Comparison of Seven Protocols To Identify Fecal ... - ACS Publications

Oct 19, 2004 - Virginia Polytechnic Institute and State University,. Blacksburg, Virginia 24061, College of William and Mary,. Virginia Institute of M...
0 downloads 11 Views 268KB Size
Download PDF
Environ. Sci. Technol. 2004, 38, 6109-6117

Comparison of Seven Protocols To Identify Fecal Contamination Sources Using Escherichia coli D O N A L D M . S T O E C K E L , * ,† MELVIN V. MATHES,‡ KENNETH E. HYER,§ CHARLES HAGEDORN,| HOWARD KATOR,⊥ JERZY LUKASIK,# TARA L. O’BRIEN,O TERRY W. FENGER,O MANSOUR SAMADPOUR,] KRISTON M. STRICKLER,[ AND BRUCE A. WIGGINS× U.S. Geological Survey, 6480 Doubletree Avenue, Columbus, Ohio 43229, U.S. Geological Survey, 11 Dunbar Street, Charleston, West Virginia 25301, U.S. Geological Survey, 1730 East Parham Road, Richmond, Virginia 23228, Virginia Polytechnic Institute and State University, Blacksburg, Virginia 24061, College of William and Mary, Virginia Institute of Marine Science, 1208 Greate Road, Gloucester Point, Virginia 23062, Biological Consulting Services of North Florida, 4641 NW 6th Street, Gainesville, Florida 32609, Marshall University School of Medicine, 1542 Spring Valley Drive, Huntington, West Virginia 25704, Institute for Environmental Health, 8279 Lake City Way NE, Seattle, Washington 98115, West Virginia Department of Agriculture, 60B Moorefield Industrial Park Road, Moorefield, West Virginia 26836, and James Madison University, Harrisonburg, Virginia 22807

Microbial source tracking (MST) uses various approaches to classify fecal-indicator microorganisms to source hosts. Reproducibility, accuracy, and robustness of seven phenotypic and genotypic MST protocols were evaluated by use of Escherichia coli from an eight-host library of knownsource isolates and a separate, blinded challenge library. In reproducibility tests, measuring each protocol’s ability to reclassify blinded replicates, only one (pulsedfield gel electrophoresis; PFGE) correctly classified all test replicates to host species; three protocols classified 4862% correctly, and the remaining three classified fewer than 25% correctly. In accuracy tests, measuring each protocol’s ability to correctly classify new isolates, ribotyping with EcoRI and PvuII approached 100% correct classification but only 6% of isolates were classified; four of the other six protocols (antibiotic resistance analysis, PFGE, and two repetitive-element PCR protocols) achieved better than random accuracy rates when 30-100% of challenge isolates were classified. In robustness tests, measuring each protocol’s ability to recognize isolates from nonlibrary * Corresponding author phone: (614)430-7780; fax: (614)430-7777; e-mail: [email protected]. † U.S. Geological Survey, Columbus. ‡ U.S. Geological Survey, Charleston. § U.S. Geological Survey, Richmond. | Virginia Polytechnic Institute and State University. ⊥ College of William and Mary. # Biological Consulting Services of North Florida. O Marshall University School of Medicine. ] Institute for Environmental Health. [ West Virginia Department of Agriculture. × James Madison University. 10.1021/es0354519 CCC: $27.50 Published on Web 10/19/2004

 2004 American Chemical Society

hosts, three protocols correctly classified 33-100% of isolates as “unknown origin,” whereas four protocols classified all isolates to a source category. A relevance test, summarizing interpretations for a hypothetical water sample containing 30 challenge isolates, indicated that false-positive classifications would hinder interpretations for most protocols. Study results indicate that more representation in known-source libraries and better classification accuracy would be needed before field application. Thorough reliability assessment of classification results is crucial before and during application of MST protocols.

Introduction Microbial source tracking (MST) is an active area of research with the potential to provide important information to effectively manage water resources. A comprehensive review of MST (1) and two more recent reviews have been produced (2, 3). MST is used to address a variety of water-quality management goals. One typical goal is to reduce loads or concentrations of fecal-indicator microorganisms such as Escherichia coli and fecal enterococci; a second, more direct, goal is to detect the presence of fecal contamination from specific animal sources. Fecal-indicator microorganisms can be associated with host sources by way of library-independent methods that take advantage of host-associated markers (4, 5) or by way of a library-dependent approach that takes advantage of differential distribution of indicator microorganisms among hosts (6, 7). MST is currently applied to fecal-source detection in experimental settings and, in total maximum daily load (TMDL) report generation, to partition loads among sources. All protocols compared in this paper have been applied using library-dependent approaches. Antibiotic resistance analysis (ARA) was used in Page Brook, VA, to evaluate the ability of best management practices (BMPs) to reduce fecal inputs from cattle to the stream (8) and in Spout Run, VA, to quantify the effects of a rural non-sewered community on water quality (9). Ribotyping has been used to assess sources of fecal contamination to three streams in Virginia (10). Application of pulsed-field gel electrophoresis (PFGE) to MST was described previously (11), and the ability by PFGE to discriminate E. coli isolated from various sources has been demonstrated (12). Repetitive DNA-element PCR (rep-PCR) for assessment of fecal sources to environmental waters has not yet been reported in the peer-reviewed literature; however, both the utility of rep-PCR (13, 14) and some limitations with respect to this goal (15) have been described. Development of these and other protocols for MST is active across the United States and in Europe. Efforts by individual researchers to develop effective protocols have been supplemented by government-sponsored evaluations in the European Union (16) and in the United States (this and other efforts, variously cited in ref 17). Results of MST studies are applied to water-management decisions that sometimes allocate substantial tax dollars toward waterquality improvement or protection plans. For example, in a small watershed in Clarke County, VA, MST by ARA indicated that the majority of enterococcus isolates originated from cattle; no human signature was found (8). State and local officials used the results to recommend implementation of BMPs for cattle, and in-stream fecal coliform populations were lowered substantially. The alternate management plan VOL. 38, NO. 22, 2004 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

6109

was to upgrade septic systems or to place homes in the watershed on an expanded community wastewater treatment system. The decision to initiate BMPs instead of upgrading the waste-treatment infrastructure was based largely on MST results. In this paper, we report a comparison of seven E. coli-based library-dependent protocols for MST based on the reproducibility, accuracy, and robustness of source associations, and we describe the relevance of these results for hypothetical environmental samples. This is only known comparison study to directly evaluate the performance of multiple methods and researchers against the same library of isolates.

