binomialRF

binomialRF implements a correlated-binomial feature-selection method for random forest classifiers to detect main-effect and multiway-interaction biomarkers in high-dimensional genomic datasets.

Key Features:

• Correlated Binomial Distribution: Models per-tree feature selection counts using a correlated binomial distribution to test selection frequency against chance.
• Theoretical Adjustment for Tree Correlation: Generalizes the binomial distribution with an additional parameter to account for correlation among Random Forest trees.
• High-Dimensional Data Handling: Addresses P > N scenarios common in genomics to enable robust biomarker detection when features far exceed samples.
• Scalability to Multiway Interactions: Efficiently analyzes second- and third-order feature interactions in high-dimensional datasets.
• R Implementation: Provided as an R-based implementation for integration with R workflows.
• Computational Performance: Reports computational speedups of approximately 5- to 300-fold compared with existing methods while maintaining competitive precision and recall.
• Validation on Simulated and Real Data: Validated using simulated datasets and real-world datasets from TCGA and UCI repositories.
• Ontology Integration (planned): Future extension aims to incorporate ontologies for pathway-level feature selection from gene expression data.

Scientific Applications:

• Biomarker Detection: Identifies biomarkers' main effects and interactions with demonstrated precision and recall in simulations and real-data validations.
• Clinical Studies: Prioritizes pathological molecular mechanisms and supports high-accuracy classification using features alone or with their statistical interactions.
• Genomics and Computational Biology: Applied to high-dimensional genomic datasets for hypothesis testing and feature prioritization in research settings.

Methodology:

Treats each Random Forest tree as a quasi-binomial stochastic process, models per-tree selection counts with a correlated binomial distribution, generalizes the binomial via an extra parameter to capture inter-tree correlation, extends analysis to multiway interactions, and is implemented and validated in R on simulated, TCGA, and UCI datasets.