A Benchmark of Modern Statistical Phasing Methods

Abstract

Modern statistical phasing methods efficiently and accurately infer haplotype phase in large genomic samples, and their performance is critical for downstream analyses. However, rigorously evaluating phasing accuracy remains challenging. Here we demonstrate the use of synthetic diploids generated from male X chromosome sequences to benchmark and compare three commonly-used phasing methods: Beagle 5.4, SHAPEIT 4, and Eagle v2.4.1. Our evaluation reveals highly correlated error rates across all methods, although we observe distinct variation among methods within error types. Specifically, Eagle v2.4.1 is characterized by a higher frequency of switch errors, whereas SHAPEIT 4 displays a higher rate of flip (double-switch) errors. These patterns are consistent across diverse populations. Additionally, we observe an enrichment of errors at both CpG sites and rare variant sites, particularly for flip errors. Validating these results in autosomal trio data, we observe similar trends between the methods and error patterns, though with an increased flip error rate, a discrepancy that may be biased by genotype error.