Experimental Section Study Area. Berkeley County, WV, was chosen as the study area because, in 1999, two bacterial indicators of fecal contamination (fecal coliforms and E. coli) were detected in 30% of sampled wells in eastern Berkeley County (18). Most of the contaminated wells are in subdivisions that use individual household septic systems for waste disposal; many of the wells supply untreated drinking water (18). The aquifer is karstic, composed of cavernous limestone (19) that can transmit groundwater rapidly from point of contamination to point of use. Determining the origin of contamination in this setting is difficult. Potential sources of fecal contamination to groundwater are numerous in Berkeley County. The human population of the county increased 28% between 1990 and 2000 to 79 202 (20). The largest population center is the sewered community of Martinsburg (population 14 972). Land use in the remainder of the county is forest, pastureland, and cropland without major centralized residential, agricultural, or animal feeding facilities. In addition to humans (Homo sapiens), the local National Resources Conservation Service (NRCS) office identified domestic animals (cattle (Bos taurus), chickens (Gallus gallus), dogs (Canis familiaris), horses (Equus caballus), and swine (Sus scrofa)) and wildlife (Canada geese (Branta canadensis) and white-tailed deer (Odocoileus virginianus)) as populations that might contribute fecal material to local groundwater (21, 22). In previous work in Berkeley County, no statistically significant relation was detected between septic-system density adjacent to wells and incidence of fecal-indicator detections in sampled wells (18). Sample Collection. The sample-collection design was to obtain a total of 100 E. coli isolates from 20 individual animals for each of 8 host species sampled. For cattle, 100 isolates were collected from beef cattle and 100 from dairy cattle to test the ability of each protocol to separate cattle by breed and husbandry practice. Samples for each host species were collected from at least four different locations and from no more than five individuals per location. Samples were collected in mid-August to early October 2001 (known-source library) and in June 2002 (challenge set). For the challenge set, only one isolate was used from each individual fecal sample to avoid including clones. The sample-collection design was generally followed, though there were minor variations in the number and distribution of samples collected. All isolates were collected from freshly dropped feces except for white-tailed deer, a wild population for which scat of indeterminate age was collected. Human fecal material was provided by local volunteers. Library size is an important consideration for the chosen protocols. Researchers were polled prior to distribution of samples to determine how many isolates per source were required for each protocol. Guidance in the contemporary literature is sparse with regard to library size, so a minimum known-source library of 630 isolates (70 per source) was chosen on the basis of logistical constraints, best judgment, and cost. Several researchers chose to evaluate a more extensive known-source library of 900 isolates (100 per 6110

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 38, NO. 22, 2004

source). The composition of the 630-isolate known-source library was from the 900-member library and was the same for every researcher. Cultivation and Confirmation. Details of cultivation and confirmation are provided in the Supporting Information. All feces samples were chilled and shipped by overnight express service to the U.S. Geological Survey (USGS) Ohio District Microbiology Laboratory in Columbus, OH (ODML). Cultivation on mTEC agar was always attempted within 24 h of sample collection (except for some delayed human samples; see Supporting Information for details), but multiple cultivation attempts were required to obtain E. coli isolates in some cases. Isolate identities were confirmed according to recommendations for recreational waters by the U.S. Environmental Protection Agency (23). The first five presumptive E. coli of eight cultivated from each feces sample that passed confirmation tests were included in the knownsource library. Protocols. Phenotypic Methods. Phenotypic methods are used to differentiate bacteria by observations of physiological variability (shape, metabolism, and growth conditions). ARA. Antibiotic resistance analysis on the full 900-isolate library was done at James Madison University (Harrisonburg, VA) using a previously reported method (24). The working hypothesis behind ARA is that various animal populations are exposed to and have developed resistance to different arrays of antibiotics; therefore, antibiotic resistance patterns may be used to differentiate fecal bacteria from the various sources. Unknowns were classified to source by discriminant analysis, as described previously, based on concentration at which growth was inhibited (25). See Supporting Information for more protocol details. CUP. Carbon utilization profiling was done on 630 members of the 900-isolate library at Virginia Polytechnic Institute and State University (Blacksburg, VA) by use of Biolog GN-2 microplates (Hayward, CA) in a protocol similar to that previously described (26). The working hypothesis behind CUP is that various animal populations have different diets; therefore, fecal bacteria have evolved in the various guts to utilize different food sources. Biolog plates are among several commercial products that allow testing of isolates for growth on individual carbon sources. Unknowns were classified by discriminant analysis based on utilization of 22 carbon sources (25). See Supporting Information for more protocol details. Genotypic Methods. Genotypic methods are used to differentiate bacteria based on observations of genetic variability. The working hypothesis behind each of the genetic methods is that specific genotypes of fecal bacteria have evolved to succeed in the gut environments of their respective hosts. Genotyping methods use various targets to detect variability with the expectation that some genetic variability is driven by host specificity. RT-HindIII. Ribotyping by use of the restriction enzyme HindIII was done on 630 members of the 900-isolate library at the Biological Consulting Services of North Florida laboratory (Gainesville, FL) according to a previously described protocol (27). Ribotyping detects variability in the ribosomal DNA operon, the 16S portion of which is commonly used to establish phylogenetic relations among bacteria. Unknowns were classified by use of the Libraries module of BioNumerics software (Applied Maths, Sint-Martens-Latem, Belgium) and a curve-based similarity coefficient. See Supporting Information for more protocol details. RT-EcoRI. Ribotyping by use of the restriction enzymes EcoRI and PvuII was done on the full 900-isolate library at the Institute for Environmental Health (Seattle, WA) by a previously described protocol (10). Ribotyping by two restriction enzymes, instead of one, may result in more bands and, therefore, more information with which to resolve more

genotypes (6). Unknowns were classified by direct visual matching with ribotypes in the known-source library. See Supporting Information for more protocol details. PFGE. Pulsed-field gel electrophoresis of macrorestriction digests using a NotI-based protocol (PFGE) was done on the full 900-isolate library by a combination of efforts at the West Virginia Department of Agriculture Laboratory (Moorefield, WV) and Marshall University (Huntington, WV). DNAfragment banding patterns resulting from whole-genome digestion are based on variability anywhere in the genome. PFGE is generally considered an extremely sensitive method of detecting genetic differences between strains (6, 28). Unknowns were classified by UPGMA cluster analysis, using curve-based similarity coefficients and accepting matches at 85% similarity. See Supporting Information for more protocol details. BOX-PCR. The protocol for repetitive-element DNA PCR (rep-PCR) with BOX primers (29, 30) was used on the full 900-isolate library at the Virginia Institute of Marine Science, College of William and Mary (Gloucester Point, VA). Various rep-PCR protocols use primers that target repeated sequences of DNA found in many bacterial genomes (BOX, REP, and ERIC sequences; see ref 31 for descriptions). DNA-fragment banding patterns resulting from amplification of regions between primers are based on variability anywhere in the genome, but only a small proportion (fragments between about 100 and 8000 base pairs) can be amplified and detected in these protocols. Genotyping by the various rep-PCR protocols is based on the portion of genetic variability, anywhere in the genome, that occurs on a fragment that can be amplified and visualized (29). Unknowns were classified by computer-assisted comparison with densitometric curves in the known-source library. See Supporting Information for more protocol details. REP-PCR. The rep-PCR protocol was done on 630 members of the 900-isolate library at the ODML (Columbus, OH) using REP primers (adapted in ref 14 from ref 29). The REP primers are a common alternative to BOX primers for rep-PCR-based genotyping (30). Unknowns were classified by computer-assisted comparison with densitometric curves in the known-source library. See Supporting Information for protocol details. Data Analysis. Protocol Evaluations. Protocols were evaluated against three criteria (reproducibility, accuracy, and robustness) and then described by an assessment of relevance. A subset of 26 challenge isolates (REPLICATE) consisted of arbitrarily selected isolates recultivated from the known-source library. The REPLICATE subset was used to evaluate reproducibility, defined in this analysis as the ability of each method to reassociate an isolate that was already in the known-source library with its source. Classification of a REPLICATE isolate as “unknown” was considered incorrect. A second subset of 150 challenge isolates (ACCURACY) was cultivated from different individual animals of the eight host species represented in the known-source library. For the purposes of this study, accuracy was defined as the ability to correctly identify the source of isolates that were collected independently from the known-source library. Summary of data only as rate of correct classification (RCC) was considered insufficient in this evaluation because of the wide range of protocols used to classify the challenge isolates. For instance, protocols with high average RCC (ARCC) could be more credibly interpreted if a large percentage of the ACCURACY isolates was classified than if a small percentage of the ACCURACY isolates was classified. To accommodate the variety of data-analysis approaches, test results for accuracy were presented as both the percentage of isolates for which classification was attempted and the correct classification rate. Failure to attempt classification, defined as classification

