Gene co-expression analysis is an attractive tool for leveraging enormous amounts of public RNA-seq datasets for the prediction of gene functions and regulatory mechanisms. However, the optimal data processing steps for the accurate prediction of gene co-expression from such large datasets remain unclear. Especially the importance of batch effect correction is understudied.
Kyoto University researchers processed RNA-seq data of 68 human and 76 mouse cell types and tissues using 50 different workflows into 7,200 genome-wide gene co-expression networks. They then conducted a systematic analysis of the factors that result in high-quality co-expression predictions, focusing on normalization, batch effect correction, and measure of correlation. The researchers confirmed the key importance of high sample counts for high-quality predictions. However, choosing a suitable normalization approach and applying batch effect correction can further improve the quality of co-expression estimates, equivalent to a >80% and >40% increase in samples. In larger datasets, batch effect removal was equivalent to a more than doubling of the sample size. Finally, Pearson correlation appears more suitable than Spearman correlation, except for smaller datasets.
Summary of this study
(A) Raw RNA-seq data was processed with 50 different combinations of normalization, batch effect correction, and correlation measures into 7,200 genome-wide sets cell type and tissue-specific co-expression predictions, which we refer to as “co-expression networks”. (B) Quality of co-expression networks was estimated based on the enrichment of functional annotations of correlated genes and regulatory motifs in their promoters. In random co-expression networks no common annotations and motifs are expected to be found among correlated genes. In contrast, in ideal networks such enrichments should be encountered frequently. Here nodes represent genes and edges co-expression. (C) Quality measures were processed into 7,200 quality scores, which were used for downstream analysis. We use 75% of cell types and tissues for regression and exploratory analysis, and the remaining 25% for validation.
A key point for accurate prediction of gene co-expression is the collection of many samples. However, paying attention to data normalization, batch effects, and the measure of correlation can significantly improve the quality of co-expression estimates.