Tunable structure priors for Bayesian rule learning for knowledge integrated biomarker discovery.

Area of Science:

• Computational biology and bioinformatics
• Machine learning for biomedical applications
• Knowledge discovery in life sciences

Background:

• Classification rule learning is crucial for knowledge discovery in biomedicine.
• Existing methods often struggle to effectively integrate prior domain knowledge.
• Bayesian rule learning (BRL) provides a framework for rule-based classification.

Purpose of the Study:

• To develop a framework, BRLp, that incorporates background domain knowledge into classification rule learning.
• To enable tunable integration of prior knowledge into Bayesian belief networks (BNs).
• To improve knowledge discovery and predictive performance in biomedical datasets.

Main Methods:

• Extended Bayesian rule learning (BRL) by incorporating informative structure priors into the Bayesian score.
• Introduced a hyperparameter (λ) to control the degree of prior knowledge incorporation.
• Evaluated BRLp using simulated data and a real-world lung cancer prognostic biomarker dataset.

Main Results:

• Increasing the hyperparameter λ in BRLp enhanced the influence of prior knowledge on model learning.
• Higher λ values led to BRLp learning the true data-generating model for simulated data, improving predictive performance (AUC).
• In lung cancer data, increased λ improved the incorporation of known biomarkers (e.g., EGFR), boosting AUC.

Conclusions:

• BRLp successfully integrates background domain knowledge and data through tunable structure priors.
• The framework demonstrates effective knowledge discovery in biomedical applications, exemplified by lung cancer biomarker analysis.
• BRLp offers a robust approach for enhancing classification rule learning by leveraging prior biological insights.