Hello everyone,
I am doing a phylogeographic/population genetic study on an island species with limited dispersal capabilities. I'm having difficulty determining between what is biologically realistic and bioinformatics error. I have ddRADseq data that came in 2 separate sequencing batches. The first batch consists of 96 samples and the second has 24 samples.
I ran a PCA for the first 96 samples and it displayed a certain clustering pattern, and when I included the remaining 24 samples the pattern remained largely the same, except the second batch samples formed tight clusters corresponding to the respective populations those samples are from (3 distinct clusters).
When I ran a PCA with just the second batch (24 samples) the clustering was almost exactly the same
I want to note that my pipeline consists of the STACKS protocol, and included additional filters from dDocent (from roughly 160,000 to 16,000 variant sites).
So my question is if there should be any further concern about batch effects considering the "investigation" I've conducted thus far.
Thank you all in advance,
Andre
Perhaps a little more detail would help. It seems based on your description the results are not far off for the two batches.
Could you provided images of what you are referring to? Could the 24 individuals just be "better" representatives of their populations? Were the batches prepared using the same protocol, sequenced on the same sequencer, etc.?
I would image the batch of 24 has higher coverage since there would be less individuals per sequencing run. Is this the case? That may explain your pattern as well: perhaps more data and better coverage in the batch of 24.
Are you using Stacks or dDocent to assemble and call? Are you using VCFtools as dDocent does to filter.
Cheers,
Justin
Hi Justin H. Bohling
To the extent of my knowledge the data was all prepared with the same protocol. And yes, the batch of 24 has much better coverage (around 4 times the coverage).
I used STACKS to assemble, call and filter. dDocent and VCFtools for additional filters. However, I did not use the entire protocol for dDocent.
- Below I attached a PCA with just the 24 samples (note the populations JVE, TOR, and STJ)
- Another from the full filtered data set
- The third I defined the populations by batch, instead of the sampling populations. This image has an additional filter from dDocent that compared the 2 batches (designated as populations). It went from 16,000 snps to ca. 4,000 snps.
Honestly, the pattern I'm seeing is biologically possible. I've heard the opinions of a few colleagues and they seem to believe I was stringent enough with my pipeline. I guess I'm looking for some outside confirmation at this point.
Andre
Thanks for providing the images Andre. I think the coverage is likely a big factor in the pattern: the populations from the higher coverage batch are clustering tighter. Along the first axis there is more of a gradient in the lower coverage batch, which could be due to natural processes such as IBD, but also could be due to lower coverage. Ultimately, it is important to frame what you see with whether the patterns make biological sense.
I commend you on making the effort to consider batch effects. Definitely something our research community (myself included) needs to be more cognoscente of.
Best of luck!
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Bioinformatics