Nucleotide String Indexing using Range Matching

Abstract: The two most common data-structures for genome indexing, FM-indices and hash-tables, exhibit a fundamental trade-off between memory footprint and performance. We present Ranger, a new indexing technique for nucleotide sequences that is both memory efficient and fast. We observe that nucleotide sequences can be represented as integer ranges and leverage a range-matching algorithm based on neural networks to perform the lookup. We prototype Ranger in software and integrate it into the popular Minimap2 tool. Ranger achieves almost identical end-to-end performance as the original Minimap2, while occupying 1.7$\times$ and 1.2$\times$ less memory for short- and long-reads, respectively. With a limited memory capacity, Ranger achieves up to 4.3$\times$ speedup for short reads compared to FM-Index, and up to 4.2$\times$ and 1.8$\times$ speedups for short- and long-reads, compared to hash-tables. Ranger opens up new opportunities in the context of hardware acceleration by reducing the memory footprint of long-seed indexes used in state-of-the-art alignment accelerators by up to 23$\times$ which results with 3$\times$ faster alignment and negligible accuracy degradation. Moreover, its worst case memory bandwidth and latency can be bounded in advance without the need to inflate DRAM capacity.