Incorporating biological prior knowledge for Bayesian learning via maximal knowledge-driven information priors.

Area of Science:

• Computational Biology
• Machine Learning
• Bioinformatics

Background:

• Phenotypic classification is challenging with small sample sizes, necessitating the use of prior knowledge.
• Optimal Bayesian classification addresses model uncertainty by directly considering feature-label distributions, outperforming classical methods.
• Prior knowledge, such as genetic pathways, can be integrated into classification models.

Purpose of the Study:

• To develop a novel methodology for constructing priors in optimal Bayesian classification.
• To address the challenge of prior construction in optimal Bayesian classification, particularly for small and unlabeled datasets.
• To enhance phenotypic classification by effectively utilizing existing biological knowledge.

Main Methods:

• Proposed a new prior construction framework based on conditional probability statements, termed the maximal knowledge-driven information prior (MKDIP).
• Developed a flexible constraint framework that accommodates potential inconsistencies and integrates diverse knowledge sources like population statistics.
• Extended optimal Bayesian classification to handle multinomial mixture models with unlabeled data.

Main Results:

• The MKDIP framework offers greater flexibility and improved inference compared to previous methods.
• Demonstrated effective application on mammalian cell-cycle and p53-related pathways, and a non-small cell lung cancer gene expression dataset.
• Showcased superior classifier design for small, unstructured datasets using the extended optimal Bayesian classification for multinomial mixtures.

Conclusions:

• The proposed general prior construction framework enhances inference by effectively utilizing prior knowledge.
• The extension to unlabeled multinomial mixtures enables superior classifier design for small, unstructured datasets.
• The underlying theory is broadly applicable to various knowledge types and small-sample scenarios beyond the demonstrated pathway analysis.