Removing certain levels of Batch before batch effect correction

Question

[english.server]

Hi all!

This post is rather lengthy, so I give a brief overview of what I intend to ask:

A) Does one need to check whether each different batch level contains at least 2 different experimental conditions of interest (ie diseased/normal)?

B) Should be the distribution of experimental condition within batches be balanced?

While analyzing a GEO dataset, I was confronted with the below sample characteristics, as Biobase::pData(my_ExpressionSet)). I think I should completely discard certain experiments before batch correction

> des # dput is given below
cell.type   disease dnr pairbatch
Fibroblast  affctd  AA1 1   1a
Fibroblast  affctd  AA1 1   2a
Fibroblast  affctd  AA2 2   1a
Fibroblast  affctd  AA2 2   2a
...........
iPSC-d-nrn  Normal  NN2 2   13b


des= structure(list(cell.type = c("Fibroblast", "Fibroblast", "Fibroblast", 
"Fibroblast", "Fibroblast", "Fibroblast", "Fibroblast", "Fibroblast", 
"Fibroblast", "Fibroblast", "Fibroblast", "Fibroblast", "Fibroblast", 
"Fibroblast", "Fibroblast", "Fibroblast", "inducedPSC", "inducedPSC", 
"inducedPSC", "inducedPSC", "inducedPSC", "inducedPSC", "inducedPSC", 
"inducedPSC", "inducedPSC", "inducedPSC", "inducedPSC", "inducedPSC", 
"inducedPSC", "iPSC-d-NPC", "iPSC-d-NPC", "iPSC-d-NPC", "iPSC-d-NPC", 
"iPSC-d-NPC", "iPSC-d-NPC", "iPSC-d-NPC", "iPSC-d-NPC", "iPSC-d-NPC", 
"iPSC-d-NPC", "iPSC-d-NPC", "iPSC-d-NPC", "iPSC-d-NPC", "iPSC-d-nrn", 
"iPSC-d-nrn", "iPSC-d-nrn", "iPSC-d-nrn", "iPSC-d-nrn", "iPSC-d-nrn", 
"iPSC-d-nrn", "iPSC-d-nrn", "iPSC-d-nrn", "iPSC-d-nrn", "iPSC-d-nrn", 
"iPSC-d-nrn", "iPSC-d-nrn"), disease = c("affctd", "affctd", 
"affctd", "affctd", "affctd", "affctd", "Normal", "Normal", "Normal", 
"Normal", "Normal", "Normal", "Normal", "Normal", "Normal", "Normal", 
"affctd", "affctd", "affctd", "affctd", "affctd", "Normal", "Normal", 
"Normal", "Normal", "Normal", "Normal", "Normal", "Normal", "affctd", 
"affctd", "affctd", "affctd", "affctd", "Normal", "Normal", "Normal", 
"Normal", "Normal", "Normal", "Normal", "Normal", "affctd", "affctd", 
"affctd", "affctd", "affctd", "Normal", "Normal", "Normal", "Normal", 
"Normal", "Normal", "Normal", "Normal"), donor = c("AA1", "AA1", 
"AA2", "AA2", "AA3", "AA4", "AN1", "AN1", "AN2", "AN2", "AN3", 
"AN4", "NN1", "NN1", "NN2", "NN3", "AA1", "AA1", "AA2", "AA2", 
"AA3", "AN1", "AN1", "AN2", "AN2", "AN3", "NN1", "NN1", "NN2", 
"AA1", "AA1", "AA2", "AA2", "AA3", "AN1", "AN1", "AN2", "AN2", 
"AN3", "NN1", "NN1", "NN2", "AA1", "AA1", "AA2", "AA2", "AA3", 
"AN1", "AN1", "AN2", "AN2", "AN3", "NN1", "NN1", "NN2"), pair = c(1L, 
1L, 2L, 2L, 3L, 4L, 1L, 1L, 2L, 2L, 3L, 4L, 1L, 1L, 2L, 3L, 1L, 
1L, 2L, 2L, 3L, 1L, 1L, 2L, 2L, 3L, 1L, 1L, 2L, 1L, 1L, 2L, 2L, 
3L, 1L, 1L, 2L, 2L, 3L, 1L, 1L, 2L, 1L, 1L, 2L, 2L, 3L, 1L, 1L, 
2L, 2L, 3L, 1L, 1L, 2L), batch = c("1a", "2a", "1a", "2a", "1a", 
"2a", "3a", "3a", "3a", "3b", "3a", "3b", "4a", "4a", "4a", "4a", 
"9a", "10a", "9a", "12a", "12a", "10a", "10a", "11a", "9a", "21a", 
"11a", "9a", "12a", "18a", "19a", "15a", "18a", "19a", "16b", 
"19a", "16a", "19b", "19a", "16a", "18b", "17a", "14d", "14a", 
"14c", "13b", "14c", "13a", "14a", "13b", "14b", "14b", "14b", 
"13a", "13b")), class = "data.frame", row.names = c(NA, -55L))

