How to correct for batch effect in microarray meta-analysis

Question

[BrunoGiotti]

Hi all,

I hope some of you can shed light on this conundrum i have been having for a while. So, i have 8 studies downloaded from GEO including macrophages derived from different mouse tissues as the table shows:

Batch     Subtype
1         BMDM
1         BMDM
1         BMDM
1         liver
1         liver
1         liver
1         lung
1         lung
1         lung
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         monocyte
1         peritoneal
1         peritoneal
1         peritoneal
1         peritoneal
1         peritoneal
1         peritoneal
1         peritoneal
1         peritoneal
1         peritoneal
2         monocyte
2         monocyte
2         monocyte
2         TAM
2         TAM
2         TAM
2         TAM
2         TAM
3         peritoneal
3         peritoneal
3         peritoneal
3         peritoneal
3         peritoneal
3         peritoneal
3         peritoneal
3         peritoneal
4         peritoneal
4         peritoneal
4         peritoneal
4         peritoneal
4         peritoneal
4         peritoneal
4         peritoneal
4         peritoneal
5         BMDM
5         BMDM
5         BMDM
5         BMDM
5         BMDM
6         BMDM
6         BMDM
6         BMDM
7         BMDM
7         BMDM
7         BMDM
7         BMDM
7         BMDM
7         BMDM
8         TAM
8         TAM
8         TAM

I had removed many samples in order to be able to run the batch correction with ComBat. In specific, i had to remove every biological group which was not represented by at least 2 batches (I think) otherwise it throws an error. However the liver and lung samples are only in batch one and still the file was ok so i wonder how these samples were treated?. Ultimately my question is what is supposed to be a sensible setting to run through a batch correction algorithm? Is this one above a sensible one or not?

Thanks a lot

[BrunoGiotti]

thanks for the answers. the samples are all of the same microaaray platform. I'm not sure anyway about all the strategy I did to be fair now. for example if I check there are no batch effect within every study but combining these together they show different average intensity which should be the way to go? could i first normalise for each study and then normalise again for all the study together? the PCA of the batch corrected arrays above in facts shows liver and lungs close together with the other monocytes for the same study meaning what you actually pointed out. Uhm...

Answer 1

[raunakms]

Are all the datasets you downloaded are of the same microarray platform or different? You need to keep in mind that there might be biological differences between different phenotypes contained in you datasets on top of the differences between various datasets at hand. COMBAT of course is one of the popular methods to correct the batch effects but may not always be useful as COMBAT sometimes might destroy the biological signal if you are not careful. As an alternative you could convert the expression values to z-score if there are some platform variability (but this does not guarantee to correct the batch effect).

[Charles Warden]

Yes - I agree. You need to make sure your batches are randomized across conditions and I would strongly recommend against trying to compare completely different platforms. If the batches are randomized, you can even just do a 2-way ANOVA. However, that may not work here (for example, it looks like liver and lung samples are only present in batch #1).