Methods To Increase Fidelity of Repetitive Extragenic Palindromic PCR Fingerprint-Based Bacterial Source Tracking Efforts

The goal of the study was to determine which similarity coefficientand statistical method to use to produce the highest rate ofcorrect assignment (RCA) in repetitive extragenic palindromicPCR-based bacterial source tracking. In addition, the use ofstandards for deciding whether to accept or reject source assignmentswas investigated. The use of curve-based coefficients CosineCoefficient and Pearson's Product Moment Correlation yieldedhigher RCAs than the use of band-based coefficients Jaccard,Dice, Jeffrey's x, and Ochiai. When enterococcal and Escherichiacoli isolates from known sources were used in a blind test,the use of maximum similarity produced consistently higher RCAsthan the use of average similarity. We also found that the useof a similarity value threshold and/or a quality factor threshold(the ratio of the average fingerprint similarity within a sourceto the average similarity of this source's isolates to an unknown)to decide whether to accept source assignments of unknowns increasesthe reliability of source assignments. Applying a similarityvalue threshold improved the overall RCA (ORCA) by 15 to 27%when enterococcal fingerprints were used and 8 to 29% when E.coli fingerprints were used. Applying the quality factor thresholdresulted in a 22 to 32% improvement in the ORCA, depending onthe fingerprinting technique used. This increase in reliabilitywas, however, achieved at the expense of decreased numbers ofisolates that were assigned a source.

INTRODUCTION

Top

Introduction

The goal of bacterial source tracking is to identify sourcesof bacterial contamination in, for example, water or food products.One of the main uses of bacterial source tracking is to identifysources of fecal pollution in natural waters. Because the bacteriumused as a tracer is part of the normal intestinal microbiota,it can be used as an indicator of the source of fecal pollution.With a DNA fingerprint library-based approach, the first stepis to establish a collection of fingerprints of isolates fromlikely sources of fecal pollution. The fingerprints are groupedinto library units according to the source of the isolate. TheDNA fingerprints from water-isolated bacteria are then comparedwith the libraries of fingerprints to determine the most-likelyorigin of the isolate. Each isolate is assigned to an origindepending on how similar its fingerprint is to those in eachlibrary unit (2, 4, 5, 10, 12, 13).

There are several coefficients that can be used to calculatesimilarity, and the choice of which similarity coefficient touse can affect the outcome of the source tracking assignment.Cosine Coefficient and Pearson's Product Moment Correlationare curve-based coefficients that use both the presence or absenceof DNA bands and the peak intensity of each band as variables,whereas Jaccard, Dice, Jeffrey's x, and Ochiai are band-basedcoefficients that consider only the presence or absence of DNAbands. Currently, there is no consensus on which coefficientresults in more accurate source assignment. Pearson's ProductMoment Correlation has been used to calculate similarities amongrepetitive extragenic palindromic (rep) PCR fingerprints (4,12) and ribotypes (3). Cosine Coefficient has been used to calculatesimilarities among rep-PCR and pulsed-field gel electrophoresisfingerprints (10). The band-based coefficients Jaccard (5) andDice (13) have been used to calculate similarities among rep-PCRfingerprints, and the latter has also been used to calculatesimilarities among ribotypes (7, 11) and among fingerprintsgenerated by the amplification of the 16S-23S intergenic spacerregion (13).

In addition to uncertainties concerning which similarity coefficientto use, there are no standards concerning acceptance of sourceassignments. When similarity coefficients are used, each environmentalisolate is assigned a source depending on which library includesthe DNA fingerprints to which its DNA fingerprint is most similar.Because each isolate is always assigned a source, without regardto the actual degree of similarity, the user must decide, usinga priori assessments of the accuracy of the method, whethereach source assignment is likely to be correct.

