Characterizing the concentration of Cryptosporidium in Australian surface waters for setting health-based targets for drinking water treatment

In line with the World Health Organization (WHO) Guidelines for Drinking-Water Quality (GDWQ) 2011 (WHO 2011) and the Australian Guidelines for Water Recycling (AGWR) (NWQMS 2006) the next revision of the Australian drinking water guidelines will include ‘health-based targets’; that is, quantitative treatment targets for pathogens based on quantitative microbial risk assessment (QMRA). Specifically it is proposed that the required level of potable water treatment will be quantitatively related to the magnitude of the pathogen concentration in each source water.

One limiting factor for surface water supply treatment is expected to be protozoan contamination, in particular the presence of the ‘reference’ protozoan pathogen, Cryptosporidium due to its extreme resistance to inactivation by chlorine (LeChevallier & Au 2004). Cryptosporidium has been identified in a range of hosts within Australian drinking water catchments (Cox et al. 2005; Power et al. 2005; Ryan & Power 2012) and has been identified often in high concentrations in source waters worldwide, including within Australia (Signor et al. 2005; Roser & Ashbolt 2007).

Information on Cryptosporidium concentrations in Australian surface waters is most heavily clustered around the metropolitan supply areas in south-eastern Australia where major water utilities operating in Queensland, New South Wales, Australian Capital Territory, Victoria and South Australia have been collecting data as part of their raw water monitoring programmes since the 1990s. The Water Services Association of Australia (WSAA), with the support of the major water utilities in these states, instigated this project to determine whether the existing historical databases held within each of these water utilities could provide the basis for estimating treatment requirements for Australian water supplies more generally. Specific objectives of the study were to:

• compile historical Cryptosporidium data from each project partner for several surface water sources representing a range of catchment sizes, land use and climatic conditions;

• characterize the distributions of Cryptosporidium in each surface water source, including the associated uncertainty in that distribution;

• quantify the treatment targets by applying the microbial ‘health-based target’ framework recommended in the GDWQ, and as applied in the AGWR.

For each of the six project partners, source water supplies were identified for which Cryptosporidium monitoring data were available, and that represented a range of different catchment types regarding the level of protection, size and land use within the catchment. For these systems the following data were requested including: Cryptosporidium (including total uncorrected counts), sample volume, method recovery data (i.e., method, number spiked and number recovered), results from molecular genotyping, faecal indicators (including Escherichia coli (E. coli), enterococci and spores of Clostridium perfringens (C. perfringens)) and physico/chemical data (including turbidity, total organic carbon (TOC), dissolved organic carbon (DOC), temperature and pH). All data were compiled in a common Excel datatable to aid QA/QC, e.g., checks for replicates, standardization of measurement units, generate summary statistics and facilitation selection and extraction of priority (Cryptosporidium) data.

The primary objective of the statistical analyses reported here was to characterize the Cryptosporidium concentration in surface water for each of the study systems, including the magnitude and variability of concentration. Cryptosporidium monitoring data consist of discrete counts, often with many zeros, and are the product of an imperfect detection method. Statistical analysis to facilitate appropriate parameter fitting and representative uncertainty analysis needs to be appropriate for these characteristics. Statistical models were constructed to:

1. quantify Cryptosporidium concentration for each study system based on discrete count data and accounting for method recovery;

2. estimate the probability that a confirmed oocyst was human infectious (where human infectivity was defined by the molecular tests applied by the two project partners who included genotyping assays in their monitoring programme).

The models for characterizing the Cryptosporidium concentration were constructed based on work published by Teunis and co-workers (Teunis et al. 1997, 1999; Teunis & Havelaar 1999), and subsequently widely applied to QMRA for drinking water systems (Medema et al. 2006; Petterson et al. 2007; Schijven et al. 2011). Details are given in Appendix A (available online at http://www.iwaponline.com/jwh/013/282.pdf). For each system the deviance (−2 × log-likelihood) was used with a likelihood ratio test to compare the fit of constant concentration model with the variable (gamma distribution) concentration model. If the difference in the deviance was greater than 3.84 (χ², 95% level with 1 degree of freedom), then the addition of a second parameter to describe variable concentration was supported by the data. If the deviance was less than 3.84, it was concluded that there was no evidence in the data that the Cryptosporidium was variable, and the point estimate of the Poisson parameter was adequate to define source water concentrations.

