Phenotypic and Phylogenetic Identification of Coliform Bacteria Obtained Using 12 Coliform Methods Approved by the U.S. Environmental Protection Agency

Abstract

The current definition of coliform bacteria is method dependent, and when different culture-based methods are used, discrepancies in results can occur and affect the accuracy of identification of true coliforms. This study used an alternative approach to the identification of true coliforms by combining the phenotypic traits of the coliform isolates and the phylogenetic affiliation of 16S rRNA gene sequences with the use of lacZ and uidA genes. A collection of 1,404 isolates detected by 12 U.S. Environmental Protection Agency-approved coliform-testing methods were characterized based on their phylogenetic affiliations and responses to their original isolation media and lauryl tryptose broth, m-Endo, and MI agar media. Isolates were phylogenetically classified into 32 true-coliform, or targeted Enterobacteriaceae (TE), groups and 14 noncoliform, or nontargeted Enterobacteriaceae (NTE), groups. It was shown statistically that detecting true-positive (TP) events is more challenging than detecting true-negative (TN) events. Furthermore, most false-negative (FN) events were associated with four TE groups (i.e., Serratia group I and the Providencia, Proteus, and Morganella groups) and most false-positive (FP) events with two NTE groups, the Aeromonas and Plesiomonas groups. In Escherichia coli testing, 18 out of 145 E. coli isolates identified by enzymatic methods were validated as FN. The reasons behind the FP and FN reactions could be explained through analysis of the lacZ and uidA genes. Overall, combining the analyses of the 16S rRNA, lacZ, and uidA genes with the growth responses of TE and NTE on culture-based media is an effective way to evaluate the performance of coliform detection methods.

INTRODUCTION

Total coliform bacteria and Escherichia coli are considered cost-effective bacterial indicators for the protection of public health. Detectable total coliforms indicate potential contamination associated with the water distribution system, while E. coli is a good indicator of fecal contamination, with a health goal (i.e., maximum contaminant level goal [MCLG]) of zero under the Revised Total Coliform Rule (RTCR) (1). When a positive result of total-coliform testing is observed, public water systems (PWSs) are required to collect and analyze three repeat samples (1). A system assessment is required when a PWS exceeds a specific frequency of total-coliform occurrence. When an E. coli maximum contaminant level (MCL) violation incurs, a PWS must have an assessment performed by the state or a state-approved entity and must correct any sanitary defects (1). False-positive (FP) results for either total-coliform or E. coli tests impose an unnecessary burden on water utilities. On the other hand, false-negative (FN) results expose consumers to potential health threats and delay response times for effective action. Therefore, accurate detection of total coliforms and E. coli in drinking water systems is crucial for both water utilities and consumers.

To date, the U.S. EPA has approved 12 methods for the detection of coliform bacteria in drinking water: the lauryl tryptose broth (LTB; including Presence-Absence broth), m-Endo (2), MI medium (3), Coliscan C MF (4), Colilert, Colilert-18 (5), Colisure, E*Colite, m-ColiBlue 24 (6), ChromoCult coliform agar, Readycult Coliforms 100 (7), and Colitag (8) methods. The last 10 methods approved are based on enzymatic reactions, which detect total coliforms and E. coli simultaneously according to the activities of β-galactosidase (GAL) (encoded by the lacZ gene) and β-glucuronidase (GUD) (encoded by the uidA gene). Enzyme-based tests have become popular among water utilities because no confirmation tests are required for analyses showing positive responses.

Various degrees of accuracy in identification have been observed for these U.S. EPA-approved methods, with reported FP rates for the detection of total coliforms and E. coli by enzyme-based tests ranging from 2.3 to 36.4% and from 0.9 to 37.5%, respectively. The rates for FN results are 3.3 to 26.2% and 20.3 to 48.6% for total coliforms and E. coli, respectively (9–15). The two likely reasons are an inability to recover target organisms from different drinking water matrices and an inability to distinguish target strains from nontarget organisms. The former is often related to the occurrence of FN results, while the latter is the major cause of FP results. These two reasons have been evaluated in several studies (9–18). For example, Olstadt et al. (13) compared 10 enzyme-based tests for the ability to recover different concentrations of spiked coliform and noncoliform strains from three diverse water matrices.

Nevertheless, no studies have been systematically conducted to test all the approved methods for the identification of coliform bacteria against a representative library of coliform isolates. The API 20E and Vitek 2 systems (both from bioMérieux Inc., France) are commonly used to verify colonies in total-coliform or E. coli tests, yet these bioMérieux tests have been reported to classify coliform bacteria (including Klebsiella spp., Enterobacter spp., and Salmonella spp.) inaccurately (19). Phenotypic trait changes under stress conditions in water systems can result in bacterial misidentification (20–22). Currently, the definition of coliform bacteria can differ according to the detection method used. Traditionally, total coliforms have been defined based on their ability to ferment lactose to acid and gas. With the introduction of enzyme-based tests, the definition includes bacteria within the family Enterobacteriaceae that possess the GAL enzyme and E. coli bacteria that exhibit both GAL and GUD activities (23). The broader definition by enzyme-based methods incorporates more species into the coliform group, which can also contribute to misidentification when these methods are compared with traditional methods.