Currently, there are no published studies on the issue of reliabilitywhen source assignments are determined using similarity coefficients.Information is needed to better understand limitations of statisticalanalyses used for assigning sources of bacteria on the basisof DNA fingerprint patterns (1). In an attempt to evaluate thesignificance of source tracking on the basis of discriminantanalysis, Whitlock et al. (18) suggested using the percentageof misassigned isolates in each source group calculated by Jackknifeanalysis as a lower limit for significance. When such an approachis used, each isolate from a collection of isolates from a knownsource (a library) is treated, one at a time, as an unknownand then identified by comparison to the remaining isolates.The proportion of isolates assigned to the correct source relativeto the number of isolates in the library is the rate of correctassignment. Whitlock et al. (18) proposed that a source canbe implicated in water pollution only when the percentage ofenvironmental isolates assigned to it exceeds the percentageof library isolates misassigned to it by Jackknife analysis.Although this approach can be useful when libraries with goodrepresentation of the diversity of isolates in the environmentalsite being studied are used, it is less useful when librarieswith poor representation are used or in the presence of isolatescontributed by nonlibrary sources. Therefore, in computer-assisted,library-based bacterial source tracking efforts, the choiceof which similarity coefficient to use for DNA fingerprint comparisonsand what similarity threshold to use in deciding whether eachsource assignment is to be accepted are important issues thatneed further investigations to improve the reliability of sourceassignments.

In the present study, six different similarity coefficientswere compared in terms of their rates of correct assignment(RCAs). In addition, statistical options to improve the reliabilityof source assignments on the basis of the use of similaritycoefficients were investigated. These options include the choiceof how the similarity values are used and the effect of theuse of a threshold similarity value and quality factor on improvementof the reliability of source assignments.

MATERIALS AND METHODS

Top

Materials and Methods

Sample collection.
Fecal samples were collected from humans, cows, deer, dogs,chicken, and gulls (Table 1). A total of 784 fecal samples wereused in the study (Table 1). Some samples did not yield Escherichiacoli, while others did not yield Enterococcus spp. Thus, thenumbers of fecal samples and isolates were not the same forthe two indicator organisms.

View this table:
[in this window]
[in a new window]
TABLE 1. Sources and numbers of fecal samples and bacterial isolates (before and after exclusion of clonal isolates) used in the study

Eighty-four human fecal samples were used in the present study.A total of 46 rectal swab samples were collected from humanvolunteers at the University of Southern Mississippi; of these,42 were collected using a BBL CultureSwab collection and transportdevice with Cary-Blair transport medium (BD Diagnostics Systems,Sparks, Md.), and 4 were collected with sterile swabs and suspendedin fetal bovine serum containing 10% dimethyl sulfoxide (DMSO).Thirty-eight samples were collected at a local hospital in Laurel,Mississippi. Fifty-nine samples yielded E. coli, and 65 yieldedEnterococcus spp. (Table 1). Samples collected using CultureSwabswere refrigerated and processed within 24 h after collection.Samples collected in fetal bovine serum were frozen immediatelyafter collection. Hospital samples were collected in sterilecups shipped to our laboratory within 24 h and used within another24 h.

Cow, chicken, and deer samples were collected from south andcentral Mississippi. All of the fecal samples were collectedfrom individuals except for some of the cow and chicken samples.Four of the cow samples were composite samples from severalcows at the same farm. A total of 86 of the chicken sampleswere litter samples from commercial chicken farms, while 27were obtained from cloacal swabs from individual chickens. Althoughsome cow and chicken cloacal samples were collected and transportedusing CultureSwabs, the majority of cow and chicken litter samples,as well as all deer samples, were collected and frozen withoutadditives at their respective collection sites across the State.Dog fecal samples from veterinary adoption centers and humanesocieties in Hattiesburg and Gulfport, Mississippi, and gullfecal samples from beaches along the Mississippi Gulf coastwere collected using CultureSwabs.

Bacteria isolation.
Fecal samples were streaked on mTEC (Difco) and mEI plates forthe isolation of E. coli and enterococci, respectively (15).mTEC plates were incubated at 37°C for 2 to 4 h and thenat 44.5°C for 18 to 24 h. Yellow colonies were picked andconfirmed using standard microbiological methods. Isolates thatlacked phenylalanine deaminase, that produced indole from tryptophan,that were unable to utilize sodium citrate as a sole carbonsource, and that fermented glucose through a mixed-acid fermentationpathway (but not a butanediol pathway) were considered to beE. coli. mEI plates were incubated at 41°C for 24 to 36h. Colonies that formed blue halos were presumed to be enterococci.Confirmation was performed by testing each isolate for growthat 45°C and in the presence of 6.5% sodium chloride at 37°Cand for esculin hydrolysis. Among the isolates picked, 89.5and 73.1% were confirmed to be E. coli and enterococci, respectively.

