A novel min-cost flow method for estimating transcript expression with RNA-Seq

Abstract

BackgroundThrough transcription and alternative splicing, a gene can be transcribed into different RNA sequences (isoforms), depending on the individual, on the tissue the cell is in, or in response to some stimuli. Recent RNA-Seq technology allows for new high-throughput ways for isoform identification and quantification based on short reads, and various methods have been put forward for this non-trivial problem.ResultsIn this paper we propose a novel radically different method based on minimum-cost network flows. This has a two-fold advantage: on the one hand, it translates the problem as an established one in the field of network flows, which can be solved in polynomial time, with different existing solvers; on the other hand, it is general enough to encompass many of the previous proposals under the least sum of squares model. Our method works as follows: in order to find the transcripts which best explain, under a given fitness model, a splicing graph resulting from an RNA-Seq experiment, we find a min-cost flow in an offset flow network, under an equivalent cost model. Under very weak assumptions on the fitness model, the optimal flow can be computed in polynomial time. Parsimoniously splitting the flow back into few path transcripts can be done with any of the heuristics and approximations available from the theory of network flows. In the present implementation, we choose the simple strategy of repeatedly removing the heaviest path.ConclusionsWe proposed a new very general method based on network flows for a multiassembly problem arising from isoform identification and quantification with RNA-Seq. Experimental results on prediction accuracy show that our method is very competitive with popular tools such as Cufflinks and IsoLasso. Our tool, called Traph (Transcrips in gRAPHs), is available at: http://www.cs.helsinki.fi/gsa/traph/.

StringTie enables improved reconstruction of a transcriptome from RNA-seq reads

“…Note that if more than a certain percentage of the reads (by default 95%) aligned in a gene locus are multi-mapped, then StringTie will skip processing that locus. The ASG captures all possible transcripts that are consistent with the mapped reads 15,17 , where nodes in the graph correspond to contiguous regions of the genome that are uninterrupted by any spliced read alignment, and directed edges correspond to reads that align across two such nodes in the correct 5′ to 3′ order. Note that the nodes do not necessarily correspond to whole exons in the transcripts; they may be only partial exons, as illustrated by node 4 in Supplementary Figure 12 (see also node 1 in Supplementary Fig.…”

“…Although a growing number of methods have been developed to solve either the transcript identification problem (e.g., Trinity 7 , Oases 8 ), the expression quantification problem (e.g., RSEM 9 , eXpress 10 ) or both (e.g., IsoInfer 11 , Scripture 12 , Cufflinks 13 , SLIDE 14 , IsoLasso 15 , iReckon 16 , Traph 17 ), much work remains to produce consistent, highly accurate solutions. A recent study 18 of the current transcript reconstruction methods showed that even in cases where these methods identified all constituent exons of a transcript, they often failed to assemble the exons into complete isoforms.…”

“…A flow network is currently used for transcriptome assembly and quantification by only one other method, Traph 17 (note, Traph uses a fundamentally different flow network formulation; for example, it solves a minimum rather than maximum flow problem; see Supplementary Discussion for details). Figure 1 illustrates the main differences between StringTie, StringTie+SR, Cufflinks and Traph.…”

“…We evaluated StringTie when it was run without annotation, and compared its performance with Cufflinks 13 , IsoLasso 15 , Scripture 12 and Traph 17 , which along with MITIE 22 are the only transcriptome assemblers than can be run without annotation. We attempted to include MITIE, but excessive running time prevented inclusion; after one month of computation, MITIE did not produce results for any of the data sets included in this study.…”

Strain-aware assembly of genomes from mixed samples using flow variation graphs

“…Some of the challenges that have to be dealt with in viral quasispecies assembly can also be found in RNA transcipt assembly, where the goal is to reconstruct an unknown number of transcripts and predict the relative transcript abundances. Not surprisingly, many RNA transcipt assemblers define graph optimization problems similar to our flow formulation [21,22,23,24,25]. Although dealing with related problems, these methods cannot be applied in a viral quasispecies setting so easily: they require a collection of reference genomes representing all possible haplotypes as input, which is not available in our setting.…”