[Paper Review] YFitter: Maximum likelihood assignment of Y chromosome haplogroups from low-coverage sequence data

Luke Jostins, Yali Xu+5 more|arXiv (Cornell University)|Jul 30, 2014

Genetic Associations and Epidemiology10 references16 citations

TL;DR

YFitter is a maximum likelihood method that uses dynamic programming to assign Y chromosome haplogroups from low-coverage sequencing or genotype data, accurately resolving haplogroups while quantifying uncertainty via confidence sets. It achieves high accuracy and resolution, outperforming manual assignments and existing tools in both genotype and sequencing data applications.

ABSTRACT

Low-coverage short-read resequencing experiments have the potential to expand our understanding of Y chromosome haplogroups. However, the uncertainty associated with these experiments mean that haplogroups must be assigned probabilistically to avoid false inferences. We propose an efficient dynamic programming algorithm that can assign haplogroups by maximum likelihood, and represent the uncertainty in assignment. We apply this to both genotype and low-coverage sequencing data, and show that it can assign haplogroups accurately and with high resolution. The method is implemented as the program YFitter, which can be downloaded from http://sourceforge.net/projects/yfitter/

Motivation & Objective

• To address the challenge of probabilistic, automated Y chromosome haplogroup assignment from low-coverage sequencing data, where uncertainty and low coverage can lead to false inferences.
• To develop an efficient algorithm that computes likelihoods across the Y chromosome haplogroup tree, enabling maximum-likelihood assignment with resolution of uncertainty.
• To provide a scalable, automated solution for large-scale population genetics studies, replacing time-consuming manual assignments.
• To integrate uncertainty quantification into haplogroup assignment using the Akaike Information Criterion (AIC), producing confidence haplogroup sets that reflect statistical support.

Proposed method

• The method uses a phylogenetic tree of Y chromosome haplogroups, with mutations defined as branch-specific variants, and computes likelihoods using per-site read data.
• It employs downward likelihoods $ L^{ ightarrow}_{i} $, calculated from leaf nodes upward, representing the probability of observing downstream reads given absence of mutations at node $ i $.
• Upward likelihoods $ L^{ ightarrow}_{i} $ are computed from the root downward, representing the probability of observing non-descendant reads given the presence of mutations at node $ i $.
• The full likelihood $ L_i $ for each node is defined as the maximum of its descendants’ likelihoods, ensuring the most recent common ancestor with maximum likelihood is selected.
• The maximum-likelihood haplogroup is the most recent ancestor among all nodes with the highest likelihood, and the confidence set includes all haplogroups within 8.685 phred-scaled log units of the maximum likelihood.
• The method is implemented in C++ as YFitter, which accepts data in phyloXML format and integrates with samtools or PLINK for input processing.

Experimental results

Research questions

• RQ1 Can a dynamic programming algorithm efficiently compute maximum-likelihood haplogroup assignments from low-coverage sequencing data while accounting for uncertainty?
• RQ2 How does the accuracy and resolution of YFitter compare to manual or existing automated methods on genotype and sequencing data?
• RQ3 To what extent does the confidence haplogroup set, based on AIC, reflect robustness and reduce false inferences in low-coverage scenarios?
• RQ4 Can YFitter resolve fine-scale haplogroup structure from low-coverage data with performance comparable to high-coverage or genotype-based assignments?

Key findings

• In genotype data from the Genomes Unzipped project, YFitter achieved consistent haplogroup assignments across 9 individuals, with only minor discrepancies due to nomenclature differences between ISOGG2010 and YCC2008 trees.
• For low-coverage 1000 Genomes Project data, 285 out of 286 maximum-likelihood haplogroup assignments were fully consistent with HapMap-based manual assignments.
• Of the 285 consistent assignments, 203 had higher resolution from sequencing data, 71 had equal resolution, and only 11 had lower resolution, demonstrating improved resolution from sequencing.
• When using confidence haplogroups, no inconsistencies were observed, with 199 assignments having higher resolution and 75 equal resolution compared to genotype-based assignments.
• The confidence haplogroup was only slightly less resolved than the maximum-likelihood assignment, indicating robust uncertainty quantification with minimal information loss.
• YFitter successfully resolved fine-scale haplogroup structure from low-coverage data, outperforming genotype-based methods in resolution while maintaining high accuracy.