Fluorescence excitation–emission matrix (EEM) spectroscopy is often used to determine the levels of trihalomethane (THM) precursors in natural organic matter. However, humic substances are known to quench the fluorescence of amino acids and proteins. To date, none of the EEM-based models for predicting THM formation potential (THMFP) have explicitly accounted for these quenching effects. Thus, we investigated the importance of correcting for fluorescence quenching during THMFP prediction. Fluorescence titration experiments revealed that the correction improved the accuracy of THM prediction. EEM-based models using the corrected fluorescence intensity displayed the highest accuracy (R2 > 0.99; mean absolute error 8.1 μg/L and 13.9 μg/L for chloroform and bromoform, respectively) among models using individual parameters of EEM intensity, dissolved organic carbon (DOC), ultraviolet absorbance at 254 nm (UV254), specific UV254 (SUVA254) and differential ultraviolet absorbance at 272 nm (ΔUV272). Thus, EEM-based models require both the fluorescence intensity of a humic-like component and the corrected fluorescence intensity of a protein-like component for accurate THMFP prediction, for both chlorination and bromination processes. We also found it to be unnecessary to combine DOC with EEM intensity in terms of prediction accuracy, as long as the fluorescence quenching correction is applied.
Disinfection byproducts (DBPs) are formed by the reaction of chlorine with natural organic matter (NOM) during disinfection processes. In the case of bromide-containing water, additional DBPs are often produced because chlorination generates aqueous bromine species (HOBr/OBr−) that further react with NOM (Langsa et al. 2017). Such DBPs are a major drawback of chlorination because they are often carcinogenic (Richardson et al. 2007). DBPs include various classes of compounds, among which trihalomethanes (THMs) are a major component, accounting for approximately 10% of the total halogenated byproducts in water after chlorination (Krasner et al. 2006). Commonly regulated THMs include chloroform (trichloromethane, TCM), bromodichloromethane (BDCM), dibromochloromethane (DBCM) and bromoform (tribromomethane, TBM) (e.g. United States Environmental Protection Agency 2009).
For operational control during disinfection processes, THM concentrations are commonly monitored using gas chromatography (GC) coupled with either an electron capture detector or a mass spectrometer. Although GC affords sufficient measurement accuracy, it is expensive and has limited potential as an online monitoring technique. To supplement or even replace such costly standard methods, various predictive models for THMs have been developed based on water quality parameters indicating the levels of THM precursors (Chowdhury et al. 2009; Beauchamp et al. 2018). Bulk water quality parameters, such as dissolved organic carbon (DOC), ultraviolet absorbance at 254 nm (UV254), specific UV254 (SUVA254) and differential ultraviolet absorbance at 272 nm (ΔUV272), have been employed as the predictors in these models (Chowdhury et al. 2009; Beauchamp et al. 2018). Robust models using these bulk parameters to predict the total concentration of four THMs (TTHM) displayed coefficients of determination (R2) ranging from 0.75 to 0.85 and standard error (SE) values ranging from 69 to 328 μg/L for a relatively large data set including various water sources (Ged et al. 2015). However, these models must be further improved because their SE values are similar to or higher than typical regulatory limits for TTHM (e.g. 80 μg/L).
The chemical structures of NOM compounds govern their reactivities with chlorine and bromine during the formation of THMs (Westerhoff et al. 2004) and also the types and concentrations of THMs formed (Liang & Singer 2003). Fluorescence excitation–emission matrix (EEM) spectroscopy is an effective method for characterizing the chemical properties of NOM specimens. The EEM peak intensities are also correlated with the concentration of TTHM or TCM, which is attributed to the common aromatic structure between fluorophore and THM precursors. Consequently, TTHM and TCM concentrations are more closely correlated with EEM peak intensities than with the bulk parameters considered in other THMFP tests (Pifer & Fairey 2012; Pifer et al. 2013). Thus, the integration of EEM data has been hypothesized to improve the prediction accuracy of models using bulk parameters for THM prediction (Peleato & Andrews 2015; Awad et al. 2016).