Two project partners provided data related to the human infectivity of isolated oocysts. For these systems, all confirmed oocyst counts were pooled and analysed to determine whether the sample contained the target genetic sequence. Statistically, a negative result was interpreted to mean that the sample contained zero human infectious oocysts; whereas a positive result indicated that the sample contained one or more human infectious oocysts.

To explore the uncertainty associated with the predicted parameter values (μ, λ, ρ, p), a Markov Chain Monte Carlo (MCMC) procedure was applied. For a detailed explanation of the MCMC techniques and applications see Gilks et al. (1996) and Gelman et al. (2014). Briefly, the Metropolis–Hastings algorithm was used to obtain a sample of the posterior distribution for the modelled parameters. This sample of parameter pairs was then used to construct credible intervals for the cumulative density function (CDF) for concentration; the upper 95% credible interval on the mean and the upper 95% credible interval on the upper 95% of concentration were quantified.

The predicted mean Cryptosporidium concentration was plotted against several potential predictive variables: mean E. coli concentration, mean enterococci concentration, mean C. perfringens spore concentration, mean turbidity, mean TOC, mean DOC, system size classification, system catchment protection classification and presence or absence of a reservoir.

For each source water, the predicted Cryptosporidium concentration was used to quantify the required level of treatment in order to achieve the annual tolerable risk target of 1 × 10⁻⁶ disability-adjusted life year (DALY). The method applied is illustrated in Figure 1. Using the QMRA framework, for a particular inflow pathogen concentration, the level of treatment required in order to meet the health-based target was then quantified. The approach relies on fixed input assumptions for the following variables.

The total predicted exposure to pathogens depends upon the volume of unboiled water consumed per person per day. The amount of water consumed varies between individuals. In the context of setting guideline treatment limits, a reference value of 1 L per person per day is selected (consistent with GDWQ (WHO 2011)).

There are several dose–response models published for Cryptosporidium. The values recommended in water quality guidelines have evolved to reflect new scientific information and statistical modelling studies. Table 1 summarizes the assumptions applied in the WHO drinking water guidelines (3rd and 4th editions, e.g., WHO 2011) and the AGWR (National Water Quality Management Strategy 2006). The 3rd edition of the WHO drinking water guidelines adopted a lower value of r = 0.004 based on the only results available at that time. Subsequent studies undertaken for the so-called TAMU and UCP Cryptosporidium isolates (and analysed using a hierarchical model by Messner et al. (2001)) indicated that the infectivity may be at least 10-fold higher. Further work undertaken combining these results with those for the Moredun isolate (and using a different statistical modelling approach as reported in Medema et al. (2009)) suggested a further 10-fold increase in oocyst infectivity.

To investigate the sensitivity of the quantified treatment targets to the assumed oocyst infectivity, the calculations were undertaken separately using the assumed value of the AGWR (NWQMS 2006) and the WHO GDWQ 4th edition (WHO 2011).

Not all infected individuals will develop symptoms of illness. The GDWQ (WHO 2011) and AGWR (NWQMS 2006) have assumed that the probability of illness, given infection, is 0.7. This value was therefore applied in the calculations.

The DALY weighting for Cryptosporidium applied in the GDWQ (WHO 2011) is 1.5 × 10⁻³ DALY per case of illness. The Australian DALY study (Leder et al. 2012) suggested a more suitable DALY weighting for Australia would be 2.46 × 10⁻³ DALY per case of illness. To investigate the sensitivity of the quantified treatment targets to the assumed DALY weight, the calculations were undertaken separately using the WHO DALY weight, and the Australian DALY weight.

Using the above assumptions, the treatment targets were calculated for the predicted mean Cryptosporidium concentration (for both IFA +ve and, when available, DAPI +ve oocysts) and the upper 95th credible quantile on the predicted mean. For those systems with variable oocyst concentrations, the treatment target for the upper 95th quantile of concentration was also quantified using both a point value approach (Equation (3)), and the more appropriate bootstrapped approach (Equation (4)).

Raw water quality data were obtained from all six project partners across 25 study sites. The Cryptosporidium-related parameters obtained from each partner are summarized in Table 2, and the data sets for the study systems are summarized in Table 3. A quantification of method recovery is essential for the concentration prediction (Petterson et al. 2007). For those systems that could not provide recovery data, a conservative value of 30% was assumed. In some cases, it was not clear whether Cryptosporidium counts were IFA +ve or DAPI-confirmed counts. In those cases they were assumed to be DAPI-confirmed. Two project partners reported genotyping of DAPI-confirmed oocysts; however, the specific methods used, and the implications regarding the Cryptosporidium species and genotypes identified by these methods, are not known.

