PaperlayerView exact revision on arXiv

Source-linked AI summary

LUMPY: A probabilistic framework for structural variant discovery

Ryan M. Layer, Ira M. Hall, Aaron R. Quinlan

arXiv:1210.2342v2q-bio.GN

TL;DR

Structural variation data has lagged behind SNP data because SV discovery is fundamentally more difficult, including the challenge of distinguishing bona fide variants from alignment-related false positives. LUMPY uses a general probabilistic framework to combine multiple SV detection signals and is more sensitive than existing tools across SV types and coverage levels.

  • Problem

    Structural variation data has lagged behind SNP data because SV discovery is fundamentally more difficult, including distinguishing bona fide variants from false positives caused by alignment.

  • Method

    LUMPY represents SV breakpoints probabilistically and combines multiple SV detection signals without modifying the underlying SV detection algorithm's logic.

  • Results

    LUMPY is more sensitive than existing discovery tools across all SV types and coverage levels.

  • Takeaways & Limitations

    LUMPY provides a general framework for SV discovery and supports identifying low-frequency variants within heterogeneous tumor samples.

  • Takeaways & Limitations

    Further performance improvements are expected from incorporating raw read-depth data.

Abstract

from arXiv · show

Comprehensive discovery of structural variation (SV) in human genomes from DNA sequencing requires the integration of multiple alignment signals including read-pair, split-read and read-depth. However, owing to inherent technical challenges, most existing SV discovery approaches utilize only one signal and consequently suffer from reduced sensitivity, especially at low sequence coverage and for smaller SVs. We present a novel and extremely flexible probabilistic SV discovery framework that is capable of integrating any number of SV detection signals including those generated from read alignments or prior evidence. We demonstrate improved sensitivity over extant methods by combining paired-end and split-read alignments and emphasize the utility of our framework for comprehensive studies of structural variation in heterogeneous tumor genomes. We further discuss the broader utility of this approach for probabilistic integration of diverse genomic interval datasets.

Background

Structural variation comprises diverse, relatively large genomic alterations with substantial potential to affect gene structure or dosage. SV discovery from DNA sequencing remains difficult because breakpoint regions compromise alignment accuracy and existing methods often integrate few signals.

  • SV biology: Structural variation includes deletions, duplications, insertions, inversions, translocations, and complex rearrangements affecting relatively large genomic segments.These variants are generally larger than 100 bp and are less common than SNPs.
  • SV biology: SVs have greater functional potential than smaller-scale variation because they are more likely to alter gene structure or dosage.
  • Detection challenge: SV detection and genotyping lag behind SNP analysis because SVs vary in size, architecture, and genomic context, while novel junctions compromise read alignment accuracy.
  • Detection challenge: SVs produce read-depth, read-pair, and split-read alignment signals, but these diverse signals are difficult to integrate and most algorithms use only one.
  • Motivation: Integrating multiple signals can improve sensitivity, particularly for low-coverage data and heterogeneous cancer samples where variants may occur in only a subset of cells.
  • Motivation: Even at high coverage, combining signals can increase specificity by supporting more stringent criteria for reporting variant calls.

Results

LUMPY is a probabilistic framework that integrates multiple SV signals and evidence sources through breakpoint probability representations. Across coverage levels, SV types, and heterogeneous-sample settings, it generally improved sensitivity while maintaining equivalent or lower false discovery rates.

  • Framework: LUMPY integrates multiple SV detection signals, including read alignments and prior evidence, and can incorporate future evidence sources through breakpoint probability mappings.
  • NA12878 evaluation: LUMPY was consistently more sensitive than other algorithms across nearly all coverage levels and SV types, with DELLY detecting three more translocations at 20X.
  • NA12878 evaluation: At 2X and 5X coverage, LUMPY detected 32.4% and 87.2% of deletions, compared with 7.4% and 49.8% for GASVPro.
  • NA12878 evaluation: At 5X coverage, LUMPY was 35.5 times more sensitive than PINDEL for translocations, detecting 69.1% versus 2%.
  • NA12878 evaluation: At 10–50X coverage, LUMPY sensitivity ranged from 88.8% to 99.6% across SV types, exceeding the reported ranges for GASVPro, DELLY, and PINDEL.
  • NA12878 evaluation: LUMPY’s FDR remained below 2% except in the highest-coverage cases, while its overall reported range was 0% to 7.2%.
  • Heterogeneous samples: At 10X coverage and 20% SV allele frequency, LUMPY detected 31% of SVs versus 9.2%, 10.9%, and 6.2% for GASVPro, DELLY, and PINDEL.
  • Framework: The framework combines read-pair and split-read evidence within one discovery process, producing a super-additive benefit relative to either signal alone.