However, the inclusion of EEM data in predictive models for THMs remains challenging, primarily owing to the large volumes of data generated. To address this problem, researchers have explored various interpretation methods for EEM data, such as peak picking, fluorescence regional integration (FRI) and parallel factor analysis (PARAFAC), in an effort to obtain the greatest prediction accuracy (Peleato & Andrews 2015). However, the limitations of the EEM technique itself have often been overlooked, not only in the application of DBP prediction but also in other applications involving NOM components (Baghoth et al. 2011; Hur & Cho 2012). One major limitation is that the fluorescence intensity does not reveal the actual concentrations of amino acids (AAs) and proteins owing to fluorescence quenching (Wang et al. 2015), even though AAs and proteins are also DBP precursors in addition to humic substances (Bond et al. 2012). These components are always present in water sources (Meng et al. 2013; Lavonen et al. 2015) and the inter-component interactions between humic substances and AAs or proteins are responsible for the fluorescence quenching. Approximately 35%–52% of the fluorescence intensity of AAs can be quenched by the interaction with humic substances (Wang et al. 2015), which may significantly reduce the accuracy of EEM-based models for DBPs. However, this limitation has never been considered during the application of EEM data in predictive models for DBPs and its influence on prediction accuracy remains unknown.
In contrast to previous studies, which directly used EEM data in their predictive models, we aimed to investigate whether correcting for inter-component fluorescence quenching could improve the performance of EEM-based models for predicting the THMFP. We conducted two fluorescence titration experiments, namely, titration of the protein bovine serum albumin (BSA) against a known concentration of humic substances and titration of humic substances against a known concentration of BSA, to investigate the influence of correcting for fluorescence quenching during THMFP prediction. The importance of correcting for fluorescence quenching was then evaluated by comparing the fitness of EEM-based models using the measured and corrected fluorescence intensities as well as that of other models based on bulk parameters. The importance of this correction was investigated for two aqueous halogens (chlorine and bromine) to obtain a fundamental understanding of the individual contribution of correcting for fluorescence quenching in the case of each halogen.
Suwannee River natural organic matter (SWNOM, 2R101N) was obtained from the International Humic Substances Society (IHSS) and was used to represent the end-member of humic substances, and BSA (≥96%, Sigma Aldrich) was used as a representative protein. A protein was considered as the end-member in this study rather than other protein-related species such as AAs owing to the occurrence of proteins as the major fraction in water sources (Aiken 2014). The preparation of the NOM stock solutions is described in Section S1 of the Supplementary Material (available with the online version of this paper). Two fluorescence titration experiments were conducted according to the procedure described by Wang et al. (2015). In the first experiment, various concentrations of SWNOM (0 to 15 mg/L) were added to a fixed concentration of BSA (10 mg/L). In the second experiment, various concentrations of BSA (0 to 15 mg/L) were added to a fixed concentration of SWNOM (10 mg/L). The final volume of the titrated samples was fixed at 200 mL in all cases. To ensure complete interaction, the titrated samples were shaken for 24 h using a mechanical shaker at room temperature in the dark. We confirmed that the shaking procedure itself did not lead to any loss of fluorescence intensity for either SWNOM or BSA. The concentrations of the components in the titrated solutions were chosen to afford a similar level of fluorescence intensity as wastewater-impacted source water.