The results of the statistical analysis including maximum likelihood model parameters for the Poisson (no variability, Equation (A3)) statistical model and the gamma distribution (Equation (A7)) statistical models for IFA +ve oocysts are included in Table 4. The mean (expected value) of the Cryptosporidium concentration (μ for the Poisson model, ρ.λ for the gamma model including the upper 95th quantile of uncertainty on this mean from the uncertainty sample generated using MCMC simulation) and the upper 95th quantile of variability of Cryptosporidium concentration (also including the upper 95th quantile of uncertainty on this upper 95th quantile from the uncertainty sample generated using MCMC simulation), are also reported. The same analysis was also undertaken for DAPI-confirmed oocysts (results not shown).

The IFA +ve oocysts data set (Table 4), for systems 1A, 1B, 1C, 3A and 3B, the decrease in deviance achieved by the gamma distribution model was less than 3.84 (χ²_95,1) and hence the additional parameter could not be justified. For these systems, the data set did not support the hypothesis that the Cryptosporidium concentration was variable, and the Poisson distribution was adequate to describe the data. For the remaining systems the assumption of a variable Cryptosporidium concentration was statistically justified by the data set, and CDFs were constructed. Two examples are illustrated in Figure 2. In general, the uncertainty interval about the predicted distribution of Cryptosporidium reduced with increasing data set size, and with increasing concentration.

Uncertainty analysis for some systems produced unstable results. For system 1D, the uncertainty analysis was highly unstable, evidenced by very broad uncertainty intervals (Figure 2(a)). This data set consisted of 15 IFA +ve oocysts detected in 37 samples; 13 IFA +ve oocysts were identified in one sample, with two other single detections, and the remaining samples were negative. These data suggest a highly variable concentration; however, due to the relatively small sample size and the frequency of zeros, quantifying the magnitude of that variability was highly uncertain. For system 3C, the model fit for the gamma distribution model was unstable, and parameter estimates could not be optimized. For this data set of 371 samples, 23 oocysts were detected in 23 separate samples indicating very limited variability in concentration and hence the gamma distribution model was inappropriate. For system 4Eiv, the uncertainty analysis was unstable and did not converge on a stationary distribution for ρ and λ. For this system, of 69 samples, only two samples were positive, with one of these samples containing three oocysts and the other four oocysts. This data set provided insufficent information for predicting the uncertainty in the parameter values of the gamma distribution describing variability in Cryptosporidium concentration.

For project partners 2 and 3, the DAPI-confirmed oocyst counts were identical to the IFA +ve. For these systems the numbers reported in Tables 5 and 6 are identical. For project partners 1, 4 and 5, the DAPI-confirmed concentrations were 0–0.08 log₁₀; 0.12–0.95 log₁₀ and 0.13–0.29 log₁₀ lower than the IFA +ve concentrations, respectively.

The maximum likelihood estimators of the parameter p (probability of an oocyst being human infectious) for each of the systems with genotyping data are summarized in Table 5. Estimated probability of human infectivity ranged from less than 2 to 57%. The probability of finding a human infectious oocyst was approximately an order of magnitude higher for systems managed by project partner 5 in comparison to project partner 4. It is essential to note that the definition of ‘human infectious’ is reliant upon the method used for identification. The methods used by project partners 4 and 5 are not known, and are most likely different.

The concentration of microbial indicators E. coli, enterococci and C. perfringens are plotted against the predicted Cryptosporidium (IFA +ve) concentration in Figure 3. The clearest predictive trend was with mean E. coli concentration. The concentrations of physicochemical parameters turbidity, TOC and DOC are plotted against predicted mean Cryptosporidium concentration in Figure 4. None of the three parameters indicated a clear predictive trend. The predicted mean Cryptosporidium concentration is plotted against the level of catchment protection, the catchment scale and the presence or absence of a reservoir in Figure 5. The mean concentration of Cryptosporidium was highest for small, unprotected catchments. The presence of a reservoir led to a lower predicted mean Cryptosporidium concentration.

