Emerging patterns in the global distribution of dissolved organic matter fluorescence

The spectra responsible for natural dissolved organic matter (DOM) ﬂ uorescence in 90 peer-reviewed studies have been compared using new similarity metrics. Numerous spectra cluster in speci ﬁ c wavelength regions. The emerging patterns suggest that most ﬂ uorescence spectra are not tied to biogeochemical origin, but exist across a wide range of di ﬀ erent environments.


Introduction
Fluorescence spectroscopy can characterize chromophores in dilute mixtures in the ultraviolet-visible (UV-vis) wavelength range, 1 and is a popular tool for the untargeted analysis of chemically complex samples such as solutions containing dissolved organic matter (DOM). 2,3 The analysis of uorescence emission excitation matrices (EEMs) most oen assumes that observed uorescence is due to the superposition of several distinct but overlapping uorescence spectra. Thus, EEMs show broad emission spectra whose emission maxima increase towards longer excitation wavelengths ( Fig. 1). Such mixtures can be described by "picking" the uorescence intensities at predened wavelength pairs that have been assigned names such as "peak C", "peak A", or "peak T" (Fig. 1, black letters). 4 Alternatively, the uorescence intensities in wavelength regions of an EEM can be summed up, as in the uorescence regional integration (FRI) approach (Fig. 1, white letters). 5 However, such approaches may sum up the uorescence of multiple uorophores.
In contrast, trilinear models such as parallel factor analysis (PARAFAC) decompose EEMs into underlying statistical components, producing a set of corresponding excitation and emission spectra that, when multiplied by the correct weights (scores) and summed, can reproduce every EEM in the original dataset. Because uorescence properties of chromophores are generally tied to molecular structures and respond to changes on the molecular level, PARAFAC components may also reveal information about the chemical structures responsible for their properties. 6 However, the chemical origin of DOM uorescence and the true number of distinguishable phenomena remains unknown. Thus, despite the high similarity of EEMs across environments, individual PARAFAC models are developed for new datasets with no assumptions made about the presence of particular spectra. 7,8 To aid the interpretation of PARAFAC components and identify the potential emergence of global patterns, a community-driven database called OpenFluor was established in Fig. 1 Typical emission excitation matrix (EEM) of dissolved organic matter (Öre estuary, Sweden). White inserts refer to the five regions defined through fluorescence regional integration (FRI). 5 Black inserts refer to wavelengths pairs at which fluorescence intensities are "picked". 4 2014. 9 As of the publication of this article, OpenFluor contains over 100 entries and more than 500 PARAFAC component spectra mainly describing DOM uorescence in natural and engineered systems (Fig. 2). Although OpenFluor is widely used to evaluate similarities between pairs of spectra, systematic trends of PAR-AFAC spectra in the database have not been reported.
In order to identify similar uorescence component spectra, sensitive metrics of similarity are required. For DOM uorescence, the Tucker congruence coefficient (TCC) has been adopted. 10 However, experience shows that the TCC does not adequately address the generally high resemblance of spectra, particularly if they are Gaussian-shaped. The reliable identication of patterns in large databases requires more sensitive algorithms.
Here, we develop a new metric called shi-and shape sensitive congruence (SSC) and use it to identify patterns in the global occurrence of DOM uorescence spectra according to data in OpenFluor. SSC is based on Tucker's congruence but is more sensitive to subtle differences in peak positions and peak broadness. Additionally, we establish a new approach to quantify similarity between uorescence models.

Experimental section
OpenFluor Ninety published uorescence datasets were extracted from the OpenFluor database (http://www.openuor.org). Each dataset represented a single PARAFAC decomposition (model) of an independently sampled dataset and its associated metadata.
Models contained a set of three to eight excitation and emission spectra that together represent all the underlying uorescence components that varied independently in that study. All spectra (N ¼ 478) were interpolated to identical wavelength increments and ranges (270-440 nm excitation, 300-540 nm emission), smoothed, and normalized by their Frobenius norm. Positions of the primary emission peak and rst excitation peak (when plotted against wavenumber) were used to calculate the Stokes shi. 11 Full details on the processing of spectra can be found in the ESI (section S1 †).

