A memory-efficient algorithm to obtain splicing graphs and de novo expression estimates from de bruijn graphs of RNA-seq data

With the advance of high-throughput sequencing, it is feasible to study entire transcriptomes through the use of de novo sequence assembly algorithms. A popular strategy of these algorithms is to first construct an intermediate de Bruijn graph structure that represents the transcriptome. An additional step is performed to obtain predicted transcripts from the graph. Since the de Bruijn graph contains all branching possibilities, we develop a memory-efficient algorithm to recover alternative splicing information from the graph. Although many of the smaller components of the de Bruijn graph can already represent alternatively spliced variants of genes, the graph still contains big tangles that need to be resolved. We remove complicated cycles in the de Bruijn graph, and extract acyclic components as splicing graphs. In almost all cases, each splicing graph represents alternatively spliced variants of only one gene. To allow the evaluation of differential expression levels of genes and splices when no prior information is available, we develop a de novo RPKM measure (number of Reads Per Kilobase of node per Million reads) within each node of a splicing graph independently for each library. Since there is no information about exons in a de novo assembly, we use reads that appear in the assembly instead of mapped reads, and we consider each node in a splicing graph instead of each exon. We implement the algorithm as a postprocessing module from the output of the Velvet assembler. We validate our algorithm by performing Drosophila transcriptome assembly using publicly available RNA-Seq libraries. Under a range of conditions, our algorithm recovers sequences and splice junctions more accurately than Oases or Trans-ABySS and is always less memoryintensive than Oases.