binda

binda implements multi-class discriminant analysis using binary predictors for variable selection and dichotomization of continuous proteomic mass spectrometry data to identify cancer biomarkers in blood samples.

Key Features:

• Binary Discriminant Analysis: Implements binary discriminant analysis to classify samples based on dichotomized predictors.
• Variable Selection and Dichotomization: Performs data-adaptive thresholding to dichotomize continuous protein expression values and select informative features.
• Relative Entropy Ranking: Ranks dichotomized features using a relative entropy measure to quantify information gain for class separation.
• Generalization of Peak Probability Contrast: Extends the 'peak probability contrast' approach of Tibshirani et al. (2004) to both two-group and multi-group settings.
• Computational Efficiency: Operates with low computational cost suitable for large-scale datasets.
• Predictive Performance: Demonstrated prediction accuracy comparable to random forests in comparative studies on large-scale drug discovery test datasets.
• Biological Relevance: Effective at identifying biologically relevant and statistically predictive marker peaks in mass spectrometry studies, including a pancreas cancer dataset where novel markers were uncovered.

Scientific Applications:

• Proteomic Differential Expression Analysis: Applied to proteomic mass spectrometry data to distinguish differentially expressed protein peaks.
• Cancer Biomarker Monitoring: Used to identify potential biomarkers from blood samples for cancer detection and monitoring.
• Drug Discovery: Applied to large-scale datasets to identify protein markers relevant to drug discovery efforts.

Methodology:

binda applies data-adaptive thresholding to convert continuous protein expression into binary features, performs binary discriminant analysis, and ranks dichotomized features by relative entropy.