Computational flow cytometry of planktonic populations for the evaluation of microbiological-control programs in district cooling plants

Biofouling poses a serious concern for the district cooling (DC) industry. Current industry practises for monitoring biofouling continue to rely on culture-based methods for microbial enumeration, which are ultimately flawed. Computational flow cytometric (cFCM) analyses, which offer enhanced reproducibility and streamlined analytics versus conventional flow cytometry were applied to samples taken from 3 sites in each of 3 plants over a 5-week sampling program. We asked whether the application of cFCM to monitoring planktonic community dynamics in DC plants could be able to provide sufficient information to enhance microbiological-control strategies at site and inform about plant performance impacts. The use of cFCM enabled the evaluation of biocide dosing, deep cleaning treatment efficiencies and routes of microbial ingress into the studied systems. Additionally, inherent risks arising from the reintroduction of microbiological communities into recently cleaned WCT basins from contaminated cooling waters were identified. However, short-term dynamics did not relate with plant performance metrics. In summary, the insights offered by this approach can inform on plant status, enable evaluations of microbial loads during biofouling mitigation programs and, ultimately, enhance industry management of the biofouling process.


Supplementary information Supplementary Methods
The following formulae were used to calculate the intact cell counts (ICCs). These were used to account for differences in total cell counts across paired subsamples (i.e. PI and SG stained counterpart samples) as described in Supplementary Fig. S4. Each step is detailed alongside an example for a typical 25µL C site (condenser outlet) sample with the following data PI positive count in PI stained sample (Cpi ⊂ Spi) 5,064 Total event count in PI stained sample (CT ⊂ Spi) 20,647 SG positive count in SG stained sample (Csg ⊂ Ssg) 20,856 Total event count in SG stained sample (CT ⊂ Ssg) 21,639 First, the estimated proportion of membrane-compromised cells (Propmcc) are computed. As differences in total event counts (CT) between subsample pairs must be accounted for, Propmcc is determined as the ratio between the proportion of PI-positive events over the proportion of SG-positive events. For this, the following formula (color-coded according to S. Fig.4) is used as presented in equation (1). (1): Where: Cpi ⊂ Spi is the count of PI-positive events (Cpi) of a PI-stained sample (Spi), CT ⊂ Spi is the total event count (CT) of Spi.
Csg ⊂ Ssg is the count of SG-positive events (Csg) of a SG-stained sample (Ssg), CT ⊂ Ssg is the total event count (CT) of Ssg.
Second, the estimated proportion of intact cells (Propicc) is determined as a remainder of the total cell population (i.e. 1) following the subtraction of the proportion of membrane-compromised cells Propmcc as shown in equation (2). (2): Third, the intact cell count (ICC) can then be determined as a factor of the estimated proportion of intact cells in samples multiplied by the count of SG-positive events in an SG-stained sample, Csg ⊂ Ssg as shown in equation (3). (3): As the analysed volume of sample was 25µL, the ICCs were then multiplied by 40 to adjust microbial populations to conventional per mL cell counts.

Supplementary Results
Quantitatively relating microbial loads to DCP performances is complicated by confounding factors and may require long term monitoring Whilst biofouling is a well-known issue for DCPs 1-3 , quantitative analyses of its impact on plant performance are limited in the literature. We therefore sought to assess whether short-term ICCs could inform on biofoulingrelated impacts on plant performance metrics from the SCADA systems at site.
The relationships between ICCs and chiller-level performance metrics (kW/TR, approach temperature and ΔT) were also examined (data not shown). However, no significant relationships could be identified in the time frame examined due to the low number of data points per chiller (chillers are frequently switched to maximise lifespan). A longer follow up of at least one complete cycle between deep clean events (~6 months) may provide enough data points for each chiller to draw such conclusions. We therefore focussed on plant-scale electrical and water performance metrics to build regression models against ICCs ( Supplementary Fig. S6).
ICCs were not a predictor of plant water performance for this study. For RB and YI, electrical performance was also not related to ICC levels.  S4). However, it should be noted that this observation was not confirmed statistically. Nevertheless, it is clear that MUW use should be considered when monitoring biofouling populations in similar recirculating systems.
Importantly, several well established factors for plant performance, including, approach temperature, chiller delta temperatures, TIC, TOC, conductivity and chiller performances also showed little or no correlation with plant performances overall during the study period. We suspect that this highlights the interplay of factors that ultimately give rise to plant performance overall. The potential interaction between multiple factors on observed ICCs and performance metrics, does raise a challenge for quantitatively relating biofouling to plant operations using planktonic counterparts in the community. Therefore, it is likely necessary to build up a longterm history of ICC readings at site in order to uncover the quantitative relationships between microbial loads and plant performances.
Given that YI was observed to have the most extensive biofouling in the WCT basins, and also used a lower frequency of biocide pulses per day (Table 1), the lower cell counts (relative to KC) was an unexpected finding ( Fig. 3b). However, this plant used considerably more water than RB or KC. Therefore, at plants where the MUW consumption and blowdown volumes are high (such as YI) these operational practices can be expected to help in the control of microbial loads within the system, but at the cost of decreasing water performance of the plant (Fig. 2a). Whereas, at KC, the water usage is lower, and the levels of ICCs are primarily driven by biocide dosing and deep cleaning (Sections 3.4.2 and 3.4.3). Supplementary Fig. S4 | Illustrative examples of how differences in total event count can be handled to arrive at reasonable approximations of viable cells when using a single staining approach for cell viability.

