lra: the Long Read Aligner for Sequences and Contigs

It is computationally challenging to detect variation by aligning long reads from single-molecule sequencing (SMS) instruments, or megabase-scale contigs from SMS assemblies. One approach to efficiently align long sequences is sparse dynamic programming (SDP), where exact matches are found between the sequence and the genome, and optimal chains of matches are found representing a rough alignment. Sequence variation is more accurately modeled when alignments are scored with a gap penalty that is a convex function of the gap length.

Because previous implementations of SDP used a linear-cost gap function that does not accurately model variation, and implementations of alignment that have a convex gap penalty are either inefficient or use heuristics, we developed a method, lra, that uses SDP with a convex-cost gap penalty.

We use lra to align long-read sequences from PacBio and Oxford Nanopore (ONT) instruments as well as de novo assembly contigs. Across all data types, the runtime of lra is between 52-168% of the state of the art aligner minimap2 when generating SAM alignment, and 9-15% of an alternative method, ngmlr. This alignment approach may be used to provide additional evidence of SV calls in PacBio datasets, and an increase in sensitivity and specificity on ONT data with current SV detection algorithms. The number of calls discovered using pbsv with lra alignments are within 98.3-98.6% of calls made from minimap2 alignments on the same data, and give a nominal 0.2-0.4% increase in F1 score by Truvari analysis. On ONT data with SV called using Sniffles, the number of calls made from lra alignments is 3% greater than minimap2-based calls, and 30% greater than ngmlr based calls, with a 4.6-5.5% increase in Truvari F1 score. When applied to calling variation from de novo assembly contigs, there is a 5.8% increase in SV calls compared to minimap2+paftools, with a 4.3% increase in Truvari F1 score.