Since the determination of FP and FN results depends on the definition used, this study proposes an alternative approach that groups coliform bacteria into cohesive clusters based on their phylogeny affiliations and then overlays individual clusters with their phenotypic responses on different U.S. EPA-approved media for drinking water testing. In this way, FP and FN results can be evaluated at a level equivalent to the genus or subgenus classification, rather than at the level of the total coliform group as a whole. The 16S rRNA gene is a well-established phylogenetic molecular marker and gives results more accurate than phenotype-based results for the identification of coliform bacteria isolated from drinking water systems. In addition, sequences of the lacZ and uidA genes, encoding the GAL and GUD enzymes, respectively, can provide additional resolution to support the identification of each total-coliform or E. coli cluster. These three key genes can be used simultaneously to define isolates from both phylogenic and functional perspectives and can serve as a robust approach for the evaluation of cultivation-based coliform methods.

To systematically test this approach, coliform isolates obtained by 12 different U.S. EPA-approved methods (approximately 100 isolates per method) were characterized by the three target genes and their responses to the original isolation medium and to LTB, m-Endo, and MI media. This enabled us to evaluate each method in terms of sensitivity, specificity, and FP and FN rates in a cluster-specific and definition-independent manner. The findings can be used to help water utilities and testing labs to better detect total coliforms and E. coli, allow the U.S. EPA to select a representative library of strains to test proposed new methods, and help manufacturers to improve medium formulations.

MATERIALS AND METHODS

Isolate collection and phenotypic testing.

Approximately 700 bacterial cultures were obtained from nine water utilities and testing laboratories across the United States, with detailed isolation information. All isolates were shipped to the laboratory at UIUC with ice packs. Once received, the cultures were streaked onto Trypticase soy agar (TSA; BD Diagnostic Systems, Franklin Lakes, NJ) to confirm viability and purity. These strains were isolated mainly by the m-Endo, MI, Colilert, Colilert-18, and Colisure methods. Among them were 121 isolates from finished water or distribution systems, the target sites of the RTCR.

The remaining bacterial strains were isolated from different source water samples shipped to the laboratory overnight from 11 water utilities. Water samples were diluted and were tested following the protocols provided by the manufacturers. Individual colonies were picked using membrane filtration-based methods after incubation. For liquid media, 50 and 200 μl of the liquid were spread onto TSA plates. Single colonies were picked from plates after 24 h of incubation at 35°C. All the isolates were reinoculated onto TSA or m-Endo agar three times to ensure purity. To eliminate human preference and potential interference, all methods except for the LTB, m-Endo, and MI methods were designated M-1 to M-9 during the entire experimental process.

Individual isolates were tested against their original isolation medium, LTB (gas production), m-Endo agar (metallic sheen generation), or MI medium (fluorescence and blue color change). At the same time, 1 to 5 g of wet biomass from each culture was individually obtained after centrifugation to produce cell pellets. The cell pellets were used for genomic DNA (gDNA) extraction. In addition, each bacterial isolate was preserved in a 15% (vol/vol) glycerol solution and was stored at −80°C.

gDNA extraction.

For the majority of the isolates, gDNA was obtained directly by heat shock methods (24). A tip of the biomass pellet was picked using sterile toothpicks, placed in 50 μl Tris-EDTA buffer, and subjected to heating at 95°C for 10 min. The suspension was then cooled to 4°C and was kept at −20°C until use. Alternatively, genomic DNA from the isolates was extracted according to the protocol described by Schmidt et al. (25).

PCR.

The 16S rRNA genes of individual bacterial cultures were PCR amplified using a primer set specific to the domain Bacteria, comprising 11F (5′-GTT TGA TCC TGG CTC AG-3′) and 1492R (5′-GGY TAC CTT GTT ACG ACT T-3′) (26). Each PCR mixture (volume, 25 μl) contained 20 to 30 ng of gDNA in 1× Bullseye Taq 2.0 master mix (Midwest Scientific, St. Louis, MO) and 25 nM forward and reverse primers. The reaction mixture was subjected to 95°C for 5 min, followed by 25 cycles of thermal amplification, consisting of denaturation (95°C for 30 s), annealing (55°C for 45 s), and extension (72°C for 60 s). Amplicons were purified before sequencing by using the Wizard SV Gel and PCR Clean-Up system (Promega, Madison, WI) or a MultiScreen PCR_μ96 plate (Millipore Corp., Bedford, MA).

To PCR amplify lacZ genes, a database containing lacZ gene sequences was used to design primers for lacZ genes from known cultures. The primer set designed for lacZ gene amplification consisted of lacZ3153F (5′-AAG ATC ARG AYA TGT GGC G-3′) and lacZ3995R (5′-CAT GCC GTG BGT YTC RAT-3′). lacZ gene sequences available from GenBank (http://www.ncbi.nlm.nih.gov/GenBank/) were retrieved (as of December 2010) and were input into ARB software (27) to form a database. Primers were designed by using the database established. To amplify uidA genes, a primer set consisting of UAL1228 (5′-ATG TTA CGT CCT GTA GAA AC-3′) and UAR3426 (5′-TTG TTT GCC TCC CTG CTG CG-3′) was used (28). PCR amplification for the lacZ and uidA genes was conducted similarly to PCR for the 16S rRNA gene.

Sequencing.

