**DOI:**10.1128/JCM.00659-10

## ABSTRACT

Pulsed-field gel electrophoresis (PFGE) and multiple-locus variable-number tandem-repeat analysis (MLVA) are used to assess genetic similarity between bacterial strains. There are cases, however, when neither of these methods quantifies genetic variation at a level of resolution that is well suited for studying the molecular epidemiology of bacterial pathogens. To improve estimates based on these methods, we propose a fusion algorithm that combines the information obtained from both PFGE and MLVA assays to assess epidemiological relationships. This involves generating distance matrices for PFGE data (Dice coefficients) and MLVA data (single-step stepwise-mutation model) and modifying the relative distances using the two different data types. We applied the algorithm to a set of *Salmonella enterica* serovar Typhimurium isolates collected from a wide range of sampling dates, locations, and host species. All three classification methods (PFGE only, MLVA only, and fusion) produced a similar pattern of clustering relative to groupings of common phage types, with the fusion results being slightly better. We then examined a group of serovar Newport isolates collected over a limited geographic and temporal scale and showed that the fusion of PFGE and MLVA data produced the best discrimination of isolates relative to a collection site (farm). Our analysis shows that the fusion of PFGE and MLVA data provides an improved ability to discriminate epidemiologically related isolates but provides only minor improvement in the discrimination of less related isolates.

Salmonellosis is one of the most common food-borne diseases in the United States (5). Consequently, it is important to understand how *Salmonella* strains disseminate within and between reservoirs and environments. Many molecular typing tools have been used for this purpose (11). Of these methods, pulsed-field gel electrophoresis (PFGE) is considered by many to be the gold standard for strain typing, and variable-number tandem-repeat (VNTR) assays are powerful alternative or complementary typing tools (3, 22). Both methods offer a high degree of genetic resolution for strain typing, depending on several factors.

PFGE involves separating chromosomal DNA macro-restriction fragments by size, and strains are discriminated based on the resulting band pattern observed after electrophoresis has been completed. It is one of the most reproducible and highly discriminatory typing techniques and has been widely and successfully used for a variety of *Salmonella enterica* serovars (12, 15); for many situations, PFGE is capable of discriminating between closely related strains. In addition, the use of the assay to analyze different serovars does not require a great deal of modification, as might be required with procedures that are dependent on PCR. Difficulties arise when strains are very closely related (i.e., poor discrimination [18, 27]) or when bands comigrate in the gel or identically sized bands represent completely different fragments of chromosomal DNA and thereby produce spurious matches (6). These complications are more pronounced when a large number of bands are generated by the restriction digest (4). In addition, while band patterns convey a crude degree of genetic relatedness, a large number of independent restriction digests would be needed to infer an accurate phylogenetic relationship (6).

Multiple-locus variable-number tandem-repeat analysis (MLVA) is a PCR-based technique that relies on the amplification of chromosomal or plasmid DNA that encompasses short tandem repeats of a DNA sequence. The tandem repeats are prone to higher-than-background mutation rates due to DNA strand slippage during replication (23), and thus, the amplified fragments will vary in length depending on the number of repeats harbored at a given locus. Different fragment lengths are tallied either as the total length (base pairs) or the estimated number of repeat units, and each discretely sized fragment is considered a unique “allele” for the locus under investigation. Because of the relatively high mutation rate, strains can accumulate distinctive allele patterns within a relatively short period of time (5). Furthermore, the technique can be multiplexed and automated and is conducive to rapid and relatively high-throughput strain typing. MLVA assays are relatively robust (5, 15-17) and, while not perfect, they can provide phylogenetic information even with a limited number of loci (13, 18). While access to a sequenced genome dramatically speeds the ability to establish new assays (3), it is not a requisite to assay development. The primary limitations of the technique include the potential need for a new set of loci for every species or serovar under investigation and the fact that some loci are very “unstable” and can “disappear” from some strains or lineages; this produces the equivalent of an uninformative “null” allele. Mutation rates can also vary between loci (5, 24, 25); if ignored, this factor can introduce bias into comparative analyses.

