Statistical analysis of a Bayesian classifier based on the expression of miRNAs

Area of Science:

• Biomarkers
• Cancer Diagnostics
• Bioinformatics

Background:

• MicroRNA (miRNA) levels are explored as diagnostic and prognostic tools for various cancers.
• Reliable cancer classifiers require addressing overlooked aspects like miRNA expression distribution and statistical uncertainty.

Purpose of the Study:

• Analyze miRNA expression distributions and statistical uncertainty in classifier development.
• Develop a Bayesian classifier for distinguishing pulmonary adenocarcinoma from squamous cell carcinoma using specific miRNA expressions.

Main Methods:

• Investigated the distribution of miRNA expression triplicates and their averages.
• Developed a Bayesian classifier using miR-205, miR-21, and snRNA U6 expression data.
• Exploited correlations between miRNAs and normalized miRNA expression to enhance classifier performance.

Main Results:

• Proved that miRNA expression triplicates and their averages are well-described by normal distributions.
• Demonstrated that correlations between miRNAs can enhance classifier performance.
• Identified methods for outlier detection using chi-square and Student's t-tests based on normal distributions.

Conclusions:

• Normal distribution of miRNA expression allows for optimal Bayesian classifier setting and performance evaluation.
• Outlier samples can be identified by analyzing variability and displacement from population means.
• The study provides a statistically robust framework for miRNA-based cancer classification.