ComBat is an R package for removing batch effects from data. This is a python version that matches the output from the ComBat function in SVA (https://www.bioconductor.org/packages/release/bioc/html/sva.html). This code is completely copied from the ComBat function in that package.
To test, run this R code (requires sva and bladderbatch from bioconductor):
Rscript R-combat.R
Then, from the same directory, run
python combat.py
you can then run this python code to see the differences:
import pandas as pa
p = pa.read_table('py-batch.txt', index_col=0)
r = pa.read_table('r-batch.txt', index_col=0)
print (p - r).max().max()
This outputs 3.9423421307e-05 on my machine. Indicating that that is the largest difference between the 22,283*57 values generated by the R version and those generated in this version.
In the example above, the combat function runs in < 1 second in python and about 15 seconds in R.
On an identical dataset, of 30K rows * 190 samples, this python version finishes in 10.008s
as measured by unix time
.
The R version takes 4m0.681s with output identical to 3 decimal places. This is a speed-up
of about 24x.
The speed improvement seems to be larger for larger datasets.
The python version is usable as a module, the function has the signature:
combat(dat, batch, mod, numCovs=None)
which is the same as the R function except the non-parametric version is not supported.
- dat is the expression/methylation data.
- batch is a list containing the batch variable
- mod is the model matrix (can use patsy for this from python)
- numCovs is a list like ["age", "height"], that gives the column name or number of numeric variables in batch (otherwise they will be converted to factors).
Johnson WE, Rabinovic A, Li C (2007). Adjusting batch effects in microarray
expression data using Empirical Bayes methods. Biostatistics 8:118-127.
Jeffrey T. Leek, W. Evan Johnson, Hilary S. Parker, Andrew E. Jaffe
and John D. Storey (). sva: Surrogate Variable Analysis. R package
version 3.4.0.