Abstract
Background The human transcriptome annotation is regarded as one of the most complete of any eukaryotic species. However, limitations in sequencing technologies have biased the annotation toward multi-exonic protein coding genes. Accurate high-throughput long read transcript sequencing can now provide stronger evidence for genes that were previously either undetectable or impossible to differentiate from sequencing noise such as rare transcripts, mono-exonic, and non-coding genes.Results We analyzed Sequel II Iso-Seq sequencing data of the Universal Human Reference RNA (UHRR) using the Transcriptome Annotation by Modular Algorithms (TAMA) software. We found that the convention of using mapping identity to measure error correction performance does not reflect actual gain in accuracy of predicted transcript models. In addition, inter-read error correction leads to the thousands of erroneous gene models. Using genome assembly based error correction and gene feature evidence, we identified thousands of potentially functional novel genes.Conclusions The standard of using inter-read error correction for long read RNA sequencing data could be responsible for genome annotations with thousands of biologically inaccurate gene models. More than half of all real genes in the human genome may still be missing in current public annotations. We require better methods for differentiating sequencing noise from real genes in long read RNA sequencing data.

This publication has 21 references indexed in Scilit: