1. Introduction

BST Methods
When reviewing the published literature on BST to date, it quickly becomes apparent that there is one non-molecular method (antibiotic resistance analysis, or ARA) and three molecular methods (ribotyping or RT, pulsed-field gel electrophoresis or PFGE, and polymerase chain reaction or PCR) to choose from. While procedures for RT and PFGE are relatively similar in studies that have used them, there are several substantially different variations in reported PCR methods. Also, other non-molecular methods such as carbon source utilization (BIOLOG System) and cell wall analysis of fatty acid methyl esters (FAME, Sherlock-MIDI System) should be available in the near future.

2. Classification of Known and Unknown Source Isolates

Antibiotic Resistance Analysis
All reports to date using ARA have employed discriminant analysis (DA) to obtain average rates of correct classification (ARCC), and ARCC have been reported in the range of 50% to 90% or higher. Harwood et al., (2000) reported 34% to 88% ARCC with ARA on 4,619 enterococcal isolates, and 50% to 95% ARCC with ARA on 6,144 fecal coliform isolates. Their known-source isolate collection was from a large geographical area in Florida and the lower ARCC for some of the sources reflected that geographic diversity. Wiggins et al., (1999) reported ARCC from 54% to 91% with ARA on 3,032 enterococcal isolates, and demonstrated that the ARCC could be increased substantially by using a larger number of antibiotics and concentrations. Bower (2000) used ARA on 830 enterococcal isolates from a coastal watershed in Oregon and obtained correct classification rates from 73% for human isolates to 89% for dairy cattle isolates. Graves (2000) reported 87% to 94% ARCC on 2,012 enterococcal isolates in a small Virginia watershed (5,800 ha) and found the majority were from livestock and wildlife, with a smaller signature that was human in origin. Bowman et al., (2000) performed ARA on 1,880 enterococcal isolates from a large watershed in Virginia (72,000 ha) and found that both humans (2.1% to 56.2% of isolates over sites and months) and wildlife (4.2% to 70.8% of isolates over sites and months) contributed to fecal pollution in addition to the suspected source, livestock.

Molecular Methods
Parveen et al., (1999) ribotyped 238 E. coli isolates and reported an 82% ARCC (using DA) when the isolates were classified between human and nonhuman categories. Hartel et al., (1999) ribotyped 119 E. coli isolates using a RiboPrinterTM but could not differentiate between isolates from three sources (two streams and cow manure). Samadpour and Chechowitz (1995) ribotyped 589 E. coli stream isolates in a 29 month watershed study and were able to match ribotype patterns (against those in their library) for 71% of the isolates, but did not disclose how ribotyping was performed or how the data was analyzed. Dombeck et al., (2000) used a matching band algorithm and reported similarity coefficients of 78% to 100% on E. coli using PCR and repetitive DNA sequences identified with custom primers (154 total isolates, average of 22 per known source). Simmons et al., (2000) used chi-square analysis of PFGE band patterns to match 51% of 439 E. coli isolates from a stream in an urban watershed, and classified the majority of isolates as being from wildlife (especially raccoons) and dogs. Carson et al., (2001) ribotyped 287 E. coli isolates from eight known sources and reported a range of correct classification rates from 48.7% for horses to 95.7% for chickens. The ARCC for all isolates was 73.6%. The ARCC for separating human isolates from nonhuman isolates was 97.1%, and the ARCC for goose vs. turkey vs. chicken was 95.9%.

3. Methodology Comparisons

Numbers of Isolates
Source classification on small numbers of isolates is currently one of the shortcomings of molecular methods, as compared to ARA. This may change in the future. The PCR method of Dombeck et al., (2000) may be very suitable for assaying larger numbers of isolates, using a molecular technique, than has been previously reported (M. Sadowsky, personal communication). With ARA, technicians and/or students can be quickly taught to perform the procedure on several hundred isolates per week. In polluted waters that yield thousands of fecal coliforms or enterococci per sample, some method is needed that best allows source determinations on a representative subset of the fecal population, whatever that subset might be (Hagedorn, et al., 1999). When statistical procedures are used to determine sampling size, such procedures usually indicate that 5 to 10% of the sampled population needs to be assayed. To date, ARA appears to be the best method available for rapid source identification on the large numbers of isolates that are needed to obtain a statistically valid sample size. Samadpour and Chechowitz (1995) performed ribotyping on an average of 16 E. coli isolates per sample, where the fecal coliform populations averaged several hundred to several thousand colony forming units (CFU)/100 ml per sample. With 16 isolates per sample, representing just 0.6% to 2.8% of the sampled population at heavily contaminated sites, the results can only reflect those sources that are predominant in the sample.

