BowSaw: Inferring Higher-Order Trait Interactions Associated With Complex Biological Phenotypes

Area of Science:

• Computational biology
• Bioinformatics
• Systems biology

Background:

• Machine learning aids in interpreting biological complexity from large datasets (genomic, transcriptomic, metagenomic).
• Understanding how variables contribute to model output offers mechanistic insights beyond black-box approaches.

Purpose of the Study:

• To develop a novel suite of algorithms, BowSaw, for interpreting machine learning models.
• To identify combinations of variables (rules) frequently used in classification by random forests.
• To generate testable biological hypotheses from complex biological data.

Main Methods:

• Developed BowSaw, a suite of algorithms leveraging random forest structure.
• Applied BowSaw to simulated data to assess rule recovery accuracy under noise.
• Utilized BowSaw on Human Microbiome Project data for microbial association analysis.

Main Results:

• BowSaw accurately recovered complex Boolean rules from simulated data, even with high noise.
• Identified novel, high-order combinations of microbial taxa associated with Crohn's disease in the Human Microbiome Project dataset.

Conclusions:

• BowSaw offers a new method for extracting mechanistic insights from machine learning models.
• The approach facilitates hypothesis generation for complex diseases like Crohn's disease.
• Leveraging random forest structures enhances the interpretability of biological data analysis.