TL;DR
RNA sequencing (RNA-seq) is a powerful tool, but it's plagued by technical biases that can skew results. This paper introduces IVT-seq, a novel method using in vitro transcribed RNA to precisely quantify these biases. They found significant within-transcript coverage variability and unpredictable "hotspots" of coverage, with rRNA depletion being a major culprit, suggesting caution in exon-level analysis.
What Problem Does This Paper Address?
RNA sequencing (RNA-seq) has revolutionized our ability to study gene expression, enabling detailed analysis of transcription, splicing, and more. However, the process of preparing RNA samples for sequencing (library preparation) involves multiple steps, each with the potential to introduce technical biases. These biases can lead to inaccurate measurements of gene expression, making it difficult to distinguish true biological differences from technical artifacts. While some biases are known (like GC content bias), a comprehensive understanding of how different library preparation methods and their components contribute to these biases, especially at the within-transcript level, has been lacking. This makes interpreting RNA-seq data, particularly for complex analyses like alternative splicing, challenging.
Methods
The core innovation here is IVT-seq (in vitro transcription sequencing). The authors created a large pool of over 1000 full-length human RNA transcripts in vitro from cDNA clones. Because these transcripts are synthesized in a controlled environment and their sequences are known, any observed variations in sequencing coverage are attributable to technical biases, not biological factors like splicing.
They then subjected these IVT RNA pools to various standard RNA-seq library preparation protocols, including:
- PolyA selection: Isolates mRNA based on its poly-A tail.
- Total RNA sequencing: Includes all RNA types.
- rRNA depletion: Removes highly abundant ribosomal RNA.
They also generated a plasmid DNA-seq library from the original cDNA clones to serve as a baseline, free from RNA-specific biases. By comparing the sequencing data from these different approaches, they could dissect the contributions of various steps and components to the observed biases. Computational analyses were used to quantify coverage variability and identify sequence features associated with these biases.
Datasets
The primary "dataset" used for evaluation was a pool of over 1000 in vitro transcribed (IVT) human RNA transcripts. These were derived from the Mammalian Gene Collection (MGC). To mimic real biological samples and assess the impact of different RNA ratios, these IVT transcripts were mixed with mouse liver total RNA at various concentrations.
The authors also generated simulated RNA-seq data using the BEERS software package to serve as a control for fragmentation and sequencing randomness.
Key Insights and Ideas
- RNA-seq is inherently biased: Even with "perfect" RNA molecules (IVT), standard library preparation methods introduce significant variability in how different parts of a transcript are sequenced.
- rRNA depletion is a major culprit: This step, crucial for removing abundant rRNA, introduces substantial and often unappreciated biases in coverage.
- Within-transcript variation is extreme: It's not just about which transcripts are detected, but how they are detected. Over 50% of transcripts showed more than a 2-fold difference in coverage along their length.
- "Hotspots" of unpredictable coverage exist: Over 6% of transcripts had regions with highly variable and unpredictable coverage across different samples, making accurate quantification difficult.
- Sequence features matter, but are modulated by library prep: While sequence characteristics like hexamer entropy and GC content influence coverage, their impact is significantly altered by the specific RNA selection method used.
- Exon-level analysis might be problematic: The observed biases, particularly the unpredictable coverage regions, raise concerns about the reliability of exon-level quantification for tasks like alternative splicing detection.
Key Results
- ~50% of IVT transcripts showed > 2-fold differences in within-transcript coverage.
- ~10% of IVT transcripts showed > 10-fold differences in within-transcript coverage.
- > 6% of IVT transcripts contained regions of high, unpredictable sequencing coverage (huncs).
- rRNA depletion was identified as a significant contributor to coverage variability.
- PolyA selection also introduced bias, though rRNA depletion was found to be equally or more impactful.
- Lower hexamer entropy (less sequence complexity) was strongly associated with higher coverage variability.
Limitations and Future Work
- Underestimation of hunc regions: The authors acknowledge that their stringent filtering criteria likely led to an underestimate of the true number of high coverage variability regions.
- Unclear cause of hunc regions: While they ruled out mapping artifacts and direct interactions with background RNA, the exact cause of these unpredictable coverage regions remains elusive, possibly involving complex trans interactions.
- Future work: The authors hope that making their IVT-seq data publicly available will spur the development of new experimental and analytical methods to better account for and mitigate these biases. They also suggest that IVT-seq itself can be broadly applied to assess new RNA-seq technologies and protocols, not just for Illumina sequencing but for other platforms as well. They emphasize the need for further understanding of these biases to fully realize the potential of digital gene expression analysis.
Summary updated