unknown, was considered a null response when evaluating accuracy. Thus, the RCC for each source category was calculated using only those isolates for which a classification was attempted. A third subset of 24 isolates (RINGER), which originated with hosts different from the eight species represented in the known-source library, was used to evaluate robustness. Possible sources for RINGER isolates included cats, raccoons, llamas, and other warm-blooded animals in the watershed. The RINGER isolates have not been identified to participating laboratories (to allow for future work with this data set). Robustness is defined for the purposes of this study as the ability of each protocol to identify isolates that came from sources not represented in the library; in other words, to detect the unexpected. Identification of RINGER isolates as anything other than unknown was considered an incorrect answer for species-level classification (eight-way). Four protocols (ARA, CUP, RT-HindIII, and BOX-PCR) used approaches that did not validate each classification and, therefore, could not classify any isolates as unknown. The remaining three protocols used either visual or computerassisted direct matching and evaluated each resulting classification for validity (RT-EcoRI, PFGE, and REP-PCR). To illustrate the relevance of these results to application of MST tools, the true host source and the classification provided by each protocol were compiled for the challenge library. Hypothetical samples contained 10 unknown isolates from each of three sources, chosen by randomly resampling the library of unknowns, with replacement. The results that each protocol would have reported for multiple trials were summarized by calculation of the arithmetic mean and the standard deviation. The summaries depict the range of classifications that each protocol would have calculated for water samples contaminated with equal concentrations of E. coli from three fecal sources. Classifications of Unknowns. Each researcher participating in this study submitted the results of his or her own protocol for classification of the challenge isolates without prior knowledge of which were REPLICATE, ACCURACY, or RINGER isolates. Researchers were asked to classify unknowns according to three strategies: species-level classification (eight-way split among source species); management-level classification (three-way split among humans, domesticated animals {cattle, chickens, dogs, horses, swine}, and wildlife {Canada geese and white-tailed deer}); or human/nonhuman classification (two-way split). Several researchers chose to submit no classification in cases where the result for an isolate was uncertain (classification as unknown). For MST protocols that were less-well developed (CUP, PFGE, BOX-PCR, and REP-PCR) or where the researcher desired to reevaluate results (RT-EcoRI), multiple submissions (up to four) were accepted. Submissions were scored upon receipt, but the composition of the challenge library was not disclosed to the researchers. Only the submission considered by each researcher to best represent the protocol is discussed in this paper. Statistical Analysis. Evaluation of results at different levels of resolution (two, three, and eight categories) sometimes resulted in imbalanced representation among categories. When possible, hypotheses were formulated to allow significance testing. Significance testing of the hypothesis that an observed rate of correct classification (RCC) was the same as the expectation for random classification was done by use of the odds-ratio test (32). The odds-ratio test calculates the probability that an observed proportion is equal to an expected proportion given a number of observations. The expectation for RCC under random classification was estimated as the inverse of the number of classification categories. Significance testing of the hypothesis that the number of true isolate classifications was the same as the VOL. 38, NO. 22, 2004 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

6111

TABLE 1. Number and Percentage of REPLICATE Isolates Correctly Reclassified into the Correct Source Categoriesa phenotypic methods ARA

CUP

genotypic methods RT-HindIII

RT-EcoRI

PFGE

BOX-PCR

REP-PCR

total correct total attempted overall percent

6 26 23%

Species-Level Classification (Eight-Way Split) 5 3 14 25 23 26 20% 13% 54%

24 24 100%

16 26 62%

11 23 48%

human correct human attempted human percent nonhuman correct nonhuman attempted nonhuman percent

1 3 33% 19 23 83%

Human/Nonhuman Classification (Two-Way Split) 0 0 2 3 3 3 0% 0% 67% 22 20 13 22 20 23 100% 100% 57%

3 3 100% 21 21 100%

1 3 33% 23 23 100%

0 3 0% 19 20 95%

a Attempted refers to the number of REPLICATE isolates that were successfully cultivated by the analyzing laboratory; correct refers to the number of REPLICATE isolates correctly classified, and percent was calculated as correct divided by attempted. ARA, antibiotic resistance analysis; CUP, carbon utilization profile; RT-HindIII, ribotyping using HindIII; RT-EcoRI, ribotyping using EcoRI and PvuII; PFGE, macrorestriction with NotI followed by pulsed-field electrophoresis; BOX-PCR, rep-PCR using BOX primer; REP-PCR, rep-PCR using REP primers

number of false isolate classifications was done by a onesided t-test (32).

Results and Discussion The design objective of classifying cattle by husbandry practice (dairy or beef) was discontinued after each of the seven research groups, working with the same E. coli isolates, submitted their classifications. It was apparent that no differences would be detected between beef and dairy cattle and that use of eight species-based classes simplified reporting. Classification of the 200 challenge isolates used up to three strategies with differing resolution: the eightway split (species-level classification), the three-way split (management-level classification to human, domestic animals, and wildlife) and the two-way split (human nonhuman classification). After classification, challenge isolates were summarized as three subsets (REPLICATE, ACCURACY, and RINGER) to allow comparison of the seven tested MST protocols based on reproducibility, accuracy, and robustness, respectively. The REPLICATE challenge subset consisted of 26 blinded isolates recultivated from the original known-source library. Because these isolates were already in the known-source libraries held by each researcher, reclassification of a REPLICATE isolate as unknown was considered incorrect. Results of reproducibility tests for eight-way (species-level) and two-way (human/nonhuman) classification are summarized in Table 1. Results in Table 1 are summarized across the eight-way split because each host had nearly equal representation, but results for the two-way split are presented by category because of imbalance (many more nonhuman isolates than human isolates). Only PFGE correctly classified all of the REPLICATE isolates into the correct species-level category. RT-EcoRI, BOX-PCR, and REP-PCR each classified between 48 and 62% of the REPLICATE isolates correctly; ARA, CUP, and RTHindIII correctly classified fewer than 25% of REPLICATE isolates. Correct classification rates were higher for the twoway split, in part because only 3 of the 26 REPLICATE isolates (12%) came from human sources. Correct classification of nonhuman sources was higher than or equal to correct classification of human-source isolates for all protocols except RT-EcoRI. One cause for incorrect reclassification of replicates may have been sharing of subtypes among individuals in different source groups. Sharing of subtypes between hosts was observed in this library and other reports (15, 27). REPLICATE isolates belonging to shared subtypes can be classified to the 6112

