markerpen (version 0.1.2)

pca_pen: Penalized Principal Component Analysis for Marker Gene Selection

Description

This function solves the optimization problem $$\min\quad-\mathrm{tr}(SX) + \lambda p(X),$$ $$s.t.\quad O\preceq X \preceq I, \quad X \ge 0, \quad\mathrm{and}\quad \mathrm{tr}(X)=1,$$ where \(O\preceq X \preceq I\) means all eigenvalues of \(X\) are between 0 and 1, \(X \ge 0\) means all elements of \(X\) are nonnegative, and \(p(X)\) is a penalty function defined in the article (see the References section).

Usage

pca_pen(
  S,
  gr,
  lambda,
  w = 1.5,
  alpha = 0.01,
  maxit = 1000,
  eps = 1e-04,
  verbose = 0
)

Value

A list containing the following components:

projection

The estimated projection matrix.

evecs

The estimated eigenvectors.

niter

Number of iterations used in the optimization process.

err_v

The optimization error in each iteration.

Arguments

S

The sample correlation matrix of gene expression.

gr

Indices of genes that are treated as markers in the prior information.

lambda

Tuning parameter to control the sparsity of eigenvectors.

w

Tuning parameter to control the weight on prior information. Larger \(w\) means genes not in the prior list are less likely to be selected as markers.

alpha

Step size of the optimization algorithm.

maxit

Maximum number of iterations.

eps

Tolerance parameter for convergence.

verbose

Level of verbosity.

References

Qiu, Y., Wang, J., Lei, J., & Roeder, K. (2020). Identification of cell-type-specific marker genes from co-expression patterns in tissue samples.

Examples

set.seed(123)
n = 200  # Sample size
p = 500  # Number of genes
s = 50   # Number of true signals

# The first s genes are true markers, and others are noise
Sigma = matrix(0, p, p)
Sigma[1:s, 1:s] = 0.9
diag(Sigma) = 1

# Simulate data from the covariance matrix
x = matrix(rnorm(n * p), n) %*% chol(Sigma)

# Sample correlation matrix
S = cor(x)

# Indices of prior marker genes
# Note that we have omitted 10 true markers, and included 10 false markers
gr = c(1:(s - 10), (s + 11):(s + 20))

# Run the algorithm
res = pca_pen(S, gr, lambda = 0.1, verbose = 1)

# See if we can recover the true correlation structure
image(res$projection, asp = 1)