16S rRNA gene sequencing of amplicons was performed using a BigDye Terminator kit (Applied Biosystems, Foster City, CA) with ABI 3730XL capillary sequencers at the Keck Biotechnology Center, UIUC. Three primers, including 11F, 1492R, and 907R (5′-CCG YCA ATT CMT TTR AGT TT-3′) (29), were used to obtain nearly full length 16S rRNA gene sequences. Sequences were analyzed using Sequencher, version 4.9 (Gene Codes Corp., Ann Arbor, MI). Alignment was first performed using the Greengenes online tool (http://greengenes.lbl.gov/) (30), and the aligned sequences were then imported into ARB software (27) using the greengenes236469 database (released in November 2008). Reference sequences were selected according to the List of Prokaryotic names with Standing in Nomenclature (www.bacterio.net) (31). The phylogenetic affiliations of individual cultures were determined with a phylogenetic tree constructed with both reference sequences and isolate sequences. Closely related sequences were grouped into a cohesive group, first at the species level, with a known bacterium (e.g., Enterobacter aerogenes). If this was not possible, the group was formed at the level of a known genus, such as Escherichia. If the sequences could not form a cohesive cluster within a known genus, more than one subcluster was proposed within a genus with a sequence similarity higher than 97%, as used for “species” definition.

Representative lacZ gene amplicons and the uidA genes of all isolates identified within the Escherichia group by enzyme-based tests were sequenced. The lacZ and uidA genes were sequenced using the same chemistry as that described above for the 16S rRNA gene. The primers used for the lacZ gene were lacZ3153F and lacZ3995R, and those for the uidA gene were UAL1228, UAR3426, and UAR2447 (5′-CGA CCA AAG CCA GTA AAG TAG AA-3′). The sequences obtained were aligned using MUSCLE (32).

Prediction of correct monitoring results.

The conditional probability of the presence or absence of coliforms in drinking water systems given a positive or negative result by culture-based methods was estimated using Bayes' theorem (equations 1 and 2) based on the occurrence rates of total coliform bacteria and the sensitivity and specificity of individual coliform-testing methods (33). The occurrence rate is defined as the natural baseline level of total-coliform occurrence in water samples taken for Total Coliform Rule (TCR) compliance. Different combinations of sensitivity and specificity (i.e., 50%, 75%, 90%, 95%, 99%, and 99.5%) were used for the calculation of method effectiveness.

$$\text{P}(\text{H}|\text{POS})=\frac{\text{P}(\text{POS}|\text{H})\text{P}\left(\text{H}\right)}{\text{P}(\text{POS}|\text{H})\text{P}\left(\text{H}\right)+\text{P}(\text{POS}|\text{H}\prime )\text{P}(\text{H}\prime )}$$ (1)

$$\text{P}(\text{H}\prime |\text{NEG})=\frac{\text{P}(\text{NEG}|\text{H}\prime )\text{P}(\text{H}\prime )}{\text{P}(\text{NEG}|\text{H}\prime )\text{P}(\text{H}\prime )+\text{P}(\text{NEG}|\text{H})\text{P}\left(\text{H}\right)}$$ (2)

In these equations, H stands for the presence of total coliforms in a water sample; H′ stands for the absence of total coliforms in a water sample; POS stands for positive results by culture-based testing methods; NEG stands for negative results by culture-based testing methods; P(H) is the total-coliform occurrence rate; P(H′) is equal to 1 − P(H); P(H|POS) is the probability of a positive response by culture-based methods that is truly caused by total coliform bacteria; P(POS|H) is the probability of total coliform bacteria giving positive responses by culture-based methods and is equal to sensitivity; and P(H′|NEG) as the probability of a negative response by culture-based methods that is caused by the absence of total coliform bacteria in the water sample.

Nucleotide sequence accession numbers.

The assembled sequences of 16S rRNA, lacZ, and uidA genes have been deposited in GenBank under accession numbers KR189031 to KR190434, KR424088 to KR424292, and KR424293 to KR424443, respectively.

RESULTS

Composition of isolates detected by all coliform-testing methods.

In total, 1,404 bacterial isolates were isolated by use of LTB, m-Endo agar, and 10 different enzyme-based methods (see Fig. S1 in the supplemental material). The number of isolates ranged from 90 for the Colisure test to 204 for the Colilert test. Among all the isolates, 1,186 (84.5%) were identified as “coliforms” based on positive responses (i.e., color change or fluorescence production) to their original isolation media, and the remaining isolates (15.5%; n = 218) were identified as “noncoliforms.” The positive response rate differed among the 12 methods, ranging from 71.4% for ChromoCult coliform agar to 98.4% for MI medium. It was difficult to collect isolates exhibiting negative responses by the MI, Colilert, Colisure, Coliscan, or m-ColiBlue 24 method, possibly due to the inhibitory compounds present in the media.

Phylogenetic diversity of coliform isolates based on 16S rRNA gene sequences.

A phylogenetic tree was constructed with 16S rRNA gene sequences obtained in this study and with reference species (Fig. 1A; see also Table S1 in the supplemental material) (sequence length, mostly >1,250 bp). The term TE (targeted Enterobacteriaceae) was used to designate true coliform bacteria from a phylogenetic perspective. The phylogenetic classification of groups belonging to TE was based on that reported by Leclerc et al. (23). Leclerc et al. classified coliform bacteria within the Enterobacteriaceae by ortho-nitrophenyl-β-galactoside (ONPG) reaction, fecal origin, and occurrence in water. Genera that could occur in water with positive ONPG reactions or that were of fecal origin were thus considered TE in this study. The rest of the isolates, or noncoliform bacteria that could interfere with culture-based tests, were referred to as nontargeted Enterobacteriaceae (NTE). All the sequences were grouped into 32 TE clusters and 14 NTE clusters (Fig. 1A). Clearly, TE comprised a wide variety of bacterial groups within the family Enterobacteriaceae as a cohesive cluster and could be differentiated from NTE clusters, including the Plesiomonas cluster in the family Enterobacteriaceae. Within TE, two branches separated the common TE clusters from the uncommon clusters comprising Providencia, Proteus, and Morganella spp. The largest sequence variation between two given clusters within TE was 6.8%. This indicated a high sequence similarity of the16S rRNA gene among TE clusters and the ability of the 16S rRNA gene to effectively distinguish TE from NTE.