rep-PCR and BOX-PCR.
rep-PCR (8, 17) was performed using a modified method of Rademakerand DeBruijn (12). Isolates were grown at 37°C in brainheart infusion for 12 to 16 h. Cells harvested from 0.5 and1.0 ml of broth for E. coli and enterococci, respectively, werewashed twice with 0.5 ml of sterile deionized water. The resultingpellets were resuspended in deionized sterile water (0.5 and0.25 ml for E. coli and enterococci, respectively) and storedfrozen at –20°C until use as a template for PCR. DNAamplification reactions were performed with a 10-µl reactionmixture that consisted of 1 µl of cell suspension and9 µl of PCR master mix. The BOX-PCR (9, 16) master mixcontained 2 µM primer (BOX A1R [CTA CGG CAA GGC GAC GCTGAC G]), 1 mM deoxynucleoside triphosphates, 4.5 mM MgCl₂, 1xbuffer provided by the manufacturer of the DNA polymerase, and0.4 units of JumpStart Taq DNA polymerase (Sigma, St. Louis,Mo.). Thermal cycling started with 2 min at 95°C followedby 35 cycles of 94°C for 3 s, 92°C for 30 s, 50°Cfor 1 min, and 65°C for 8 min. A final extension step wasperformed at 65°C for 8 min after completion of the 35 cycles.The REP-PCR (6, 9, 14) master mix contained 3 µM of eachof two primers (REP 1R [III ICG ICG ICA TCI GGC] and REP 2I[ICG ICT TAT CIG GCC TAC]), 1 mM deoxynucleoside triphosphates,2.5 mM MgCl₂, 1x buffer, and 0.4 units of JumpStart Taq DNApolymerase. The thermal cycling protocol for REP-PCR was thesame as that for BOX-PCR except that 40°C was used insteadof 50°C for primer annealing.

Jackknife analysis.
The effect of having clonal isolates in fingerprint librarieson RCAs was examined by performing Jackknife analysis both beforeand after their removal by use of BOX and REP fingerprints ofenterococcal and E. coli isolates. Clonal isolates were definedin the present study as isolates with identical fingerprintsobtained from the same sample. Removal of clonal isolates wasperformed for Jackknife analysis only.

Jackknife analysis was also used to compare the RCAs generatedusing six similarity coefficients (Cosine Coefficient, Pearson'sProduct Moment Correlation, Jaccard, Dice, Jeffrey's x, andOchiai). The fingerprints used were produced by BOX-PCR usingenterococcal isolates with clonal isolates removed. All Jackknifeanalyses were performed using BioNumerics version 3.0 (AppliedMaths, Sint-Martens-Latem, Belgium). Pattern optimization (i.e.,the percentage of pattern shift within which the software looksfor the best match) was set at 5%, and band tolerance (i.e.,the maximum gel migration difference for any pair of bands tobe considered matching) was set at 2% with a 2% gradual toleranceincrease towards the bottom of the gel.

Discriminant analysis and multivariate analysis of variance.
Enterococcal isolates from human, cow, deer, dog, chicken, andgull fecal samples were classified into groups according totheir sources. Discriminant analysis was used to show the separationbetween these predefined groups on the basis of their BOX fingerprints.The first and second discriminants were plotted on the x andy axes, respectively, generating a two-dimensional plot showingthe separation of isolates from six sources. Multivariate analysisof variance was performed accounting for the covariance structureto evaluate the significance of discriminant analysis. The Pvalue indicated the probability of obtaining equivalent separationresults among isolates of different sources due to random classificationof isolates. The probability of obtaining the same level ofdiscrimination, assuming that all isolates were obtained froma homogeneous population (i.e., the effect of grouping by sourcewas insignificant), is indicated by the Wilkinson's likelihoodfor normal distribution (L). Low P and L values indicate significantdiscrimination by source group.

Identification libraries and the blind test.
The RCAs using each of the six similarity coefficients listedabove were also determined using a blind test. First, librariescontaining rep-PCR fingerprints of enterococcal and E. coliisolates from each known animal source were constructed. Eachlibrary consisted of five units, one for each known source:human, cow, deer, dog, and chicken. These libraries were thenused as reference to determine the most likely animal sourceof new isolates in the blind test. The enterococcal fingerprintlibrary contained 762 isolates (67 human, 141 cow, 99 deer,103 dog, and 352 chicken). All were analyzed by BOX-PCR, butonly 458 (40 human, 118 cow, 84 deer, 71 dog, and 145 chicken)were analyzed by REP-PCR. The E. coli library contained 514isolates (65 human, 136 cow, 39 deer, 142 dog, and 132 chicken),and all were analyzed by both BOX- and REP-PCR.