THMFP tests were conducted for all of the samples obtained from the fluorescence titration experiments, and phosphate buffer solution was used as a negative control. These water samples were divided into two series for chlorination and bromination experiments. The tests were conducted according to the 22nd edition of Standard Methods for the Examination of Water and Wastewater (APHA/AWWA/WEF 2012) using a headspace-free reactor. An aqueous stock solution of chlorine (5,000 mg Cl2/L) was prepared from sodium hypochlorite (>5%, Kanto Chemical), and an aqueous stock solution of bromine (5,000 mg Br2/L) was prepared from pure liquid bromine (99.9%, Sigma Aldrich). To perform the THMFP tests, these stock solutions were separately added to the reactor containing 100 mL of a water sample. Two series of THMFP tests were conducted, using each of the two aqueous halogen solutions at a final concentration of 1 mmol/L, which was selected based on the halogen demand corresponding to a maximum DOC concentration of 6.67 mg C/L. The reaction mixtures were incubated for 24 h under controlled laboratory conditions at 25 °C and pH 7.0 ± 0.2. For each sample, we analysed the EEM, DOC, UV254, SUVA254, ΔUV272, halogen concentration and THM concentrations both before and after the test. Analytical procedures and the instruments used for water quality parameters, halogen concentration and THM concentrations are presented in Sections S2 and S3 of the Supplementary Material (available online).
Fluorescence measurements and calculation of unquenched fluorescence intensity
Three-dimensional fluorescence spectra were measured at controlled room temperature (25 °C) using a spectrofluorometer equipped with a xenon lamp as the light source (RF5300, Shimadzu, Japan) and a quartz cuvette (1 cm path length). The EEM spectra of the samples and phosphate buffer solutions (blanks) were scanned over an excitation (Ex) range of 240–500 nm with an increment of 5 nm and an emission (Em) range of 280–550 nm with an increment of 2 nm. The raw EEM data were pretreated according to the procedure described by Murphy et al. (2010) to correct for the inner filter effect and fluorescence signal fluctuation (Raman unit calibration). The pretreatment was performed using MATLAB software (MathWorks Inc., USA) with the fdomcorrect script of the FDOMcorr1.6 toolbox, as described by Murphy et al. (2010). Following this pretreatment, the fluorescence intensities were analysed by determining the change in fluorescence intensity for five specific Ex/Em pairs (Table 1) representing different components present in NOM (Coble 1996). These Ex/Em pairs were selected based on the peak locations of PARAFAC components that are typically present in various standard NOM samples from the IHSS (Ateia et al. 2017).
|Component||Peak name||This study||Reference peak regions (Coble 1996)|
|Excitation (nm)||Emission (nm)||Excitation (nm)||Emission (nm)|
|Component||Peak name||This study||Reference peak regions (Coble 1996)|
|Excitation (nm)||Emission (nm)||Excitation (nm)||Emission (nm)|
THMFP predictive models
Mathematical models were developed using simple and multi-linear regression to examine the relationships between the water quality parameters (EEM, DOC, UV254, SUVA254 and ΔUV272) and THM concentrations. In the case of EEM, stepwise multi-linear regression (forward and backward, p < 0.05) was applied to determine the suitable combinations of Ex/Em pairs, including all of the peaks listed in Table 1, with both measured and corrected fluorescence intensities. Model fitness was evaluated using the fitness indicators of R2 and mean absolute error (MAE). In addition, the models were compared with the EEM-based model recently reported by Awad et al. (2016). In terms of prediction accuracy, the contribution of the addition of DOC was determined using multi-linear regression. The DOC was added to the EEM-based models both before and after the correction for fluorescence quenching, as well as the models based on the other water quality parameters, to cover both non-aromatic and aromatic organic matter fractions for THMFP prediction.
