Prefix-free parsing for building large tunnelled Wheeler graphs

TL;DR

This paper introduces a novel method combining Wheeler graphs and prefix-free parsing to efficiently build compact, space-saving indexes for large, repetitive genomic datasets, enabling practical pangenomic references.

Contribution

It presents a new approach that uses prefix-free parsing to accelerate and reduce memory usage in constructing Wheeler graph-based indexes for large genomic collections.

Findings

Faster construction of Wheeler graphs with less memory.

Effective compression of large repetitive texts.

Enabling practical pangenomic indexing.

Abstract

We propose a new technique for creating a space-efficient index for large repetitive text collections, such as pangenomic databases containing sequences of many individuals from the same species. We combine two recent techniques from this area: Wheeler graphs (Gagie et al., 2017) and prefix-free parsing (PFP, Boucher et al., 2019). Wheeler graphs (WGs) are a general framework encompassing several indexes based on the Burrows-Wheeler transform (BWT), such as the FM-index. Wheeler graphs admit a succinct representation which can be further compacted by employing the idea of tunnelling, which exploits redundancies in the form of parallel, equally-labelled paths called blocks that can be merged into a single path. The problem of finding the optimal set of blocks for tunnelling, i.e. the one that minimizes the size of the resulting WG, is known to be NP-complete and remains the most computationally challenging part of the tunnelling process. To find an adequate set of blocks in less time, we propose a new method based on the prefix-free parsing (PFP). The idea of PFP is to divide the input text into phrases of roughly equal sizes that overlap by a fixed number of characters. The original text is represented by a sequence of phrase ranks (the parse) and a list of all used phrases (the dictionary). In repetitive texts, the PFP of the text is generally much shorter than the original. To speed up the block selection for tunnelling, we apply the PFP to obtain the parse and the dictionary of the text, tunnel the WG of the parse using existing heuristics and subsequently use this tunnelled parse to construct a compact WG of the original text. Compared with constructing a WG from the original text without PFP, our method is much faster and uses less memory on collections of pangenomic sequences. Therefore, our method enables the use of WGs as a pangenomic reference for real-world datasets.