Create normalized counts

[epurdom]

Look at what MAST does to get residuals

[fperraudeau]

Below are snapshots of the parts of the MAST paper regarding the residuals. The take home message is that they use the standardized deviance residuals for visualization (heatmaps and PCA). In the hurdle model, Z_g and Y_g are defined conditionally independent for each gene. Thus, tests with asymptotic χ2 null distributions (LRT or Wald tests) can be summed and remain asymptotically χ2. In the zinbwave, we could separately compute the likelihood for the model without covariates and for the model with covariates. What I have in mind is that we could:

1. fit zinbwave without covariates (just the intercept), get W from the fit,
2. use clusterExperiment on W (for one K or several K), get the cluster labels,
3. fit again zinbwave with the cluster labels,
4. perform DE analysis with the LRT (asymptotic χ2 null distributions) and use the (standardized?) deviance residuals for visualization (heatmaps, PCA,...).