FIG 1.

(A) Phylogenetic tree based on 16S rRNA gene sequences obtained in this study and on reference sequences. The number of reference sequences is given in parentheses before the name of the genus or group. (B) The number of isolates belonging to each group is given to the right of each bar.

Figure 1B reveals the frequency of occurrence of each cluster with respect to all coliform-testing methods. The most abundant TE cluster was Enterobacter group VI, followed by the Escherichia cluster, Citrobacter group II, Klebsiella group I, Serratia group II, Enterobacter group IX, and Enterobacter group X. Nine clusters were observed to contain no more than seven isolates each (0.5% of total isolates), suggesting that these clusters were rare TE groups during coliform testing. They were the Enterobacter group I, Citrobacter group I, Trabulsiella, Pantoea, Erwinia, Enterobacter group III, Enterobacter group IV, Enterobacter aerogenes, and Yersinia clusters. Among the 14 NTE groups, Aeromonas, Pseudomonas, Plesiomonas, and Acinetobacter clusters were dominant. All the other NTE clusters were less frequently detected.

Sequences of lacZ and uidA genes as additional target genes for TE and E. coli.

Sequences of the lacZ and uidA genes, which encode GAL and GUD, respectively, in enzyme-based testing can provide further information on the identity of TE and E. coli. lacZ sequences could be clustered into 12 major groups, with sequence dissimilarity ranging from 13.1% to 36.7% (Fig. 2A). However, the clustering topology is different from that observed with the 16S rRNA gene sequences. This inconsistency suggests that the lacZ gene is likely not conserved and that its sequence could be influenced by a combination of horizontal gene transfer among the species and loss of common ancestor genes (34, 35). Among the 205 sequences obtained, 42.0% were related to Enterobacter/Pantoea lacZ genes (86 sequences), followed by the Escherichia/Shigella/Klebsiella lacZ-like group (20.1%; 43 sequences), and the Escherichia/Klebsiella/Enterobacter lacZ-like group (12.7%; 26 sequences) (see Table S2 in the supplemental material).

FIG 2.

Neighbor-joining trees of lacZ (A) and uidA (B) gene sequences obtained in this study. In panel A, for each cluster, the number of reference sequences belonging to the cluster is shown, followed by a plus sign and the number of sequences obtained in this study that belong to the cluster.

Figure 2A further indicates that the lacZ gene alone is not sufficient to distinguish TE from NTE. This can be exemplified by the high lacZ gene similarities between two Plesiomonas isolates (NTE) and one Klebsiella group I isolate (TE) in isolate group II. Furthermore, the low sequence similarity among lacZ genes made it difficult to design an effective primer set for the entire TE group. As a result, isolates from which lacZ gene amplicons could not be obtained can still be TE, as observed with Serratia group I and with the Providencia, Proteus, and Morganella groups (see Table S3 in the supplemental material). Therefore, methods based on lacZ gene amplification have limitations in accurately identifying TE and should be used only together with conserved marker genes, such as the 16S rRNA gene.

Figure 2B reveals that the uidA gene sequences are highly similar among all the isolates in the Escherichia cluster, except for three isolates that gave no uidA PCR amplicons. Among the three isolates, one (N391_S5) is related to Escherichia hermannii, as indicated by its 16S rRNA and lacZ gene sequences. The reason for the failure of the other two isolates to yield PCR amplicons is not clear. One reason could be sequence variations not targeted by the primer set used. The remaining sequences obtained could be divided into three groups. The sequence homology difference within individual groups ranged from 2.8% for uidA groups I and II to 8.1% for uidA group III, indicating that sequences within group III had large gene variations. Within group I, four E. coli O157:H7 uidA gene sequences from the public database are included. It is hard to distinguish E. coli from Shigella sonnei on the basis of uidA gene sequences, since they share almost identical sequences (97.5 to 99.6%).

Comparing the phenotypic traits of isolates with their genotypic traits.

The performance of coliform-testing methods was evaluated by combining the response of each individual isolate on or in its original isolation medium with its phylogenetic identity. The results were evaluated in terms of sensitivity, specificity, FP rate, and FN rate (Table 1). Sensitivities were above 85.0% for all 12 methods except LTB (71.8%). Due to the limited number of NTE isolates obtained, specificity and FP rates were calculated for only four methods. With the available data, the m-Endo and E*Colite methods had specificities equal to or higher than 90.0%, and ChromoCult coliform agar had the lowest specificity. Three methods had higher calculated FP rates than FN rates, and m-Endo and E*Colite media tended to perform better for TE detection than the rest of the methods (sensitivity, >85.0%; specificity, >90.0%). Furthermore, the Colitag, m-ColiBlue 24, Coliscan, and MI media also exhibited high sensitivities (>95.0%). However, their specificities could not be properly evaluated, because few or no NTE isolates were obtained.