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 38, NO. 22, 2004

wrong host species even if the data are reproduced accurately. In the case of methods using discriminant analysis, another explanation is that patterns from various sources did not form distinct clusters; therefore, the original isolate in the library also may have been incorrectly classified by means of the cross-validation procedure. The ACCURACY challenge subset consisted of 150 isolates collected independently from the known-source library. Results of this test are summarized in Table 2. At a minimum, each protocol was expected to perform better than random classification of ACCURACY isolates into categories. Random classification would result in expected RCC of 12% for the eight-way split, 33% for the three-way split, and 50% for the two-way split. Individual protocols exceeded random correct classification rates for some sources and failed to exceed for others. Five of the seven protocols achieved ARCC better than the random expectation at one or more levels of classification; however, CUP and RT-HindIII did not exceed the random expectation in ARCC (p > 0.05) at every tested level. These results indicate a host-specific signal in the known-source library as analyzed by at least five of the seven protocols. Despite the finding that five of seven protocols met the minimum expectation, the ARCC for the ACCURACY subset are substantially lower than were anticipated on the basis of values reported in the literature. A 64% ARCC was reported in a six-way classification using fecal coliforms with a protocol similar to ARA (33). In our study, ARA had an ARCC of 27% for eight-way classification. An ARCC of 93% was reported for a two-way split using CUP on enterococci (26). In our study, CUP on E. coli had an ARCC of 55%, not significantly different from the random expectation of 50% (p < 0.05). An ARCC of 81% was reported for a two-way split using E. coli with a protocol similar to RT-HindIII (27). In our study, RTHindIII had an ARCC of 49% for the two-way split, also not significantly better than the random expectation. Classification was attempted for 29% of unknowns in a previous report of the RT-EcoRI protocol using a local known-source library (about 1000 isolates; 10). In this study, RT-EcoRI attempted classification for 8 of 144 (6%) ACCURACY isolates tested. Little was previously known about performance of PFGE for this type of analysis, although, in another methodscomparison study, PFGE and ribotyping gave nearly identical results (6). A protocol similar to BOX-PCR was reported to have ARCC of 63% for a seven-way split (originally reported in ref 14; later revised by ref 34). The BOX-PCR protocol in this study had an ARCC of 22% for the eight-way split. On the basis of prior results (14), REP-PCR was expected to have accuracy similar to but somewhat lower than BOX-PCR; in

TABLE 2. Percentage of ACCURACY Isolates Classified by Each Protocol into the Correct Categoriesa phenotypic methods source category

isolates collected

humans cattle chickens dogs horses swine canada geese white tailed deer sum or average

17 33 21 14 16 18 17 14 150

humans domestic wildlife sum or average humans nonhumans sum or average

isolates tested

ARA

genotypic methods RT-HindIII

CUP

classified

correct

classified

correct

classified

15-17 28-33 19-21 13-14 15-16 17-18 16-17 10-14

100% 100% 100% 100% 100% 100% 100% 100% 100%

18% 6% 0% 79%*** 19% 72%*** 12% 14% 27% ***

100% 100% 100% 100% 100% 100% 100% 100% 100%

12% 24%* 5% 23 0% 24 0% 20% 13%

17 102 31 150

15-17 94-102 27-31

100% 100% 100% 100%

24% 62%*** 32% 39%

100% 100% 100% 100%

12% 62%*** 30% 34%

17 133 150

15-17 121-133

100% 100% 100%

24% 86%*** 55%

100% 100% 100%

12% 98%*** 55%

correct

RT-EcoRI classified

correct

PFGE

BOX-PCR

REP-PCR

classified

correct

classified

correct

classified

correct

40% 16% 58% 36% 27% 17% 25% 15% 29%

67%** 40% 18% 0% 25% 100%** 75%** 0% 41%***

100% 100% 100% 100% 100% 100% 100% 100% 100%

31%* 21% 52%*** 7% 6% 11% 24% 21% 22%**

67% 64% 63% 69% 75% 83% 50% 82% 69%

60%*** 22% 17% 11% 0% 33%* 50%* 11% 26%***

humans cattle chickens dogs horses swine Canada geese white tailed deer sum or average

40% 29% 21% 30%

67% 82%*** 67% 72%***

100% 100% 100% 100%

31% 71%*** 42% 48%***

67% 70% 63% 67%

60% 82%*** 41% 61%***

humans domestic wildlife sum or average

40% 27% 33%

67% 91%*** 79%***

100% 100% 100%

31% 95%*** 63%***

67% 69% 68%

60% humans 94%*** nonhumans 77%*** sum or average

Species-Level Classification (Eight-Way Split)

VOL. 38, NO. 22, 2004 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

100% 100% 100% 100% 100% 100% 100% 100% 100%

6% 13% 19% 8% 20% 11% 18% 7% 13%

6% 6% 10% 8% 0% 0% 12% 0% 5%

100% 100%* 50% 100% na na 100%* na 90% ***

Management-Level Classification (Three-Way Split) 100% 100% 100% 100%

6% 63%*** 29% 33%

6% 5% 7% 6%

100% 100%** 100% 100%***

Human/Nonhuman Classification (Two-Way Split) 100% 100% 100%

6% 92%*** 49%

6% 5% 6%

100% 100%** 100%**

a Classified refers to the percentage of tested isolates for which classification was attempted, and correct refers to rate of correct classifications (RCC). Significance of (H:RCC ) random expectation) is denoted as follows: *, p < 0.05; **, p < 0.01; ***, p < 0.001. ARA, antibiotic resistance analysis; CUP, carbon utilization profile; RT-HindIII, ribotyping using HindIII; RT-EcoRI, ribotyping using EcoRI and PvuII; PFGE, macrorestriction with NotI followed by pulsed-field electrophoresis; BOX-PCR, rep-PCR using BOX primer; REP-PCR, rep-PCR using REP primers; na, not applicable because no isolates were classified to the source category

