The paper has many comparisons with Trinity and Oases. We wrote a lot about those two transcriptome assemblers two years back. It is unfortunate that Daniel Zerbino was forced to stop developing Oases further, because the assembler was pretty good except for its humongous memory needs.
Getting back to SOAPdenovo-Trans, those looking for ideas for future development should jump straight to discussion section.
Sequence assembly on real world datasets has always required a lot of minor algorithmic developments to produce the best results. There is no one magic idea that solves all of the problems. In this spirit, SOAPdenovo-Trans combined insights from both Trinity and Oases, merged them with ideas developed for the genome version of SOAPdenovo2, and then added a few insights of our own, to produce an algorithm that is demonstrably superior to the previous. This however is unlikely to be the last word in transcriptome assembly. We tried one of the reference-based assemblers, Cufflinks, and recovered even more full-length transcripts than SOAPdenovo-Trans. It is unclear just how much of this improvement can be replicated without recourse to a reference genome, but the results suggest that there is information in these datasets that, perhaps, with the right algorithm can be recovered.
For example, a multiple k-mers strategy may improve transcriptome assembly. Current multiple k-mers assembly strategies generally fall into one of two categories - (a) After running different values of k-mer assembly separately, merge these assemblies into one final set. This strategy may construct a more complete transcript set but may also introduce redundancy. (b) Iterate different k-mer de Bruijn graph assemblies during contig construction. This strategy potentially makes the best use of reads and paired-end information. Whether or not it is worth the effort to develop such algorithms depends in part on continuing progress in sequencing technology, since if the promised improvements in read lengths materialize, the nature of the problem will change radically.
Related Seqanswers threads -
The comment by dongilbert is helpful.
Find here a summary of my uses of your RNA transcript assemblers, comparing with what I see as 3 good and improving programs for this: Velvet/Oases, Trinity and SOAPdenovo-Trans.
Very briefly, three de-novo assemblers tested here are closely ranked, and ranking depends on the particular species and data set used.
Locust insect: Velvet/O > Trinity > SOAPTrans
Cacao plant: SOAPTrans > Trinity > Velvet/O » Cufflinks
Daphnia waterflea: Velvet/O > SOAPTrans > Trinity » Cufflinks
SOAPTrans in particular can assembly better, quicker, with less memory use than the other two. It can also fail inexplicably, or do worse than the others.
My recommendation is to try these three, see which works for you and if possible use them all and extract the best subset by some gene evidence criteria (like homology, high coding ratio, …).
Many people reported odd behavior (i.e. failing inexplicably, as mentioned by dongilbert).
While Oases found massive sequences that have possible alternative splice products, SOAPdenovo-trans did not find a single one. I used 12 different k-mers from 19 to 89, e 1,3,5 and d 1,3,5 with all combinations. I allowed up to 10 alternative splicing products.