Yes, I think that you can. When you use a reference you are defining a population ("universe", "genome", "transcriptome") to compare your sample ("expressed gene in leaves", "metabolism related genes"...). The only think that you need to keep in mind is that your expression values will be referenced to your reference population and they can be compared with other samples that are using the same reference. It is common that after an transcriptome de-novo assembly you filter short scaffolds/contigs and contaminations, so really you are not using "all the transcripts".

Thanks for your answer. That was what I meant the whole assembly.

But I'm still a little confused that the expression value can be affected by the reference size. Is it because when whole assembly is used, there is multiple assemblies for one gene because of assembling errors, there is alternative splicing, and multiple homologous transcripts for one gene? But aren't the reads allowed to be mapped o more than just one assemblies?

I guess, as Aureliano says, in a way we all use only a subset as a reference. I generally mask all abundantly expressed and varying genes like rRNA, tRNA etc. for RNAseq analysis so that it does not bias my results . My answer is based on my experience with Cufflinks (not denovo). The expression values of course are very much dependent on the reference you use, as the total mapped reads on your reference transcripts is the denominator for your FPKM values for a gene. For a typical mRNAseq experiment, I would recommend using a global reference (all expressed genes) for the following reasons

1) It is more robust and sensitive as it normalizes your data over the entire range of mRNA expression you see in your library.

2) Your downstream analysis can be more flexible. For example, say you analysed differential expression of gene A with genes A to F as reference, you might later want to know its correlation with differential expression of gene X, which now are not comparable.

3) As you mentioned, also the case of reads mapping to multiple areas in the genome, though I would not worry about alternative splicing because the expression values are distributed across the isoforms only within a particular locus and are not affected by the reference.

The expression is affected, not for the number of genes/transcripts of the reference, by the number of mapped reads (generally we use RPKM that normalize the expression values per million of mapped reads), so more sequences that you have in your reference, more chances that a read will have to be mapped (supposing that your reference contains the transcripts represented in the sample).

When you have a de-novo assembly where one gene is represented by multiple sequences (for example alternative splicings when you are using Trinity) a read can map in all of them producing a bias in the expression. For example, if you have a gene X with three exons A, B and C and two alternative splicings AB and AC. The B and C reads will map specifically to B and C exons, but the A reads will map EQUALLY to both transcripts (AB, AC). Usually mappers will assign these reads at random (http://bowtie-bio.sourceforge.net/bowtie2/manual.shtml#randomness-in-bowtie-2), so it can introduce a bias. In the example, if AB has 200 reads (A=100, B=100) and AC has 20 (A=10 and C=10) the mapping will introduce a bias because it will map B=100, C=10 and A=100+10=110 divided between two splicings = 55 each, so AB=155 and AC=65. So the solution (for now, for de-novo assemblies, cufflinks knows about alternative splicings and assign expression measures according that: http://cufflinks.cbcb.umd.edu/faq.html#cuffdiff) is to use scripts to measure the expression of the gene and not of the transcripts (Gene X in this case an not AB and AC trancripts). I think that Trinity has an script to do it if you maintain the names (compX_cY_seqZ) in the sequences (http://trinityrnaseq.sourceforge.net/analysis/abundance_estimation.html). Still genes with multiple copies from a polyploidization event or a tandem duplication could be a problem.

For genes that have been assembled in two different pieces (for example gene Y with three exons, ABC have been assembled in two transcripts AB and C) you will see two separate expression values, but because the expression is normalized by the length of the transcript it should not be major difference (except because you will have two genes, probably with the same functional annotation, with similar expression)

I think that Mr. Ramaprasad made good observations why it could be better to use the whole dataset for downstream comparisons.

Thank you so much, Mr. Bombarely and Mr. Ramaprasad. I have much better understanding now. Though a small subset could be used as a ref, it is more robust to use the entire assemblies and easier for downstream comparisons. Alternative splicing could bring bias because of the random assigning of the aligner.

Mr Bombarely, about the same example you gave for the alternative splice, although I understand now how the AB and AC expression values are biased, it is still a little different from I thought how the aligner worked.

This is what I was thinking. The aligner will map all the reads of A to both AB and AC, so AB=110+100=210, and AC=110+10=120, then it will be divided with the transcript length and the total reads.

I am not sure, AB and AC are separate sequences. The aligner only can assign reads to AB or AC (as primary alignment) but not both at same time. In your dataset you have only 110 A reads in total. A random assignment will give ~55 A reads to AB and ~55 A to AC. The expression tools doesn't use the secondary alignments (as I know).

I forget to say that reads that maps in two (or more) places at the same time can be filtered using -m option in bowtie (for example bowtie -m 1 means report unique reads with unique alignments)

Can -m also be set to more than 1? But theoretically one read should just occur once and be mapped once. Then why there is such an option in Bowtie to allow mapping to multiple places?

Yes, -m can be set to more than 1. A read can be mapped in multiple places because there are gene duplications and gene families with some regions that can be almost identical. The mapping algorithm for these tools doesn't give you best map of all the genome/transcriptome, it gives you the best map of all the maps that it has found with the search parameters (for more details take a look to the -k and -a options in bowtie).

Thank you, sir