Semi-supervised discovery of differential genes.

Area of Science:

• Bioinformatics
• Statistical Genetics
• Computational Biology

Background:

• Traditional methods for gene significance evaluation assume well-controlled experimental conditions and available sample labels.
• Clinical studies often face challenges in obtaining precise condition labels due to practical constraints, necessitating alternative analytical approaches.
• Unsupervised and semi-supervised learning offer viable alternatives when complete sample labeling is infeasible.

Purpose of the Study:

• To adapt statistical gene significance evaluation for scenarios with limited or no condition labels.
• To enhance the detection of differentially expressed genes in unsupervised and semi-supervised settings.
• To improve the performance of gene discovery by incorporating unlabeled samples.

Main Methods:

• Development of a latent variable model for gene expression analysis.
• Application of the optimal discovery procedure (ODP) framework to the latent variable model.
• Comparison of two ODP implementations differing in latent variable handling through simulation studies.

Main Results:

• The latent variable model successfully extends gene significance scoring to unsupervised and semi-supervised cases.
• Sharing estimated model parameters across multiple tests within the ODP framework improves the detectability of significant genes.
• Sharing latent variable estimations between different ODP implementations was found to be effective in increasing the detection rate of true positive genes.
• Utilizing unlabeled samples in gene rating significantly improved active gene detection in real-world gene discovery applications.

Conclusions:

• The optimal discovery procedure (ODP) framework is robust and effective for statistical hypotheses involving latent variables.
• Sharing estimations of latent variables across multiple tests further enhances the performance and reliability of the ODP framework.
• The proposed methods offer a valuable tool for gene discovery in complex biological datasets where sample labeling is incomplete or absent.