Assessment of component congruence
To sensitively assess the similarity between PARAFAC spectra, we developed the shi-and shape sensitive congruence coefficient as a modication of the Tucker congruence coefficient. 10,12 The SSC penalizes mismatches in uorescence peak area and maxima in comparisons between two spectra follows: The two penalty terms a and b accounted for differences between two spectra x and y with regards to shape (b) and peak wavelength position (a). Further details on the calculation of a and b and their sensitivity to differences between spectral shapes are shown in Fig. 3 and given in the ESI (section S2 †).
The SSC was developed primarily to address a lack of sensitivity in TCC that is evident when comparing emission spectra, since these are unimodal and supercially more similar than multipeak excitation spectra. For example, emission peaks that differ by more than 15 nm in peak wavelength oen have high Tucker congruence. Thus, we used the SSC for comparing emission spectra while excitation spectra were compared using the previously established TCC.

Assessment of model similarity
Each model in OpenFluor contains the excitation and emission spectra of several underlying uorescence components reported  Evaluation of fluorescence emission spectrum similarity of salicylic and ferulic acid by TCC and SSC. DS Peak refers to the difference in the peak integrals and Dl max refers to the difference in peak wavelength. The two penalty terms a, b quantify differences in peak position and peak area between both spectra.
to vary in relative concentration in a published study. An objective of the current study was to determine the similarity between collections of uorescence components (i.e. between models). Here, we assessed the question of model similarity by focusing on the degree of spectral similarity between the components in different models. The following system was devised for comparing models with different numbers of uorescence components: if a model X with n components was compared to another model Y with n + m components, then the best n comparisons were considered while the worst m comparisons were omitted: F was either the average TCC (excitation spectra) or average SSC (emission spectra), and f X,Y was the similarity score between two components in model X and Y. F Ex and F Em were subsequently compared for all models and the most congruent model was identied as having the highest average of F Ex and F Em .

Results and discussion
Assessment of spectral similarity The conventional Tucker congruence coefficient was less sensitive than the newly developed shi-and shape sensitive congruence (ESI section S2 †). SSC uses the conventional TCC to assess overall shape similarity, but additionally quanties differences in emission peak position (term a in eqn (1)) and peak areas (term b in eqn (1)). Fig. 3 shows how the conventional TCC and its general interpretation (good match > 0.95) 10,12 would lead to the conclusion that the uorescence emission spectra of salicylic and ferulic acid are interchangeable, while the corresponding SSC is far lower due to differences in peak integral and peak wavelength position. SSC is thus a sensitive metric to assess the similarity of uorescence emission spectra with generally high resemblance. For TCC, 0.95 has been proposed as a threshold value for good similarity, whereas achieving the same score of SSC would require a better match between two spectra. In scenarios where the high quality of spectral matches is central to an interpretation, SSC would provide a more stringent assessment compared to TCC. One such scenario is model validation, where two models derived on independent halves of an overall data set are compared. Another scenario is the comparison of PARAFAC spectra between studies or with spectra of pure substances, where claims of identity have a signicant impact on the (biogeochemical) interpretation of components. Reporting high scores for SSC rather than TCC would provide superior support for interpretations in these cases.

