mRMRe: Parallelized Minimum Redundancy, Maximum Relevance (mRMR)

Nicolas Jay, Simon Papillon‐Cavanagh, Catharina Olsen, Gianluca Bontempi, Bo Li, Benjamin Haibe-KainsOpenalex Unknown 2012 · DOI · OpenAlex

Abstract

Computes mutual information matrices from continuous, categorical and survival variables, as well as feature selection with minimum redundancy, maximum relevance (mRMR) and a new ensemble mRMR technique. Published in De Jay et al. (2013) <a href="https://doi.org/10.1093%2Fbioinformatics%2Fbtt383" target="_top">doi:10.1093/bioinformatics/btt383</a>.

Introduction

In high-throughput biological fields like genomics, researchers frequently encounter the "curse of dimensionality." Datasets often contain thousands of features, such as gene expression levels, but relatively few samples. Identifying which features are truly informative for a specific outcome—like drug sensitivity or patient survival—is a significant challenge.

Standard feature selection methods often prioritize "relevance," focusing on how well an individual feature correlates with the target. However, biological features are often highly redundant. Selecting multiple genes that provide the same information is inefficient. The Minimum Redundancy, Maximum Relevance (mRMR) algorithm addresses this by selecting features that are highly relevant to the target but have low redundancy with each other. The mRMRe package provides a high-performance, parallelized implementation of this algorithm, uniquely supporting mixed data types and ensemble selection for improved robustness.

Unified Mutual Information Framework

A core strength of mRMRe is its ability to handle disparate data types—continuous, categorical, and survival data—within a single mathematical framework. Most existing tools require extensive pre-processing or are limited to a single data type. mRMRe overcomes this by employing specific correlation estimators tailored to the data:

  • Continuous: Pearson, Spearman, or Kendall correlations.
  • Categorical: Cramer’s V.
  • Survival: The Concordance Index (C-index).
The `correlate` function parameters illustrate the package's support for diverse statistical methods and data types, including survival analysis via the C-index.

By using the C-index, survival data (which includes "censored" observations) can be treated as a standard variable for mutual information calculations. This allows for seamless feature selection for "time-to-event" outcomes.

High-Performance Implementation

Feature selection is computationally intensive, especially when calculating large mutual information matrices. mRMRe leverages OpenMP to parallelize these calculations and the subsequent search for optimal feature subsets. This C-level parallelization provides significant speedups, making it feasible to process large-scale genomic matrices on modern multi-core systems.

Furthermore, the package introduces Ensemble mRMR. Feature selection can be unstable; minor changes in the input data often lead to different feature sets. By performing selection on multiple bootstrapped versions of the data, mRMRe identifies a consensus set of features that are more likely to be biologically relevant rather than statistical artifacts.

Benchmark Datasets

The package documentation highlights the cgps dataset to demonstrate its utility. This dataset is a subset of the Cancer Genome Project and includes gene expression data for 1,000 genes across 200 cancer cell lines, paired with drug sensitivity (IC50) values for Camptothecin.

The `cgps` dataset is a primary example used in the package to demonstrate genomic marker identification for drug response.

Key Insights and Results

Beyond simple feature selection, mRMRe offers tools for network inference and putative causality analysis. The causality function utilizes a co-information lattice algorithm to suggest directionality in feature-target relationships, moving beyond simple correlation.

Research by De Jay et al. (2013) confirms that this implementation is not only faster but also more versatile than its predecessors. Its ability to handle survival objects directly is a major advantage for clinical researchers.

Limitations

While powerful, mRMRe has limitations. Exhaustive searches in extremely high-dimensional spaces (e.g., whole-genome sequencing) remain computationally demanding. Additionally, the core logic is primarily designed to capture linear or monotonic relationships; complex non-linear interactions may require different estimators. Finally, "putative causality" scores are statistical estimates that require experimental validation to confirm actual biological mechanisms.

Summary updated

Abstract

Computes mutual information matrices from continuous, categorical and survival variables, as well as feature selection with minimum redundancy, maximum relevance (mRMR) and a new ensemble mRMR technique. Published in De Jay et al. (2013) &lt;<a href="https://doi.org/10.1093%2Fbioinformatics%2Fbtt383" target="_top">doi:10.1093/bioinformatics/btt383</a>&gt;.
BibTeX
@inproceedings{jay2012mrmre,
  author = {Nicolas Jay and Simon Papillon‐Cavanagh and Catharina Olsen and Gianluca Bontempi and Bo Li and Benjamin Haibe-Kains},
  title = {{mRMRe: Parallelized Minimum Redundancy, Maximum Relevance (mRMR)}},
  booktitle = {Openalex Unknown},
  year = {2012},
  doi = {10.32614/cran.package.mrmre},
}
Figure