Learning sparse inverse covariance matrices in the presence of confounders