See https://github.com/theislab/diffxpy/issues/35, this is currently done by fitting separate models per group.