TABLE 1.

Performance of 12 coliform detection methods

Method	No. of isolatesa that were:				Sensitivity (%)b	Specificity (%)c	FP rate (%)d	FN rate (%)e
	TP	FP	FN	TN
LTB	74	2	29	0	71.8	N/Af	N/A	28.2
E*Colite	71	2	11	18	86.6	90	10	13.4
ChromoCult coliform agar	21	44	3	23	N/A	34.3	65.7	N/A
Readycult	96	0	12	16	88.9	N/A	N/A	11.1
Colisure	80	2	6	2	93	N/A	N/A	7
m-Endo	95	3	7	30	93.1	90.9	9.1	6.9
Colilert	181	9	12	2	93.8	N/A	N/A	6.2
Colilert-18	81	4	5	18	94.2	81.8	18.2	5.8
Colitag	78	4	3	9	96.3	N/A	N/A	3.7
m-ColiBlue 24	108	0	4	0	96.4	N/A	N/A	3.6
Coliscan	102	9	1	5	99	N/A	N/A	1
MI	113	7	1	1	99.1	N/A	N/A	0.9

^aClassified on the basis of responses in original isolation media. TP (true positive), TE showed positive responses; FP (false positive), NTE showed positive responses; TN (true negative), NTE showed negative responses; FN (false negative), TE showed negative responses.

^bSensitivity indicates the ability of a method to correctly identify TE and is calculated as (number of TP results)/(number of TP + FN results).

^cSpecificity refers to the ability of a method to correctly identify NTE and is calculated as (number of TN results)/(number of FP + TN results).

^dThe FP rate is calculated as (number of FP results)/(number of FP + TN results) and is equal to 1 − specificity.

^eThe FN rate is calculated as (number of FN results)/(number of TP + FN results) and is equal to 1 − sensitivity.

^fN/A, not applicable. If there were not enough isolates for the calculation of sensitivity, specificity, the FP rate, or the FN rate (i.e., if the sum of FP and TN results was <20 or the sum of TP and FN results was <80), the results are not shown in the table.

The phylogenetic clustering was further used to evaluate whether culture-based methods could achieve consistent accuracy across all the TE and NTE subgroups (Fig. 3). Among all the TE clusters, five FN results were identified within the Providencia, Serratia group I, Enterobacter group VI, Morganella, and Proteus clusters, suggesting a low overall FN rate. Within the NTE groups, Aeromonas and Plesiomonas isolates tended to give more FP results. Aeromonas isolates could cause FP responses on ChromoCult, Coliscan, MI, and m-Endo media, and Plesiomonas isolates on Colilert and Colilert-18 media. Most of these observations were supported by the lacZ and uidA gene analysis, predicting that Serratia group I, Providencia, Proteus, and Morganella isolates could give FN results, whereas Plesiomonas isolates were likely responsible for FP results. In addition, the number of TE groups detected by a given method could range from 7 (ChromoCult coliform agar) to 23 (Colilert), suggesting a wide range of variation among all coliform-testing methods in recovering TE or detecting their presence.

FIG 3.

Numbers of isolates showing positive (POS) and negative (NEG) responses in original isolation media within each TE and NTE group.

Effectiveness of enzyme-based E. coli detection methods.

A total of 145 isolates detected by 10 different enzyme-based methods that were phylogenetically clustered within the Escherichia group were compared based on the responses in or on their original media and MI medium and the presence or absence of uidA genes (Fig. 4). One hundred fifteen (79.3%) isolates were identified as E. coli using the latter three methods. Eighteen strains (12.4%) were identified as giving FN results in or on their original isolation media. Most of them occurred with MI medium as the original isolation medium (seven isolates), and none with the E*Colite, Colitag, or Colisure method. The discrepancies in response between the original isolation media and MI medium were further observed with 14 isolates detected by eight of the nine enzyme-based methods.

FIG 4.

Evaluation of 10 enzyme-based methods for the detection of E. coli. Each cell represents one culture isolated by a certain method that was identified by its 16S rRNA genes as belonging to the Escherichia group (e.g., 11 cultures originally isolated on MI medium; 26 cultures originally isolated by the Colilert method). The response of each isolate in or on the original isolation medium was compared with the response on MI medium and the presence or absence of uidA genes.

Of the nine results FP for E. coli, four belonged to the Yersinia cluster (two obtained using the Colisure test, one using the Colilert-18 test, and one using the Colitag test), and two belonged to Enterobacter group VI (one each obtained by using the Colilert-18 and Readycult tests). Of the remaining three, obtained using Colisure medium, two were in Citrobacter group II and one in Raoultella group I. FP GUD activity was common with Yersinia strains (36) but rare with Enterobacter and Citrobacter strains (37). Among all the media, the Colisure medium tended to have the most results FP for E. coli (five isolates).

Accuracy of detection of isolates from drinking water distribution systems.

A total of 121 isolates from drinking water distribution systems, which are the target sites of the RTCR, were further analyzed (see Fig. S3 in the supplemental material). Isolates from the distribution systems were affiliated with 16 of the 32 TE clusters. The most abundant group was Enterobacter group VI (59 isolates), followed by Citrobacter group II (12 isolates) and Enterobacter group II (9 isolates). The isolates from the distribution systems accounted for 18.2%, 11.8%, and 81.8% of all the isolates within each of these groups, respectively. Based on the frequency of occurrence, only Trabulsiella isolates (n = 6) appeared to be specific to drinking water distribution systems. However, Trabulsiella spp. are usually reported to be present in human stool, vacuum cleaner contents, and termite guts and are not specific to drinking water distribution systems (38, 39).