Variability of BST Methods
For anyone considering using bacterial source tracking methodology, one goal should be to combine non-molecular methods (such as ARA) with molecular methods to cross-validate both approaches and to assess where one method might be more suitable than the other. Some investigators have suggested that molecular methods are more accurate than non-molecular techniques, and ARA has been criticized as being too variable a characteristic for reliable source identification. Published reports to date have not yet established that molecular methods are more reliable or accurate than ARA as a fecal sourcing methodology. To assess method variability, ARCC need to be determined on isolates from the same region over some substantial period of time. In a two-year study using ARA in the Page Brook watershed in Virginia, there were no substantial reductions in ARCC for any of the known sources that were included in the library developed for that watershed when comparing isolates collected both at the start of the project and those collected from the same locations over one year later (Hagedorn et al., 1999). While molecular methods may be more accurate in correctly classifying the specific type of animal (e.g. cows, sheep, deer, waterfowl, etc.), such specific identifications may not always be the best approach, or even needed (see following section).

4. BST and TMDL Projects

Our approach of classifying isolates based on human vs. wildlife vs. livestock has been very useful to regulatory officials in Virginia where ARA has been used in seven TMDL watershed projects to date. Harwood et al., (2000) reported that regulatory officials in Florida were also satisfied with ARA results that could determine if a human signature was present and then divide animal sources between livestock and wildlife. Samadpour and Checkowitz (1995) reported that lack of landowner cooperation was a serious obstacle to obtaining access to property for known source sample collection. The 3-way classification used in our studies has proven to be a non-confrontational approach that has been readily accepted by the public in the participation component of the TMDL process (McClellan et al., 2000). Landowner cooperation was obtained for every farm and property in the Spout Run watershed where access was desired (Graves, 2000), and the same level of cooperation was achieved earlier for the Page Brook study (Hagedorn et al., 1999). This approach does not "point fingers" at any individual property owner and is an important consideration as landowner cooperation and participation in TMDL projects is largely voluntary. Also, the 3-way classification (dogs or pets could be added as a fourth category for more urban watersheds, and livestock removed if necessary) allows source classification to be used in the modeling component of the TMDL process where load reduction allocations could be assigned to sources in a watershed based on the proportionality of source classification results. As well as using BST methods to confirm the presence or absence of sources in a watershed, source tracking results could also be used to adjust loads from different sources in traditionally calibrated models and, once sufficient BST data is available, then water quality models could be entirely calibrated based on the contributions of fecal coliforms from individual sources (McClellan et al., 2000).

5. What is Needed?

Most ARA studies to date have focused on the enterococci while virtually all of the molecular methods have used E. coli or fecal coliforms. There is a need for both indicators as many coastal states use an enterococcus standard for marine waters, while fecal coliforms are more widely used for freshwaters. However, if regrowth of E. coli is substantial, especially in more sub-tropical environments, then the enterococci may be the preferred indicator for freshwaters as well. There may be situations where neither traditional indicator is appropriate, and the Bacteroides-Prevotella group used in BST by Bernhard and Field (2000a and b) should be considered as well. BST needs are (for each indicator):

1. Comparative method studies where a collection of known source isolates is used by different investigators (BST method of their choice) to build libraries.
2. Data analysis of libraries where the statistical methodologies allow suitable comparisons of correct classification rates between methods to be made.
3. Studies on shared stream samples to determine how similar the identified sources and proportionality of those sources are for different BST methods.
4. Geographic evaluation to determine if there are limitations within any of the BST methods. Developing the known source isolate collection in #1 (above) from different regions should be appropriate to provide an answer.
5. Stability evaluations where known source isolates are collected over some multi-year time frame and added to the libraries in #1 (above) to determine the impact of time on results obtained with the different BST methods.