Batch correction when capturing meta-program from multiple samples

[Jasonqte]

Hi,

Nice work! I have a question that when I calculate the meta-program, do I need to do batch correction first. For example, I first perform NMF on multiple samples individually. The meta-program is obtained by calculating the correlation between the programs that decompose from each sample. If it does not involve aggregating multiple samples first, is it necessary to consider the batch effect before performing NMF?

[dylkot]

Hi Jason,

If there are strong batch signals within the individual samples, that might cause the programs to correlate less well across samples. In that case, there can sometimes be a benefit to batch correcting each sample first. Are there any batch signals within an individual sample for you? For example, was the same sample sequenced in multiple runs? In that case there may be some benefit.

However, often times there aren't significant batch effects at the individual sample level and so running the batch effect per sample may remove signal without providing much benefit.

Hope that helps!