Application to large sparse design matrix

[GabrielHoffman]

I have a lasso problem with ~10K samples and ~20K features where most of the entries in the design matrix x are zero. I can store them in a dgCMatrix sparse matrix and use glmnet to fit a lasso while taking advantage of the sparse design matrix. This is can be over 100x faster than using a standard matrix object.

Is there a way to extend coef, fixedLassoInf and estimateSigma to handle sparse matrices??
This would involve dealing with the centering and scaling of features, but it seems doable.

Also, fixedLassoInf that is very slow in the high dimensional setting in both the sparse and non-sparse case. If I do a pre-filtering step, will the p-values and FDR control still be accurate?

[swzCuroverse]

I have a similar issue where I am using glmnet with a sparse X matrix.