blast: trans chimera cleanup using blastx against model protein database; Oases MK filter: filter loci from Oases single k-mer assemblies by number of transcripts per locus at k = 21, 31, 41 and 51, with k = 61 not subject to filtering by number of transcripts per locus, before combining them. Oases MN: Oases-M merging single k-mer assemblies of 21, 31, 41, 51 and 61; MW: Oases-M merging single k-mer assemblies of 19–71, with increment of 2; SOAPdenovo-Trans contigs: combining contigs from SOAPdenovo-Trans single k-mer assemblies of 21, 31, 41, 51 and 61; Trans-ABySS MK: Trans-ABySS merging single k-mer assemblies of 21, 31, 41, 51 and 61; Trinity pickH: only keeping the transcript with the highest read coverage for each subcomponent; Trinity removeL: when there are two or more transcripts per subcomponent, remove the one with the lowest read coverage.