Clearly, PFGE and MLVA offer different technical and interpretive advantages and disadvantages, but it is important to emphasize that the nature of the methodological and interpretive errors is independent between the techniques. For example, errors due to comigration of bands for PFGE are independent of band size estimation errors for MLVA because differences in MLVA band size are not detectable using PFGE and macro-restriction fragments are generally independent of tandem-repeat sequences. Provided that most of the experimental variation from these two methods is uncorrelated (i.e., independent), it is possible to combine results from the two methods to produce improved estimates (8), and this premise underlies the current study.

Our objective was to determine whether combining the information obtained from both PFGE and MLVA assays produces more rigorous and discriminatory analyses of bacterial pathogens, such as *Salmonella*. Two sets of *Salmonella enterica* isolates were used in this study; one set included serovar Typhimurium isolates from a wide range of sampling dates, locations, and host species while the other set included a group of serovar Newport isolates collected over a limited geographic and temporal scale for a single host species. The results of the different typing methods were assessed by comparison with those of phage-typing assays (serovar Typhimurium) and with known epidemiological relationships (serovar Newport). To interpret the MLVA data, we employed a metric that incorporates a stepwise-mutation model, and to interpret the PFGE data, we employed Dice coefficients to construct distance matrices. Our analysis shows that the fusion of the two typing methods provides an improved ability to discriminate between isolates when PFGE and MLVA separately provide partial but incomplete discrimination between strains with a high degree of probable genetic similarity.

## MATERIALS AND METHODS

*Salmonella* strains.Two sets of isolates were used for this study. Set A included 37 *S. enterica* serovar Typhimurium strains that were previously collected from different animal hosts and from different locations and time periods (see the figures for strain designations and descriptors). Because these isolates were epidemiologically unrelated, we assumed that they encompassed a high degree of genetic variability. Set B included 63 *S. enterica* serovar Newport isolates, mostly collected from cattle in Washington State, and this set was assumed to represent less genetic diversity because of the restricted geographic representation, inclusion of multiple isolates from individual farms, and a limited temporal scale. *Salmonella* serovar Typhimurium strains were phage typed at the National Microbiology Laboratory, Canadian Science Center for Human and Animal Health, Winnipeg, Manitoba. Serovar Newport isolates were tested for antibiotic resistance using a disc diffusion method (2) according to Clinical and Laboratory Standards Institute (NCCLS) guidelines (19, 20). Northwestern isolates from cattle were tested for susceptibility to a panel of antimicrobial drugs that included ampicillin (10 μg), chloramphenicol (30 μg), gentamicin (10 μg), kanamycin (30 μg), streptomycin (10 μg), tetracycline (30 μg), triple-sulfa (a combination of sulfadiazine, sulfamethazine, and sulfamerazine) (250 μg), trimethoprim-sulfamethoxazole (1.25 μg to 23.75 μg), ceftazidime (30 μg), amoxicillin-clavulanic acid (20/10 μg), and nalidixic acid (30 μg) (BD Diagnostics, Sparks, Maryland). Northeastern isolates were tested with the same panel except that a sulfisoxazole disc (250 μg) was substituted for the triple-sulfa disc.

Pulsed-field gel electrophoresis.We followed a standard PFGE protocol for *Salmonella enterica*, using an XbaI restriction digest (21). Briefly, genomic DNA was digested in agarose plugs with the restriction enzyme XbaI, and the resulting DNA fragments were gel separated using a CHEF-DR II (Bio-Rad, Hercules, CA) apparatus. The electrophoresis conditions included an initial pulse time of 2.2 s, final pulse time of 63.8 s, running temperature of 14°C, and run time of 18 to 20 h at 6 V/cm. PFGE gels were stained with ethidium bromide and visualized using a UV transilluminator. Gel images were analyzed using Bionumerics version 4.6 (Applied Maths, Sint-Martens-Latem, Belgium). Estimated band sizes were exported from Bionumerics for the current study.

