Typical experimental design advice for expression analyses using RNA-seq generally assumes that single-end reads provide robust gene-level expression estimates in a cost-effective manner, and that the additional benefits obtained from paired-end sequencing are not worth the additional cost. However, in many cases (e.g., with Illumina NextSeq and NovaSeq instruments), shorter paired-end reads and longer single-end reads can be generated for the same cost, and it is not obvious which strategy should be preferred. Using publicly available data, Harvard University researchers test whether short-paired end reads can achieve more robust expression estimates and differential expression results than single-end reads of approximately the same total number of sequenced bases.
At both the transcript and gene levels, 2 × 40 paired-end reads unequivocally provide expression estimates that are more highly correlated with 2 × 125 than 1 × 75 reads; in nearly all cases, those correlations are also greater than for 1 × 125, despite the greater total number of sequenced bases for the latter. Across an array of metrics, differential expression tests based upon 2 × 40 consistently outperform those using 1 × 75.
Spearman’s rank correlations for kallisto-derived transcripts per million (TPM) between the gold standard paired-end 2 × 125 strategy and alternative strategies
Violin plots of (a) transcript and (b) gene-level inference. Comparison of correlations with 2 × 125 between 2 × 40 and 1 × 75 for (c) transcript and (d) genes, and between 2 × 40 and 1 × 125 for (e) transcript and (f) genes. For c–f, symbol colors correspond to SRA accessions, and points above the red dotted line are samples where estimates of expression from 2 × 40 is more highly correlated with the gold standard than the contrasted single-end strategy
Researchers seeking a cost-effective approach for gene-level expression analysis should prefer short paired-end reads over a longer single-end strategy. Short paired-end reads will also give reasonably robust expression estimates and differential expression results at the isoform level.