MXM

MXM performs feature selection, cross-validation, and Bayesian network analysis to identify statistically equivalent predictive feature subsets for classification, regression, and survival analysis.

Key Features:

• Statistically Equivalent Signatures (SES) algorithm: Identifies multiple predictive feature subsets that are statistically equivalent in performance.
• Constraint-based Bayesian network inspiration: The SES algorithm is inspired by constraint-based learning of Bayesian networks for feature selection and structure inference.
• Extension of max-min parent children algorithm: Extends previous algorithms such as the max-min parent children algorithm to enable detection of statistically equivalent signatures.
• Multiple predictive subsets: Detects several distinct feature subsets that achieve close to maximal predictive accuracy rather than returning a single presumed-best subset.
• Comparison with LASSO: Provides comparative analyses showing SES achieves predictive accuracy comparable to LASSO on publicly available datasets.
• Support for analysis types: Implements functionality for classification, regression, and survival analysis on large datasets.
• R package implementation: The SES algorithm is implemented as a function within the MXM R package.

Scientific Applications:

• Feature selection and biomarker discovery: Identification of multiple statistically equivalent signatures for predictive modeling and biomarker selection.
• Classification: Selection of predictive features for categorical outcome models.
• Regression: Selection of predictive features for continuous outcome models.
• Survival analysis: Feature selection applicable to time-to-event (survival) data.
• Bayesian network structure learning: Constraint-based analysis for Bayesian network inference and causal-structure exploration.

Methodology:

Implements the SES algorithm as an extension of the max-min parent children algorithm based on constraint‑based Bayesian network learning, employs cross‑validation and comparative benchmarking against LASSO on publicly available datasets, and is provided as a function in the R package MXM.