Multiple-locus variable-number tandem-repeat analysis.For *S. enterica* serovar Typhimurium isolates, four of five previously described VNTR loci were employed for MLVA (STTR5, STTR6, STTR9, and STTR10pl), following a published protocol (18) with the addition of a VNTR locus identified in-house (STTR11) (STTR11-Forward, GATAAGCCGTACTGTTCAGG, and STTR11-Reverse, TACTCCTTTGTGGTCTACGC). For the *S. enterica* serovar Newport strains, two Typhimurium loci (STTR5 and STTR6) (18) and four published Newport-specific loci were employed (NewportA, NewportB, NewportM, and NewportL) (6). The PCRs for both serovars were conducted as previously described (6, 20). Capillary electrophoresis was carried out at the Washington State University Genomics Core using a 3730 DNA analyzer with Pop-7 polymer (Applied Biosystems). The resulting electropherograms were analyzed using GeneMarker software (Softgenetics LLC, State College, PA).

Data analysis.Dice similarity coefficients were calculated using Bionumerics (Applied Maths, Austin, TX) from PFGE data to generate the distance matrix, and the unweighted-pair group method with arithmetic mean (UPGMA) algorithm was used to construct a dendrogram. For the MLVA data, the total length of the tandem repeats was divided by the estimated size of each repeat to obtain the number of tandem repeats for each locus and each strain. There were five available loci with tandem repeats for *S. enterica* serovar Typhimurium, but data from one locus were not used because it was from a plasmid locus and <73% of 37 bacterial isolates were positive for this locus. For *S. enterica* serovar Newport, there were six loci, all of which were used. Because passage experiments indicate that VNTR mutations are usually composed of a single step (5, 24), we modeled our data using a single-step stepwise-mutation model (SMM). Based on this statistical model, we estimated the distance *S* (the total number of single steps) between two lineages (two different isolates) and their most recent common ancestor (MRCA) using the number of tandem repeats, *X _{L}
*, of the two lineages. The distances,

*S*, for all lineage pairs were then used to construct the distance matrix, and UPGMA was used to obtain a dendrogram.

If μ is the rate of stepwise mutations per generation and if we assume that the gain or loss of a repeat is equally probable, then the following conditional probabilities, *P*, characterize the single-step SMM: *P*(*X _{t}
*

_{+1}=

*i*+ 1 |

*X*=

_{t}*i*) =

*P*(

*X*

_{t}_{+1}=

*i*− 1 |

*X*=

_{t}*i*) = μ/2,

*P*(

*X*

_{t}_{+1}=

*i*|

*X*=

_{t}*i*) = 1 − μ, and

*P*(‖

*X*

_{t}_{+1}−

*X*‖ ≥ 2 |

_{t}*X*=

_{t}*i*) = 0, where

*i*denotes the number of tandem repeats at distance

*t*and

*t*is an integer number between zero and infinity. Based on these conditional probabilities, the probability of the distance

*t*is given by (26)

*P*(

*t*|

*n*) =

_{0}, …, n_{k}*N*(

*t*)

*/D*(

*t*), where

*N*(

*t*) =

*e*

^{− (}

^{λ}

^{+ 2μ}

^{n}

^{)}

^{t}Π

_{j}^{k}_{=0}[

*I*(2μ

_{j}*t*)]

*and*

_{j}^{n}*D*(

*t*) = ∫

_{0}

^{∞}

*e*

^{− (}

^{λ}

^{+ 2μ}

^{n}

^{)}

^{t}Π

_{j}

^{k}

_{= 0}[

*I*(2μ

_{j}*t*)]

_{j}

^{n}

*dt*. The equation for

*P*assumes that the mutation rate μ is constant for all loci, which is approximately true for our

*S. enterica*serovar Typhimurium data;

*n*denotes the locus number where the subscript

_{m}*m*= 0, 1, 2,…,

*k*is the difference between the number of tandem repeats for two lineages, and

*m = k*is the maximum number of differences;

*n*is the number of loci used; λ is a parameter associated with the distance to a MRCA (in this work, λ = 0.0002 was found to give satisfactory results [26]); and

*I*is the

_{j}*j*th-order modified Bessel function of type 1. The distance

*S*is the value of

*t*with the maximum probability.