The impact of two input assumptions of the estimated treatment requirements the dose–response model and the DALY weighting, was tested. Comparison of the candidate dose–response parameters indicated that a change from the values applied in the AGWR (NWQMS 2006) (r = 0.059) to GDWQ (WHO 2011) (r = 0.2) would lead to an increase in the required level of treatment of 0.6 log₁₀. For the treatment target calculations the GDWQ (WHO 2011) assumption was applied as it was assumed to represent more recent scientific evidence, has been put together by a credible group, and has undergone extensive international peer review. Comparison of the DALY weighting assumptions demonstrated that a change from the GDWQ (WHO 2011) (1.5 × 10⁻³) weighting to the Australian DALY weighting (2.46 × 10⁻³) would lead to an increase in the required level of treatment of 0.2 log₁₀. For the treatment target calculations, the Australian DALY weighting was applied as it was assumed to be locally relevant.

Table 6 summarizes the results of the analysis to determine the appropriate statistic of Cryptosporidium concentration that should be used for quantifying the treatment targets. For a selection of systems, the table shows the annual DALY for treated water, assuming a default 3 log₁₀ treatment using: the mean concentration; a point estimate of the upper 95th percentile of concentration (Equation (3)) and a bootstrapping approach (Equation (4)). When the point estimate of the upper 95th percentile was used, the quantified risk was up to 1 log₁₀ higher than the mean risk. This, however, is an overestimation of the reality since the upper 95th percentile of concentration does not occur every day of the year. The bootstrapping approach accounts for the daily variability in concentration, and the difference from the mean is much lower (up to 0.3 log₁₀ for the most variable system (4Eiv), and ≈0.1 log₁₀ for other systems). The arithmetic mean was therefore considered suitably conservative and applied for the remainder of the treatment target calculations.

The treatment targets for the study systems based on IFA +ve oocyst concentration and DAPI-confirmed oocysts concentration are reported in Table 7 and illustrated in Figure 6. The best estimates for the required level of treatment to achieve the health-based target for the study systems varied between 1.4 and 6.1 log₁₀ relying on IFA +ve counts and 1.4 and 5.8 log₁₀ relying on DAPI-confirmed counts.

In general, the lowest levels of treatment were required for intermediate-to-large protected systems with storage reservoirs (e.g., 3B, 3A and 2D), and the highest levels of treatment were required for small unprotected systems drawing directly from a small weir within the river or stream (e.g., 5Ai, 5Aii and 5C).

Some important exceptions were noted, for example, system 4A (a large protected system with storage reservoir and very low mean E. coli concentration) was classified as requiring a relatively high level of treatment at 3.9 log₁₀. This value would be reduced to 2.9 log₁₀ if only DAPI-confirmed oocysts were counted. In this protected system 16 oocysts were detected from 620 L of analysed sample, of which two oocysts were DAPI-confirmed. Neither of these oocysts were positive by the genotype testing. Similarly, system 4B (a large protected system with low mean E. coli concentration) was classified as requiring 4.5 log₁₀ treatment, which would be reduced to 4.2 log₁₀ based on DAPI-confirmed oocysts. While this system is protected, 51 IFA oocysts and 31 DAPI-confirmed oocysts were detected from a total of 680 L sampled, indicating that the total Cryptosporidium load could not be assumed to be insignificant. Based on the genotyping test, 4.8% of these oocysts would be expected to be human infectious. If only these ‘human infectious’ oocysts were counted, the treatment target would be reduced to 2.4 log_10. Furthermore, some systems that are known to be impacted were classified as requiring lower levels of treatment. For example, system 1B (large unprotected system drawn directly from the river, with a high E. coli concentration) would require only 3.3 log₁₀ as the limited data set found only two oocysts (IFA +ve and DAPI-confirmed) in their total volume sampled of 314 L.

In this study, existing monitoring data for Cryptosporidium from six project partners were compiled and used to characterize the Cryptosporidium concentration in 25 surface water sources in Australia drawing from catchments of different sizes, land use characteristics and climatic regions. These concentrations were then used within the QMRA framework recommended by WHO (2011) for drinking water treatment, and which have already been applied to develop health-based targets for recycled water treatment in Australia via the AGWR (NWQMS 2006).

Relying on historical monitoring data gathered not for this study, but for a wide range of purposes, over many years, from a range of laboratories and using a variety of methods, presented very significant challenges when attempting to meet the purposes of the present study. First, there was no uniformity between utilities with respect to enumeration methodology and data reporting protocols. The use of quantitative recovery controls, reporting of IFA +ve and/or DAPI – or DIC-confirmed oocysts, determination of sample volume (total volume and volume filtered), and the meaning of infectivity assays, varied both between and within utilities and was not standardized. The data supplied did not routinely come with explanations of methodology and how measurements should be interpreted. Interpretation of the data provided for the study, which was fundamental to the appropriate quantification of the Cryptosporidium concentration in surface water and hence any quantified treatment target, required ongoing discussion between the project team, project partners and associated laboratories, with many questions still outstanding at the conclusion of the study. As a result it is strongly recommended that international guidelines for pathogen enumeration be amended to promote reporting transparency between utilities and laboratories, and to facilitate future data comparisons.

