deconvG: Base function of G-modeling deconvolution

In jingshuw/descend: Gene Expression Distribution Deconvolution for Single Cell RNA-seq

Description

Base function used by deconvSingle to deconvolve the underlying distribution. We assume X ~ F(T) where F is the noise distribution. We assume that

log(T) = offset + γ Z + ε

P(T = 0) = β_0 + β_1 Z0

The goal is the recover the distribution of exp(log(T) - offset - gamma Z), which has density g and is discretized at exp(tau) (add 0 when zero inflation is allowed). There can be some warning messages for the optimization process, which can be ignored.

Usage

deconvG(tau, X, offset, family = c("Poisson", "Negative Binomial"),
  ignoreZero = F, zeroInflate = F, Z = NULL, Z0 = NULL, c0 = 1,
  NB.size = 100, only.value = F, pDegree = 5, aStart = 1,
  bStart = 0, gStart = 0, ...)

Arguments

tau	log of the discrete points of the deconvolved distribution
X	a vector of observed counts
offset	a vector with the same length as X. See details
family	family of the noise distribution, support either "Poisson" or "Negative Binomial" with known tuning parameter
ignoreZero	whether ignore the zero count. If true, then use truncated Poisson / Negative Binomial distribution. Default is False
zeroInflate	whether add zero inflation part to the deconvolved distribution to reflect transcriptional bursting. Default is True.
Z	covariates for nonzero mean. Default is NULL.
Z0	covariates for nonzero fraction. Used only when zeroInflate is True. Default is NULL.
c0	the tuning parameter on the L2 penalty term. Default is 1. c0 will be selected automatically in deconvSingle
NB.size	over-dispersion parameter when the family is Negative Binomial: mu = mu + mu^2/size
only.value	whether not to compute the estimation statistics but only the value of the optimized lieklihood. Used for likelihood ratio test.
pDegree	the degree of the Spline matrix. Default is 5.
aStart, bStart, gStart	initial values of the density parameters, the coefficients of Z0 and coefficients of Z
...	extra parameters for the nlm function

Value

a list with elements

stats	a list of two elements. One is the mat.dist, which is the matrix of the estimated distribution. The other is mat.coef, which is the matrix of the coefficients of Z and Z0
mle	the estimated parameters of the the density function
mle.g	the estimated coefficients of Z
value	the optimized penalized negative log-likelihood value
S	the fake information proportion
cov	the covariance of the parameters
bias	the bias of the parameters
cov.g	the covaraince of the estimated density points
cov.g.gamma	the covariance between the estimated density points and the coefficient of Z
loglik	the objective function being optimized
statsFunction	the function computing the relavant statistics

Note

This is an extension of the G-modeling package

Examples

X <- rpois(1000, 0.2 * 3)
lam.max <- quantile(X[!X == 0], probs = c(0.98))/0.2
tau <- seq(0, lam.max, length.out = 51)[-1]

Z <- rnorm(1000)

result <- deconvG(log(tau), X, zeroInflate = TRUE,
                     Z = Z,
                     Z0 = Z,
                     offset = rep(log(0.2), 1000))

jingshuw/descend documentation built on Nov. 2, 2021, 4:23 p.m.