9

6113

actuality, REP-PCR achieved a 26% ARCC for the eight-way split in this study. Lack of classification accuracy is not the only source of error for these protocols. Failure to attempt classification for a large proportion of challenge isolates can also limit effectiveness because a different source distribution among unclassified isolates compared to classified isolates would bias interpretations. Among the explanations for low ARCC with this data set are variability in data collection and analysis protocols, subtype sharing among source groups at the same location, inadequate library size, and temporal variability. Variability in data collection and analysis protocols could not be evaluated in this paper because we did not have multiple researchers using the same protocol. Subtype sharing among animals of different source groups (15, 27) could be accentuated by our sample design, in which individuals from several source categories were sampled at various locations and times. In these instances, proximity may lend itself to cross-species transfer of E. coli, particularly through coprophagy. The extent of cross-species subtype sharing at single locations was evaluated by considering ACCURACY isolates from three locations at which three or four host species were sampled. Results submitted by the RT-EcoRI protocol were used because the protocol includes assignation of a subtype number to each isolate. The total number of isolates collected from the three locations was 202, among which 113 subtypes were detected by RT-EcoRI. None of the subtypes was detected in multiple host species at the same location. Direct sharing of E. coli between sources at a single location does not appear to account for the lower than expected correct classification rates in this study. Effective source tracking using known-source libraries requires that the library adequately represent diversity of source feces in the study area. Library size is one measure of representation. Limited information is available to describe the number of isolates required for a library to represent source populations of different sizes being studied via various analytical methods (see refs 24 and 35). The known-source library developed in this study included 186 feces samples. Of the 900 isolates included in the library, some are not independent because they represent clones from the same individual. The effective size of this library was estimated by assuming that indistinguishable isolates from each feces sample were clones. PFGE and RT-EcoRI, both considered highly discriminating subtyping methods (28), were used to declone the library, resulting in an effective library size of about 540 independent isolates. Library-size requirements are difficult to predict because they depend, in part, on analytical method (6) and the size of contributing populations. Many prior applications of source tracking have used libraries of comparable or lesser size to that used in this study. Dombek et al. (14) used a collection of 154 isolates representing six sources to evaluate the efficacy of a BOX-PCR protocol similar to that reported in this study. Carson et al. (13) used a library of 482 isolates representing six sources to compare the performance of protocols similar to the RT-HindIII and BOX-PCR protocols in this study. Harwood et al. (33) used a library approximately 4 times the size of that used in this study (2136 isolates) to evaluate the utility of fecal coliforms for source tracking by a method similar to our ARA protocol. Hartel et al. (36) used a collection of 568 isolates representing four sources to address geographic variability among source populations at three locations in Georgia and one in Idaho by a method similar to our RT-EcoRI protocol. It has been estimated that libraries composed of 10002000 isolates per source would be needed to adequately represent a study area (sensu ref 35), and it has been estimated that more than 100 distinct subtypes were present in the population at a single cattle feed lot (37). Despite these 6114

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 38, NO. 22, 2004

indications of the great diversity that must be represented in known-source libraries, the authors are aware of only two groups of MST studies that report using substantially larger known-source libraries than our own-those that used a commercial mega-library (as reported in ref 10) and those that used ARA-type protocols (as reported in ref 24). The commercial mega-library currently includes more than 40 000 known-source strains (38); and, for ARA using enterococci, it was reported that a library consisting of at least 2300 total isolates would be appropriate to represent multiple sources in a single watershed (24). In this study, a modest increase in the number of isolates per source from 70 to 100 did not substantially change rates of correct classification for four methods (data not shown), likely because many of the added isolates were clones already represented in the subset of 70 isolates. Evidence that the library of 540 independent isolates was inadequate to characterize the contributing sources include results of discriminant analysis using two protocols, in which the resubstitution tables had higher correct classification rates than did cross-validation tables. This pattern was suggested to indicate incomplete representation in the known-source library (24). It is likely that the known-source library assembled for this study, although consistent in size with contemporary applications of source tracking using genotypic methods, did not adequately represent the diversity of local populations. Future work may involve testing the challenge isolates against larger libraries kept by the contributing researchers. Temporal variability may also have contributed to lower than expected measures of accuracy by most methods. By design, ACCURACY isolates were collected 9 months after the known-source library. Separation of ACCURACY and known-source isolate collection in time may have reduced the effectiveness of the known-source library if E. coli populations changed because of factors such as seasonally variable diet, reproductive cycle, or temperature (35). Two of the protocols (BOX-PCR and REP-PCR) were used to test for the temporal effect by comparison of internal validation using the known-source library (cross validation) with external validation using the ACCURACY data set (challenge with an independent collection of isolates). On the basis of this comparison, correct classification rates using the ACCURACY subset could be compared with the expected correct classification rates (based on cross validation) if a test subset had been collected simultaneously with the known-source library. Classification results for the ACCURACY data set by the REP-PCR protocol are listed in Table 2. For protocol REP-PCR, the ARCC for an eight-way (species-level) classification of the ACCURACY data set (26%, Table 2) was not different from the cross-validation ARCC of 30% (p > 0.05). For protocol BOX-PCR, recalculated using the classification protocol from REP-PCR, the ARCC for the ACCURACY data set (26%) was different from the cross-validation ARCC of 49% (p < 0.001). This analysis indicates that temporal variability may have contributed to lower than expected ARCC of the ACCURACY data set for some protocols. The potential effect of temporal variability can be minimized by simultaneous collection of known-source and water-isolated E. coli but, in most cases, cannot be completely avoided because of the cost and logistics of compiling a complete library simultaneously with each set of water-isolated E. coli. The 24-isolate RINGER challenge subset represented isolates that originated with previously unseen sources such as raccoons. The ability of the three protocols to detect RINGER isolates closely mirrored their abilities to classify ACCURACY isolates. The RT-EcoRI protocol succeeded by not classifying any of the 24 RINGER isolates, although 94% of ACCURACY isolates also were not classified using the RT-EcoRI protocol. The PFGE protocol did not classify 16

(67%) of the RINGER isolates and 71% of ACCURACY isolates. The REP-PCR protocol did not classify 8 (33%) of the RINGER isolates and 31% of ACCURACY isolates. The three protocols that attempted to separate isolates from unknown sources could not be shown to do so at a higher rate than expected based on the number of other classifications attempted. What is clear is that effective rejection of RINGER isolates was accomplished by only one of the tested protocols (RT-EcoRI). Because the choice of library sources is based on prior knowledge, inability to detect the presence of unexpected sources could be an important source of error for approaches using species-level classification. The results and discussion above cover reproducibility, accuracy, and robustness on the basis of individual isolates. A further evaluation of results with the ACCURACY subset was attempted on a sample-level basis to illustrate relevance, defined as how each protocol would perform in a field application. Hypothetical test samples were built of 10 randomly selected ACCURACY isolates from each of three source categories. Data were summarized as though the isolates represented an environmental sample. The results of 30 such trials in which sample composition was limited to isolates from humans, swine, and Canada geese are shown in Figure 1. By this approach, it is apparent that no method would have reported results that would correctly identify contributing sources to the hypothetical sample. Each protocol, except RT-EcoRI, correctly indicated the presence of all three sources (human, swine, Canada geese); however, they generally also indicated the presence of most other sources. It has been suggested that a cutoff be instituted to remove the effect of low-incidence false-positive classification (6, 7). Institution of a cutoff at one isolate (3%) would allow PFGE to indicate the presence of only the correct three contamination sources, but with a 63% unknown category. A higher cutoff of 10-15% would eliminate virtually all results from several protocols, including PFGE. RT-EcoRI did not require a cutoff; the protocol correctly indicated the presence of Canada goose and human waste in the sample but missed the presence of swine and attributed the remainder (about 90%) to unknown sources. Similar analysis was done on 300 hypothetical samples, each composed of 10 randomly chosen isolates belonging to each of three randomly chosen sources, of which there are 56 possible combinations. Classification was considered false if the source was not present in the sample and true if the source was present in the sample, regardless of the true origin of each individual isolate. In this way, an overall assessment of contamination sources at the sample level could be constructed for all possible combinations of sources (Table 3, eight-way split). A minimum expectation for these methods is that the true detection rate should be higher than the false detection rate. Each protocol was able to reliably detect at least two of the eight sources, but none was able to reliably detect all eight sources at the p < 0.05 significance level. With the exception of ARA with swine, none of the protocols would have correctly detected, within one standard deviation, all 10 isolates from the true sources in any sample. Carson et al. (39) suggested that attempts to separate more than three sources are overly ambitious for a protocol similar to RT-HindIII, though in a later publication the same group reported no such difficulties for a protocol similar to BOXPCR (13). Scott et al. (27) also reported that a protocol similar to RT-HindIII was most applicable to separation of only two sources. A further set of 200 hypothetical samples was constructed from the ACCURACY subset to evaluate the human/nonhuman two-way classification. The first 100 samples were composed of 10 human and 20 nonhuman isolates, with the nonhuman isolates evenly divided between

