Recurrent functional misinterpretation of RNA-seq data caused by sample-specific gene length bias

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• RNA sequencing (RNA-seq) data normalization is crucial for removing technical biases.
• Common normalization methods fail to address a prevalent sample-specific gene length effect.
• This length bias creates an association between gene length and fold-change estimates.

Purpose of the Study:

• To identify and characterize a sample-specific length bias in RNA-seq data.
• To evaluate the impact of this bias on gene-set enrichment analysis (GSEA).
• To assess the effectiveness of different normalization methods in correcting this bias.

Main Methods:

• Analysis of numerous RNA-seq datasets to detect sample-specific length effects.
• Comparison of standard normalization methods (RPKM, TMM, RLE, quantile) against advanced methods (cqn, EDASeq).
• Evaluation of gene-set enrichment analysis (GSEA) results before and after normalization.

Main Results:

• A significant sample-specific gene length bias was detected, impacting fold-change estimates.
• Standard normalization methods did not correct this bias, leading to false positives in GSEA.
• Conditional quantile normalization (cqn) and EDASeq effectively removed the length bias, reducing GSEA false positives.
• Gene-set tests accounting for intergene correlations reduced false positives but also decreased statistical power.

Conclusions:

• Sample-specific length bias is a critical issue in RNA-seq analysis, causing functional misinterpretation.
• Conditional quantile normalization (cqn) and EDASeq are recommended for correcting this bias.
• Accounting for intergene correlations in gene-set tests can mitigate bias but requires careful consideration of power.