Comparison of normalization approaches for gene expression studies completed with high-throughput sequencing.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate RNA-Seq data normalization is vital for reliable transcriptomic study findings.
• Technical artifacts in high-throughput sequencing data necessitate robust normalization strategies.
• Existing methods for library size and across-sample normalization vary in their effectiveness.

Purpose of the Study:

• To compare widely used library size (UQ, TMM, RLE) and across-sample (SVA, RUV, PCA) normalization methods for RNA-Seq data.
• To evaluate the performance of across-sample normalization methods in estimating technical artifacts using simulated data.
• To investigate the impact of reduced degrees of freedom on differential expression analysis after normalization.

Main Methods:

• Utilized publicly available RNA-Seq data from The Cancer Genome Atlas (TCGA) cervical cancer study (CESC).
• Conducted extensive simulations to assess artifact estimation by across-sample normalization methods.
• Analyzed the effect of normalization on degrees of freedom and downstream differential expression analysis.

Main Results:

• TMM and RLE library size normalization methods yielded similar results for the CESC dataset.
• The SVA ("BE") method demonstrated superior performance in estimating latent artifacts compared to SVA ("Leek") and PCA.
• Ignoring the loss of degrees of freedom post-normalization led to inflated Type I error rates in differential expression analysis.

Conclusions:

• Recommends adjusting for library size differences and assessing technical artifacts (known and unknown) for robust RNA-Seq analysis.
• Suggests incorporating known and estimated latent artifacts into the design matrix to correctly account for the loss of degrees of freedom.
• Emphasizes performing across-sample normalization when necessary, rather than solely relying on post-processed normalized data.