Fewer FP and FN events were observed with isolates from the distribution systems than with all the isolates tested (Table 2). Isolates from the distribution systems were recovered using the Colilert, m-Endo, Colilert-18, Coliscan, and LTB methods. Among them, two FP results and one FN result were detected. One of two FP events was related to TE identification using the Colilert test and was caused by an Aeromonas sp. The other FP event was related to E. coli identification using the Colilert-18 test and was due to an isolate from Enterobacter group VI. The only FN event occurred with TE detection using LTB.

TABLE 2.

Responses of isolates from distribution systems

Isolation medium	No. of isolates	FP result and reason	FN result and reason
Colilert	94 TE (including 2 E. coli isolates), 1 NTE	TE FP, caused by an Aeromonas isolate	None
Colilert-18	9 TE	E. coli FP, caused by an Enterobacter group VI isolate	None
Coliscan	4 TE	None	None
LTB	1 TE	None	TE FN, caused by an Enterobacter group V isolate
m-Endo	9 TE, 3 NTE	None	None

Comparison of the 12 methods against LTB, m-Endo, and MI media.

The effectiveness of the 12 methods was evaluated using three culture-based methods (i.e., LTB, m-Endo, and MI media). Isolates were classified into those that gave TP, FP, FN, and TN results with their original isolation media (see Fig. S3 in the supplemental material). In the TP category (see Fig. S3A in the supplemental material), TE isolates responded mostly positively to MI and m-Endo media but showed a higher rate of negative responses to LTB. This was consistent with the detection principle behind those culture-based methods: m-Endo and MI media targeted the intermediate products of lactose fermentation, and LTB detected the end products. In the TN category (see Fig. S3B), isolates detected by all the methods also showed negative responses to LTB, m-Endo, and MI media, except for a few isolates detected by the Colilert-18, Colisure, Readycult, and E*Colite tests. The greatest number of disagreements between the original isolation media and the three testing methods occurred for isolates in the FP and FN categories (see Fig. S3C and D). MI medium gave high rates of positive results within the FP category while producing the lowest number of negative results within the FN category. Isolates that were identified as giving FN results with their original media tended to give negative responses with LTB and m-Endo agar. In summary, isolates causing FP and FN results in their original isolation media were likely to be FP and FN in other media.

DISCUSSION

Diversity of TE isolates.

The term “coliform” has been used to describe lactose-fermenting Enterobacteriaceae since the 20th century. Traditional biochemical tests for TE are based on phenotypic information, including (i) pH-based reactions that require 15 to 24 h of incubation, (ii) enzyme-based reactions that take 2 to 4 h, (iii) utilization of carbon sources, such as indole production from tryptophan and fermentation of glucose, lactose, and sucrose (40), and (iv) visual detection of bacterial growth (41). Based on these principles, commercial kits such as the API 20E and Vitek 2 systems were developed. However, phenotypic traits can often be influenced by growth and stress conditions. As a result, biochemical tests can sometimes fail to detect important members of TE, such as Klebsiella spp. and Enterobacter spp. (19). For example, the accuracy of the API 20E kit at identifying Enterobacteriaceae is approximately 80 to 90% for cultures obtained from clinical laboratories (41).

Our findings and recent studies support the notion that 16S rRNA genes can serve as a more-robust approach than traditional biochemical tests for identifying TE to the genus or species level (42, 43). By use of 16S rRNA gene-based phylogeny, TE could be successfully classified into at least 32 clusters and could be differentiated from NTE clusters (Fig. 1A). Although the lacZ gene has often been used in PCRs to detect the presence of TE, the sequence homology of lacZ genes among TE was not well characterized (34, 44–47). By combining the findings obtained from lacZ gene sequences with those from the 16S rRNA gene, we can effectively predict clusters in Fig. 1A that cause FN and FP results in TE detection. TE clusters that caused FN results include Providencia, Morganella, Proteus, and Serratia group I clusters. The first three clusters are lacZ negative based on the currently available genome sequences, and the last, which is closely related to Serratia marcescens, possesses a lacZ gene with low sequence homology to those of known TE (72% with E. coli lacZ). NTE causing FP events included Plesiomonas, Aeromonas, and Vibrio isolates, which can produce the GUD enzyme but have lacZ gene sequences distantly related to that of E. coli (∼55% sequence similarity).

The representativeness of TE obtained in this study was compared with that in past studies (9, 10, 12, 18, 48), where 16S rRNA gene sequences (by BLAST searches) or biochemical reactions observed with the Vitek 2 or API 20E system were used for bacterial identification and taxonomy (Fig. 5). In each of the past studies, 60 to 270 bacterial strains were classified into 14 to 33 bacterial species based on the investigators' classification methods, and the results corresponded to 12 to 16 bacterial clusters as defined in this study. The most commonly observed clusters were the Escherichia, Enterobacter group VI, Enterobacter group II, Citrobacter group II, Enterobacter aerogenes, Klebsiella group I, Serratia group II, and Aeromonas clusters. The results were similar to our findings except for Enterobacter group II and Enterobacter aerogenes. The differences are possibly due to a lower number of bacterial isolates evaluated in previous studies, resulting in failure to detect these two groups with lower occurrence rates. Alternatively, the application of different bacterial identification methods in the previous studies may have resulted in the differences.

