We are strongly interested in assembly a good transcriptome of reference for a non-model organism and build a local database. We have sequenced the same individual with Illumina (150 millions of pair-end reads) and PacBios IsoSeq v3 (2 SMRT cells, one for shorter transcripts, shorter than 5kb and other for longer transcripts, up to 5kb).

To process long-reads, I have followed the PacBio IsoSeq pipeline proposed in their Github repo (https://github.com/PacificBiosciences/IsoSeq). The final result was removing 70% of the long-reads. Is that normal?

Using this data, I have assembled the transcriptome using only short reads and another combining long- and short-reads. In the end, I have not found any difference... Approx. The same N50, the same number of transcripts assembled, the rate of misassemblies... Does anyone know if PacBio data does not worth for transcriptome de novo assembly?

  • $\begingroup$ Could you clarify your question please ... " Does anyone know if PacBio data does not worth for transcriptome de novo assembly?" I think there is a spelling error which confuses the meaning. $\endgroup$ – M__ Oct 17 '20 at 0:07
  • $\begingroup$ Biostars cross-posting. $\endgroup$ – zorbax Oct 17 '20 at 0:11
  • $\begingroup$ When are you saying removing, in which stage it got removed? $\endgroup$ – user3377241 Oct 17 '20 at 5:51

A few comments:

Never use N50 as a metric especially for transcriptomes. It has some semblance of relevance for genome assembly, but all that is void for a transcriptome with inherently dynamic lengths.

At the end of your IsoSeq pipeline, you should have (ideally) full-length transcripts. Have you considered that you don't miss much in your IsoSeq data? The biggest gain using IsoSeq is that it should have fewer non-existent isoforms assembled than a De-Bruin graph approach.

Let's say you have the following locus in your genome:

5'                                                                 3'
Upstream      E1     Intron     E2    Intron     E3      Downstream

There are theoretically 7 isoforms:

  • E1
  • E2
  • E3
  • E1,E2
  • E2,E3
  • E1,E3
  • E1,E2,E3

A short read assembler has to look at evidence that connects exons to see which isoforms are represented. Let's say you have the following PE reads:

5'                                                                 3'
               -->..............<--- Read 1
                                 -->...............<-- Read 2
                 --___________-->..<-__________--- Read 3

Read 1 and 2 both have a mate in either exon, so that's good evidence that they come from a single transcript. You can also have reads like Read 3, which spans two exon-exon junctions, again good evidence. There are some problems, however, if things get more complicated: if the cumulative length of your exons is greater than the fragment length of your short reads, or you have many exons in a gene, you will never get a single read pair that confirms the existence of some transcripts. The same goes for the opposite. If you find a read pair on a single exon, that is not evidence that this single exon is a transcript in and of itself.

Here's where short read assemblers take different approaches. Some make educated guesses about the existence of some transcripts, e.g. by utilizing expression counts in an EM or dynamic programming approach, some just enumerate possibilities.

And that's where IsoSeq comes in. A single, long read from a single transcript helps a lot to add evidence and make sure your transcripts actually represent the transcripts that exist in your organism. So, often, your IsoSeq should be fewer transcripts than your short read assembly.

In essence, what you should do, is take the IsoSeq and curate those as a set of high-confidence transcripts, i.e. those that you have observed from a single read source. That's not to say that some transcripts that exist only in the short read assembly aren't real, just that you have less definitive evidence for them.

If you are mostly interested in the gene level, that doesn't make a huge difference, but once you need to look into transcript level usage/expression etc it helps a lot to be able to eliminate low confidence transcripts.

Use tools better suited for transcriptome assembly evaluation: Detonate, TransRate and (Busco) and compare your assemblies using those.

  • $\begingroup$ Hi! Thanks, I know that N50 is not the best proxy to evaluate my assembly. I've used BUSCO, transrate and also E90N50 as proxy to evaluate the differente assemblies and the result with and without long-reads is the same. may it be possible assembled all the isoforms using only short-reads? $\endgroup$ – Manuel Sánchez Mendoza Oct 18 '20 at 1:56
  • $\begingroup$ That is entirely possible! I edited my answer to highlight the real advantage of isoseq data in more detail, as imo its greatest strength is that you have evidence for the existence of transcripts from a single-read source. It's a metric that is hard to benchmark so I'm not aware of any tools to assess that besides visual checks $\endgroup$ – Bastian Schiffthaler Oct 18 '20 at 11:18

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.