Short-read genome alignment is a fundamental computational step used in many bioinformatic analyses. It is therefore desirable to align such data as fast as possible. Most alignment algorithms consider a seed-and-extend approach. Several popular programs perform the seeding step based on the Burrows-Wheeler Transform with a low memory footprint, but they are relatively slow compared to more recent approaches that use a minimizer-based seeding-and-chaining strategy. Recently, syncmers and strobemers were proposed for sequence comparison. Both protocols were designed for improved conservation of matches between sequences under mutations. Syncmers is a thinning protocol proposed as an alternative to minimizers, while strobemers is a linking protocol for gapped sequences and was proposed as an alternative to k-mers. The main contribution in this work is a new seeding approach that combines syncmers and strobemers. We use a strobemer protocol (randstrobes) to link together syncmers (i.e., in syncmer-space) instead of over the original sequence. Our protocol allows us to create longer seeds while preserving mapping accuracy. A longer seed length reduces the number of candidate regions which allows faster mapping and alignment. We also contribute the insight that speed-wise, this protocol is particularly effective when syncmers are canonical. Canonical syncmers can be created for specific parameter combinations and reduce the computational burden of computing the non-canonical randstrobes in reverse complement. We implement our idea in a proof-of-concept short-read aligner strobealign that aligns short reads 3-4x faster than minimap2 and 15-23x faster than BWA and Bowtie2. Many implementation versions of, e.g., BWA, achieve high speed on specific hardware. Our contribution is algorithmic and requires no hardware architecture or system-specific instructions. Strobealign is available at https://github.com/ksahlin/StrobeAlign.
Burrows-Wheeler transform acceleration based on CUDA
