Impact of compounding error on strategies for subtyping pathogenic bacteria

Abstract

Comparative-omics will identify a multitude of markers that can be used for intraspecific discrimination between strains of bacteria. It seems intuitive that with this plethora of markers we can construct higher resolution subtyping assays using discrete markers to define strain "barcodes." Unfortunately, with each new marker added to an assay, overall assay robustness declines because errors are compounded exponentially. For example, the difference in accuracy of strain classification for an assay with 60 markers will change from 99.9% to 54.7% when average probe accuracy declines from 99.999% to 99.0%. To illustrate this effect empirically, we constructed a 19 probe bead-array for subtyping Listeria monocytogenes and showed that despite seemingly reliable individual probe accuracy (>97%), our best classification results at the strain level were <75%. A more robust strategy would use as few markers as possible to achieve strain discrimination. Consequently, we developed two variable number of tandem repeat (VNTR) assays (Vibrio parahaemolyticus and L. monocytogenes) and demonstrate that these assays along with a published assay (Salmonella enterica) produce robust results when products were machine scored. The discriminatory ability with four to seven VNTR loci was comparable to pulsed-field gel electrophoresis. Passage experiments showed some instability with ca. 5% of passaged lines showing evidence for new alleles within 30 days (V. parahaemolyticus and S. enterica). Changes were limited to a single locus and allele so conservative rules can be used to determine strain matching. Most importantly, VNTRs appear robust and portable and can clearly discriminate between strains with relatively few loci thereby limiting effects of compounding error.

FIG. 1.

Expected proportion of correctly classified strains where assays consist of 5, 10, or 20 probes and assuming probes are independent and accuracy estimates are equal between all probes. For example, when probe accuracy is 95% for each of 20 probes, the expected proportion of correctly classified strains will be 36%. When probe accuracy differs between probes (σ² > 0), then the actual correct classification rate will be less than shown in this figure.

FIG. 2.

(Left) Dice coefficient based UPGMA dendrogram for NotI PFGE profiles of 29 Vibrio parahaemolyticus isolates from various sources (Table S4).(Right) Categorical coefficient based UPGMA dendrogram using data from seven variable number of tandem repeat (VNTR) loci (Table S4).Both dendrograms were generated using Bionumerics software; X-axis represents arbitrary units of genetic distance.