Haplotype-resolved diploid genome inference on pangenome graphs

Recent algorithmic advancements have shown how to utilize pangenome graphs in combination with the haplotype reconstruction framework of Li and Stephens to accurately reconstruct a haplotype from a reference pangenome graph and a set of input reads. However, significant work remains in developing techniques that utilize a pangenome graph to obtain a pair of phased haplotypes called a diploid pair. ResultsWe introduce new problem formulations and scalable algorithms for inferring phased diploid genomes from a pangenome graph and a set of input reads. We implement them in our tool DipGenie. The key idea is to jointly optimize genotyping and phasing along global paths through the pangenome graph, guided by a biologically motivated recombination budget that constrains inferred haplotypes to plausible mosaics of reference haplotypes. We evaluate DipGenie on real Illumina short-read data from the highly polymorphic MHC region in 22 leave-one-out diploid experiments, benchmarking against three tools that also operate on graph structures: VG, which samples haplotypes directly from the pangenome graph, and PanGenie + Beagle and Paragraph + Beagle, which derive local graphs from a VCF panel for per-site genotyping and delegate phasing to a statistical method. At full coverage, DipGenie achieves a geometric mean switch error rate (SER) of 0.86%, which is 5.7x lower than PanGenie + Beagle (4.88%), 7.9x lower than VG (6.77%), and 13.2x lower than Paragraph + Beagle (11.35%). For structural variant calling, DipGenie leads with a geometric mean F1-score of 0.571, compared to 0.470 (PanGenie + Beagle), 0.450 (VG), and 0.379 (Paragraph + Beagle). These advantages hold at every coverage level tested. Availability and Implementationhttps://github.com/gsc74/DipGenie.