Optimal Bayesian Transfer Learning for Count Data.

Area of Science:

• Biomedical data analysis
• Computational biology
• Machine learning in genomics

Background:

• Limited omics data hinders accurate predictive model design in biomedicine.
• Omics data often share underlying biological pathways and disease mechanisms.
• Leveraging data from relevant source domains can improve predictions in data-limited target domains.

Purpose of the Study:

• To develop Bayesian transfer learning methods for next-generation sequencing (NGS) data analysis.
• To improve prediction accuracy and reliability in target domains with scarce labeled data.
• To propose an Optimal Bayesian Transfer Learning (OBTL) classifier.

Main Methods:

• Formulating transfer learning within a fully Bayesian framework.
• Defining relatedness via joint prior distributions of source and target domain model parameters.
• Utilizing the Negative Binomial model for overdispersed RNA-seq count data and proposing the OBTL classifier to minimize expected classification error.

Main Results:

• The proposed Optimal Bayesian Transfer Learning (OBTL) classifier effectively transfers knowledge from source to target domains.
• The OBTL classifier demonstrates improved predictive performance on both synthetic and real-world cancer genomics data.
• Bayesian transfer learning provides a robust approach for omics data analysis with limited sample sizes.

Conclusions:

• Bayesian transfer learning offers a powerful framework for enhancing predictive modeling with omics data.
• The OBTL classifier effectively addresses challenges posed by data scarcity and overdispersion in RNA-seq data.
• This approach holds significant potential for advancing precision medicine and biomedical research through improved data utilization.