Fully moderated t-statistic in linear modeling of mixed effects for differential expression analysis

Area of Science:

• Genomics
• Statistical Bioinformatics
• Computational Biology

Background:

• Gene expression profiling with few replicates causes high variability in gene variance estimates.
• Existing moderated t-test methods, often using linear models with fixed effects, struggle with complex correlation structures.
• There is a need for moderated methods applicable to linear models with mixed effects for robust differential expression analysis.

Purpose of the Study:

• To implement and evaluate a fully moderated t-statistic method for linear models with mixed effects.
• To address limitations of current methods in handling complex correlation structures in gene expression data.
• To improve the power and accuracy of differential expression testing.

Main Methods:

• Developed a fully moderated t-statistic method within a hierarchical Bayes framework.
• Smoothed both residual variances and variance estimates of random effects.
• Compared the proposed method against two existing moderated methods.

Main Results:

• The proposed method effectively controls the expected number of false positives at the nominal level.
• Existing moderated methods failed to control false positives adequately.
• Demonstrated improved performance in differential expression analysis under complex correlation structures.

Conclusions:

• The novel approach offers variance shrinkage for differential expression testing.
• The method successfully handles complex correlation structures.
• Provides improved statistical power and reliable control of false positives.