Fusion algorithm.Dendrograms constructed from PFGE or MVLA data can differ substantially. Consequently, if we assume that both types of data contain useful information as well as error, it is possible that better results can be obtained by combining the data. In fact, it is known that (i) if two different algorithms used to describe a set of samples give different results, (ii) if the error for each result is less than the error associated with randomly generated results, and (iii) if the errors for both are uncorrelated, then a combination of these algorithms will give better results than either of the two algorithms alone (8). For our problem, we have different data (PFGE and MLVA) from the same sets of samples. We can safely assume that the error for both the MLVA and PFGE algorithms is less than the error for random clustering, because both PFGE and MLVA can recapitulate epidemiological relationships (6, 14). Furthermore, we can assume that the errors are uncorrelated because the PFGE and MLVA assays measure differences that arise from different genetic mechanisms. Consequently, because the two methods provide different results, it is probable that combining the PFGE and MLVA data will provide a more comprehensive picture of the underlying population structure of these strains.

Several strategies can be used to combine different types of data. One strategy is to treat each data type independently and produce two independent dendrograms that are then combined to form a single dendrogram. While conceptually simple, this approach weights all sources of information equally, which may not be an optimal approach. Another strategy is to combine the data sets together before generating a dendrogram, but it may be difficult to combine the data if they differ in type (e.g., discrete and continuous), and even if this is accomplished, a suitable approach for evaluating the combined data may not exist. An alternative approach is to process each type of data using an algorithm that is appropriate for that data type and then fuse the results at some midpoint in the process; this is the approach we employed.

We begin with two distance matrices, one for the PFGE data set and one for the MLVA data set, as described above. One distance matrix is used to construct a dendrogram, and a threshold is selected (see below) to define distinct clusters from this dendrogram. If two strains in the second distance matrix are grouped together in one of the clusters formed from the first distance matrix, the distance between them is reduced (see below); if these two strains are not grouped together, then the distance in the second matrix is left unchanged. After all pairwise distance values are adjusted based on the clusters from the first matrix, a final dendrogram is generated. Both sets of data can be used in alternating roles: PFGE data are used to create the clusters while MLVA data are used to create a second distance matrix to be modified, and MLVA data are used to create the clusters while PFGE data are used to create a second distance matrix to be modified.

For the fusion algorithm described above, values for two parameters must be chosen. The first is the threshold value, *thr*, which divides the initial dendrogram into distinct clusters. The second is the degree that each distance value should be reduced in the modified distance matrix. If *D* is the distance matrix to be modified with elements *d*(*i*,*j*) and *D** is the modified distance matrix, the elements of *D** are given in terms of *d*(*i*,*j*) by the equation *d**(*i*,*j*) = *d*(*i*,*j*) if bacterial samples *i* and *j* are not in the same cluster and by the equation *d**(*i*,*j*) = *r*·*d*(*i*,*j*) if bacterial samples *i* and *j* are in the same cluster, with 0 < *r* ≤ 1. Thus, the weight parameter *r* dictates how much the distance value will be reduced. The choice of these two parameters, *thr* and *r*, changes the results, and there is no obvious way of knowing what values are the best to use. Our approach was to use the entire range of values for the threshold *thr*, i.e., 0.05 to 1, and several ranges of values for the weight parameter *r*. For the former, this corresponds to a range from having each strain form its own cluster (*thr* = 0.05) to the opposite extreme where all strains are grouped into a single cluster (*thr* = 1). For each set of ranges, we created a dendrogram using UPGMA with the modified distance matrix; from the resulting set of dendrograms, we constructed a “generalized tree” using Consense from the software package PHYLIP (version 3.68) with the default parameters (10).

