fast_estimate_ziber: Fast parameter estimation of zero-inflated bernoulli model

In YosefLab/scone: Single Cell Overview of Normalized Expression data

Fast parameter estimation of zero-inflated bernoulli model

Description

This function implements Newton's method for solving zero of Expectation-Maximization equation at the limit of parameter convergence: a zero-inflated bernoulli model of transcript detection, modeling gene expression state (off of on) as a bernoulli draw on a gene-specific expression rate (Z in 0,1). Detection conditioned on expression is a logistic function of gene-level features. The bernoulli model is modeled numerically by a logistic model with an intercept.

Usage

fast_estimate_ziber(
  x,
  fp_tresh = 0,
  gfeatM = NULL,
  bulk_model = FALSE,
  pos_controls = NULL,
  rate_tol = 0.01,
  maxiter = 100,
  verbose = FALSE
)

Arguments

x	matrix. An expression data matrix (genes in rows, cells in columns)
fp_tresh	numeric. Threshold for calling a positive detection (D = 1). Default 0.
gfeatM	matrix. Numeric gene level determinants of drop-out (genes in rows, features in columns)
bulk_model	logical. Use median log-expression of gene in detected fraction as sole gene-level feature. Default FALSE. Ignored if gfeatM is specified.
pos_controls	logical. TRUE for all genes that are known to be expressed in all cells.
rate_tol	numeric. Convergence treshold on expression rates (0-1).
maxiter	numeric. The maximum number of steps per gene. Default 100.
verbose	logical. Whether or not to print the value of the likelihood at each iteration.

Value

a list with the following elements:

• W coefficients of sample-specific logistic drop-out model

• Alpha intercept and gene-level parameter matrix

• X intercept

• Beta coefficient of gene-specific logistic expression model

• fnr_character the probability, per gene, of P(D=0|E=1)

• p_nodrop 1 - the probability P(drop|Y), useful as weights in weighted PCA

• expected_state the expected value E[Z] (1 = "on")

• loglik the log-likelihood

• convergencefor all genes, 0 if the algorithm converged and 1 if maxiter was reached

Examples

mat <- matrix(rpois(1000, lambda = 3), ncol=10)
mat = mat * matrix(1-rbinom(1000, size = 1, prob = .01), ncol=10)
ziber_out = suppressWarnings(fast_estimate_ziber(mat,
   bulk_model = TRUE,
   pos_controls = 1:10))

YosefLab/scone documentation built on Feb. 19, 2025, 4:22 p.m.