Isolates used as blind samples were obtained from feces of animalsin the same general population as those used to obtain isolatesfor rep-PCR fingerprint library construction. A total of 131enterococcal isolates (28 human, 29 bovine, 27 deer, and 47chicken) and 130 E. coli isolates (19 human, 43 bovine, 17 deer,and 51 chicken) were analyzed using BOX-PCR for the blind test.A total of 96 enterococcal isolates (12 human, 28 bovine, 23deer, and 33 chicken) and the same 130 E. coli isolates wereanalyzed using REP-PCR for the blind test. Source assignmentswere made using Cosine Coefficient to calculate similarity matrices,and average RCAs were compared using both maximum and averagesimilarity options.

Setting "similarity value" and "quality factor" thresholds.
A similarity threshold was determined for each indicator organism-fingerprintingtechnique combination. Cosine Coefficient was used to calculatesimilarity matrices. The threshold was determined by dividingthe sum of the average similarity values of the correctly andthe incorrectly assigned isolates by 2. In other words, thethreshold was the midpoint between the average similaritiesof the correctly and incorrectly assigned isolates. When thismethod was used, the similarity threshold values were 90, 90.8,and 89.2% for enterococcal BOX, REP, and BOX-REP combined fingerprints,respectively. The combined fingerprints were generated electronicallyusing BOX and REP fingerprints. The threshold values for E.coli BOX, REP, and combined fingerprints were 91.5, 90.1, and87.7%, respectively.

A quality factor was also used as a threshold for determiningthe reliability of source assignments. A quality factor is generatedby BioNumerics for each unknown as the unknown is assigned toan animal source. This value is calculated by dividing the averagepairwise similarity of all fingerprints in the source groupby the average pairwise similarity of the unknown with eachof the library's component isolates. Assignments with a qualityfactor of 1.0 or less (B or better) were accepted, while thosewith a quality factor of more than 1.0 (C, D, or E) were consideredunidentifiable.

RESULTS AND DISCUSSION

Top

Results and Discussion

The effect of clonal isolates on Jackknife RCAs.
The presence of clonal isolates artificially inflates RCAs calculatedby Jackknife analysis. The overall RCA (ORCA) among 1,584 enterococcalisolates, including clonal isolates, was 92%. However, the ORCAof the remaining 1,020 isolates after removing clonal isolateswas 82% (Table 2).

View this table:
[in this window]
[in a new window]
TABLE 2. Comparison of the RCAs of enterococcal isolates obtained before and after excluding clonal isolates

The decrease in RCA as a result of the removal of clonal isolatesdiffered among animal sources, ranging from a 7% decrease amongdeer and chicken isolates to 25% for gull isolates. The decreasein RCA did not correlate with the number of clonal isolatesremoved from each source (Fig. 1). For example, the human anddog groups contained 46 and 18% clonal isolates, respectively,but removing clonal isolates from either group resulted in a10% decrease in the RCA. On the other hand, the cow and gullgroups contained 33% clonal isolates each, and removing clonalisolates resulted in a 9 and 25% decrease in the RCA, respectively.The observation that the RCA of gull isolates was the most susceptibleto artificial inflation due to the presence of clonal isolatesmay have resulted from their lowest RCA among isolates of othersources (see below). This may be due to the scavenger feedingstrategy of gulls. Gulls were frequently observed feeding andbathing at a sewage treatment lagoon near the study site. Ourhypothesis is that by feeding on food or drinking water alreadycontaminated by enterococci from other animal sources, gullisolates can be more easily confused with those from other animalsources, hence, their low RCA. Therefore, providing perfectmatches to some of the isolates, by including clonal isolates,would artificially inflate the RCA of gull isolates to the greatestdegree.

View larger version (34K):
[in this window]
[in a new window]
FIG. 1. The percentages of clonal isolates among enterococci isolated from different animal sources and the decreases in RCA that resulted from their removal.

