A Dirichlet-Multinomial Bayes Classifier for Disease Diagnosis with Microbial Compositions

Area of Science:

• Microbiome research
• Computational biology
• Disease diagnostics

Background:

• Microbial community dysbiosis is linked to human diseases.
• Microbial compositions show potential as disease biomarkers.
• Accurate classification of microbiome data is crucial for diagnostics.

Purpose of the Study:

• To develop a novel Bayes classifier for microbiome data analysis.
• To improve disease diagnosis accuracy using microbial compositions.
• To create a tool for identifying disease-associated microbial taxa.

Main Methods:

• Modeled microbial compositions using Dirichlet-multinomial distributions.
• Estimated distribution parameters via maximum likelihood estimation.
• Calculated posterior probabilities using Bayes' theorem and incorporated prior disease prevalence information.

Main Results:

• The developed Dirichlet-multinomial Bayes classifier (DMBC) demonstrated superior classification accuracy compared to existing Bayesian and random forest methods.
• DMBC features automatic selection of informative microbial taxa, enhancing classification performance.
• The method maintained or improved accuracy when analyzing species-level taxa.

Conclusions:

• DMBC offers a powerful alternative for disease diagnosis using microbiome data.
• Incorporating prior disease prevalence improves diagnostic probability estimation.
• The developed R package provides a freely accessible tool for microbiome-based diagnostics.