I am running SomaticSignatures 2.6.0 using somatic calls from 2 sources (~40 samples, internal; ~200 samples external source). I am not using a 'group' to generate the motif matrix, eg:

>sca_mm = motifMatrix(sca_motifs, normalize = TRUE)

I am trying to correct for batch effect, but this produces some negative numbers in sca_mm which causes problems for nmf decomposition:

>meta<-read.table("meta.test", header=TRUE, sep="\t")
>model_null = model.matrix(~ 1, meta)
>sca_mm_batch = ComBat(sca_mm, batch = meta$study, mod = model_null)
>gof_nmf= assessNumberSignatures(sca_mm_batch, n_sigs, nReplicates = 5, nmfDecomposition)
Error: NMF::nmf - Input matrix x contains some negative entries.

My 'meta' dataframe is columns of 'sampleNames' and 'study' and other metadata; each sample is classified as either coming from the internal or external data source.

Are negative entries expected or am I not applying the batch effect correction appropriately? Is this telling me something about my data?

Thanks

ComBat cleans genomic data by fitting a shrunken regression model and returning a carefully formulated set of residuals. But since they are residuals from a regression model some of the values can be negative. This isn't something you can fix with ComBat unless you do something like log transform your raw values (if that transformation is appropriate), apply ComBat, and then exponentiate the values again. This will lead to all positive values, but may have major unintended consequences (for example, you are now modeling batch effects on a multiplicative scale which may or may not be appropriate). 

If there is a way to include covariates like batch int he assessNumberSignatures function that would be the more appropriate way to proceed. 

Best

Jeff