Discussion

LUMPY is a general probabilistic framework that integrates diverse SV signals through paired breakpoint probability distributions. The framework is more sensitive than existing tools across SV types and coverage levels, while supporting priors, evolving evidence sources, and uncertainty-aware genomic interval comparisons.

  • Performance: LUMPY is more sensitive than existing discovery tools across all SV types and coverage levels, with increased sensitivity demonstrated in real and simulated human genome datasets.The paper also reports equivalent or lower false discovery rates and greater sensitivity than competing algorithms across nearly all coverage levels.
  • Framework: LUMPY integrates disparate SV signals by representing predicted breakpoint intervals as paired probability distributions.Supported signals include paired-end and split-read alignments, assembled-contig alignments, read-depth measurements, and copy-number variants from segmentation or array data.
  • Framework: The framework can incorporate any SV detection signal that maps to a breakpoint probability distribution, without modifying the underlying detection logic.Adding a new signal requires a module that converts it into a paired probability distribution.
  • Performance: Previously discovered SVs can serve as priors to enhance discovery sensitivity, although substantial gains may require more comprehensive and accurate SV catalogs.This flexibility is demonstrated using the NA12878 genome.
  • Broader utility: Probability distributions propagate positional uncertainty, improve breakpoint localization, and enable final distributions to be compared across studies or reused as priors.The same uncertainty-aware interval-comparison strategy may apply to genomic features whose exact coordinates are unknown.
  • Limitations and future improvements: Including read-depth information is expected to improve performance for duplication and deletion variants, with raw read-depth potentially providing larger gains than segmented outputs.At present, segmented copy-number outputs can be converted into breakpoint probability distributions for LUMPY.

Materials and Methods

LUMPY represents heterogeneous breakpoint evidence in a common probabilistic framework, allowing multiple evidence classes and sources to be integrated for SV discovery. It maps evidence to breakpoint intervals, merges corroborating predictions, and produces calls when user-defined evidence criteria are met.

  • Framework design: LUMPY accommodates multiple classes of breakpoint evidence within a flexible framework for breakpoint prediction.The framework supports evidence from different samples, evidence subclasses, and prior information about known variant positions.
  • Evidence sources: The demonstrated framework combines paired-end and split-read evidence and can incorporate prior variants, read-depth calls, and future evidence sources through new modules.New modules map additional SV detection signals to paired probability distributions without changing the underlying detection logic.
  • Breakpoint representation: Breakpoints are represented by left and right genomic intervals with probability vectors describing the relative likelihood of each endpoint position.The breakpoint tuple also records corroborating evidence and breakpoint variety, with variety based on evidence orientation.
  • Breakpoint merging: Corroborating breakpoint predictions with the same variety are merged, using a liberal mean-based interval procedure to reduce sensitivity to early outliers.Spurious alignments can pull merged intervals away from the true breakpoint, while later trimming removes weakly supported regions.
  • SV calling: SV calls are made when breakpoint clusters meet a user-defined minimum evidence threshold, and endpoint distributions are combined and trimmed by support.The framework sums aligned probability vectors across evidence and retains positions in a high-support percentile, such as the top 99.9 percent.
  • Framework design: An abstract breakpoint evidence type standardizes how each evidence instance detects breakpoints, assigns SV variety, and maps signals to breakpoint intervals.The interface includes is_bp, get_v, and get_bpi functions.

Figure Legends

The figures compare LUMPY’s multi-signal workflows with other SV tools across variant types, coverage levels, allele frequencies, and real or simulated data. Across these evaluations, LUMPY generally improves sensitivity, particularly under low-evidence conditions, while maintaining low false discovery rates.

  • Figure 1: LUMPY integrates read-pair, split-read, read-depth, and prior-evidence signals across single- and multi-sample workflows.The framework supports both three-signal workflows from one sample and single-signal workflows across multiple samples.
  • Figure 2: LUMPY’s false discovery rate remained below 2% in all but the highest-coverage cases, despite some cases with many false positives.Higher false-positive counts were counterbalanced by higher true-positive counts, producing similar FDR in those cases.
  • Figure 3: Combining paired-end and split-read signals produced a super-additive sensitivity gain for LUMPY at low SV allele frequencies.At 5% allele frequency, GASVPro and DELLY showed unchanged or only modestly improved sensitivity when adding a second signal.
  • Figure 5: LUMPY’s performance remained favorable across ROC thresholds, indicating that the advantage was not restricted to an arbitrary evidence threshold.ROC curves evaluated deletion predictions in NA12878 at 5X and 50X coverage using shared input alignments and truth-set criteria.
Loading 1210.2342v2…