FIG 5.

Distributions of isolated coliform bacteria found in this study and in previous studies by Kämpfer et al. (18), Covert et al. (9), Freier and Hartman (48), Brenner et al. (10), and Bernasconi et al. (12). Isolates were clustered using the groups defined in this study. Each circle represents the relative abundance of a particular group found in a study.

Bacterial isolates responsible for FP and FN results in the detection of TE and E. coli.

In TE detection, Aeromonas and Plesiomonas isolates are the main causes of FP results observed here and in previous studies (9, 12, 13, 18, 49–52). In addition, a number of NTE bacteria could interfere with TE detection. Within the class Gammaproteobacteria, they included Vibrio vulnificus (53), Vibrio cholerae (50, 54), Pseudomonas pickettii, Pseudomonas vesicularis, Pseudomonas putida, Pseudomonas maltophilia (55, 56), and Pseudomonas aeruginosa (49). Within the class Alphaproteobacteria, they included Agrobacterium radiobacter, Sphingomonas paucimobilis (56), and Ochrobactrum anthropi (18). A few strains from the Firmicutes and Bacteroidetes were reported as well, including Aerococcus spp. (56), Streptococcus spp. (55), Sphingobacterium spp. (56), and Flavobacterium spp. (9). Some of these species were also detected in this study (Fig. 1B). However, they generally cause far fewer problems than Aeromonas and Plesiomonas spp., on which future improvement in coliform testing media should be focused.

In E. coli detection, positive responses are reported based on GUD activity. However, this enzyme activity has been detected with non-E. coli species, including members of the Enterobacteriaceae (i.e., Escherichia vulneris [28, 57] and Shigella [40 to 67%] [28, 58, 59], Salmonella [17 to 29%] [58, 60], Yersinia [61–63], and Citrobacter, Enterobacter, Edwardsiella, and Hafnia [37, 61, 62, 64, 65] spp.) and bacteria outside the family Enterobacteriaceae (i.e., Clostridium spp., Streptococcus spp., Staphylococcus spp. [66], Bacteroides spp. [67], Aerococcus spp., Bacillus spp. [56], and Corynebacterium spp. [68]). In this study, Yersinia isolates are the main cause of FP results in E. coli detection. Considering that some Yersinia species are human pathogens, this may not be an unfavorable outcome.

This study observed more FN events than FP events during E. coli testing. It is possible that some of the FN results are caused by Shigella spp. (i.e., S. sonnei and Shigella boydii), which are hard to distinguish from E. coli by using 16S rRNA and uidA gene sequences due to high sequence homology (47, 69). The FN result can also be caused by GUD-negative but uidA-positive E. coli strains. The best-known example is E. coli O157:H7, a pathogen of fecal origin and of great public health concern (70, 71), which does not generate a color or fluorescence change by culture-based coliform methods (72). Thus, more efforts are needed to validate those E. coli detection methods, possibly by testing them against the same E. coli library.

Effectiveness of current total coliform-testing methods.

A wide range of FP and FN rates have been reported (9–15, 48) (Table 3). In TE detection, LTB (or similar methods) generally gave much higher FP rates (>22%) than enzyme-based media (∼10%). However, high FP rates were also reported with the Colilert-18 test (36.4%) and MI medium (37.5%). The highest FN rates for TE and E. coli occurred with MI medium (26.2%) and the Colilert test (48.6%), respectively. In E. coli detection, the reported FN rates (∼20%) were much higher than the overall FP rates (∼10%). In contrast, interpreting the Colisure, Coliscan, and Colitag test as methods with high performance requires caution due to the limited data available.

TABLE 3.

Percentages of FP and FN results by U.S. EPA-approved methods

Medium	% FN results		% FP results		Reference
	Total coliforms	E. coli	Total coliforms	E. coli
LTB	N/Aa		73.0	68.0b	12
			21.9	18.3b	15
m-Endo (Endo LES) agar	N/A		15.4		15
			32.5		38
Colilert	15.9	48.6	3.4	0.0	14
	0.0				13
	20.5		8.0		9
			4.9	0.9	15
Colilert-18	3.3				13
			15.0	13.0	12
		11.0	36.4	10.3	11
			2.3	0.0	15
Colisure	0.0				13
			8.8	18.7	15
m-ColiBlue 24	23.0				13
			27.0	6.0	12
Readycult	15.9	18.9	3.4	12.5	14
	20.0				13
ChromoCult coliform agar	15.0	20.3	3.4	12.5	14
			6.9	5.6	15
Coliscan	0.0				13
			4.9	8.1	15
E*Colite	20.0				13
MI agar	4.3		6.9	4.3	10
	26.2	20.3	3.4	37.5	14
	0.0				13
			12.1	3.2	15
Colitag	0.0				13

^aN/A, not applicable.

^bTested on Lactose TTC agar with Tergitol 7, a medium similar to LTB.

Our study (Table 1) suggested that several coliform-testing methods could be of concern to water utilities. In TE testing, ChromoCult coliform agar and LTB seemed to give high FP and FN results, respectively. In E. coli detection, the Colisure test ended to give more FP results, whereas MI medium and the Colilert test could underestimate the occurrence of E. coli. Overall, methods that have high sensitivity in TE detection and few FN and FP results in E. coli testing can be good detection methods. The Readycult, Colilert-18, Colitag, m-ColiBlue 24, and Coliscan tests appeared to meet the criteria.