# see whether both diseased/normal conditions are repsented in each batch: point A above
table(des$batch, des$disease) 
          affctd Normal
  10a      1      2
  11a      0      2
  12a      2      1
  13a      0      2
  13b      1      2
  14a      1      1
  14b      0      3
  14c      2      0
  14d      1      0
  15a      1      0
  16a      0      2
  16b      0      1
  17a      0      1
  18a      2      0
  18b      0      1
  19a      2      2
  19b      0      1
  1a       3      0
  21a      0      1
  2a       3      0
  3a       0      4
  3b       0      2
  4a       0      4
  9a       2      2
#try to remove baches lacking arrays from either condition
>bad.batch = apply (table(des$batch, des$disease),1,min) == 0 
>table(des$batch, des$disease) [! bad.batch,]
      affctd Normal
  10a      1      2
  12a      2      1
  13b      1      2
  14a      1      1
  19a      2      2
  9a       2      2

> keep=rownames(table(des$batch, des$disease) [! bad.batch,])

> data4further.analysis= des[des$batch %in% keep,]

> data4further.analysis   
    cell.type disease donor pair batch
17 inducedPSC  affctd   AA1    1    9a
18 inducedPSC  affctd   AA1    1   10a
19 inducedPSC  affctd   AA2    2    9a
20 inducedPSC  affctd   AA2    2   12a
21 inducedPSC  affctd   AA3    3   12a
22 inducedPSC  Normal   AN1    1   10a
23 inducedPSC  Normal   AN1    1   10a
25 inducedPSC  Normal   AN2    2    9a
28 inducedPSC  Normal   NN1    1    9a
29 inducedPSC  Normal   NN2    2   12a
31 iPSC-d-NPC  affctd   AA1    1   19a
34 iPSC-d-NPC  affctd   AA3    3   19a
36 iPSC-d-NPC  Normal   AN1    1   19a
39 iPSC-d-NPC  Normal   AN3    3   19a
44 iPSC-d-nrn  affctd   AA1    1   14a
46 iPSC-d-nrn  affctd   AA2    2   13b
49 iPSC-d-nrn  Normal   AN1    1   14a
50 iPSC-d-nrn  Normal   AN2    2   13b
55 iPSC-d-nrn  Normal   NN2    2   13b

and then proceed to sva::ComBat for batch effect removel or model.matrix %>% limma's lmFit and ... ,considering batch.

Answer 1

[Ram]

The point B you mention is of the highest importance when dealing with Batch effects. Experiment design that spreads conditions evenly across batches enables downstream analysis to account for batch effects by "removing" them or treating them as covariates.

If that's not the case, batch correction can remove genuine biological differences as it is unable to distinguish them from other differences of interest. Either way, in my opinion, accounting for batch effect as a covariate is better than using ComBat to "remove" them as I've heard ComBat tends to over-compensate. Check out this paper: Methods that remove batch effects while retaining group differences may lead to exaggerated confidence in downstream analyses

[english.server]

Thank you for your response and for reading my (long) post. I would appreiate if you also share your opinion on point A above, which I reiterate/paraphrase here:

A) Does one need to check whether each different batch level contains all different experimental conditions of interest (eg diseased/normal)? That is do you recommend ommiting rows 11a and 13a containing 0s in this table:

  batch        number.of.arrays.for.diseased       number.of.arrrays.for.Normal.controls
  10a               1                               2
  11a               0                               2
  12a               2                               1
  13a               0                               2

[Ram]

This can only be answered with the exact statistical question in mind. If a batch contains samples of one type only, any observations made on samples in that batch cannot be of high confidence because it is not trivial to deconvolute the underlying cause of the observations. It could be biology or it could be batch effects or a combination of both.

[english.server]

Thank you for your response. I wondered if you know of a reference or the name of this situation so that I myself search aticles/books.

[Ram]

Search for anything batch effect related and most papers/articles will stress on the importance of experiment design.

[english.server]

Thank you again, RsmRS.