Genomic variant identification methods alter Mycobacterium tuberculosis transmission inference

Abstract

20Pathogen genomic data are increasingly used to characterize global and local transmission patterns of important 21 human pathogens and to inform public health interventions. Yet there is no current consensus on how to measure 22 genomic variation. We investigated the effects of variant identification approaches on transmission inferences for 23 M. tuberculosis by comparing variants identified by five different groups in the same sequence data from a clonal 24 outbreak. We then measured the performance of commonly used variant calling approaches in recovering 25 variation in a simulated tuberculosis outbreak and tested the effect of applying increasingly stringent filters on 26 transmission inferences and phylogenies. We found that variant calling approaches used by different groups do 27 not recover consistent sets of variants, often leading to conflicting transmission inferences. Further, performance 28 in recovering true outbreak variation varied widely across approaches. Finally, stringent filters rapidly eroded the 29 accuracy of transmission inferences and quality of phylogenies reconstructed from outbreak variation. We 30 conclude that measurements of genetic distance and phylogenetic structure are dependent on variant calling 31 approach. Variant calling algorithms trained upon true sequence data outperform other approaches and enable 32 inclusion of repetitive regions typically excluded from genomic epidemiology studies, maximizing the 33 information gleaned from outbreak genomes. 34 35 42 settings, it is unknown where and between whom the majority of transmission occurs 2-4 and therefore where to 43 focus interventions. Patterns of M. tuberculosis genetic and genomic variation are frequently used to identify 44 potential recent transmission events. M. tuberculosis isolates that share a genotype (RFLP, spoligotype, or MIRU-45 VNTR) 5-7 , or whose whole genome sequences are within a given genetic distance [8][9][10][11] , are considered clustered and 46 potentially epidemiologically linked. Phylogenies inferred from outbreak variation may reveal patterns of 47 relatedness within and between clusters 11-13 . Finally, transmission trees integrate epidemiological and phylogenetic 48 information to capture probable transmission histories, chains of who infected whom 14,15 . Predicted transmission 49 links have been used to infer the likely location and/or timing 16,17 of transmission, to identify risk factors for 50 transmission and high risk populations 18 , to distinguish between acquired (primary) and transmitted drug 51 resistance 19 , and to declare an outbreak over 20 . 52 3 Transmission inferences in molecular epidemiology for M. tuberculosis and other pathogens rely on the 53 high-quality measurement of genetic variation from sequence data. However, there is no consensus on how to 54 measure pathogen genomic variation, and studies frequently employ different sequence quality control measures, 55 mapping algorithms, variant callers, and variant filters 21 . Further, the performance of variant calling...

Experiences from 4 Years of Organization of an External Quality Assessment for Mycobacterium tuberculosis Whole-Genome Sequencing in the European Union/European Economic Area

“…1 ). In addition, the inclusion of two synthetic FASTQ files generated from an edited reference sequence ( 22 ) and differing by three single SNPs, one double nucleotide change, and one 3-bp insertion (sample 12 and 13 in Table 3 ) were reported by participating laboratories as differing by a number of SNPs ranging from one to seven, which is in line with previous findings ( 21 ). These minor differences between pipelines, although having only minor effects when detecting potentially epidemiologically linked isolates, may have implications when inferring transmission chains as the natural accumulation of mutations in M. tuberculosis is extremely slow ( 9 ).…”

“…Nonetheless, some laboratories did report structurally high SNP distances for closely related isolates, even if the isolates were correctly reported as potentially related. This is likely due to incomplete filtering of noise in the SNP-calling algorithm, often caused by not excluding all poorly mapped reads, used by different laboratories, which makes the comparison of precise SNP distances provided by the EQA participants potentially misleading ( 21 ).…”

New insights into the mycobacterial PE and PPE proteins provide a framework for future research

“…First of all, while generally all pe and ppe genes are excluded from bioinformatic datasets, this approach is overly stringent in most cases. Even short‐read sequencing techniques can reliably map almost all pe and ppe genes thanks to paired‐end technologies and increased read lengths, if only pe_pgrs and ppe‐mptr genes/transcripts are excluded (Miran and Farhat – personal communication, Holt et al , ; Ates et al , ; Walter et al , ). Furthermore, knowing the subgroup of a PE/PPE can be an excellent starting point to hypothesize the most‐likely route of secretion and may even suggest functions, or redundancy.…”

Genomic Epidemiology Analysis of Infectious Disease Outbreaks Using TransPhylo

“…Exporting a simulated outbreak 6. This simulation protocol of TransPhylo is especially useful to simulate outbreaks and benchmark how accurate methods of inference (either Basic Protocol 3 of TransPhylo or some other method) are likely to be when applied to real datasets (Ness et al, 2019;Stimson et al, 2019;Walter et al, 2019). In this case, it is important to make sure that the parameters used for the simulation are realistic for the pathogen of interest in the real data.…”