Yanagi: Fast and interpretable segment-based alternative splicing and gene expression analysis

Mohamed K. Gunady, Stephen M. Mount, Héctor Corrada BravoBMC Bioinformatics 2019 · DOI

TL;DR

This paper introduces Yanagi, a novel approach that segments transcriptomes into smaller, non-overlapping units. This allows for faster and more interpretable analysis of gene expression and alternative splicing by decoupling pseudo-alignment from transcript quantification, a common bottleneck in RNA-seq analysis.

What Problem Does This Paper Address?

RNA sequencing (RNA-seq) is a powerful tool for understanding gene expression and its changes. A crucial step is quantifying transcript abundance, which is often coupled with pseudo-alignment (a fast method for mapping reads to a reference). This coupling, however, limits the direct application of pseudo-alignment to downstream analyses like alternative splicing or differential gene expression without an additional, often computationally intensive, transcript quantification step. This is particularly problematic because alternatively spliced transcripts share many sequence regions, leading to ambiguous read mappings. Existing methods struggle with incomplete transcript annotations, which can lead to inaccurate estimations of splicing events.

Methods

Yanagi tackles this by introducing transcriptome segmentation. The core idea is to break down the transcriptome into maximal, L-disjoint segments. These segments are designed to be non-overlapping beyond a specified length L, which is typically chosen based on the sequencing read length. This segmentation allows for:

  • Decoupling Pseudo-alignment and Quantification: Yanagi uses pseudo-alignment tools (like Kallisto or Salmon) to align reads to these segments, generating "segment counts" (SCs). These SCs can then be directly used for downstream analyses without needing to quantify full transcript abundances first.
  • Maximal Segments: The algorithm generates "maximal" segments, meaning they are extended as much as possible while maintaining their disjointness. This property helps to reduce variability in segment counts and improve stability.
  • Segment-based Analysis: The SCs are then used for gene-level differential expression and alternative splicing analyses. The paper demonstrates how to adapt existing methods like DEXSeq for this purpose.

Datasets

The authors used the following datasets for their experiments and evaluations:

  • Simulated Data:
    • SwitchTx dataset: Used for simulating differential transcript usage across conditions in fruit fly and human genomes.
  • IncompTx dataset: A novel simulation designed to mimic incomplete transcript annotations, where a new isoform is created by combining two distant splicing events. This dataset was used to evaluate the impact of annotation incompleteness on splicing analysis.
  • Experimental Data:
    • Drosophila melanogaster male fly head RNA-seq dataset (GEO accession number GSM2108304): Used to compare Yanagi's approach with other methods in a real-world scenario with known annotation complexity.

Key Insights and Ideas

  • Breaking the Quantification Bottleneck: The most significant insight is that segment counts, derived from pseudo-alignment, can serve as a direct statistic for downstream analyses, bypassing the need for transcript quantification. This offers a substantial speed and efficiency gain.
  • Robustness to Incomplete Annotations: Yanagi's segment-based approach is shown to be more robust to incomplete transcript annotations compared to isoform-based methods. By focusing on local splicing events captured by segments, it can provide more accurate PSI (percent spliced in) estimates even when the full set of isoforms is not annotated.
  • Interpretability: Segment counts are directly tied to genomic locations and splicing structures, making the results more interpretable for biologists compared to abstract equivalence classes (ECs) used in some other methods.
  • Maximality Property: The "maximal" nature of segments helps to reduce the variability of segment counts, leading to more stable and reliable downstream analyses.

Key Results

  • Reduced Multimapping: Using segments for alignment significantly reduces the number of multi-mapped reads compared to aligning to the full transcriptome, simplifying downstream analysis. For example, using segments achieved around a 10-fold decrease in multimappings.
  • Compression Rates: Segment libraries offer substantial compression rates for the reference transcriptome. For instance, with L=100, Yanagi achieved 54% and 35% compression for fruit fly and human genomes, respectively, in terms of sequence bases.
  • Comparable Performance: Yanagi's segment-based approach achieved comparable performance to established count-based methods (like rMATS for splicing analysis) and isoform-based methods (like SUPPA) in differential gene expression and alternative splicing analyses, while being computationally more efficient.
  • Improved PSI Accuracy with Incomplete Annotations: In simulations with incomplete annotations, Yanagi's PSI estimates showed a much higher correlation (0.92) with the true PSI values compared to SUPPA (0.6), highlighting its advantage in such scenarios.

Limitations and Future Work

  • Current Version Focus: The current version of Yanagi generates L-disjoint segments from genes independently. The authors suggest that future versions could process multiple genes simultaneously to tackle ambiguities arising from overlapping genes and paralogs.
  • Transcript Quantification Extension: The authors propose that segment counts could be extended to enable simpler transcript quantification, potentially with improved bias modeling.
  • Novel Transcript Discovery: Investigating the use of segments for discovering novel transcripts is suggested as a future direction.
  • Parameter L Selection: While the paper discusses the impact of parameter L, further research could explore more automated or adaptive methods for selecting the optimal L for different experimental setups.

Summary updated

BibTeX
@article{gunady2019yanagi,
  author = {Mohamed K. Gunady and Stephen M. Mount and Héctor Corrada Bravo},
  title = {{Yanagi: Fast and interpretable segment-based alternative splicing and gene expression analysis}},
  journal = {BMC Bioinformatics},
  year = {2019},
  doi = {10.1186/s12859-019-2947-6},
}
Figure