Nevertheless, it is critical to have a consistent set of parameters with which to evaluate how effectively coliform-testing methods could detect total coliform bacteria present in drinking water systems. The ultimate goal is to reduce the health risks associated with the occurrence of total coliforms and/or E. coli in drinking water systems. Therefore, risk assessment based on Bayes' probabilistic model is used to determine the cutoff of the parameters. The confidence level or probability of correct detections was estimated based on the natural total coliform occurrence rate and the 5% compliance level. When a negative response is obtained by coliform-testing methods, the probability (expressed as a percentage) of a TN result [P(H′|NEG], or the true absence of total coliform bacteria in a water sample, is at least 97.3% based on all the combinations of sensitivities and specificities listed in Table 4. Table 4 further suggests that the probability (expressed as a percentage) of a TP result [P(H|POS)], or a positive response obtained by a culture-based method that is truly caused by total coliform bacteria, is greatly influenced by the specificity levels. However, even with a sensitivity of 90% and a specificity of 99%, only 42.1% of the positive results are expected to be TP events under the natural coliform occurrence rate. This observation is crucial for method evaluation and improvement. While the occurrence of FN results could pose a concern to public health, FP events are more problematic, because coliform bacteria rarely occur in properly treated water, and the probability of encountering FP results is much higher than the probability of encountering FN results. To get a ≥90% probability that a 5% TP occurrence is really due to TE in drinking water systems, a method achieving at least 95% sensitivity and 99.5% specificity is needed.

TABLE 4.

Probability of TP as P(H|POS) and TN as P (H′|NEG) for various combinations of method sensitivity and specificity

Sensitivity (%)	Specificity (%)	Probability (%) of:
		True-negative result [P(H′|NEG)] based on:		True-positive result [P(H|POS)] based on:
		Total-coliform occurrence rate (0.8%)a	Compliance level (5.0%)a	Total-coliform occurrence rate (0.8%)	Compliance level (5.0%)
50	95	99.6	97.3	7.5	34.5
75	95	99.8	98.6	10.8	44.1
90	95	99.9	99.4	12.7	48.6
50	99	99.6	97.4	28.7	72.5
75	99	99.8	98.7	37.7	79.8
90	99	99.9	99.5	42.1	82.6
95	50	99.9	99.5	1.5	9.1
95	75	99.9	99.7	3	16.7
95	90	100	99.7	7.1	33.3
95	99	100	99.7	43.4	83.3
95	99.5	100	99.7	60.5	90.9
99	50	100	99.9	1.6	9.4
99	75	100	99.9	3.1	17.2
99	90	100	99.9	7.4	34.3
99	99	100	100	44.4	83.9
99	99.5	100	100	61.5	91.2

^aThe natural baseline level of total-coliform occurrence in a TCR water sample, P(H), is 0.8% according to U.S. EPA data for community water systems (73), and 5.0% is the compliance level of the TCR.

Based on the analyses presented above, a set of criteria can be used—for example, by the U.S. EPA—to screen and select methods with good performance. To illustrate this concept, the following criteria are used: (i) calculate the sensitivity and specificity of a method, obtained using approximately 80 to 100 isolates, of which 80% comprise TE; (ii) approve methods using a specificity of ≥99.5% and a sensitivity of ≥95%, which correspond to a P(H|POS) of 90.9% and a P(H′|NEG) of 99.7% under the 5% compliance level of the TCR (Table 4), respectively. If one applied these criteria to the 12 EPA-approved methods tested in this study, none of the methods could meet these criteria. Although high specificity and sensitivity were observed with the E*Colite and m-Endo methods, they are lower than the criteria designed. No conclusive answer could be given for the remaining methods, because insufficient NTE isolates were collected to enable determination of their specificity. To overcome this problem, a library of 100 isolates consisting of approximately 80 TE and 20 NTE could be selected from this study for testing against the remaining methods.

Likewise, a reasonable number of Escherichia isolates that give TP, FP, and FN responses can be selected from this study and used to test the 10 enzymatic methods for their abilities to identify E. coli correctly. In this study, 145 E. coli strains out of 1,164 isolates were identified based on the 16S rRNA and uidA genes. Among them, 127 are classified as yielding TP results, 9 as yielding FP results, and 18 as yielding FN results. Assuming that there is no significant difference in results among all E. coli testing methods, the sensitivity, specificity, FP rate, and FN rate could be calculated as 87.6%, 99.1%, 0.88%, and 12.4%, respectively. These observations suggest that the current methods perform better in E. coli testing than in total coliform detection.

In summary, total coliforms and E. coli have been used as bacterial indicators for more than a century, but the specificity of identification is still not satisfactory. Advances in sequencing technology provide a new perspective on the concept of total coliforms, their phylogeny, and functionalities. The three target genes (i.e., 16S rRNA, lacZ, and uidA) utilized in this study enable cluster-wise evaluation of culture-based coliform detection methods. They could not only identify each isolate but also provide a genetic-level explanation of the occurrence of FN and FP results. A representative library of isolates can be selected from the databases established in this study and used to test the effectiveness of both existing and future coliform detection methods. With regard to the currently approved methods, manufacturers can improve their medium formulations to enhance specificity, thus increasing the accuracy of all the total coliform and E. coli testing methods.