Resolving Conflicting Predictions from Multimapping Reads

Abstract

The first step in the analysis of data produced by ultra-high-throughput next-generation sequencing technology is to map short sequence "reads" to a reference genome, if available. Sequencing errors, repeat regions, and polymorphisms may lead a read to align to multiple locations in the genome reasonably well. While ignoring such multimapping reads, or some of their alignments, will reduce the sensitivity of almost any type of downstream analysis (e.g., detecting structural variants), erroneous mappings will typically yield false positive predictions. Here we propose a framework that aims to identify true predictions among a large set of candidate predictions by selecting for each read a unique mapping that collectively imply conflict-free predictions. We formulate this problem as the maximum facility location problem, for which we propose LP-rounding heuristics. We provide a theoretic guarantee on the quality of the solution and demonstrate the utility of our algorithm in resolving conflicting deletions implied by simulated reads mapping ambiguously to Craig Venter's genome model and Illumina sequencing reads of the well-studied NA12878 individual.

Identification and Validation of Reference Genes in Clostridium beijerinckii NRRL B-598 for RT-qPCR Using RNA-Seq Data

“…Therefore, the number of reads in the sample remained the same. Multi-mapping reads present a problem for downstream analyses as they reduce sensitivity (Canzar et al, 2016). Although several strategies and specialized tools were proposed to count multi-mapping reads (Roberts et al, 2013;Zytnicki, 2017), for our purposes, specificity was the principal issue.…”

Cautionary Tales of Inapproximability

“…Importantly, the hardness of MFLP (and thus sequence alignment in this form) can be demonstrated by reduction from maximum coverage (Canzar et al, 2016). This reduction also preserves approximation bounds.…”

“…A practical limitation of theoretical modeling is highlighted by our analysis of the sequence alignment problem (Canzar et al, 2016). As the hardness result for this problem is at least 0:63, one might think all hope is lost.…”

“…We have recently shown that sequence alignment can be modeled as an instance of the maximum facility location problem (MFLP) (Canzar et al, 2016). In brief, given a bipartite graph of putative mappings from the donor genome (set of clients, C) to the reference genome (set of facilities, F), the weight w u‚ v of an edge from a client, u 2 C, to facility, v 2 F, is equivalent to our confidence in that deletion.…”