Integrative recipe runs voom + integration in spilterlize, even though one cannot use that for log-normalized counts

[bednarsky]

• I just ran the integrative recipe

• the preprocessing with spilterlize also does VOOM, also for integration
  

  MrBiomics/config/CorcesINT/CorcesINT_spilterlize_integrate_config.yaml

  Lines 65 to 84 in 815151a

  	### VOOM
  	# limma::voom -> https://rdrr.io/bioc/limma/man/voom.html
  	# use case: show the user how the data actually looks inside of a subsequent limma-voom differential analysis approach
  	# to skip leave normalize.method empty i.e., ""
  	voom_parameters:
  	calcNormFactors_method: "none" # "none"; options: any other supported method from edgeR::CalcNormFactors (parameters from above are used)
  	normalize.method: "quantile" # "none"; options: "scale", "quantile", "cyclicloess"
  	span: 0.5 # 0.5
  	##### INTEGRATE #####
  	### limma's removeBatchEffect
  	# limma::removeBatchEffect -> https://rdrr.io/bioc/limma/man/removeBatchEffect.html
  	# use case: remove unwanted (batch) effects from data, while desired effects are preserved, prior to downstream analyses e.g., dimensionality reduction with PCA/UMAP
  	# to skip leave desired empty i.e., []
  	removeBatchEffect_parameters:
  	norm_methods: ["TMM","VOOM","upperquartile"] # list of normalization outputs to be integrated; have to be configured above; options: "TMM","TMMwsp","RLE","upperquartile","CPM","RPKM","CQN","VOOM"
  	desired: ["cell_type"] # list of the column names of desired variation in annotation
  	unwanted_categorical: ["modality", "donor"] # list of maximal 2 column names of categorical variables of unwanted variation in annotation e.g., batch; leave empty [] if none
  	unwanted_numerical: [] # list of the column names of numerical variables of unwanted variation in annotation; leave empty [] if none

• for me, VOOM + integration looked best, so I wanted to continue with that (didn't re-read the dea_limma config instructions)

• However, also setting calcNormFactors_method to 0 in the annotation for dea_limma leads to Error in is(counts, "DGEList") : object 'dge' not found, because the dge variable is never defined
  https://github.com/epigen/dea_limma/blob/70fe01d1e56ec9223b542dfe1133fb28cc30202b/workflow/scripts/limma.R#L106-L109
  https://github.com/epigen/dea_limma/blob/70fe01d1e56ec9223b542dfe1133fb28cc30202b/workflow/scripts/limma.R#L138-L143

• Would be better to not use VOOM in spilterlize in the integrative recipe to avoid confusion?

• Do I understand correctly that one shouldn't use VOOM here?

[sreichl]

No. For integrative analysis, we process and transform the data quite heavily to remove a specific effect we later want to test for. All outputs (apart from counts and filtered) of the spilterlize_integrate module are by default log-transformed. To use them in differential analysis we have to use limma-trend(not limma-voom, which requires raw counts) in the dea_limma module. Even if you have used VOOM for normalization in spilterlize_integrate you have to use limma-trendfor DEA.

Bottom line: It does not matter what happened before to the data (to get rid of the systemic differences). The output in the end is log-transformed, and this implies usage of limma-trend for differential testing.

[sreichl]

PS: I will clarify this part in the integrative analysis recipe.