Fit LASSO Regression for Multi-Gene Biomarker Discovery

Applies cross-validated LASSO logistic regression to identify a sparse set of genes that jointly predict a binary phenotype. Unlike univariate tests, LASSO accounts for the joint contribution of all genes simultaneously, automatically penalizing redundant predictors.

Usage

fit.lasso(expression.matrix, phenotype.vector, alpha = 1, nfolds = 10)

Arguments

expression.matrix

Numeric matrix of gene expression values as returned by extract.expression()$expression. Rows are probes, columns are samples. The matrix is transposed internally so samples become rows as required by glmnet.

phenotype.vector

Factor or integer vector of binary phenotype labels, one per sample (e.g., 0 for control, 1 for disease). Must be the same length as the number of columns in expression.matrix.

alpha

Numeric. Elastic net mixing parameter. 1 (default) gives pure LASSO; 0 gives Ridge; values between 0 and 1 give elastic net.

nfolds

Integer. Number of cross-validation folds. Default is 10.

Value

A cv.glmnet object. Key elements:

lambda.min

Lambda with minimum cross-validated error.

lambda.1se

Largest lambda within 1 SE of the minimum (sparser model).

glmnet.fit

The full sequence of fitted models across lambda values.

Extract selected genes with coef(fit, s = "lambda.1se"); probes with non-zero coefficients are the LASSO-selected biomarkers.

Details

LASSO is suited to high-dimensional genomic data where the number of genes far exceeds samples. It shrinks most coefficients to exactly zero, retaining only genes with independent predictive value. This complements run.limma.de() - while limma ranks genes by individual differential expression, LASSO identifies the minimal subset with non-redundant joint predictive signal.

Examples

# \donttest{
# Synthetic example — small expression matrix with binary outcome
set.seed(42)
expr.mat <- matrix(rnorm(200), nrow = 20, ncol = 10)
rownames(expr.mat) <- paste0("probe", 1:20)
colnames(expr.mat) <- paste0("sample", 1:10)
phenotype.binary <- c(0, 0, 0, 0, 0, 1, 1, 1, 1, 1)
lasso.fit <- fit.lasso(expr.mat, phenotype.binary)
#> Warning: one multinomial or binomial class has fewer than 8  observations; dangerous ground
#> Warning: one multinomial or binomial class has fewer than 8  observations; dangerous ground
#> Warning: one multinomial or binomial class has fewer than 8  observations; dangerous ground
#> Warning: one multinomial or binomial class has fewer than 8  observations; dangerous ground
#> Warning: one multinomial or binomial class has fewer than 8  observations; dangerous ground
#> Warning: one multinomial or binomial class has fewer than 8  observations; dangerous ground
#> Warning: one multinomial or binomial class has fewer than 8  observations; dangerous ground
#> Warning: one multinomial or binomial class has fewer than 8  observations; dangerous ground
#> Warning: one multinomial or binomial class has fewer than 8  observations; dangerous ground
#> Warning: one multinomial or binomial class has fewer than 8  observations; dangerous ground
#> Warning: one multinomial or binomial class has fewer than 8  observations; dangerous ground
#> Warning: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per fold
coef(lasso.fit, s = "lambda.1se")
#> 21 x 1 sparse Matrix of class "dgCMatrix"
#>             lambda.1se
#> (Intercept)          .
#> probe1               .
#> probe2               .
#> probe3               .
#> probe4               .
#> probe5               .
#> probe6               .
#> probe7               .
#> probe8               .
#> probe9               .
#> probe10              .
#> probe11              .
#> probe12              .
#> probe13              .
#> probe14              .
#> probe15              .
#> probe16              .
#> probe17              .
#> probe18              .
#> probe19              .
#> probe20              .
# }