A second challenge with the interpretation of historical monitoring data was the ambiguity with respect to monitoring protocols, in particular under what conditions the samples were taken. In most cases, project partners would sample randomly (for example, every Tuesday), and may then take additional samples when triggered by specific conditions. Statistical interpretation of these data are therefore compromised as the overall data set is not truly representative of a purely random sample, nor is it representative of a clear set of defined conditions. Project partner 5 included the clearest designation within their database regarding whether a sample was considered to be ‘event’ affected or ‘routine’. However, on further investigation, the definition of an event sample was broad and had changed over time within the organization as sampling protocols were refined. Retrospective analysis of ‘event’ versus ‘routine’ samples would therefore not represent a meaningful comparison between what, in the context of this study, would be inferred as high flow and baseline flow conditions.

An ongoing question with the development of treatment targets is the appropriateness of using the total number of IFA +ve oocysts for quantifying the Cryptosporidium concentration given that the method cannot distinguish between viable and non-viable oocysts, nor between human infectious and non-human infectious oocysts. From the results reported, treatment targets based on DAPI-confirmed oocysts were only ≈0.2 log₁₀ lower than those based on IFA +ve (except for system 4A, where the difference was 1 log₁₀). However, DAPI confirmation is a conservative marker for oocyst viability since it merely indicates the presence of intact internal contents and does not measure ability to cause an infection. In addition, while the probability of an oocyst being human infectious was quantitatively predicted in this study based on reported data, the true meaning of genotyping assays was not possible to interpret because the methods were not reported and were not consistent between project partners. The historical monitoring data therefore provided only limited scientific basis for addressing these issues of viability and human infectivity. Further, the continuing uncertainty about human infectivity and viability of Cryptosporidium meant that, at present, the data support much higher levels of treatment than might in fact be needed. It would be interesting to do a cost–benefit analysis on the relative price of determining viability and infectivity versus providing an additional 1–3 log₁₀ worth of treatment.

Predicting the concentration of Cryptosporidium in the water body from the oocyst count data required more complex statistical models than are usually applied to water quality data in order to provide the most defensible possible predictions of treatment requirements, given the above limitations. The models used in this study have been widely applied for the purpose of quantifying Cryptosporidium concentration for QMRA (Teunis & Havelaar 1999; Medema et al. 2006; Schijven et al. 2011) and represent current best practice. Nevertheless, as evidenced in this study, real data sets frequently provide limited information for fitting model parameters and can lead to high uncertainty intervals, particularly when pathogen concentration is low and the data sets consist of large proportions of zeros or non-detects.

Rainfall events have been linked with elevated loading of Cryptosporidium loading in surface water (Kistemann et al. 2002; Signor et al. 2005) and with waterborne disease (Thomas et al. 2006; Cann et al. 2013). Accounting for rainfall loading events, particularly in smaller catchments with highly variable flows, and probably quality, is an important issue for the quantification of treatment requirements that remains unaddressed. The historical data used in this study were inadequate for that purpose because the monitoring data were not collected to reflect the diversity of conditions within the study catchments. In addition, water supplies that are protected by storage reservoirs typically had lower Cryptosporidium concentrations and hence required lower levels of treatment. However, it may be that rainfall-induced peaks had been missed by the monitoring programme, a problem illustrated by past events at Lake Burragorang (Hawkins et al. 2000). These factors, and in particular how they should be incorporated within the classification of treatment targets, need further consideration.

The statistic for describing Cryptosporidium concentration as an input to the QMRA model was investigated by comparing the mean concentration with a point estimate of the 95th percentile, and the more appropriate bootstrapped 95th percentile of annual risk. Using a point estimate of the 95th percentile in the risk model is overly conservative for quantifying drinking water risks since it assumes that the 95th percentile occurs every day of the year. The annual risk calculated from a bootstrapped sample (n = 365) of daily risk demonstrated a small difference between the mean and 95th percentile (≈0.1 log₁₀). The difference between the arithmetic mean and the upper 95th percentile of the bootstrapped annual risk is reduced with increasing number of exposures per year. The AGWR uses the upper 95th percentile of pathogen concentration in sewage as a point value input to the QMRA calculations; using the mean in the drinking water guidelines would be a deviation from this national precedent. It is, however, reasonable to adopt a different approach for drinking water exposures since consumption is assumed to occur every day. In the case of wastewater recycling exposures, frequencies can vary and may only involve a single exposure event. For less frequent exposures, using the upper 95th percentile as a point value for quantifying the annualized risk is more appropriate than the mean, and is considered to be appropriately conservative.