The presence of clonal isolates inflates the RCA in Jackknifeanalysis because members of a clone produce almost identicalfingerprints. In Jackknife analysis, isolates are divided intogroups depending on their sources. Then, individual isolatesare sequentially removed, one at a time, from the collectionof isolates under study, treated as unknowns, and assigned tosources on the basis of fingerprint similarity. When clonalisolates are present, the isolate treated as the unknown isalways assigned to the library source where another member ofthe same clone is present. Because Jackknife analysis can beused to validate the reliability of the source tracking methodor used to determine isolate overlap among potential sources,it is important for users to be aware of this potential sourceof error. It is also important to know, however, that this potentialsource of error is concerned specifically with Jackknife analysisand not with library-based source tracking.

The effect of using different similarity coefficients on RCAs.
A comparison of Jackknife RCAs of 1,020 enterococcal isolateson the basis of their BOX fingerprints indicated that the highestRCAs were obtained using curve-based similarity coefficients.The ORCA was 82% with both Pearson's Product Moment CorrelationCoefficient and Cosine Coefficient. When band-based coefficientswere used, the ORCA was 78% for each of the four coefficients.Although the RCAs differed among animal sources with each similaritycoefficient, they were less variable using curve-based coefficients.The standard deviations for the ORCAs were 11.2 and 12.0% usingPearson's Product Moment Correlation Coefficient and CosineCoefficient, respectively, but ranged from 13.9 to 14.7% forthe four band-based coefficients (Table 3). These results suggestthat the use of curve-based coefficients is preferred over theuse of band-based coefficients for source tracking in our studyarea with BOX fingerprinting. Additional data, from other studyareas and with other DNA fingerprinting protocols, are neededto ascertain whether the superiority of curve-based coefficientsis a general rule.

View this table:
[in this window]
[in a new window]
TABLE 3. Comparison of the RCAs of enterococcal isolates source assigned using six similarity coefficients^a

The effect of using maximum versus average similarity on RCAs.
The ORCAs obtained with the blind test were consistently higherusing maximum similarity for source assignments than using averagesimilarity (Table 4). This was the case regardless of the targetorganism (E. coli or Enterococcus spp.) or the fingerprintingtechnique used (BOX-PCR or REP-PCR). Although the ORCAs obtainedusing maximum similarity were consistently higher, the individualRCAs among animal sources were frequently but not always higherwhen maximum similarity was used. Among the 16 pair-wise comparisons(4 animal sources x 2 fingerprinting methods x 2 bacterial indicators),average similarity yielded higher RCAs in five cases (humanand cow enterococci analyzed by BOX-PCR, human and deer enterococcianalyzed by REP-PCR, and chicken E. coli analyzed by BOX-PCR).A striking difference between using average and maximum similarityis the large gap in the RCAs for some animal sources. For example,among cow enterococci analyzed using REP-PCR, the RCA was only11% using average similarity but was 71% using maximum similarity.Among deer E. coli isolates, the RCA was 0% using average similaritybut 35% using maximum similarity.

View this table:
[in this window]
[in a new window]
TABLE 4. Comparison of the RCAs obtained by use of average versus maximum similarity as a basis for bacterial source assignments

When maximum similarity, an epidemiological approach, is used,an isolate is assigned to the source group containing the best-matchingisolate in the identification library. When average similarity,a population genetics approach, is used, an isolate is assignedto the source group with which it shares the highest averagesimilarity. Consequently, the use of maximum similarity wouldbe advantageous when the within-source group fingerprints aredivergent, while the use of average similarity would be advantageouswhen the fingerprints within source groups are closely related;our results imply that the former is the case for the sourceanimal populations in our study area. The implied genetic diversitywas reflected in the complex cluster analysis dendrogram ofthe isolates. Isolates from one source often clustered withisolates from a different source (data not shown). This overlapis apparent in the two-dimensional plot generated using discriminantanalysis (Fig. 2). The relative contributions of discriminantsto total discrimination in descending order from discriminant1 through 5 were 42, 21, 18, 11, and 7%. The P value associatedwith each of the discriminants was 0.001%, indicating that thelikelihood of obtaining the same level of discrimination byuse of random grouping of isolates is remote. The L values fordiscriminants 1 through 5 were 0.0904, 0.2166, 0.3656, 0.5849,and 0.8053, respectively. In other words, the probability thatisolates from all sources belong to a homogeneous populationcalculated using the first discriminant to distinguish groups(i.e., the probability that grouping by source by use of thefirst discriminant is insignificant) equals 9%. These resultsindicate that the grouping of enterococcal isolates by animalsource is significant. However, the loose clustering of isolatesin the discriminant analysis plot (Fig. 2) shows diversity amongfingerprints which supports the use of maximum similarity insource tracking.

