Citation

If you use BSGS in your research, please cite: Paplomatas, P., Kyriakou, Z., Anagnostopoulos, K., & Vrahatis, A. (2025). Bayesian subset-based gene selection for biomarker discovery in high-dimensional data. Academia Molecular Biology and Genomics, 2(2). https://doi.org/10.20935/AcadMolBioGen7733

BSGS: Bayesian Subset-based Gene Selection

A Stochastic Multi-Criteria Approach for High-Dimensional Biomarker Identification

Introduction

The Bayesian Subset-based Gene Selection (BSGS) framework is an advanced, customizable tool designed for identifying biomarkers in high-dimensional datasets. By integrating Bayesian subset sampling, Lasso-based preselection, and multi-criteria scoring, BSGS delivers a robust methodology for biomarker discovery.

This R package is specifically crafted to enhance user experience and reproducibility in biomarker research, making it easy to apply and adapt for different datasets.

---

Key Features

• Lasso-based Preselection: Efficiently reduces the feature space by selecting genes with high importance.
• Bayesian Subset Sampling: Iteratively updates posterior probabilities to refine gene scores.
• Customizable Scoring: Combines statistical metrics (e.g., SNR, variance, correlations) and SHAP values for multi-criteria ranking.
• Visualization: Automatically generates distribution plots for gene scores.
• Reproducibility: Offers clear workflows for reproducible research.

---

Installation

To get started, install the BSGS package directly from GitHub:

# Install the devtools package if not already installed
install.packages("devtools")

# Install BSGS from GitHub
devtools::install_github("PaplomatasP/BSGS")

# Load the BSGS package
library(BSGS)

# Load the included example dataset
data(LUAD_MYE_MRS_GSE97168)

# Explore the dataset
str(LUAD_MYE_MRS_GSE97168)

# Convert the dataset to a data frame
full_data <- as.data.frame(LUAD_MYE_MRS_GSE97168)

# Add target labels
full_data$target <- as.factor(LUAD_MYE_MRS_target)

# Run the BSGS framework
bsgs_results <- BSGS(
  data = full_data,
  label_col = "target",
  iterations = 1000,
  alpha = 0.3,
  min_genes = round(floor(sqrt(length(colnames(full_data[,-1])))) / 2),
  max_genes = 5 * round(floor(sqrt(length(colnames(full_data[,-1])))), 2),
  n_genes = 150,
  plot_distribution = TRUE,
  save_plot = TRUE,
  file_path = "./bsgs_results.png"
)

# View selected genes
print(bsgs_results$final_genes)

# View gene scores
print(bsgs_results$final_scores)

Gene Score Distribution Plot

When plot_distribution = TRUE is set in the BSGS function, a gene score distribution plot is displayed to the user. This plot visualizes the scores of all genes analyzed, allowing the user to identify meaningful thresholds for selecting the final genes based on their scores. The user is prompted to provide a threshold value, and genes with scores above this threshold are retained.

Below is an example of the gene score distribution plot:

Insights and Observations

• Selecting a threshold above Q3: Based on our experiments, retaining genes with scores above the third quartile (Q3) ensures a robust and highly informative subset of genes. This threshold balances reducing dimensionality with maintaining meaningful biomarkers.

• Convergence of Q1, Q2, Q3, and the mean: When the quartiles (Q1, Q2, Q3) and the mean converge, as shown in the plot, it indicates that the desired number of iterations has been reached. This convergence ensures stability in the gene scores and highlights the robustness of the BSGS method.

If plot_distribution = FALSE, you can retain the desired number of genes with the best scores by using the parameter n_genes.