FIGURE 1. Results generated by each protocol for a hypothetical sample composed of 10 E. coli each of human, swine, and Canada goose origin. Bars represent one standard deviation around the mean results for 30 trials. Black bars are true classifications, and gray bars are false classifications. Note compressed y-axis for RT-EcoRI and PFGE. ARA, antibiotic resistance analysis; CUP, carbon utilization profile; RT-HindIII, ribotyping using HindIII; RT-EcoRI, ribotyping using EcoRI and PvuII; PFGE, macrorestriction with NotI followed by pulsed-field electrophoresis; BOX-PCR, rep-PCR using BOX primer; REP-PCR, rep-PCR using REP primers. two randomly selected sources. Results for these trials were used to calculate rates of true classification for humans and nonhumans (Table 3, two-way split). A perfect true score would be 10 human isolates and 20 nonhuman isolates. The second 100 samples were composed of 30 nonhuman isolates, 10 from each of three randomly selected sources. The rate of human classification for these samples was used as the false classification rate as human source in Table 3. The false nonhuman-classification rate could not be calculated for multiple hypothetical samples of 30 human isolates because VOL. 38, NO. 22, 2004 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

6115

TABLE 3. Number of ACCURACY Isolates Correctly (True) or Incorrectly (False) Classified by Each Protocol to Source Categoriesa phenotypic methods CUP

genotypic methods RT-HindIII

RT-EcoRI

source category

ARA

humans (true) humans (false) cattle (true) cattle (false) chickens (true) chickens (false) dogs (true) dogs (false) horses (true) horses (false) swine (true) swine (false) Canada geese (true) Canada geese (false) white-tailed deer (true) white-tailed deer (false) unknowns

3.1 ( 1.7*** 1.7 ( 1.4 1.0 ( 1.1*** 0.6 ( 1.0 0.7 ( 1.0** 1.0 ( 1.3 16.9 ( 3.1*** 13.1 ( 3.1 3.8 ( 1.9 3.6 ( 2.1 8.1 ( 2.1*** 2.2 ( 1.5 2.4 ( 1.8 2.6 ( 1.8 2 ( 1.4*** 0.7 ( 1.0 0

Species-Level Classification (Eight-Way Split) 1.4 ( 1.2*** 1.8 ( 1.4 0.6 ( 0.8*** 0.4 ( 0.7 2.0 ( 1.5 0 6.8 ( 2.9 2.8 ( 1.7 0.8 ( 0.9*** 6.5 ( 2.5 2.6 ( 2.1 0 0.9 ( 1.0 3.8 ( 1.8* 0.5 ( 0.8*** 1.7 ( 1.4 3.3 ( 2.0 0 5.0 ( 2.5* 2.9 ( 2.3 0.7 ( 0.9*** 4.4 ( 2.5 3.1 ( 2.0 0 1.2 ( 1.5 5.3 ( 2.4 0 1.6 ( 1.4 5.8 ( 2.6 0 5.0 ( 2.1*** 3.4 ( 2.0*** 0 4.0 ( 2.0 2.6 ( 1.9 0 1.5 ( 1.5 3.8 ( 2.0 1.1 ( 0.9*** 2.1 ( 1.7 3.5 ( 2.0 0 6.5 ( 2.6 4.8 ( 2.6 0 7.2 ( 2.9 6.3 ( 3.0 0 0 0 27 ( 1.8

humans (true, of 10) humans (false) nonhumans (true, of 20) unknowns

5.0 ( 2.2*** 3.9 ( 2.1 25 ( 2.3 0

PFGE

BOX-PCR

REP-PCR

2.4 ( 1.3*** 0.7 ( 0.9 1.8 ( 1.9 1.8 ( 1.8 1.6 ( 1.3*** 0.9 ( 1.1 0.5 ( 0.8 0.7 ( 0.8 0.8 ( 1.0 0.3 ( 0.7 1.8 ( 1.3*** 0.5 ( 0.8 2.1 ( 1.4*** 0.8 ( 1.0 2.4 ( 1.8 2.0 ( 1.4 20 ( 3.4

4.4 ( 1.9*** 1.8 ( 1.5 5.3 ( 2.4 5.3 ( 2.6 5.7 ( 1.9*** 3.8 ( 2.1 2.5 ( 1.6 3.5 ( 2.0 3.1 ( 1.8 3.6 ( 2.0 3.3 ( 2.2 3.1 ( 2.3 4.3 ( 2.3*** 3.2 ( 2.0 4.1 ( 2.1** 3.5 ( 1.9 0

4.1 ( 2.1*** 1.2 ( 1.2 5.0 ( 2.6 5.7 ( 2.5 2.3 ( 2.0 2.2 ( 2.0 1.6 ( 1.4 1.3 ( 1.2 1.7 ( 1.2 2 ( 1.4 1.5 ( 1.5*** 6.5 ( 2.6 2.7 ( 1.4*** 0.8 ( 1.0 1.7 ( 1.5 1.7 ( 1.5 8.1 ( 2.6

Human/Nonhuman Classification (Two-Way Split) 2.0 ( 1.6*** 1.9 ( 1.4 0.7 ( 0.8*** 2.7 ( 1.5*** 1.1 ( 1.4 2.2 ( 1.6 0 0.7 ( 0.8 27 ( 2.2 27 ( 1.7 0.8 ( 1 6.0 ( 3.3 0 0 27 ( 1.7 19 ( 3.4

4.6 ( 1.9*** 1.7 ( 1.3 25 ( 1.9 0

4.1 ( 1.6*** 1.0 ( 1.1 15 ( 3.1 8.1 ( 2.7

a In all species-level classifications, samples contained 10 isolates from 3 randomly selected sources, and in human/nonhuman trials, samples contained 10 human and 20 nonhuman isolates. Reporting format is number detected of 10 (or 20) possible, plus-or-minus one standard deviation. Significance of (H: true detects not greater than false detects) is denoted as follows: *, p < 0.05; **, p < 0.01; ***, p < 0.001. ARA, antibiotic resistance analysis; CUP, carbon utilization profile; RT-HindIII, ribotyping using HindIII; RT-EcoRI, ribotyping using EcoRI and PvuII; PFGE, macrorestriction with NotI followed by pulsed-field electrophoresis; BOX-PCR, rep-PCR using BOX primer; REP-PCR, rep-PCR using REP primers.