View larger version (45K):
[in this window]
[in a new window]
FIG. 2. A two-dimensional plot using discriminant analysis showing the separation of isolates on the basis of BOX fingerprints. The plot was generated by plotting the first discriminant (contributing 42% of total discrimination) on the x axis and the second discriminant (contributing 21% of total discrimination) on the y axis. The P value was 0.001 for both discriminants, while the L values were 0.0904 and 0.2166 for the first and second discriminants, respectively. Human isolates are shown in cyan, cow isolates are shown in green, deer isolates are shown in blue, dog isolates are shown in yellow, chicken isolates are shown in red, and gull isolates are shown in pink.

In addition to resulting in higher RCAs in general, the useof maximum similarity also resulted in less-variable RCAs amonganimal sources compared to the results seen with RCAs obtainedusing average similarity. The standard deviations of RCAs amongenterococcal isolates as determined on the basis of their REPfingerprints were 13.2 and 31.5% by use of maximum and averagesimilarity, respectively (Table 4). Among E. coli isolates,they were 24.9 and 34.2%, respectively, as determined on thebasis of their BOX fingerprints. When REP fingerprints wereused, the standard deviation of RCAs among E. coli isolateswere 24.6 and 32.1% by use of maximum and average similarity,respectively (Table 4). The only case in which the standarddeviation of RCAs was higher using maximum similarity than usingaverage similarity was in the case of RCAs of enterococci determinedusing BOX-PCR fingerprints, where they were 9.8% using averagesimilarity and 11.3% using maximum similarity. These resultssuggest that the use of maximum similarity in future DNA fingerprintlibrary-based bacterial source tracking efforts is warranted.The use of maximum similarity results in RCAs that are not onlyon average higher but also more consistent among different animalsources, regardless of the target organism (E. coli or Enterococcusspp.) or the fingerprinting technique used (BOX-PCR or REP-PCR).

The effect of using a similarity value threshold on RCAs.
By classifying as unidentified those isolates that did not meetthe similarity threshold requirement, the RCAs among isolatesassigned a source improved significantly. The ORCA among enterococcifingerprinted by BOX-PCR increased from 76 to 87% (Table 5).In addition, the number of isolates assigned incorrectly toa source decreased from 24 to 9%. The drawback, however, isthat the proportion of isolates that were assigned to a sourcedecreased from 100 to 69%, and 16% of the isolates assignedcorrectly to a source when a threshold was not used are designatedas unidentifiable when a threshold was used. Similar resultswere obtained using enterococcal REP fingerprints, where theORCA increased from 63 to 80% after the threshold was applied.When combined BOX-REP fingerprints were used, the RCAs beforeand after application of the similarity threshold were 77 and89%, respectively. With E. coli, the ORCAs increased from 65to 70% for isolates fingerprinted using BOX-PCR, from 65 to82% for isolates fingerprinted using REP-PCR, and from 69 to89% for isolates fingerprinted using combined BOX- and REP-PCRfingerprints (Table 5). However, note that these increases inORCA are achieved at a cost. A total 13 to 29% of the isolatesformerly assigned to a correct source when a threshold was notused are classified as unidentifiable when a threshold is used.

View this table:
[in this window]
[in a new window]
TABLE 5. Effect of using a similarity threshold during source assignments on the ORCA, the percentages of isolates correctly assigned, and the percentages incorrectly assigned^a

The use of a similarity threshold in deciding whether to acceptsource assignments increased the RCA for each of the six similaritycoefficients in a blind test. Its use also had a significanteffect on the proportion of isolates that were identifiablein terms of animal origin and of the proportion identified correctly.Among the 131 enterococcal isolates analyzed in blind testingby BOX-PCR, the ORCAs were 76 to 77% for the six similaritycoefficients when similarity thresholds were not used (Table6). When source assignments were accepted only when similaritythresholds were met, the ORCAs increased to 84 to 85%. Althoughthe ORCAs were equivalent among the six similarity coefficients,the proportion of isolates that were correctly assigned to ananimal source was significantly higher by use of curve-basedcoefficients than by use of band-based coefficients (Fig. 3).When the curve-based coefficients Pearson's and Cosine Coefficientwere used, 57 and 60%, respectively, of the isolates were correctlyassigned. When the band-based coefficients Jaccard, Dice, Jeffery'sx, and Ochiai were used, only 28, 25, 26, and 26%, respectively,of the total isolates were correctly assigned (Table 6 and Fig.3).