Spectral patterns in DOM uorescence
The 478 PARAFAC component spectra in OpenFluor showed a wide range of excitation and emission peak positions and Stokes shis (Fig. 4). Excitation peaks ranged from 270 to 427 nm with a median of 308 nm. Fluorescence emission peaks spanned the entire range of measured emission wavelengths (300-540 nm), whereby half of the components described uorescence emission in the ultraviolet range (<416 nm). Components had an average Stokes shi of 0.85 AE 0.31 eV. In comparison, simple uorophores have Stokes shis in the range of 0.3-1.4 eV. 13 94% of all components in OpenFluor had a Stokes shi in this range (Fig. 4). This demonstrates that, while the description of DOM uorescence by PARAFAC makes no assumptions regarding peak shapes or positions, the resulting statistical components largely have viable Stokes shis.
Numerous published uorescence components fell within the uorescence emission range of the classic DOM peaks dened in previous works (Table 1, Fig. 5). 4,14 In 66.7% of studies, a component falling within the range of peak C was identied. Components exhibiting uorescence in the ranges of peaks B, T, M, A, and D were less frequently identied (45.5, 40, 36.6, 16.7, and 8.9% of studies respectively, Table 1). These results indicate that some of the peaks identied historically through raw data analysis are also frequently identied as components in PARAFAC. This is particularly the case in the centre of the EEM (peak C), where uorescence signals are high, spectral corrections are reliable, and instruments perform well. However, PARAFAC oen identied components outside of, or in between the classic peak ranges.
On average, each component in OpenFluor showed high spectral congruence (TCC Ex > 0.95, SSC Em > 0.95) with 3.8 AE 3.6 other components in the database. This increased to 4.9 AE 3.4 if components with no matches were excluded (N ¼ 111). The most frequently matching components showed high similarity with up to 16 other entries (Fig. 5). When components were plotted according to their excitation and emission peaks and weighted by their match-frequency in the database ( Fig. 4 and  5), six distinct clusters could be identied: (1) components similar to classic peak T (l Ex /l Em 275/340); (2) components similar to peak T, but with excitation maxima closer to 300 nm; (3) components located between peak M and C (l Ex /l Em 300-330/380-430); (4) components similar to peak C (l Ex /l Em 330-360/420-455); (5) components similar to group 4, but with longer emission maxima near 480 nm; (6) components similar to peak D with emission maxima between 500 and 520 nm and a wide range of excitation maxima.
The clustering of peaks in groups 1, 2, and 4 largely agrees with known peak locations reported earlier by Coble (2007) and provides independent evidence for the frequent occurrence of uorescence peaks in these regions. However, group 3 notably diverges from these predened peaks, since this group was located between peaks M and C. The frequent identication of peaks in this region by PARAFAC and a lack of identication by peak-picking is most likely a consequence of the highly overlapping nature of uorescence components in this wavelength region. Similarly, the clustering of PARAFAC spectra between 450 and 490 nm (group 5) also does not agree with predened "picked" peaks, or the peak positions of recently proposed ubiquitous PARAFAC spectra emitting at 417 AE 10 nm (F 420 ), 445 AE 6 nm (F 450 ), or 514 AE 9 nm (F 520 ). 15 Since PARAFAC modelling depends on choosing an appropriate number of components, it is susceptible to tting models with too few or too many components. While protection against overtting is usually provided through validation methods, the identication of underspecied models is more difficult. Such models are typically chosen as a compromise due to small sample sizes, low overall signal intensity, or highly correlated abundances of the underlying spectra. 8 The use of too few components results in the description of uorescence properties of multiple components in one "averaged" component (as demonstrated in ESI section S3 †). The occurrence of components in group 5 (emission peaks near 480 nm) may be related to underspecied models having fewer than three components emitting in the visible wavelength range. Of the 90 analysed models, 53% featured two or less components with emission maxima > 420 nm, while a recent study suggests that the adequate description of DOM uorescence requires at least three components emitting in the visible range. 15 The challenge of tting the appropriate number of components to model DOM uorescence is especially critical in datasets with small sample sizes or little spectral variability. Recent studies demonstrate that such challenges can be addressed with methods that selectively inuence the underlying constituents of uorescent DOM in a single sample (through e.g. photochemistry or chromatography) in a manner that allows robust statistical descriptions ("one-sample PARAFAC"). 15,16 Future studies can benet from such methods and we hypothesize that this will result in a convergence towards more similar uorescence components across independent studies. 15

