CoDaCoRe

CoDaCoRe identifies sparse, interpretable, and predictive log-ratio biomarkers from high-throughput sequencing (HTS) compositional data to analyze relationships among compositional variables and predict biological outcomes.

Key Features:

• Sparse log-ratio biomarker selection: Produces sparse, interpretable log-ratio balances as candidate biomarkers.
• Continuous relaxation of combinatorial optimization: Converts the discrete log-ratio selection problem into a continuous relaxation for optimization.
• Gradient descent optimization: Optimizes the relaxed objective using gradient descent.
• Deep learning integration: Leverages deep learning techniques to enable optimization and model learning.
• Scalability to high-dimensional HTS/CoDa data: Designed to scale to high-dimensional compositional datasets such as metagenomic data.
• State-of-the-art predictive accuracy and sparsity: Maintains competitive predictive performance while enforcing sparsity in selected log-ratios.
• Computational speed: Operates several orders of magnitude faster than existing sparse log-ratio selection methods.
• Benchmark validation: Validated across microbiome, metabolite, and microRNA benchmark datasets.

Scientific Applications:

• Microbiome biomarker discovery: Identification of log-ratio biomarkers from microbiome HTS datasets.
• Metabolomics biomarker discovery: Discovery of predictive log-ratio features in metabolite datasets.
• microRNA biomarker discovery: Selection of interpretable log-ratio biomarkers from microRNA data.
• Compositional data analysis and predictive modeling: Analysis of compositional (CoDa) HTS data for interpretable predictive models in high-dimensional settings.

Methodology:

Applies a continuous relaxation of the combinatorial log-ratio selection problem and optimizes the resulting objective using gradient descent within a deep learning framework.