abSNP: RNA-Seq SNP Calling in Repetitive Regions via Abundance Estimation

Shunfu Mao, Soheil Mohajer, Kannan Ramchandran, David Tse, Sreeram KannanWABI 2017 · DOI

TL;DR

This paper introduces abSNP, a novel RNA-Seq SNP calling method that effectively handles repetitive genomic regions. By leveraging gene expression abundance and read mapping quality, abSNP significantly improves SNP detection accuracy in these challenging areas where traditional methods fail.

What Problem Does This Paper Address?

Identifying genetic variations, particularly Single Nucleotide Polymorphisms (SNPs), is crucial for understanding individual traits and diseases. While DNA sequencing methods like Whole Genome Sequencing (WGS) and Whole Exome Sequencing (WES) are common, RNA sequencing (RNA-Seq) offers a cost-effective alternative, especially for genes with known expression. However, existing SNP callers, primarily designed for DNA-Seq, struggle with RNA-Seq data due to splice junctions and, more importantly, the challenge of mapping reads from repetitive genomic regions. In repetitive regions, reads can map to multiple locations, making it difficult to confidently determine the origin of a variant. Traditional callers often discard these multiply mapped reads, leading to missed SNPs and reduced accuracy. This paper aims to bridge this gap by developing an RNA-Seq SNP caller that can reliably call SNPs even in these difficult repetitive regions.

Methods

The core innovation of abSNP lies in its novel approach to handling multiply mapped reads in repetitive regions. Instead of discarding them, abSNP leverages two key pieces of information derived from RNA-Seq data:

  • Gene Expression Abundance: RNA-Seq data naturally provides information about the expression levels of different genes. abSNP uses this abundance estimation (obtained using tools like RSEM) to infer which gene copy a particular read is most likely to have originated from, especially when reads map to multiple similar gene copies.
  • Mapping Quality (MAPQ) Filtering: While standard aligners provide MAPQ scores, they can be uninformative for multiply mapped reads. abSNP refines these MAPQ scores by incorporating abundance information. This iterative process helps to better estimate the probability that a read truly belongs to a specific genomic locus, allowing for more accurate filtering of low-quality alignments.

After these initial steps, abSNP employs a Bayesian SNP calling algorithm that considers the refined MAPQ scores and abundance estimates to call SNPs at each locus independently. A crucial additional step is Shadow SNP Filtering. This technique addresses the assumption that only one copy in a repetitive region harbors a SNP. It works by identifying and filtering out "shadow SNPs" – variants that are called in repetitive regions due to multiply mapped reads originating from a true SNP in a different, but similar, locus. This filtering step helps to enforce the assumption of a single SNP per locus in repetitive regions, reducing false positives.

Datasets

The authors used simulated data for their evaluation. They developed a custom RNA-Seq SNP simulator to generate datasets with controlled parameters, including:

  • A reference genome (human chromosome 15 of GRCh37).
  • Transcriptome annotations (UCSC gene annotations).
  • A specified number of SNPs (2000 per dataset).
  • Read requirements (2M 100-bp reads with 1% error rate).

This simulation approach was chosen because obtaining real-world ground truth for SNPs in multiply mapped regions is challenging, as existing methods are not designed to call them reliably. The simulated datasets allowed for a direct comparison of abSNP against GATK, a widely-used SNP caller, under controlled conditions.

Key Insights and Ideas

  • Abundance is Key for Repetitive Regions: The most significant insight is that gene expression abundance, a byproduct of RNA-Seq analysis, is a powerful signal for disambiguating reads in repetitive regions. By knowing which genes are highly expressed, we can infer the likely origin of reads that map to multiple similar copies.
  • MAPQ Refinement Matters: Simply using raw MAPQ scores is insufficient for multiply mapped reads. Refining these scores by integrating abundance information provides a more robust measure of alignment confidence, crucial for accurate SNP calling.
  • Shadow SNPs are a Real Problem: The concept of "shadow SNPs" highlights a critical challenge in repetitive regions. Without specific filtering, multiply mapped reads can propagate false SNP signals, leading to inaccurate calls. abSNP's shadow SNP filtering is essential for mitigating this.
  • RNA-Seq Data Has Unique Strengths: The paper demonstrates that RNA-Seq data, beyond just gene expression, offers unique features (like abundance variations) that can be exploited for SNP calling, especially in regions where DNA-Seq methods falter.

Key Results

The performance evaluation on simulated data showed significant improvements for abSNP over GATK, particularly in repetitive regions:

  • Overall Error Reduction: abSNP (with parameter α=1) achieved a total error rate of 102, compared to 189 for GATK.
  • Multiply Mapped SNPs: For uniquely mapped SNPs, abSNP and GATK performed similarly (around 96-97% sensitivity). However, for multiply mapped SNPs, GATK failed to make any calls, while abSNP successfully identified them.
  • Sensitivity vs. Specificity Trade-off: abSNP offers a tunable parameter (α) that allows for a trade-off between false negatives and false positives. Setting α=0 resulted in very few false positives (5) but a higher number of false negatives (187), while α=1 significantly reduced false negatives (to 122) with a modest increase in false positives.
  • Coverage Dependence: Sensitivity for both callers increased with SNP read coverage. abSNP approached 100% sensitivity at the highest coverage levels, even for multiply mapped SNPs.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Real Data Validation: The current evaluation relies on simulated data. Testing abSNP on real RNA-Seq datasets with known ground truth for repetitive regions is a critical next step, though challenging due to the lack of such benchmarks.
  • Paired-End Reads: abSNP does not currently utilize pairing information from paired-end reads, which could potentially improve performance.
  • RNA Editing: The current SNP calls are considered post-transcriptional. Incorporating RNA editing detection could refine SNP calls and quantify its impact.
  • Allele-Specific Expression: abSNP assumes equal allele contribution to abundance. While tested with differing expressions in simulation, explicitly modeling allele-specific expression could further enhance accuracy. This presents a "chicken-and-egg" problem, as accurate SNP calls are needed to quantify allele-specific expression, and vice-versa. A joint approach is suggested.
  • Integration with DNA-Seq: Combining DNA and RNA sequencing data, especially from both normal and diseased tissues, could lead to more comprehensive SNP calling.
  • Somatic Mutation Detection: abSNP could be a valuable tool for detecting somatic mutations in cancer datasets.

Summary updated

BibTeX
@inproceedings{mao2017absnp,
  author = {Shunfu Mao and Soheil Mohajer and Kannan Ramchandran and David Tse and Sreeram Kannan},
  title = {{abSNP: RNA-Seq SNP Calling in Repetitive Regions via Abundance Estimation}},
  booktitle = {WABI},
  year = {2017},
  doi = {10.4230/lipics.wabi.2017.15},
}
Figure