I've assembled RNAseq data into a transcriptoms using Trinity. There's a option to keep only the longest isoforms for each transcripts and it lead me to wonder how it deals with duplicated genes (-->transcripts). I haven't found an anwser in the original paper. Althought I reckon paralogs and isoforms are two distinct features I don't see a way to make a distinction when you only have trasncripts.
It depends. If you are performing Illumina based RNAseq, which is likely, it is definitely tricky, You should be able to do this in an ideal world because,
- a paralogue will contain SNPs, i.e. there will be nucleotide variation between paralogues, but the mRNA can of an identical length
- isoform will also have identical sequence but the length of the mRNA will vary depending on what initiation codon it used
I sort of agree however because,
- there will be 'grey-zones' where paralogues are different sizes and VCF/ SNP calling is not perfect and ...
- assembly of paralogues can be difficult, risking collapsing a paralogue into a single mRNA, but it is not impossible depending on the level of SNP variation
- assembly of isoforms would be really hard
I personally wouldn't know how point 3 is solvable, but I guess it is if your isoform is relatively small, i.e. well within the read-length of Illumina output.
If a reference genome assembly exists, then reference-guided assembly with Trinity will keep paralogs, because it treats each genomic region as a different assembly set.
If a reference genome doesn't exist, paralogs will probably result in the construction of non-representative "isoforms" containing variants present in multiple genes. Inchworm does a greedy assembly of reads, so if a chimeric transcript can be constructed from the reads (which is likely with paralogous genes), then it might end up being fed to the subsequent steps. Whether that is actually emitted at the end by Butterfly will mostly depend on how the paths through the putative assemblies are traversed, which probably depends on the tolerance of Bowtie to errors.
If paralogous transcripts are sufficiently similar that the differences are within the error tolerance for assembly, then I would expect that they'd be collapsed into a single transcript.