CANTARE

CANTARE constructs predictive regression models from network neighborhoods to integrate multi-omic data and produce interpretable outcome predictions.

Key Features:

• Pairwise regression modeling: Generates pairwise regression models across all analyte pairs from multiple omes.
• Network encoding: Encodes pairwise relationships into a network structure representing analyte interdependencies.
• Network-neighborhood predictors: Builds predictive logistic regression models centered on specific network neighborhoods.
• Interaction modeling: Supports inclusion of an interaction term for IBD in pairwise models to inform network construction.
• Quantitative directional effects: Reports quantitative and directional effect sizes for predictors to support hypothesis generation.
• Multi-omic support: Applies to datasets including gut microbiome, metabolomics, and microbial-derived enzymes.
• Model parsimony: Produces parsimonious models with predictor counts reported between 3 and 13 for top models.
• Performance metrics: Evaluates models using area under the curve (AUC) and statistical comparisons (including p-values).
• Method comparisons: Benchmarks against random forests and elastic-net penalized regressions and against correlation networks for performance assessment.
• Predictor prioritization and dynamic range: Demonstrates prioritization of predictors from multiple omes and a broader dynamic range in predicted probabilities (p = 1.35 × 10^-5; p = 0.005 reported for comparisons).

Scientific Applications:

• Disease outcome prediction: Building interpretable predictive models for inflammatory bowel disease (IBD) using multi-omic data.
• Multi-omic integration: Integrating gut microbiome, metabolomics, and microbial-derived enzyme data to identify cross-omic associations.
• Method benchmarking: Comparing network-based logistic regression models to random forests, elastic-net penalized regressions, and correlation networks using AUC and statistical tests.
• Hypothesis generation: Using quantitative directional effect sizes from models to generate testable biological hypotheses.

Methodology:

Generate pairwise regression models across all analyte pairs, encode significant pairwise relationships into a network, and construct predictive logistic regression models from network neighborhoods; optionally include an interaction term for IBD, evaluate model AUCs, and compare performance to random forests, elastic-net penalized regressions, and correlation-network-derived models using statistical tests (reported p-values).