Source-linked AI summary
Haplotype-based variant detection from short-read sequencing
Erik Garrison, Gabor Marth
TL;DR
Short-read haplotype detection requires variant models that handle multiallelic loci and non-uniform copy number. The paper generalizes a Bayesian framework for these settings and implements it in FreeBayes, providing a foundation for directly detecting longer, multi-base alleles from sequence alignments.
Problem
Existing variant detectors commonly assume biallelic loci and uniform copy number, limiting unified haplotype detection across multiallelic and variable-copy contexts.
Method
The authors generalize a Bayesian model to multiallelic loci and non-uniform sample copy numbers, implementing it as the haplotype-based detector FreeBayes.
Results
The framework provides a foundation for directly detecting longer, multi-base alleles from short-read sequence alignments.
Takeaways & Limitations
Haplotype-based detection can establish local phase while evaluating allele classes in a unified context, supporting direct short-range haplotype inference from sequencing data.
Takeaways & Limitations
The model is limited to unphased genotypes because the primary data supports phasing only within a limited context.
Abstract
from arXiv · showhide
The direct detection of haplotypes from short-read DNA sequencing data requires changes to existing small-variant detection methods. Here, we develop a Bayesian statistical framework which is capable of modeling multiallelic loci in sets of individuals with non-uniform copy number. We then describe our implementation of this framework in a haplotype-based variant detector, FreeBayes.
1 Motivation
Haplotype-based detection uses short-range phasing from sequencing traces to evaluate nearby alleles together, but existing methods must handle multiallelism and non-uniform copy number. The authors generalize a Bayesian model accordingly and implement it in FreeBayes.
- Motivation: Sequencing traces provide short-range phasing that can establish phase between proximal alleles, although current read lengths and error rates limit this approach to variants clustered within tens of bases.The passage notes that larger-scale haplotype structure still requires statistical phasing approaches.
- Motivation: Haplotype-based methods simultaneously evaluate all allele classes in one context, ensuring semantic consistency among described variants.Locally phased genotype data can also reduce the haplotype space considered during genotype imputation.
- Challenges: As haplotype length increases, the number of possible alleles increases, creating challenges for methods that detect genetic variation in a unified haplotype context.This expanding allele space motivates methods capable of modeling multiallelism.
- Challenges: Most variant detectors assume biallelism and uniform, typically diploid, copy number, limiting accurate small-variant detection in sex chromosomes, polyploid organisms, and other copy-number-variable regions.The passage identifies improper copy-number modeling as a specific source of detection error.
- Contribution: The authors generalize a Bayesian method to model multiallelic loci and non-uniform copy number across samples, implementing the model in FreeBayes.The generalization is intended to enable population-level inference methods for haplotype detection.
2 Generalizing variant detection to multiallelic loci and non-uniform copy number
The framework models multiallelic loci across samples with non-uniform copy number by combining genotype likelihoods from sequencing reads with population-informed genotype priors. It accounts for sequencing errors, unphased genotypes, and allele-frequency distributions while using optimizations for large datasets.
- Model representation: Each sample may contribute a different locus copy number, yielding M total copies, K distinct alleles, and unphased genotypes represented as allele multisets.Allele frequencies describe both the population-level allele set and each individual genotype.
- Bayesian inference: The joint posterior combines sequencing-data likelihoods for individuals with a prior over genotype combinations and allele-frequency distributions.This population-level Bayesian formulation incorporates information shared across individuals rather than evaluating each sample independently.
- Read likelihood: Read likelihoods model observations as draws from the underlying genotype and scale them by per-base and allele-quality probabilities that account for sequencing errors.Observed alleles can therefore differ from true alleles according to quality-derived error probabilities.
- Population prior: Genotype priors decompose into genotype-combination probabilities conditional on allele frequencies and the probability of observing those frequencies in the sampled population.Ewens’ sampling formula supplies the allele-frequency component under an infinite-alleles coalescent model with mutation and genetic drift.
- Unphased genotypes and scalability: Because the data support limited phasing, the model adjusts phased-genotype probabilities to represent equivalent unphased genotypes and uses optimizations for tractable large-scale inference.For fully diploid populations, the unphased adjustment reduces to 2^h, where h is the number of heterozygous genotypes.
3 Direct detection of phase from short-read sequencing
This section describes FreeBayes, a haplotype-based variant detector implementing a Bayesian framework for direct detection of longer, multi-base alleles. It assembles quality-scored haplotype observations in dynamically determined windows and jointly estimates genotypes and polymorphism probabilities.
- 3 Direct detection of phase from short-read sequencing: FreeBayes implements a Bayesian haplotype-based method for directly detecting longer, multi-base alleles from sequence alignments.The framework models multiallelic loci and provides the foundation for this implementation.
- 3.1 Haplotype observation and dynamic windowing: Dynamic reference-relative windows combine multiple proximal segregating variations into single haplotype allele observations.Windows are bounded to avoid splitting clusters of alleles into separate variant calls.
- 3.1 Haplotype observation and dynamic windowing: Haplotype observation quality is the minimum of supporting-read mapping quality and component-allele base qualities.Only reads anchored by reference-matching sequence at both window boundaries contribute observations.
- 3.3 Joint genotyping: The method estimates joint sample genotypes by convergent gradient ascent over the posterior probability distribution.The search begins with maximum-likelihood genotypes and iteratively considers nearby alternatives until convergence within a bounded number of steps.
- 3.3 Joint genotyping: Basic genotype imputation improves raw genotype quality relative to maximum-likelihood-only methods, while posterior marginalization provides marginal genotype likelihoods.FreeBayes also reports the probability that a locus is polymorphic as a quality estimate.