Adoption of the QMRA approach requires reliance upon default input assumptions for water consumption, infectivity (i.e., dose–response model), illness rates and DALY weightings. The treatment targets reported in this document have been quantified using the dose–response point value assumption (r = 0.2 (WHO 2011)) and the Australian DALY weighting (1.24 × 10⁻³, Leder et al. (2012)). Together, these two factors influence the treatment targets by 0.8 log₁₀ in comparison to previously applied defaults in the national water recycling guidelines (AGWR) and the WHO GDWQ 3rd edition. Hence, when applying the QMRA approach for defining national treatment targets, the sensitivity of the quantified targets to the default model inputs may not be insignificant, and the scientific data most relevant to the local context selected.

While the general pattern of the treatment requirements for each system met with expectations (e.g., higher treatment required for small impacted systems, less for large protected catchments), the identified anomalies represent a concern. Misclassification leading to a significant underestimation of treatment requirements is critical to avoid but, conversely, overestimation will lead to unnecessary treatment. Based on our analysis of historical data, relying on Cryptosporidium monitoring data as the sole means for classifying systems and defining categories is inadequate. Any classification system must also take into consideration catchment characteristics and land use, and the magnitude of faecal contamination. Pending research questions related to the relative persistence of oocysts in the environment, and the prevalence and identification of species and genotypes of Cryptosporidium of known infectivity to humans, limit the meaningfulness of quantitatively including these factors in the quantification of treatment targets.

In this study, the statistical analysis and interpretation of Cryptosporidium data for the setting of health-based treatment targets has been undertaken based on historical monitoring data. The following recommendations ensue from the work undertaken:

• Rather than relying on historical monitoring data, a data collection programme specifically tailored for the characterization of pathogen concentrations for the purpose of setting health-based treatment requirements needs to be defined. This programme needs to consider: the number of samples, the selection of sampling conditions, the analytical methods used, and the protocols and data reporting.

• It is recommended that international reporting guidelines for pathogen enumeration data be developed to promote transparency between utilities and laboratories, and to facilitate any future data comparisons between water utilities.

• Given the limitations of monitoring data sets for capturing the full range of variability in Cryptosporidium loading and the importance of avoiding misclassification, treatment classifications should not be defined based on Cryptosporidium enumeration data alone, but also need to take into consideration catchment knowledge, sanitary surveys and E. coli monitoring results.

• Based on the data sets analysed in this study, the arithmetic mean is a suitable statistic for defining the Cryptosporidium concentration to be used for quantifying treatment targets for drinking water exposure.

• Selection of appropriate default assumptions for the QMRA model needs to be further investigated and understood in the light of their influence upon the quantified treatment targets, and their appropriateness to the local context.

• Reasonable methods for accounting for rainfall loading events in the quantification of treatment targets are still unexplored, and need to be addressed, particularly for small flashy catchments, and water supplies protected by large storage reservoirs.

This project was led by Water Research Australia (WaterRA) through Gareth Roeszler and Dr David Halliwell on behalf of the Water Services Association of Australia (WSAA), represented by Evelyn Rodrigues, and a consortium of water utilities led by Dr Arran Canning of Seqwater. The key investigators were the named authors of this research article as well as Prof. Peter White, School of Biotechnology and Biomolecular Sciences, University of NSW; Dr Nick O'Connor, Ecos Environmental Consulting Pty Ltd; Dr Paul Monis, Australian Water Quality Centre; A/Prof. Una Ryan, Murdoch University and Dr Martha Sinclair, Monash University. The investigators gratefully acknowledge the water utilities that provided data, information, funding and technical contributions to the project, including Seqwater, Brisbane, Queensland; Melbourne Water, Victoria; Sydney Catchment Authority, New South Wales; Water Corporation, Western Australia; South Australian Water Corporation and ACTEW, Canberra, ACT. David Sheehan is thanked for his detailed review of this manuscript on behalf of WSAA.