Restricted maximum-likelihood method for learning latent variance components in gene expression data with known and...

Area of Science:

• Genomics
• Statistical Genetics
• Bioinformatics

Background:

• Random effects models are crucial for analyzing genome-wide gene expression data, particularly for correcting spurious correlations caused by hidden confounders.
• Estimating known and latent variance components simultaneously in these models is challenging, often relying on suboptimal numerical methods.

Purpose of the Study:

• To develop a more efficient and accurate method for estimating latent variance components in random effects models for gene expression data.
• To address the limitations of existing gradient-based optimizers in handling known confounding factors.

Main Methods:

• Developed a restricted maximum-likelihood (REML) method based on the analytical proof that latent variables can be orthogonal to known confounding factors.
• The method estimates latent variables via probabilistic principal component analysis on the orthogonal subspace.
• Variance-covariance parameters are estimated using a novel analytic solution.

Main Results:

• The proposed REML method achieves greater or equal likelihood values compared to gradient-based optimizers.
• Latent factors are demonstrably orthogonal to known confounding factors, preventing overlap.
• The computational runtime is reduced by several orders of magnitude, enabling analysis of larger datasets.

Conclusions:

• The REML method offers a significant advancement for random effects modeling in genomics.
• It provides a computationally efficient and statistically robust approach for identifying latent variance components in large-scale gene expression studies.
• This facilitates broader application of these powerful statistical models in genetic research.