Proper design for DESeq2 and other RNAseq general questions

Question

[Stef]

Hello, I am new to the RNAseq. I sequenced 20 libraries from Eucalyptus. They are all different tissues: early flower, late flower pollinated, late flower unpollinated, early seed capsule, late seed capsule, mature pollen, and mature leaf. I have three biological replicates for all the tissues except pollen; I only have two for pollen.

I started my analysis using Hisat2, stringtie, and ballgown. But, I have not figured out how to do multiple pairwise comparisons in ballgown yet. Any suggestions?

Now, I have moved to using DESeq2 instead of ballgown. However, I am not sure how to design the formula for DESeq2 since I only have two replicates for pollen. The design: tissue*tree gets an error because "the model matrix is not full rank". I read on a blog post to combine tree and tissue into one. However, that makes it seems like there are no replicates. I am not sure how to go forward without replicates.

Also, I am concerned that my counts are not normalized well-enough by calculating FPKM values. My reads are single-ended. Is it possible to generate RPKM values using ballgown or DESeq2?

Last, is STAR considered a better aligner than Hisat2?

Thanks! Stef

[piyushjo]

Stringtie has a prepDY.py script that takes .gtf and converts into reads. I am not entirely sure of the algorithm used, but I use those counts for DESeq2 analysis.

[Carlo Yague]

Could you paste your Design matrix ? From your post, it is unclear whether the samples came from the 3 (2) same trees or from 20 different trees. If they come from different trees, then, as h.mon suggested, you should only consider the "tissue" factor in your formula.

[Stef]

So they come from the same trees (tree one, tree two, and tree three). I apologize about not being clear:

id tree tissue age type group

eg1MT1 one early_capsule early flowering one.early_capsule

eg1MT2 two early_capsule early flowering two.early_capsule

eg1MT3 three early_capsule early flowering three.early_capsule

eg1WT1 one mature_flower mature flowering one.mature_flower

eg1WT2 two mature_flower mature flowering two.mature_flower

eg1WT3 three mature_flower mature flowering three.mature_flower

eg3MT1 one mature_capsule mature flowering one.mature_capsule

eg3MT2 two mature_capsule mature flowering two.mature_capsule

eg3MT3 three mature_capsule mature flowering three.mature_capsule

egAT1 one early_flower_pol early flowering one.early_flower_pol

egAT2 two early_flower_pol early flowering two.early_flower_pol

egAT3 three early_flower_pol early flowering three.early_flower_pol

egL1 one mature_leaf mature vegetative one.mature_leaf

egL2 two mature_leaf mature vegetative two.mature_leaf

egL3 three mature_leaf mature vegetative three.mature_leaf

egNPT1 one early_flower_unpol early flowering one.early_flower_unpol

egNPT2 two early_flower_unpol early flowering two.early_flower_unpol

egNPT3 three early_flower_unpol early flowering three.early_flower_unpol

egP1 one mature_pollen mature flowering one.mature_pollen

egP2 two mature_pollen mature flowering two.mature_pollen

[Carlo Yague]

Ok, now I understand why tissue*tree is not working. tissue*tree is equivalent to tissue + tree + tissue:tree (interaction). However, to be able to compute the interaction term, you would need at least two exact replicate of each (tissue-tree) couple. With your data, the best you can do is use a tissue + tree model, without the interaction term.

Answer 1

[h.mon]

From the description of your experiment, I think your design should be only tissue, why are you including tissue*tree? Having two replicates for pollen is not ideal, but shouldn't cause errors.

Also, I am concerned that my counts are not normalized well-enough by calculating FPKM values. My reads are single-ended. Is it possible to generate RPKM values using ballgown or DESeq2?

Your counts won't be properly normalized with both FPKM and RPKM (and, for single-ends reads, RPKM is the same as FPKM). A better within sample normalization is TPM, which ballgown calculates. For DESeq2, you don't need these normalizations, it expect raw counts as input.

Last, is STAR considered a better aligner than Hisat2?

I think so, but not better enough to warrant realigning your reads

[Stef]

Ok. I will stick to DESeq2 then because I am not sure if ballgown does RPKM. I think they only normalize with FPKM or gene coverage.

Answer 2

[swbarnes2]

20 libraries So that's 7 trees, six tissues from each, with one of the pollen samples dropping out for some reason, right?

However, I am not sure how to design the formula for DESeq2 since I only have two replicates for pollen.

That's not the problem. Your design is just "tissue". It's all you can do. You can't model differences between trees with only a single tissue sample per tree. If you'd taken 4 flowers of each type from each tree, then you could.

And pretty much no one uses FPKM anymore. DESeq2 takes raw gene counts. It will do its own normalization.