The set of dendrograms used to construct the generalized tree is a combination of two sets of data, one created when PFGE clusters are used to modify the distance matrix and the other when MLVA clusters are used. When the sets are combined, a conflict will occur if they disagree completely on the relationship between two strains. For example, the PFGE data may indicate that strains A and B always occur as a pair, while the MLVA data may indicate that strains A and C always occur as a pair. If this happens, Consense (10) will construct a tree that depends on the order of the input. To prevent such an occurrence, we “break the tie” by multiplying the values of one set of data by 0.501 and the other set by 0.499.

One other issue arises when the two different data sets are combined; their distance measures are not the same. The manner in which we resolved this was to scale the dendrogram produced by Consense using an average-distance matrix obtained by averaging the values of the normalized PFGE and MLVA distance matrices. Normalization was achieved by dividing all matrix values by the maximum distance to obtain values between 0 and 1. Then, the final distances were determined as follows. Find all leaves on a branch (e.g., branch 1 contains A and B; branch 2 contains C, D, and E; and F contains itself). Find all combinations of leaves on one branch and leaves on the other branch and obtain the distance for each combination from the average distance matrix. Use the maximum value as the distance between the two branches.

The fusion program can be downloaded at http://www.vetmed.wsu.edu/research_vmp/MicroArrayLab/ .

## RESULTS AND DISCUSSION

Comparison of genetically diverse strains of *S. enterica* serovar Typhimurium.To compare genetically diverse strains of *S. enterica* serovar Typhimurium, generalized trees were constructed using the fusion, MLVA, and PFGE algorithms described above. For the fusion algorithm, we present results from weight parameter *r* between 0 and 1. Potential ties were broken by multiplying PFGE cluster data by a factor of 0.499 and MLVA cluster data by a factor of 0.501. This weights the analysis in favor of MLVA under the assumption that there is more phylogenetically relevant information available from MLVA data than from PFGE (7).

Assessing the validity of our analysis is complicated by the lack of a gold standard with which to compare our results. Indeed, barring a complete genome sequence for each strain and suitable algorithms for assessing genetic relationships, the only potential gold standards available are multilocus sequence typing (MLST) and phenotypic characteristics. Given the probable lack of genetic variation for intraserovar MLST comparisons (9), we chose to compare the *S. enterica* serovar Typhimurium strains using susceptibility to a panel of lytic phages as a measure of relatedness. The Centre for Infections of the Health Protection Agency (Colindale, London, United Kingdom) provided 38 phages (1), and 5 additional phages were developed at the National Microbiology Laboratory (Winnepeg, Canada). Our analysis assumes that strains with similar phage susceptibilities are more closely related than strains with dissimilar phage susceptibilities. All of the strains were subjected to susceptibility testing using a panel of 31 phages. Strains that were judged untypeable with this panel were subjected to testing with an additional 12 phages. Only one strain (8745) was considered untypeable using the combined panel of 43 lytic phages (see Tables S1a and b in the supplemental material).