View this table:
[in this window]
[in a new window]
TABLE 6. Effect of using a similarity threshold on the RCAs of enterococcal isolates, the proportion of isolates assigned a source, and the proportion of isolates assigned to the correct source^a

View larger version (33K):
[in this window]
[in a new window]
FIG. 3. The effect of curve-based and band-based similarity coefficients on the percentages of enterococcal isolates assigned to an animal source and the percentages of isolates assigned correctly and incorrectly. Source assignments were made on the basis of their BOX-PCR fingerprints (n = 131).

The effect of using a quality factor threshold on RCAs.
Applying the quality threshold criterion to source assignmentsalso resulted in significant improvements in the RCAs, regardlessof whether BOX, REP, or combined BOX-REP fingerprints were used.The ORCA increased from 74 to 98% with BOX-PCR fingerprints,from 63 to 77% with REP-PCR fingerprints, and from 77 to 97%with combined BOX-REP fingerprints (Table 7). This increasein source assignment accuracy was obtained at the expense ofefficiency. The percentage of isolates assigned a source decreasedfrom 100% when all source assignments were accepted to 43, 50,and 32% for BOX, REP, and combined BOX-REP fingerprints, respectively,when a quality threshold was used. However, mistakes in sourceidentification also decreased dramatically. When a quality thresholdwas not used for BOX, REP, and combined BOX-REP fingerprints,the percentages of isolates assigned to an incorrect sourcewere 26, 37, and 23%, respectively (Table 7). When a qualitythreshold was used, these values decreased to 1, 11, and 1%,respectively.

View this table:
[in this window]
[in a new window]
TABLE 7. Effect of applying a quality factor threshold on the ORCA, the percentages of enterococcal isolates correctly assigned, and the percentages incorrectly assigned^a

Although the usefulness of a similarity value and/or a qualityfactor threshold to improve the reliability of bacterial sourceassignments appears obvious, the reduction in the proportionof isolates that can be assigned to a source can be a concern.If the proportion of isolates classified as unidentified islarge, a significant source of fecal pollution may remain hidden.This would occur if the fingerprint libraries used were of insufficientsize. The proportion of unknown isolates that can be classifiedgiven a certain similarity value or quality factor thresholdmay be useful in future studies in determining whether librariesof sufficient size have been achieved for reliable bacterialsource tracking efforts.

In conclusion, results from the present study indicate that(i) the use of curve-based coefficients (e.g., Cosine Coefficientand Pearson's Product Moment Correlation) results in higherORCAs than the use of band-based coefficients (e.g., Jaccardand Dice); (ii) the removal of clonal isolates is essentialfor the proper calculation of RCAs by Jackknife analysis; (iii)the use of maximum, as opposed to average, similarity yieldshigher ORCAs; and (iv) the application of a similarity valueor a quality factor threshold for source assignment improvesthe ORCA, but this is achieved at the expense of the total numbersof isolates assigned a source.

ACKNOWLEDGMENTS

We thank Jim Lipe and Damon Nowlin of the Mississippi Departmentof Agriculture and Commerce for coordinating the collectionof cow fecal samples, Jim Watson, Betty Roberts, Jerry Jones,and Keith Rusell of the Mississippi Board of Animal Health forthe chicken fecal samples, Christopher Alonzo and Larry Castleof the Mississippi Department of Wildlife, Fisheries, and Parksfor the deer fecal samples, and Sabrina Bryant for some of thehuman fecal samples. We also thank Joel Brumfield, KimberlyPeterson, Mary Phares, and Brian Robinson for their assistancein the laboratory.

Financial support that made the research possible was providedby the Mississippi Departments of Agriculture and Commerce andEnvironmental Quality, the U.S. Environmental Protection AgencyGulf of Mexico Program (EPA MS97449202), and the U.S. CoastalImpact Assistance Program (NOAA 17OZ2171 Project MS.R.17).

FOOTNOTES

* Corresponding author. Mailing address: Department of Biological Sciences—USM, 118 College Dr. #5018, Hattiesburg, MS 39406-0001. Phone: (601) 266-6795. Fax: (601) 266-5797. E-mail: shiao.wang{at}usm.edu.

REFERENCES

Top