Similarity between PARAFAC models
The average similarity between PARAFAC models ranged from 0.37 to 0.97 (mean 0.84) for F Em , and 0.45 to 0.99 (median 0.89) for F Ex . Twenty-four models in the database were highly similar to at least one other model (F Em and F Ex > 0.95). While this only represents a small fraction of comparisons (0.3% of 90 Â 89 comparisons between 90 models), a high model similarity may not be expected in comparisons involving multiple components since average model similarity scores are signicantly impacted by the presence of any mismatching component. Artefacts possibly introduced due to laboratory-, instrument-, user-, and environment-specic factors may serve to further increase differences between models. 17,18 Fig. 6 shows an example comparison of a six-component model of DOM uorescence in the Otonabee River. 19 In this case, the most-similar amongst the 89 remaining models was found to be a ve-component model of DOM uorescence in boreal lakes and rivers where three components matched well between both studies. 20,21 The average congruence of the ve Table 1 Percentage of models in OpenFluor with components falling within the fluorescence emission/excitation range of classic fluorescence peaks. For peaks with distinct wavelengths, a range of AE10 nm was assumed (e.g. peak B). For peaks encompassing ranges, the explicit definition was used (e.g. peak M). For peak A, all a larger range of excitation (+20 nm) was considered due to the wavelength boundaries in our study (lowest l Ex ¼ 270 nm) most similar components was 0.96 (F Em ) and 0.95 (F Ex ) and the two models generally show good agreement. Our approach to investigating model similarity may be useful to quickly identify highly congruent models in spectral databases such as OpenFluor.
Many studies in recent years have sought to use uorescent components of DOM to discriminate between chemical fractions and biogeochemical sources of DOM. Oen, the specicity of components as proxies for terrestrial or microbial fractions is inferred. If this assumption held true, uorescence models describing DOM in different biogeochemical contexts may also be distinguishable. However, since DOM is oen highly conserved, 22 and its chemical character is similar across many environments, 23 a lack of specicity that reects the ubiquitous processes governing the turnover of organic molecules in aquatic environments may be expected. Our meta-analysis of the 90 uorescence models indeed suggests this lack of speci-city (ESI section S4 †). Patterns of model similarity were not driven by DOM biogeochemistry. Instead, there was a tendency for models with similar number of components to be good matches, particularly the number of components with visible wavelength uorescence featuring emission maxima > 400 nm (Fig. 7). This indicates that the results of inter-study comparisons are likely dependent on the number of components chosen to describe a particular dataset, i.e. studies describing DOM uorescence with similar number of components tend to converge towards similar solutions. Moreover, our analysis indicated that methodological similarities (e.g. particular choices made by users involving wavelength ranges, increments, instruments, integration time) may contribute towards model similarity (Fig. S5 †). Further developments in the application of PARAFAC to DOM uorescence should therefore focus on estimating the impact of instrument settings and other parameters on the modelling outcome to improve the comparability of globally acquired datasets.

Conclusions
The newly developed shape-and shi-sensitive congruence presents a sensitive metric to assess subtle differences between  uorescence emission spectra of uorescent DOM and, in combination with the comparison of corresponding excitation spectra, reduces the risk of falsely identifying signicant spectral similarity. The SSC may in the future be useful in studies seeking to identify highly similar components of uorescent DOM or establish links between DOM uorescence and pure organic substance uorescence.
An analysis of 90 peer-reviewed PARAFAC models revealed that most PARAFAC uorescence spectra of DOM exhibit properties typical for uorophores. In the past, a central hypothesis has been that these independent spectra represent different biogeochemical fractions of DOM. 1 In contrast to this hypothesis, we identied six key regions of uorescent DOM components across numerous geographical regions. A meta analyses revealed no clear connection between uorescence spectral composition and DOM biogeochemistry. These ndings provide evidence that certain uorescence components reoccur irrespective of sample source. While methodological biases may contribute to this result, another possible interpretation is that physicochemical reactions act upon DOM to produce a set of highly conserved uorescence spectra. 15 Further investigation of this hypothesis is warranted, and methodological developments are required to utilize information about reoccurring uorescence species in DOM.
The attribution of DOM uorescence to specic chemical compounds remains largely unachieved. 13 Therefore, deeper insights into the biogeochemistry of DOM fractions require linking uorescence to complementary chemical information. Studies in recent years have demonstrated the benets of DOM multidetector analysis to the interpretation of uorescence. [24][25][26] This approach combines the strengths of individual detection methods and thus may provide further advances in the characterization of DOM.

Conflicts of interest
There are no conicts to declare.