In the recent article, 'A benchmark for RNA-seq quantification pipelines', we generated a simulated human RNA-seq dataset that combined real paired-end reads from GEUVADIS samples aligned by TopHat2 to chr1 and simulated reads from transcripts on chr2. After publication, Zsuzsanna Etches and Mikael Christensen noticed that some of the read pairs from chr1 were duplicates of the same sequence (both reads in a pair had the identical sequence). We had unintentionally included multiple alignments when extracting chr1 reads from the GEUVADIS alignment files. When multiple alignments were available for one of the read pairs and the other read was missing (e.g. it had more than 20 alignments and was in the unmapped.bam file), then such an alignment would be written by bedtools bamtofastq to both FASTQ files. We have since generated a new simulated dataset, in which we extract only unique, proper pair alignments from chr1 to combine with the original simulated reads from transcripts on chr2. We have run the evaluation pipeline on the new 'cleaned' simulated dataset and the results do not change qualitatively (see figure below). The original FASTQ files and tables from the paper are still available on the webtool labelled as "ORIGINAL SIMULATION", along with the new FASTQ files and tables labelled as "SIMULATION". We recommend for testing purposes to use the new simulation dataset.
http://rafalab.rc.fas.harvard.edu/rnaseqbenchmark
