Choice of reference-guided sequence assembler and SNP caller for analysis of Listeria monocytogenes short-read sequence data greatly influences rates of error

Arthur W. Pightling, Nicholas Petronella, Franco Pagotto

The influences that different programs and conditions have on error rates of single-nucleotide polymorphism (SNP) analyses are poorly understood. Using Illumina short-read sequence data generated from Listeria monocytogenes strain HPB5622, we assessed the performance of four SNP callers (BCFtools, FreeBayes, UnifiedGenotyper, VarScan) under a variety of conditions, including: (1) a range of sequencing coverages; (2) use of four popular reference-guided assemblers (Burrows-Wheeler Aligner, Novoalign, MOSAIK, SMALT); (3) with and without read quality trimming and filtering; and (4) use of different reference sequences. At 8-fold coverage the proportions of true positive calls ranged from 0.22 to 25.00 % when reads were aligned to a nearly identical reference (0.000096 % distant). Calls made when reads were aligned to a non-identical reference (0.85 % distant) were from 92.54 to 98.88 % accurate. At 79-fold coverage accuracies ranged from 3.95 to 20.00 % with the nearly identical reference and 93.80-98.75 % with the non-identical reference. Read preprocessing significantly changed the numbers of false positive calls made, from a 65.24 % decrease to a 54.55 % increase. The combinations of reference-guided sequence assemblers and SNP callers greatly influenced not only the numbers of true and false positive sites but also the proportions of true positive calls relative to the total numbers of calls made. Furthermore, the efficacy of different assembler and caller combinations changed dramatically with the different conditions tested. Researchers should consider whether identifying the greatest numbers of true positive sites, reducing the numbers of false positive calls, or achieving the highest accuracies are desired.