Profile-ML dispersion at fixed coefficients, blocked over genes

For each gene, the negative binomial dispersion that maximises the likelihood at the mean exp(W alpha + offset), the coefficients held fixed: one profile search per gene and no Newton, by the same bisection on the score that polishNB()'s batched engine uses. A gene whose optimum sits on a bound of psi.range (an under-dispersed or near-empty gene), or whose search is not finite, keeps its input dispersion rather than a boundary value stored as an estimate, the rule polishNB() applies. A caller that re-polishes the mean at a held dispersion (polishNB(warm = TRUE)) uses this to put the dispersion back at the reported mean.

Usage

nbProfilePsi(
  Y,
  W,
  alpha,
  psi,
  psi.range = c(0.001, 1000),
  block.size = NULL,
  BPPARAM = BiocParallel::SerialParam(),
  offset = NULL
)

Arguments

Y

a genes x cells matrix of counts (dense, sparse or DelayedArray; densified one gene block at a time).

W

a cells x p numeric design matrix, or an nbBlockDesign().

alpha

a genes x p matrix of coefficients.

psi

the per-gene dispersions (length nrow(Y), or one value for every gene), kept for a gene whose optimum is on a bound.

psi.range

the search interval for the dispersion.

block.size, BPPARAM

gene blocking and dispatch, as in polishNB().

offset

NULL, a per-cell log-scale offset (length ncol(Y)) or a genes x cells matrix, added to every linear predictor. The dispersion is profiled at the mean it gives, so the offset must be the one the coefficients were fitted with.

Value

a numeric vector of dispersions, one per gene.

See also

polishNB().

Examples

set.seed(1)
W <- cbind(1, rnorm(200))
Y <- t(replicate(3, rnbinom(200, mu = exp(1 + 0.3 * W[, 2]), size = 4)))
nbProfilePsi(Y, W, matrix(c(1, 0.3), 3, 2, byrow = TRUE), rep(0.5, 3))
#> [1] 0.2917504 0.2388419 0.2547512