RESULTS AND DISCUSSION
The Ex/Em pairs were used to track the changes in fluorescence intensity during the fluorescence titration experiments. In this method, the overlap of fluorescence spectra between the two end-members was neglected because it contributed less than 5% of the fluorescence peak intensity in each case. At a fixed BSA concentration, the presence of SWNOM reduced the fluorescence intensities of BSA at both of the Ex/Em pairs corresponding to protein-like components (i.e. peaks T and B; Figure 1(a)). The fluorescence intensities of BSA at the Ex/Em pairs of peaks T and B were quenched in the presence of SWNOM with average quenching ratios (Q) of 0.89 ± 0.01 and 0.69 ± 0.01, respectively. The higher value of Q for peak T than for peak B was possibly attributable to the greater initial fluorescence intensity of BSA at the Ex/Em pair of peak T in the absence of SWNOM. In the second titration experiment at a fixed SWNOM concentration, the addition of BSA did not quench the fluorescence intensities of SWNOM at any of the Ex/Em pairs corresponding to humic-like components (i.e. peaks A, C and M; Figure 1(b)). These results are in accordance with a previous report that described the quenching of the fluorescence of AAs and proteins by humic substances and the lack of quenching of the fluorescence of humic substances by AAs and proteins (Wang et al. 2015). Further details regarding the fluorescence characteristics of the end-members and their mixtures as well as an example of an EEM spectrum of natural water are presented in Section S4 of the Supplementary Material (available with the online version of this paper).
In the THMFP tests, TCM was the only THM detected in the chlorination experiments and it was present in the samples at concentrations in the range of 130–480 μg/L. In the bromination experiments, TBM was the dominant product among the THMs and occurred in the range of 350–1,326 μg/L, while the concentrations of other species were <0.5% of the TBM concentration. The concentrations of both TCM and TBM increased upon increasing the concentration of either BSA or SWNOM. Therefore, both of the end-members reacted with both chlorine and bromine to serve as THM precursors.
Influence of correcting for fluorescence quenching on model fitness
The use of the corrected fluorescence intensities improved the model fitness in terms of both R2 and MAE, as indicated by the dashed lines shown in Figure 2(a) and 2(b). For the combination of the fluorescence intensities at peaks A and T, the use of the corrected fluorescence intensity of peak T (i.e. peak Tc) for TCM and TBM prediction increased the value of R2 by 25.0% and 0.5%, respectively, and decreased the value of MAE by 45.5% and 30.0%, respectively. For the combination of the fluorescence intensities at peaks A and B, the use of the corrected fluorescence intensity of peak B (i.e. peak Bc) for TCM and TBM prediction increased the value of R2 by 13.0% and 4.2%, respectively, and decreased the value of MAE by 59.5% and 79.9%, respectively. The use of the corrected fluorescence intensities also reduced the deviation between the actual and modelled concentrations of THMs in the case of the models showing the best fitness for TCM and TBM prediction (Figure 3(a) and 3(b)). When this correction was not applied, the predicted TCM concentrations exhibited a greater deviation from the 1:1 line than the predicted TBM concentrations. The different magnitudes of the deviation for TCM and TBM prediction were presumably attributable to the greater quenching ratio for the fluorescence intensity at peak T compared with that at peak B upon the interaction with SWNOM, as discussed earlier. Furthermore, the EEM-based model using the corrected fluorescence intensities displayed the highest accuracy, with R2 > 0.99 and MAE values of 8.1 μg/L and 13.9 μg/L for TCM and TBM, respectively, compared with the models based on individual water quality parameters such as EEM peaks, DOC, UV254, SUVA254 and ΔUV272 (Figure 2(a) and 2(b)). The examples of the data set used for calculation of R2 and MAE of each individual water quality parameter are shown in Figure 3(c) and 3(d) for TCM and TBM, respectively. Therefore, correcting for fluorescence quenching is necessary for improving the prediction accuracy of THMs, regardless of whether chlorination or bromination is being considered.
Interestingly, the stepwise regression resulted in the selection of different Ex/Em pairs corresponding to protein-like components to indicate the TCM and TBM precursors, possibly indicating the specific preferences of chlorine and bromine to react with different parts of the protein molecule. The selection of the fluorescence intensity at peak T rather than that at peak B for TCM prediction was probably attributable to the greater ability of tryptophan to produce TCM (three-fold higher yield in μg TCM/mg C) compared with tyrosine (Hong et al. 2009). With respect to the selection of peak B rather than peak T for TBM prediction, further studies are required to explain this observation by comparing the production of TBM from each AA and bromine. In addition, chlorine tended to react more with humic substances than with protein, as indicated by the higher value of the coefficient for the fluorescence intensity at peak A than that at peak Tc. In contrast, bromine tended to react more with protein than with humic substances, as indicated by the higher value of the coefficient for the fluorescence intensity at peak Bc than that at peak A. A similar tendency of chlorine and bromine with different EEM peaks was also reported by Awad et al. (2016).
