Effective sequence similarity detection with strobemers

Abstract

k-mer-based methods are widely used in bioinformatics for various types of sequence comparisons. However, a single mutation will mutate k consecutive k-mers and make most k-mer-based applications for sequence comparison sensitive to variable mutation rates. Many techniques have been studied to overcome this sensitivity, for example, spaced k-mers and k-mer permutation techniques, but these techniques do not handle indels well. For indels, pairs or groups of small k-mers are commonly used, but these methods first produce k-mer matches, and only in a second step, a pairing or grouping of k-mers is performed. Such techniques produce many redundant k-mer matches owing to the size of k. Here, we propose strobemers as an alternative to k-mers for sequence comparison. Intuitively, strobemers consist of two or more linked shorter k-mers, where the combination of linked k-mers is decided by a hash function. We use simulated data to show that strobemers provide more evenly distributed sequence matches and are less sensitive to different mutation rates than k-mers and spaced k-mers. Strobemers also produce higher match coverage across sequences. We further implement a proof-of-concept sequence-matching tool StrobeMap and use synthetic and biological Oxford Nanopore sequencing data to show the utility of using strobemers for sequence comparison in different contexts such as sequence clustering and alignment scenarios.

Flexible seed size enables ultra-fast and accurate read alignment

“…The main idea of the seeding approach presented here is to first compute open syncmers (21) from the reference sequences, then link the syncmers together using the randstrobe method (22) with two strobes. The study introducing strobemers (22) described strobemers as linking together strobes in ‘sequence-space’, i.e ., over the set of all k-mers. Since syncmers represent a subset of k-mers from the original sequence, computing randstrobes over this subset of strings is very fast; it suffice to compare a smaller set of syncmers to produce the next strobe, while still having a similar range on the upper and lower window bounds on the original sequence.…”

“…This means that s 1 , s 2 , and s ′ are syncmers, and we will let [ w min , w max ] refer to the lower and upper number of syncmers downstream from s 1 where we will sample s 2 from. A second modification to the strobemers as described in (22) is that we store the strobemer hash value from two strobes s 1 and s 2 as H ( s 1 , s 2 ) = v ( s 1 ) / 2 + v ( s 2 ) / 2. The hash function H is symmetric ( h ( v ( s 1 ), v ( s 2 )) = h ( v ( s 2 ), v ( s 1 ))) and together with canonical syncmers it produces the same hash value if the strobemer is created from forward and reverse complement direction.…”

“…The hash function H is symmetric ( h ( v ( s 1 ), v ( s 2 )) = h ( v ( s 2 ), v ( s 1 ))) and together with canonical syncmers it produces the same hash value if the strobemer is created from forward and reverse complement direction. It is stated in (22) that a symmetrical hash function is undesirable for mapping due to unnecessary hash collisions. However, when masking highly repetitive seeds as commonly performed in aligners (17), it turns out that a symmetrical hash function helps to avoid sub-optimal alignments, and we will now describe why.…”

“…Assume we would use an asymmetric hash function, such as v ( s 1 )/2 + v ( s 2 )/3 proposed in (22). Also assume that strobemer seeds ( s 1 , s 2 ) and ( s 2 , s 1 ) are both found in forward orientation the reference due to, e.g ., inversions.…”

“…Here we show that syncmers and strobemers can be used in combination in what becomes a high-speed indexing method, roughly corresponding to the speed of computing minimizers. Our technique is based on first subsampling k -mers from the reference sequences by computing canonical open syncmers (21), then, producing strobemers (22) formed from linking together syncmers occurring close-by on the reference using the randstrobe method. A consequence is that instead of using a single seed (e.g., k =21 as default in minimap2 for short-read mapping), we show that we can link together two syncmers as a strobemer seed and achieve similar accuracy to using individual minimizers.…”