Supplementary
a -Theoretical ideal for a sample that has been analysed on the flow cytometer, either unstained, stained with SYBR green (SG) or stained with propidium iodide (PI). In this example, the sample possesses the following true characteristics (the unknowns that we seek to determine through the analysis): 1,000 events per 25 µL (the analysed volume), 50% of which are cells (500 cell events) of which 25% are non-viable membrane-compromised cells (125 cell events). In this theoretical ideal scenario, the total event counts (CT) should equal 1,000 events for each subsample (unstained, SGstained and PI-stained). The unstained subsample is used to generate a gate to define the events that are SG-or PIpositive (green and red dashed lines, respectively). Here, the SG-stained subsample should show 500 SG-negative events (for the 50% abiotic events in this example) and 500 events in the SG-positive region (for the 50% cells in this example).
The PI-stained sample would show 875 events in the PI-negative region to the left of the gate (i.e. the 50% abiotic events plus 75% of the cell events which are intact and viable) and 125 events in the PI-positive region to the right of the gate (the 25% of cells, which are membrane-compromised). In this situation, the PI-positive non-viable cell count, can simply be subtracted from the SG-positive cell count to give the viable cells per 25 µL to arrive at the correct viable cell count. However, in reality, the stochastic nature of subsampling leads to variations in CT between partner subsamples, which would lead to incorrect cell counts if PI-positive events were subtracted from SG-positive event counts if these differences are not accounted for. The 3 formula shown here were therefore used as a means to account for differences in CT and a working example of these formula is given for this ideal scenario. b -A theoretical set of biological replicates of samples, with the same characteristics as the sample illustrated in a, but representing a real-life scenario where CTs differ between subsamples. Here examples show severe differences in CT and deviation from the expected CT of 1,000 events. For these results, the ICC estimates are shown in blue (for the simple approach) and purple (using the formula presented in a). When the results of these estimates are plotted against the true count in this set of scenarios, we can see that the spread of the ICCs is roughly halved (CV of 11.6% vs 19.2%) and that the mean (x) of the ICCs using these formula is a fairly good approximation (350, a 6.66% error) of the true ICC value (375) despite the high variance in CT between subsamples. In this theoretical example, the CV between PI and SG stained sample readings is 22.1% whereas for the cooling tower dataset (with 414 paired SG and PI-stained subsamples) was 20.5%, we therefore expect the ICCs to be >94% accurate with this approach.
Supplementary Fig. S5 | Density histograms of representative frames for unstained samples indicating the results of computational gating in these control samples.
Green and red lines indicate the positions of the SG gate and PI gate (respectively). Percentages shown indicate the proportion of events that are SG or PI positive. Autofluorescence was observed for a small percentage of events (1-5%) which are taken to be the result of autofluorescent photosynthetic microbes in these waters.
Supplementary Fig. S6 | ICCs in relation to DCP water and electrical performances.
Plant performances (electricity -left panel and water -right panel) are presented in relation to intact cell counts for samples from WCT basins (B sites) and condenser outlet waters (C sites). Linear regressions (showing the 95% confidence interval) are shown for each of the plant sites, those with an adjusted R 2 <0.2 are greyed out. Asterisks represent significance level where *** -P < 0.001, * -P <0.05 and "." -P <0.1. Data points are coloured according to plant except those in grey that represent those outlier KC samples, which were taken from WCT basins which had undergone deep cleaning but not been reconnected to the system (and were therefore excluded from the regression analyses).

Supplementary Script
The following code can be used in R or RStudio to run the major processing and analysis for the cFCM used from the DCP water dataset. The code will first prepare the environment for analysis, process the .fcs files (containing the FCM data), apply the gating strategy, return key populations statistics and annotate cFCM data with site metadata. Some example plots along the way are given to allow the user to check for proper FCM data transformation and gating.