Dealing with very unbalanced class sizes in differential expression

Question

[murphy.charlesj]

I' am looking at some TCGA data and want to compare the expression profiles of tumors with a certain fusion (4 samples) vs those without it (~500 samples). I don't think current algorithms for differential expression will work with such unbalanced classes (e.g. DESeq2). It seems there may be some strategies I can take from the machine learning community (e.g. under sampling the larger class). Do you have any recommendations or can point me to any papers?

Thanks!

[Ron]

One way could be taking the mean of Expression values of all Samples(500 samples,without it), to get a single Mean Expression value per gene. And similarly take the mean of Expression values of 4 samples,and divide to get a log2foldchange for differential expression

[murphy.charlesj]

My advisor suggested something similar, so I'll definitely give it a try. Thanks.