Pooling Samples

[TheSeqGeek]

I have seen some papers on pooling and apparently it's a statical mouth watering conversation so let me explain how this "pooling" would be done before anyone raises hell.


The experiment is RNA-seq to determine differentially expressed genes between wild type bacterial strain and a mutant. Call them WT and M

You need three biological replicates for both WT and M. Let's talk about just WT for now because it is the same for M.

What you do is take three colonies from WT and grow overnight cultures. Call them WT1, WT2, WT3.

Next day take the three overnight culture and reinnoculate them but this time make two culture for each overnight culture so you end up with six total where only three are from different colonies.

Call them WT1(a), WT1(b), WT2(a), WT2(b), WT3(a), WT3(b)

Grow them to a desired OD and then pool WT1(a+b), WT2(a+b), WT3(a+b) so you end up with three biological replicates coming from two independent growths.

What are your thoughts on making biological replicates this way?

[bastianwur]

Stupid question: Why not just sequence all 6 samples, instead of pooling half of them?
(guess money)

[gringer]

Money is an odd reason unless coverage is an issue. You can barcode samples and multiplex in the same run, and the sequencing will cost the same as sequencing a single sample. Library preparation will be a little more expensive, but not much.

[bastianwur]

Thought a bit about the problem (keep in mind I barely know anything about statistics).
You do multiple samples to detect + correct errors.
1 error in 3 will allow you to discard the wrong sample.
Or 2 out of 6.
With this pooling, the following situations can arise:
1 wrong in 6. -> 1 good + 1 bad will get pooled. You'll have a harder time figuring out what's exactly wrong with the pooled sample, since it's not clearly wrong.
2 wrong in 6.
-> a) 2 bad ones get pooled. Same situation than 1 of 3.
-> b) 2 bad get pooled with 2 good. You end up discarding the good sample, since it's clearly different from the 2 others.

So, my opinion: Don't do it, makes work harder and increases the chance of wrong results.

[DonDolowy]

I would definitely sequence at lower depth and sequence all of them for the same amount of money as if you were to pool them and sequence deeper (taken that you are just interested in gene expression).

[TheSeqGeek]

I guess my reasoning was this.

If I sequence three replicates that will show the variability between the replicates

If I pool a bunch together to make three replicates that came from pools of other replicates this would cut down on variability and I would be sequencing the average rather than absolute from one snapshot.

In my way I was thinking you could have multiple snapshots that have been averaged which can then be sequences.

I don't know if this is clear or not though... But anyway something like this


You have three biological replicates to be sequenced that were collected from different colonies and different days.

Each replicate came from 3 or more growths of that biological replicate

If it were a tree diagram you would only see three stems but under those stems were 9 more stems that came to make the three.

[gringer]

You can always pool afterwards in-silico by combining reads from multiple samples. What you can't do is un-pool samples that have been pooled (or barcoded with the same index).

We needed to pool 15 mice in our experiments to get enough RNA for our sequencing provider, but would have really liked to be able to do a single mouse per index and investigate how pooling would affect results. Pooling removes your ability to assess variation in expression between samples, restricting you to assessing variation between populations of samples.

[TheSeqGeek]

You can always pool afterwards in-silico by combining reads from multiple samples. What you can't do is un-pool samples that have been pooled (or barcoded with the same index).

We needed to pool 15 mice in our experiments to get enough RNA for our sequencing provider, but would have really liked to be able to do a single mouse per index and investigate how pooling would affect results. Pooling removes your ability to assess variation in expression between samples, restricting you to assessing variation between populations of samples.

that's true and I agree completely... If this was a free experiment I would have run 100 samples per condition