collectResiduals: Residual hooks and collection methods
In MAST: Model-based Analysis of Single Cell Transcriptomics

Description Usage Arguments Value Functions Total residual types Partial residuals See Also Examples

After each gene is fit, a hook function can optionally be run and the output saved. This allows extended computations to be done using the fitted model, without keeping it in memory. Here this is used to calculate various residuals, though in some cases they can be done using only the information contained in the ZlmFit-class.

collectResiduals(x, sca, newLayerName = "Residuals")

discrete_residuals_hook(x)

continuous_residuals_hook(x)

combined_residuals_hook(x)

deviance_residuals_hook(x)

fitted_phat(x)

partialScore(x, effectRegex)

`x`	`ZlmFit`-class
`sca`	`SingleCellAssay` object to which the residuals should be added
`newLayerName`	`character` name of the assay layer
`effectRegex`	a regular expression naming columns of the design corresponding to Z_0. Generally these should be the treatment effects of interest.

copy of sca with new layer

discrete_residuals_hook: Hook to get the discrete residuals, ie, difference between expected probability of expression and observed
continuous_residuals_hook: Hook to get the continuous residuals, ie, residuals for conditionally positive observations. If an observation is zero, it's residual is defined to be zero as well.
combined_residuals_hook: Hook to get the combined residuals, ie, Y-E(U)*E(V)
deviance_residuals_hook: Standardized deviance residuals hook. Computes the sum of the standardized deviance residuals for the discrete and continuous models (scaled to have unit variance). If the observation is zero then only the discrete component is used.
fitted_phat: Hook to return p_hat, the predicted probability of expression.
partialScore: Compute Y_i-E(V_i|X_i, Z_0)E(U|X_i, Z_0), where Z_0 is a treatment effect (being left in) and X_i is a nuisance effect (being regressed out).

Each component of the model contributes several flavors of residual, which can be combined in various fashions. The discrete residual can be on the response scale (thus subtracting the predicted probability of expression from the 0/1 expression value). Or it can be a deviance residual, revealing something about the log-likelihood.

It's also possible to consider partial residuals, in which the contribution of a particular covariate is added back into the model.

zlm

data(vbetaFA)
svbeta <- subset(vbetaFA, ncells==1)
svbeta <- svbeta[freq(svbeta)>.4,]
window <- function(x1) lapply(assays(x1), function(x2) x2[1:3, 1:6])
#total residuals of the response
z1 <- zlm(~ Stim.Condition, svbeta, hook=discrete_residuals_hook)
window(collectResiduals(z1, svbeta))
z2 <- zlm(~ Stim.Condition, svbeta, hook=continuous_residuals_hook)
window(collectResiduals(z2, svbeta))
z3 <- zlm(~ Stim.Condition, svbeta, hook=combined_residuals_hook)
window(collectResiduals(z3, svbeta))
#partial residuals
colData(svbeta)$ngeneson <- colMeans(assay(svbeta)>0)
z5 <- zlm(~ Stim.Condition + ngeneson, svbeta)
partialScore(z5, 'Stim.Condition')