Probabilistic estimation of short sequence expression using RNA-Seq data and the “positional bootstrap”

Hui Xiong, Leo J. Lee, Hannes Bretschneider, Jiexin Gao, Nebojša Jojić, Brendan J. FreybioRxiv 2016 · DOI

TL;DR

This paper introduces the "positional bootstrap," a novel method to quantify uncertainty in RNA-Seq expression estimates, particularly for short sequences where traditional methods falter. By accounting for biases in read coverage, this technique significantly improves the accuracy of analyses like alternative splicing, outperforming existing methods by up to tenfold.

What Problem Does This Paper Address?

Estimating gene or transcript expression from RNA-Seq data relies on the assumption that sequencing reads are generated uniformly across the transcript. While this works reasonably well for long transcripts, it leads to significant errors when analyzing short sequences (e.g., less than 100 bp), which are crucial for studying phenomena like splice junctions and microRNAs. These errors stem from various biases in RNA-Seq coverage, including sequence content, experimental conditions, and sample preparation, which are often poorly understood and difficult to model directly. Existing methods either make strong, potentially inaccurate assumptions about these biases or simply ignore them, leading to unreliable expression estimates for short sequences.

Methods

The core innovation is the "positional bootstrap." Instead of trying to correct for known biases with potentially flawed models, this method quantifies the uncertainty introduced by all biases, including unknown ones. The approach treats the number of reads mapped to each genomic position as individual data points. The positional bootstrap then generates multiple "bootstrap" datasets by resampling these positions with replacement. For each bootstrap dataset, expression is estimated using standard methods, assuming bias-free read generation. By analyzing the distribution of estimates across these bootstrap samples, the method quantifies the uncertainty in the original estimate. This allows for more robust confidence measures and better prioritization of estimates. The paper also outlines Bayesian frameworks for expression and splicing level estimation that incorporate this positional bootstrap for improved accuracy.

Datasets

The authors utilized two publicly available RNA-Seq datasets for their analysis:

  • Illumina Bodymap dataset: A 75bp dataset.
  • Kaessmann laboratory human dataset: A 76bp dataset.

These datasets were derived from poly-A selected mRNA of healthy individuals and included four overlapping tissue types: brain, heart, kidney, and liver. The data was mapped to 671,448 exon-exon junctions. The paper does not introduce new benchmark datasets but rather uses these existing ones to demonstrate the effectiveness of their method.

Key Insights and Ideas

  • The "uniform coverage" assumption is a major bottleneck for short sequences: This assumption, while convenient, breaks down significantly for short transcripts, leading to substantial errors in expression estimation.
  • Quantifying uncertainty is more robust than correcting unknown biases: Instead of trying to model complex and often uncharacterized biases, the positional bootstrap directly measures their impact on expression estimates.
  • Sequencing artifacts are pervasive and impact short sequences disproportionately: The analysis reveals that biases are present across different sample preparations and sequence content, and these biases are not necessarily uniform.
  • The positional bootstrap provides a more realistic confidence measure: By accounting for the variability introduced by sequencing artifacts, the positional bootstrap yields confidence estimates that are more reliable, especially for the most confident estimates.

Key Results

  • The positional bootstrap increases the accuracy of estimates for alternative splicing levels, tissue-differential alternative splicing, and tissue differential expression by a factor of up to 10.
  • The method demonstrates that even with a high number of mapped reads, sequence-dependent biases can still significantly affect expression estimates.
  • Comparisons with a "standard method" (likely referring to common existing approaches) show that the positional bootstrap yields significantly fewer errors in estimating PSI (Percent Spliced In) for highly confident exons.
  • The positional bootstrap method shows a lower error rate in detecting tissue-dependent splicing differences and junction expression differences compared to the standard method.

Limitations and Future Work

  • Computational cost: While the paper mentions a Python implementation, the computational cost of generating many bootstrap samples might be a consideration for extremely large datasets. However, the paper also suggests that for certain tasks, the distribution of ψ can be approximated in constant time after the bootstrap samples are generated.
  • Generalizability: The paper suggests that the positional bootstrap idea can be applied to many other problems involving short reads, such as DNA sequencing, other RNA-Seq tasks, CLIP-Seq, and ChIP-Seq, as long as the estimator depends on measurements from genomic positions. This indicates a broad potential for future applications.
  • Integration with existing methods: The authors note that their method can be readily combined with other bias-correcting methods to capture residual biases. This suggests future work could explore synergistic combinations of techniques.

Summary updated

Abstract

Abstract When estimating expression of a transcript or part of a transcript using RNA-seq data, it is commonly assumed that reads are generated uniformly from positions within the transcript. While this assumption is acceptable for long transcript sequences where reads from many positions are averaged, it frequently leads to large errors for short sequences, e.g ., less than 100 bp. Analysis of short sequences, such as when studying splice junctions and microRNAs, is increasingly important and necessitates addressing errors in short-sequence expression estimation. Indeed, when we examined RNA-seq data from diverse studies, we found that large errors are introduced by variations in RNA-seq coverage due to sequence content, experimental conditions and sample preparation. We developed a technique that we call the positional bootstrap, which quantifies the level of uncertainty in expression induced by non-uniform coverage. Unlike methods that attempt to correct for biases in coverage, but do so by making strong assumptions about the form of those biases, the positional bootstrap can quantify the noise induced by all types of bias, including unknown ones. Results obtained using independently generated RNA-seq datasets show that the positional bootstrap increases the accuracy of estimates of alternative splicing levels, tissue-differential alternative splicing and tissue differential expression, by a factor of up to 10. A Python implementation of the algorithm to quantify splicing levels is freely available from github.com/PSI-Lab/BENTO-Seq.
BibTeX
@misc{xiong2016probabilistic,
  author = {Hui Xiong and Leo J. Lee and Hannes Bretschneider and Jiexin Gao and Nebojša Jojić and Brendan J. Frey},
  title = {{Probabilistic estimation of short sequence expression using RNA-Seq data and the “positional bootstrap”}},
  year = {2016},
  doi = {10.1101/046474},
  howpublished = {bioRxiv},
}
Figure