Title:BLEND: A Fast, Memory-Efficient, and Accurate Mechanism to Find Fuzzy Seed Matches

Abstract:Comparing genomic sequences is a fundamental yet costly step in most genomic analyses. It requires identifying similar genomic sequence pairs sharing a sufficient number of short subsequences, called seeds. The length and number of seed matches between sequences directly impact the sensitivity, performance, and memory footprint of today's read mappers. Existing attempts optimizing seed matches suffer from either 1) increasing the use of the costly sequence alignment step due to finding a large number of sequence pairs for alignment or 2) limited sensitivity when finding seed matches. Our goal is to efficiently and effectively identify potentially similar sequences using a small number of seed matches, while tolerating variations within each seed. To this end, we introduce BLEND, a fast, memory-efficient, and accurate mechanism to find approximate (i.e., fuzzy) seed matches. BLEND 1) finds the minimizer k-mers and extends them to increase the length of seeds and 2) utilizes the SimHash technique to enable generating the same hash values for highly similar seeds, which allows matching fuzzy seeds with a single-look up.
We show the benefits of BLEND when used in two important genomics applications: read overlapping and read mapping. For read overlapping, BLEND enables a 2.6x-63.5x (on average 19.5x) faster and 0.9x-9.7x (on average 3.6x) more memory-efficient implementation than the state-of-the-art tool, Minimap2. We observe that BLEND finds better quality overlaps that lead to more accurate de novo assemblies compared to Minimap2. For read mapping, BLEND provides 0.7x-3.7x (on average 1.7x) speedup compared to Minimap2. Source code is available at this https URL.