there were only 17 human-origin isolates in the ACCURACY subset. In all cases except RT-HindIII, the true human-classification rate was higher than the false human-classification rate for the single-tailed hypothesis (p < 0.001). In no case, however, did one standard deviation around the mean result include the true number of isolates in a sample. The small number of human isolates detected indicates that most protocols would underestimate the importance of human contributions to water contamination in this study area. Only protocols PFGE and RT-EcoRI reported a sample composition for the two-way split that was close to the correct 2:1 nonhuman:human ratio (2.2:1 and 1.2:1, respectively, calculated from Table 3). Despite numerous reports in the literature that indicate efficacy of microbial source tracking methods to identify sources of fecal contamination to environmental waters, the results of this investigation clearly show limitations in tested protocols that might hinder application to goals such as source-allocation in a study area. Within-host similarity among E. coli isolates, the cohesiveness of the host-specific signal, was not sufficiently strong to accurately classify a large proportion of the ACCURACY isolates (as would be predicted from ref 40). The host-specific signal does however exist (sensu ref 41): four of the seven protocols exceeded random expectation at all levels, and a fifth had ARCC that exceeded the random expectation at one level. Two approaches for improvement of these protocols would be to (i) better represent contributing host feces in the known-source E. coli library and (ii) develop better approaches to classify E. coli to host species. When results of isolate-level analysis were applied to hypothetical water samples, ARA, CUP, RT-HindIII, BOXPCR, and REP-PCR would have made inaccurate reports of contributing sources. RT-EcoRI and PFGE would have reported contributing sources with reasonable accuracy; however, interpretation of the reports by RT-EcoRI and PFGE would be severely hampered by classification of more than 6116

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 38, NO. 22, 2004

half of the isolates as unknown. These difficulties may be addressed for each protocol with larger library size, an analysis that was beyond the scope of our study. This paper represents a single attempt to evaluate seven protocols for application to microbial source tracking as used in current practice. This follows other reports of the limitations of MST based on the RT-HindIII protocol (27) and multiple methods comparison (17). The results of this study indicate that current protocols for isolate subtyping may be insufficient to accomplish many goals of MST; however, extensive investigations continue by researchers to standardize and improve protocols (42). Given the results described in this paper and those by other investigators, it is crucial that thorough reliability assessment of the chosen protocol be done before and during attempts to apply these protocols to a given environmental setting.

Acknowledgments The authors thank Roger Boyer of the Potomac Headwaters Resource Conservation and Development Project Office; Rebecca MacLeod, District Conservationist of the Natural Resources Conservation Service and her staff; and the many residents of Berkeley County who provided access to their farms and animal shelters. This research was done with financial support by West Virginia Department of Environmental Protection; West Virginia Department of Health and Human Resources; Virginia Department of Environmental Quality; and Berkeley County Commission. In-kind support was provided by the West Virginia Department of Agriculture. The use of firm, trade, or brand names in this paper is for identification purposes only and does not constitute endorsement by the U.S. Geological Survey.

Supporting Information Available Details of sample collection and experimental protocols. This material is available free of charge via the Internet at http://pubs.acs.org.

Literature Cited (1) Sinton, L. W.; Finlay, R. K.; Hannah, D. J. Distinguishing human from animal faecal contamination in water: a review. N. Z. J. Mar. Freshwater Res. 1998, 32, 323-348. (2) Scott, T. M.; Rose, J. B.; Jenkins, T. M.; Farrah, S. R.; Lukasik, J. Microbial source tracking: current methodology and future directions. Appl. Environ. Microbiol. 2002, 68, 5796-5803. (3) Simpson, J. M.; Santo Domingo, J. W.; Reasoner, D. J. Microbial source tracking: state of the science. Environ. Sci. Technol. 2002, 36, 5279-5288. (4) Field, K. G.; Bernhard, A. E.; Brodeur, T. J. Molecular approaches to microbiological monitoring: fecal source detection. Environ. Monit. Assess. 2003, 81, 313-326. (5) Field, K. G.; Chern, E. C.; Dick, L. K.; Fuhrman, J.; Griffith, J.; Holden, P. A.; LaMontagne, M. G.; Le, J.; Olson, B.; Simonich, M. T. A comparative study of culture-independent, libraryindependent genotypic methods of fecal source tracking. J. Water Health 2003, 1 (4), 181-194. (6) Myoda, S. P.; Carson, C. A.; Fuhrmann, J. J.; Hahm, B. K.; Hartel, P. G.; Yampara Iquise, H.; Johnson, L. A.; Kuntz, R. L.; Nakatsu, C. H.; Sadowsky, M. J.; Samadpour, M. Comparison of genotypicbased microbial source tracking methods requiring a host origin database. J. Water Health 2003, 1 (4), 167-180. (7) Harwood, V. J.; Wiggins, B.; Hagedorn, C.; Ellender, R. D.; Gooch, J.; Kern, J.; Samadpour, M.; Chapman, A. C. H.; Robinson, B. J.; Thompson, B. C. Phenotypic library-based microbial source tracking methods: efficacy in the California collaborative study. J. Water Health 2003, 1 (4), 153-166. (8) Hagedorn, C.; Robinson, S. L.; Filz, J. R.; Grubbs, S. M.; Angier, T. A.; Reneau, R. B. J. Determining sources of fecal pollution in a rural Virginia watershed with antibiotic resistance patterns in fecal streptococci. Appl. Environ. Microbiol. 1999, 65 (12), 55225531. (9) Graves, A. K.; Hagedorn, C.; Teetor, A.; Mahal, M.; Booth, A. M.; Reneau, R. B. J. Antibiotic resistance profiles to determine sources of fecal contamination in a rural Virginia watershed. J. Environ. Qual. 2002, 31, 1300-1308. (10) Hyer, K. E.; Moyer, D. L. Sources and Patterns of Fecal Coliform Bacteria in Three Streams in Virginia, 1999-2000; U.S. Geological Survey: Reston, VA, 2003; p 89. (11) Simmons, G. M.; Waye, D. F.; Herbein, S.; Myers, S.; Walker, E. In Estimating Nonpoint Fecal Coliform Sources in Northern Virginia’s Four Mile Run Watershed; Younos, T., Poff, J., Eds.; Abstracts of the Virginia Water Research Symposium 2000; Virginia Water Research Symposium: Blacksburg, VA, 2000; pp 248-267. (12) Kariuki, S.; Gilks, C.; Kimari, J.; Obanda, A.; Muyodi, J.; Waiyaki, P.; Hart, C. A. Genotype analysis of Escherichia coli strains isolated from children and chickens living in close contact. Appl. Environ. Microbiol. 1999, 65 (2), 472-476. (13) Carson, C. A.; Shear, B. L.; Ellersieck, M. R.; Schnell, J. D. Comparison of ribotyping and repetitive extragenic palindromicPCR for identification of fecal Escherichia coli from humans and animals. Appl. Environ. Microbiol. 2003, 69 (3), 1836-1839. (14) Dombek, P. E.; Johnson, L. K.; Zimmerley, S. T.; Sadowsky, M. J. Use of repetitive DNA sequences and the PCR to differentiate Escherichia coli isolates from human and animal sources. Appl. Environ. Microbiol. 2000, 66 (6), 2572-2577. (15) McLellan, S. L.; Daniels, A. D.; Salmore, A. K. Genetic characterization of Escherichia coli populations from host sources of fecal pollution by using DNA fingerprinting. Appl. Environ. Microbiol 2003, 69 (5), 2587-2594. (16) TOFPSW. http://www.ub.edu/microbiologia/TOFPSW.htm (accessd June 2, 2004). (17) Griffith, J. F.; Weisberg, S. B.; McGee, C. D. Evaluation of microbial source tracking methods using mixed fecal sources in aqueous test samples. J. Water Health 2003, 1 (4), 141-151. (18) Mathes, M. V. Relation of Bacteria in Limestone Aquifers to Septic Systems in Berkeley County, West Virginia; U.S. Geological Survey: Reston, VA, 2000; p 54. (19) Gorman, J. L.; Pasto, J. K.; Crocker, C. D. Soil Survey of Berkeley County, West Virginia; U.S. Department of Agriculture, National Resources Conservation Service: Washington, DC, 1966; p 154. (20) U.S. Census Bureau. USA QuickFacts; available at http://quickfacts.census.gov/qfd/states/54000.html; 2001 (accessed September 22, 2003). (21) MacLeod, R. NRCS, personal communication, 2000. (22) MacLeod, R. NRCS, personal communication, 2003. (23) U.S. Environmental Protection Agency. Improved Enumeration Methods for the Recreational Water Quality Indicators: Enterococci and Escherichia coli. Office of Science and Technology: Washington, DC, 2000; p 27.

