Haplotype-aware sequence alignment to pangenome graphs

Abstract

Modern pangenome graphs are built using haplotype-resolved genome assemblies. During read mapping to a pangenome graph, prioritizing alignments that are consistent with the known haplotypes has been shown to improve genotyping accuracy. However, the existing rigorous formulations for sequence-to-graph co-linear chaining and alignment problems do not consider the haplotype paths in a pangenome graph. This often leads to spurious read alignments to those paths that are unlikely recombinations of the known haplotypes. In this paper, we develop novel formulations and algorithms for haplotype-aware sequence alignment to an acyclic pangenome graph. We consider both sequence-to-graph chaining and sequence-to-graph alignment problems. Drawing inspiration from the commonly used models for genotype imputation, we assume that a query sequence is an imperfect mosaic of the reference haplotypes. Accordingly, we extend previous chaining and alignment formulations by introducing a recombination penalty for a haplotype switch. First, we solve haplotype-aware sequence-to-graph alignment in O(|Q| | E| |[H]|) time, where Q is the query sequence, E is the set of edges, and [H] is the set of haplotypes represented in the graph. To complement our solution, we prove that an algorithm significantly faster than O(|Q| | E| |[H]|) is impossible under the Strong Exponential Time Hypothesis (SETH). Second, we propose a haplotype-aware chaining algorithm that runs in O(|[H]| N log |[H]|N) time after graph preprocessing, where N is the count of input anchors. We then establish that a chaining algorithm significantly faster than O(|[H]|N) is impossible under SETH. As a proof-of-concept of our algorithmic solutions, we implemented the chaining algorithm in the Minichain aligner (https://github.com/at-cg/minichain). We demonstrate the advantage of the algorithm by aligning sequences sampled from human major histocompatibility complex (MHC) to a pangenome graph of 60 MHC haplotypes. The proposed algorithm offers better consistency with ground-truth recombinations when compared to a haplotype-agnostic algorithm.