Hi there:

I have hundreds of samples with reads counted on genebody or 2kb sliding windows. I'd like to use these matrix to build binary classification and prediction models using elasti net.

I plan to do the following things:

First, do normalization (logCPM) in total samples to check wether batch exists then split the matrix into training set and test set.

Second, use limma to find different genes or windows in training set.

Third, select important features in different genes and windows with elasticnet in training set to construct binary classification models and make predictions in test set.

However, in step1, t-SNE plots showed that there was obvious batch effect. limma could handle 'batch' by including batch as a covariate in a design formula for the purpose of differential expression analysis. If I include batch in design formula while doing differential expression analysis in training set, what value should I use for model construction and prediction? Could I use results of removeBatchEffect() to build model?

Thanks, Xinlei Hu