Elastic Net Feature Selection

Source: https://github.com/aipoch/medical-research-skills

Elastic Net Feature Selection

When to Use

• Use this skill for binary case-vs-control classification on bulk expression matrices.
• Use it when you need elastic net logistic regression feature selection, coefficient paths, and cv.glmnet-based lambda selection.
• Use custom labels such as Tumor and Normal only when the group file still contains exactly two outcome levels.

Out of Scope

• Survival or Cox modeling
• Multiclass outcomes
• Single-cell data
• Non-expression tables

Out-of-scope enforcement:

• If the group file contains any label outside the requested case_group and control_group, the command stops with SKILL_INVALID_DATA instead of silently dropping samples.
• If either requested class is missing after validation, the command stops with SKILL_INVALID_DATA.

When to Read External Files

Usage

Rscript scripts/main.R \
  --input_file ./expression_matrix.csv \
  --group_file ./groups.csv \
  --feature_file ./genes.csv \
  --case_group case \
  --control_group control \
  --alpha auto \
  --alpha_grid 0,0.25,0.5,0.75,1 \
  --nfolds 5 \
  --lambda_choice lambda.min \
  --standardize TRUE \
  --timeout_seconds 600 \
  --output_dir ./output/ \
  --seed 42

Arguments

Input Format

Expression Matrix (input_file)

Genes as rows, samples as columns, CSV format with gene IDs in the first column.

,Sample01,Sample02,Sample03
TNMD,0.0349,0.0533,1.3889
DPM1,4.8627,5.4208,5.6370

Group File (group_file)

CSV with sample IDs and binary group labels.

sample,group
Sample01,case
Sample02,control
Sample03,case

Feature File (feature_file)

Optional plain text or single-column CSV file with one feature per line.

TNMD
DPM1
SCYL3

Output Files

Workflow

Step 1: Validate Input

• WHEN preparing input files for a first run or regression test, READ: tests/data/
• Check file existence
• Reject empty input files
• Detect sample and group columns in the group file
• Reject group files that contain labels outside the requested binary comparison
• Validate sample matching between expression matrix and group file

Step 2: Prepare Modeling Matrix

• Restrict samples to the requested case and control groups
• Intersect the optional feature list with matrix row names
• Build a sample-by-feature numeric matrix for glmnet
• Drop zero-variance features before modeling

Step 3: Run Elastic Net

• WHEN deciding between alpha, lambda.min, and lambda.1se, READ: references/algorithm.md
• If alpha=auto, evaluate the candidate alpha_grid with the same cross-validation folds
• Fit the regularization path with glmnet
• Run cv.glmnet to estimate the optimal lambda
• Extract coefficients at lambda.min or lambda.1se
• Apply runtime timeout and capture non-fatal warnings

Step 4: Export Results

• WHEN you need exact invocation patterns or output inspection commands, READ: references/cli-guide.md
• Save tuning tables and selected features
• Generate coefficient path and cross-validation plots
• Record session information for reproducibility

Methods

Elastic Net Logistic Regression

Elastic net combines lasso (L1) and ridge (L2) penalties through alpha, enabling sparse feature selection while stabilizing correlated predictors.

Cross-Validation

cv.glmnet evaluates the lambda path and reports both lambda.min and the more conservative lambda.1se.

Automatic Alpha Selection

When alpha=auto, the skill reuses the same cross-validation folds across all values in alpha_grid, compares the minimum cross-validated error for each candidate, and selects the best alpha before reporting coefficients and lambda-based outputs.

If the chosen alpha is 0, the model is ridge rather than sparse elastic net. In that case, selected_features.csv is written empty to avoid mislabeling dense ridge coefficients as selected features; use model_coefficients.csv for coefficient ranking instead.

Feature Selection Rule

Selected features are the coefficients whose absolute value exceeds a small numerical tolerance at the chosen lambda, excluding the intercept term.

If the chosen alpha is 0, the workflow writes an empty selected_features.csv because ridge coefficients are dense by design and should not be mislabeled as sparse selected features.

Examples

Recommended First Run

Rscript scripts/main.R \
  -i expression_matrix.csv \
  -g groups.csv \
  -f genes.csv \
  -a auto \
  --alpha_grid 0,0.25,0.5,0.75,1 \
  -o output/first_run

Fixed-Alpha Baseline

Rscript scripts/main.R \
  -i expression_matrix.csv \
  -g groups.csv \
  -f genes.csv \
  -a 0.5 \
  -o output/fixed_alpha

More Conservative Selection

Rscript scripts/main.R \
  -i expression_matrix.csv \
  -g groups.csv \
  -l lambda.1se \
  -o output/lambda_1se

Error Handling

Common Errors

IF error persists, READ: references/troubleshooting.md

Testing

Test with Sample Data

# Check help
Rscript scripts/main.R --help

# Run with bundled test data
Rscript scripts/main.R \
  -i tests/data/expression_matrix.csv \
  -g tests/data/groups.csv \
  -f tests/data/genes.csv \
  -a auto \
  --alpha_grid 0,0.5,1 \
  -o tests/output \
  -n 5 \
  -t 600

Validation Commands

# Inspect selected features (may be header-only if auto-alpha selects ridge)
cat tests/output/selected_features.csv

# Check plots exist
ls -la tests/output

Implementation Checklist

• CLI parsing with optparse
• set.seed() for reproducibility
• requireNamespace() dependency checks
• Runtime package loading with library()
• Session info recording
• Timeout control with setTimeLimit()
• Console warning handling
• Out-of-scope label enforcement
• gc() snapshot reporting
• File reading instructions in SKILL.md
• Modular script structure
• Test data provided
• Error handling with SKILL_* codes
• Scripts in scripts/ directory
• References in references/ directory

Last updated: 2026-04-20 | Version: 1.0.0