(24) Wiggins, B. A.; Cash, P. W.; Creamer, W. S.; Dart, S. E.; Garcia, P. P.; Gerecke, T. M.; Han, J.; Henry, B. L.; Hoover, K. B.; Johnson, E. L.; Jones, K. C.; McCarthy, J. G.; McDonough, J. A.; Mercer, S. A.; Noto, M. J.; Park, H.; Phillips, M. S.; Purner, S. M.; Smith, B. M.; Stevens, E. N.; Varner, A. K. Use of antibiotic resistance analysis for representativeness testing of multiwatershed libraries. Appl. Environ. Microbiol. 2003, 69 (6), 3399-3405. (25) Wiggins, B. A.; Andrews, R. W.; Conway, R. A.; Corr, C. L.; Dobratz, E. J.; Dougherty, D. P.; Eppard, J. R.; Knupp, S. R.; Limjoco, M. C.; Mettenburg, J. M.; Rinehardt, J. M.; Sonsino, J.; Torrijos, R. L.; Zimmerman, M. E. Use of antibiotic resistance analysis to identify nonpoint sources of fecal pollution. Appl. Environ. Microbiol. 1999, 65 (8), 3483-3486. (26) Hagedorn, C.; Crozier, J. B.; Mentz, K. A.; Booth, A. M.; Graves, A. K.; Nelson, N. J.; Reneau, R. B., Jr. Carbon source utilization profiles as a method to identify sources of faecal pollution in water. J. Appl. Microbiol. 2003, 94 (5), 792-799. (27) Scott, T. M.; Parveen, S.; Portier, K. M.; Rose, J. B.; Tamplin, M. L.; Farrah, S. R.; Koo, A.; Lukasik, J. Geographical variation in ribotype profiles of Escherichia coli isolates from humans, swine, poultry, beef, and dairy cattle in Florida. Appl. Environ. Microbiol. 2003, 69 (2), 1089-1092. (28) Tenover, F. C.; Arbeit, R. D.; Goering, R. V. How to select and interpret molecular strain typing methods for epidemiological studies of bacterial infections: a review for healthcare epidemiologists. Infect. Control Hosp. Epidemiol. 1997, 18, 426-439. (29) Rademaker, J.; de Bruijn, F. J. Characterization and classification of microbes by REP-PCR genomic fingerprinting and computer assisted pattern analysis. In DNA Markers: Protocols, Applications, and Overviews; Caetano-Anolles, G., Gresshoff, P. M., Eds.; Wiley-Liss, Inc.: New York, 1997; pp 151-172. (30) Versalovic, J.; Schneider, M.; de Bruijn, F. J.; Lupski, J. R. Genomic fingerprinting of bacteria using repetitive sequence-based polymerase chain reaction. Methods Mol. Cell. Biol. 1994, 5, 25-40. (31) Versalovic, J.; de Bruijn, F. J.; Lupski, J. R. Repetitive sequencebased PCR (rep-PCR) DNA fingerprinting of bacterial genomes. In Bacterial Genomes: Physical Structure and Analysis; de Bruijn, F. J., Lupski, J. R., Weinstock, G. M., Eds.; Chapman and Hall: New York, 1997; pp 437-456. (32) Rosner, B. A. Fundamentals of Biostatistics; PWS-KENT Publishing Company: Belmont, CA, 1990; p 655. (33) Harwood, V. J.; Whitlock, J.; Withington, V. Classification of antibiotic resistance patterns of indicator bacteria by discriminant analysis: use in predicting the source of fecal contamination in subtropical waters. Appl. Environ. Microbiol. 2000, 66 (9), 3698-3704. (34) Sadowsky, M. Written communication, 2003. (35) Jenkins, M. B.; Hartel, P. G.; Olexa, T. J.; Stuedemann, J. A. Putative temporal variability of Escherichia coli ribotypes from yearling steers. J. Environ. Qual. 2003, 32, 305-309. (36) Hartel, P. G.; Summer, J. D.; Hill, J. L.; Collins, V. J.; Entry, J. A.; Segars, W. I. Geographic variability of Escherichia coli ribotypes from animals in Idaho and Georgia. J. Environ. Qual. 2002, 31, 1273-1278. (37) Yang, H. H.; Vonopal, R. T.; Grasso, D.; Smets, B. F. High diversity among environmental Escherichia coli isolates from a bovine feedlot. Appl. Environ. Microbiol. 2004, 70 (3), 1528-1536. (38) Institute for Environmental Health. http://www.iehinc.com/mst/home.htm. (39) Carson, C. A.; Shear, B.; Ellersieck, M. R.; Asfaw, A. Identification of fecal Escherichia coli from humans and animals by ribotyping. Appl. Environ. Microbiol. 2001, 67 (4), 1503-1507. (40) Gordon, D. M.; Bauer, S.; Johnson, J. R. The genetic structure of Escherichia coli populations in primary and secondary habitats. Microbiology 2002, 148, 1513-1522. (41) Souza, V.; Rocha, M.; Valera, A.; Eguiarte, L. E. Genetic structure of natural populations of Escherichia coli in wild hosts on different continents. Appl. Environ. Microbiol. 1999, 65 (8), 3373-3385. (42) Stewart, J. R.; Ellender, R. D.; Gooch, J. A.; Jiang, S.; Myoda, S. P.; Weisberg, S. B. Recommendations for microbial source tracking: lessons from a methods comparison study. J. Water Health 2003, 1 (4), 225-231.

Received for review December 23, 2003. Revised manuscript received August 23, 2004. Accepted September 2, 2004. ES0354519 VOL. 38, NO. 22, 2004 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

6117