A full Bayesian hierarchical mixture model for the variance of gene differential expression

Area of Science:

• Genomics
• Bioinformatics
• Statistical Genetics

Background:

• High throughput microarray experiments often have few replicates, leading to inaccurate gene variance estimates.
• Standard log2 transformation may not fully address heteroscedasticity, and gene-specific variances are unstable with limited data.
• Accurate gene variability estimation is crucial for reliable gene expression comparisons and identifying differentially expressed genes.

Purpose of the Study:

• To develop a more robust method for estimating gene expression variance, particularly when replicate numbers are limited.
• To address the limitations of constant variance assumptions and unstable gene-specific variance estimates.
• To improve the reliability of statistical analyses in high throughput gene expression studies.

Main Methods:

• A Bayesian mixture model was proposed to classify genes based on the similarity of their variances.
• Genes within the same latent class share a common variance estimate, pooled from all replicates within that class.
• The model was applied to a dataset of 9216 genes with four replicates per condition, resulting in four latent classes.

Main Results:

• The Bayesian mixture model successfully classified genes into latent classes based on variance similarity.
• Genes in the same latent class shared variance estimates derived from a larger effective number of replicates.
• An example dataset yielded four distinct latent classes, demonstrating the model's ability to capture variance heterogeneity.

Conclusions:

• The mixture variance model offers a flexible and realistic approach to estimating gene expression variance with limited replicates.
• Utilizing latent class variances provides more robust p-values compared to constant or gene-specific variance estimates.
• This method enhances the reliability of differential gene expression identification in microarray studies.