Contribution of DOC in predictive models
The addition of DOC to the EEM-based models with the correction of fluorescence quenching resulted in the same value of R2 and reduced the MAE value by only 3% compared with the models without DOC, for both TCM and TBM prediction. This result was in contrast to the previous study by Awad et al. (2016), which reported that the combination of DOC with UV254, rather than with EEM, afforded the best accuracy for THM prediction. This difference was probably caused by the use of EEM data without correcting for fluorescence quenching in the previous models. Nevertheless, the addition of DOC to the EEM-based predictive model after correcting for fluorescence quenching showed a minor improvement.
The addition of DOC to the TCM predictive models using UV254, SUVA254 and ΔUV272 improved the fitness of these models. However, the model fitness of these combinations remained lower than that of the model using only DOC (R2 = 0.96, MAE = 21.4 μg/L). Thus, the EEM-based model using the combination of the fluorescence intensities at peaks A and Tc was still the best method for TCM prediction (R2 = 0.99, MAE = 8.1 μg/L). For TBM prediction, the addition of DOC to UV254 or ΔUV272 improved the model fitness in terms of both R2 and MAE, whereas the addition of DOC to SUVA254 increased R2 but did not decrease MAE. Nonetheless, the model using the combination of the fluorescence intensities at peaks A and Bc afforded superior model fitness for TBM prediction (R2 = 0.99, MAE = 13.9 μg/L) compared with the combined models using DOC and UV254 (R2 = 0.96, MAE = 21.25 μg/L), DOC and ΔUV272 (R2 = 0.99, MAE = 31.14 μg/L) and DOC and SUVA254 (R2 = 0.96, MAE = 21.26 μg/L). Improvements in model accuracy for THM prediction have been observed upon combining DOC with other predictors (Ged et al. 2015), although our results suggest that correcting for fluorescence quenching leads to a greater improvement in model accuracy than the addition of DOC to other parameters.
Fluorescence EEM spectroscopy is often applied to determine the levels of THM precursors in NOM. We investigated the influence of correcting for fluorescence quenching on the accuracy of predictive models for THMs. We found that EEM-based models required both the fluorescence intensity of a humic-like component and the corrected fluorescence intensity of a protein-like component to accurately predict the levels of THMs, for the reactions with both aqueous chlorine and bromine. Correction of the fluorescence intensity of the protein-like components improved the accuracy of THM prediction, and EEM-based models exhibited the highest accuracy among models using individual water quality parameters (i.e. EEM peaks, DOC, UV254, SUVA254 and ΔUV272) as the indicator of THM precursors. In addition, the combination of DOC with the corrected EEM data was found to reduce the MAE by only 3% compared with the models using only the corrected EEM data as the indicator of THM precursors. Thus, the addition of DOC to the EEM peaks was unnecessary in terms of improving the prediction accuracy if the correction for fluorescence quenching was applied. Overall, this study quantitatively investigated the importance of correcting for fluorescence quenching in the development of predictive models for THMs. These results support the potential application of fluorescence measurements for predicting THMs during online monitoring. Further investigations are necessary to elucidate the importance of correcting for fluorescence quenching under various chlorination conditions (e.g. pH, contact time, temperature) and develop correction methods for the fluorescence quenching of unknown samples.
This study was financially supported by the JSPS Core-to-Core Program: Asia–Africa Science Platforms ‘Establishment of Asian Model for Research and Education on Urban Water Resource Management’ and SATREPS (JST/JICA). The authors would like to thank the Chiba Prefecture Pharmaceutical Association Inspection Center for their support with THM analysis and Mr Daniel Chang for English proofreading.