Latent Variable Model For Aligning Barcoded Short-Reads Improves Downstream Analyses

Recent years have seen the emergence of several “third-generation” sequencing platforms, each of which aims to address shortcomings of standard next-generation short-read sequencing by producing data that capture long-range information, thereby allowing us to access regions of the genome that are inaccessible with short-reads alone. These technologies either produce physically longer reads typically with higher error rates or instead capture long-range information at low error rates by virtue of read “barcodes” as in 10x Genomics’ Chromium platform. As with virtually all sequencing data, sequence alignment for third-generation sequencing data is the foundation on which all downstream analyses are based. Here we introduce a latent variable model for improving barcoded read alignment, thereby enabling improved downstream genotyping and phasing. We demonstrate the feasibility of this approach through developing EMerAld— or EMA for short— and testing it on the barcoded short-reads produced by 10x’s sequencing technologies. EMA not only produces more accurate alignments, but unlike other methods also assigns interpretable probabilities to the alignments it generates. We show that genotypes called from EMA’s alignments contain over 30% fewer false positives than those called from Lariat’s (the current 10x alignment tool), with a fewer number of false negatives, on datasets of NA12878 and NA24385 as compared to NIST GIAB gold standard variant calls. Moreover, we demonstrate that EMA is able to effectively resolve alignments in regions containing nearby homologous elements— a particularly challenging problem in read mapping— through the introduction of a novel statistical binning optimization framework, which allows us to find variants in the pharmacogenomically important CYP2D region that go undetected when using Lariat or BWA. Lastly, we show that EMA’s alignments improve phasing performance compared to Lariat’s in both NA12878 and NA24385, producing fewer switch/mismatch errors and larger phase blocks on average. EMA software and datasets used are available at .