Each dendrogram was divided into six clusters, A to F, and the strains within a cluster were compared according to their phage susceptibilities with the expectation that, on average, strains with greater genetic similarity will have fewer phage susceptibility mismatches. Because lytic phage susceptibility is unlikely to have a 1:1 correspondence with genetic similarity, we arbitrarily selected a cutoff where isolates were considered “more similar” if they had ≤7 differences in the lytic phage panel or “less similar” if they had >7 differences in lytic phage susceptibility. For this putatively diverse set of isolates, the correspondence between lytic phage results, fusion (Fig. 1), MLVA (Fig. 2), and PFGE (Fig. 3) was small but measurable (Table 1). The fusion results included one or three fewer phage susceptibility mismatches relative to the MLVA and PFGE results, respectively. We also examined the ability of the three approaches to distinctly classify unique phage types within the same clusters (Fig. 1 to 3). With a discrimination index calculated as one minus the average proportion of unique phage types per cluster, the fusion algorithm had an improved but not statistically significantly different discrimination index (0.34 ± 0.13 [average ± standard error of the mean]) compared with those of PFGE (0.25 ± 0.12) and MLVA (0.20 ± 0.10). Thus, for this relatively diverse set of isolates, we did not find a compelling benefit to merging PFGE and MLVA data using the fusion algorithm, although the fusion algorithm does give better statistical performance than random allocation of phage types to clusters (0.10, standard deviation = 0.03, *n* = 1,000).

Comparison of genetically similar *S. enterica* serovar Newport strains.To compare genetically similar *S. enterica* serovar Newport strains, generalized dendrograms were constructed using the fusion algorithm, PFGE-only data, and MLVA-only data with the same parameters discussed in the previous section (Fig. 4, 5, and 6). As with the serovar Typhimurium analysis, we lacked a gold standard for assessing the performance of our results relative to the true genetic relationships, but we did have epidemiologically relevant information as inferred from the location (farm) where the isolates were collected. We assumed that isolates collected from the same farm were more likely to be genetically similar relative to isolates collected at different farms. Consequently, we examined all cases in the dendrogram where serovar Newport isolates were indistinguishable and calculated a discrimination index as one minus the average proportion of distinct farms within these “monophyletic” groups. The number of monophyletic groups ranged from 6 (fusion and PFGE) (Fig. 4 and 6) to 9 (MLVA) (Fig. 5), with a significantly better discrimination index for the fusion algorithm (0.76 ± 0.05) than for MLVA (0.31 ± 0.12) and PFGE (0.39 ± 0.09) (*P* = 0.004; analysis of variance). Based on this assessment procedure, it is clear that combining the data from PFGE and MLVA analyses provided a greater degree of isolate discrimination as a function of the farm where the isolates were collected. While the antimicrobial resistance profiles cannot be used quantitatively to validate the fusion results, the clustering of the susceptible strains by the fusion algorithm (Fig. 4) relative to the clustering by the MLVA (Fig. 5) and PFGE (Fig. 6) algorithms provides further evidence that the fusion algorithm provides more biologically consistent classifications.

Clearly, combining data from PFGE and MLVA can serve to buffer the extremes in the level of genetic variation measured by these two methods alone, but the degree of benefit may be a function of the genetic similarity among the isolates under consideration. In cases in which isolates represent a relatively diverse and epidemiologically unrelated group of bacteria, combining PFGE and MLVA data may provide a minor benefit. In contrast, for closely related isolates, as judged by their collection from identical farm locations in this study, the combination of PFGE and MLVA data are likely to provide a higher degree of discrimination at the farm level. The advantage of increased discrimination for closely related isolates should extend to other cases of epidemiologically related strains, including situations where there is a need to trace food-borne disease outbreaks and infections in clinical settings.

## ACKNOWLEDGMENTS

K. N. K. Baker provided invaluable technical assistance. Martin Wiedmann, Cornell University, kindly provided northeastern *S*. Typhimurium isolates.

This project has been funded in part with federal funds from the National Institute of Allergy and Infectious Diseases, National Institutes of Health, Department of Health and Human Services, under contract no. NO1-AI-30055, and by the Agricultural Animal Health Program, College of Veterinary Medicine, Washington State University, Pullman, WA. Scholarship support for D.M. was provided by the Carl M. Hansen Foundation.

## FOOTNOTES

- Received 29 March 2010.
- Returned for modification 18 June 2010.
- Accepted 17 August 2010.

- Copyright © 2010 American Society for Microbiology