Three humic brown-water lakes (DOC concentration range 21.2–23.7 mg C L⁻¹; Specific UV absorbance at 254 nm (SUVA₂₅₄) range 3.3–4.1 L mg C⁻¹ m⁻¹), Lumpen, Ramsjön, and Övre Långsjön, were sampled in a forested area of east-central Sweden in October 2013. The catchments of the lakes are dominated by boreal forest (~20% of peat in the catchment of the Lake Lumpen) and underlain by the calcareous moraine and granitic bedrock. Mean annual precipitation in this area is 600–700 mm and mean annual temperature is 4–6°C (1961–1990, Swedish Meteorological and Hydrological Institute; http://www.smhi.se/klimatdata).

The water was collected from the surface (oxic epilimnion, 0−0.5 m depth), filtered through 0.2 μm membrane filters (Supor, Pall) and stored in the dark at 4°C. Portions of filtered water were irradiated with UV light (300−400 nm range) for 144 h, and about 30% of initial DOC was photochemically mineralized as a result (see also Mostovaya et al., 2016). Thus, the subsequent dark decomposition experiment was provided with treatments containing non-manipulated and UV-manipulated lake water.

The experimental setup consisted of non-manipulated and UV-manipulated lake water treatments, duplicated for each lake (12 experiments in total). The DOM incubations were conducted on 0.2 μm filtered water, inoculated with a 64 μm plankton net filtered aliquot from the respective lake (5% of the volume), with addition of inorganic nutrients (480 μg N L⁻¹ as KNO₃ and 100 μg P L⁻¹ as Na₂HPO₄). Lake water was incubated in pre-acid washed (10% HCl), pre-combusted (4 h at 550°C) 40 ml glass vials sealed with PTFE-lined silicone septa, at 20°C and in the dark. All vials were submersed in water to minimize the gas exchange, and incubations lasted for 120 days. The length of the experiment was motivated by the fact that humic DOM can decay very slowly (e.g., Koehler et al., 2012). At 12 different time points of the experiment individual vials were sacrificed for measuring the DOC concentration and taking the samples for mass spectrometry analysis. The DOC concentration was measured on a Sievers 900 TOC Analyzer (General Electric Analytical Instruments). Water for mass spectrometry analysis was stored in pre-combusted (4 h at 450°C) 2 ml glass vials with PTFE-lined silicone septa at 4°C. Prior to usage, the septa and screw caps were pre-soaked in methanol for 24 h and repeatedly rinsed with MilliQ water (Millipore).

Mass spectrometry analysis was performed with a 15-T Fourier transform ion cyclotron resonance mass spectrometer (FT-ICR-MS; Solarix, Bruker Daltonics) with an electrospray ionization source operating in negative mode. Lake water samples were diluted with ultrapure water to a concentration of 10 mg C L⁻¹ and mixed with methanol (HPLC grade, Sigma-Aldrich) in a proportion of 2:1 (final concentration 6.7 mg C L⁻¹). The samples were injected without further treatment, i.e., without performing the solid phase extraction, at 360 μL h⁻¹. Mass spectra were collected over 400 scans, with ion accumulation time of 0.5 s, and within a range of 150–2,000 m/z. Each sample was internally calibrated with a reference mass list generated from North Equatorial Pacific Intermediate Water (NEqPIW) (Green et al., 2014) using the Bruker Daltonics Data Analysis software package. Molecular formulae containing the elements C, H, O, N, S, and P were assigned to peaks with a signal to noise ratio greater than four and based on the following criteria: O ≤ C; O > (2P + S); N ≤ 4; S ≤ 4 and P ≤ 1. The S-, P-, and N > 1-containing formulae did not show any specific behavior (see section Formula-Specific Apparent Decay Coefficient k_exp) and, since higher errors are associated with the assignments containing S, P, and N > 1, these formulae were excluded from further analysis to avoid excessive complexity. Unassigned peaks were disregarded, as they were mainly noise or isotopologue (e.g., ¹³C) peaks. Known contaminant peaks and peaks found in procedural blanks, as well as the peaks with unusually high intensity, were removed from the dataset. For each lake, only peaks found in both replicates of each experimental treatment were considered. For each decomposition experiment, time series of total intensity were compared to the corresponding bulk DOC loss (Figure S1). This was done to check the consistency of the data, as we expected linear relationship between total intensity and bulk DOC. One of the replicate incubations (lake Övre Långsjön, non-manipulated water, replicate 2) and four individual samples from different incubations fell out of the linear trend, possibly due to the short-term instrument or software problems, (Figure S1) and, thus, were removed from the subsequent analysis. Further, data from each lake-treatment-replicate combination (eleven in total, since one incubation time series was removed) was processed separately. A detection limit was applied based on dynamic range. The dynamic range of each individual sample was calculated as a ratio of maximum to minimum peak intensity found in the sample. Detection limit was then calculated for each sample, by dividing maximum intensity in this sample by the lowest dynamic range found among the samples. Intensities below the detection limit were removed. This left in consideration 1,654−2,045 formulae across the range of samples.

In each sample the intensities were normalized to the total intensity and multiplied to account for dilution and to estimate the equivalent intensity in the undiluted sample. In other words, the samples were normalized internally for total intensity and then externally to account for DOC concentration. This was done to correctly estimate the relative change in concentration over the course of incubation (Lechtenfeld et al., 2014; Seidel et al., 2015; Hawkes et al., 2016b; Mostovaya et al., 2017). The signal intensity we refer to further in the text (e.g., in the section Formula-Specific Apparent Decay Coefficient k_exp) corresponds to DOC normalized intensity in undiluted sample.

Double bond equivalence (DBE), reflecting the degree of unsaturation, was calculated for each formula as DBE = 1 + ½(2C – H + N) and modified aromaticity index (AI_mod), reflecting the existence of aromatic and condensed aromatic structures, was calculated as AI_mod = [1 + C − ½O − ½(N + H)]/(C − ½O − N), where C, H, O, and N refer to a number of respective atoms per molecule. The formulae for DBE and AI_mod are adjusted from (Koch and Dittmar, 2006, 2016) as S- and P-containing formulae were not included in our analysis.

Molecular data was visualized in van Krevelen diagrams, i.e., the plots of formula-specific H/C vs. O/C ratios (Kim et al., 2003). Compound groups were assigned based on AI_mod and H/C ratio, with AI_mod > 0.66 indicating the presence of combustion derived polycondensed aromatics (Koch and Dittmar, 2006), 0.50 < AI_mod ≤ 0.66 delineating plant polyphenols and condensed aromatics with few aliphatic side chains (Šantl-Temkiv et al., 2013; Kellerman et al., 2015), AI_mod ≤0.5 & H/C < 1.5 indicating highly unsaturated and phenolic compounds (Šantl-Temkiv et al., 2013), and H/C ≥ 1 indicating aliphatic compounds (Riedel et al., 2016). We also calculated nominal oxidation state of carbon (NOSC) as NOSC = 4−[(4C+H−3N−2O)/C]. This formula is modified from Riedel et al. (2012), since S-containing compounds were not included in our analysis.

For each formula, a single-exponential model (Olson, 1963) was fit into the 12-point time series of relative change in signal intensity following the equation

where k_exp is an apparent exponential decay coefficient (day⁻¹) and I_t/I₀ is the relative change in intensity over time (Mostovaya et al., 2017). The choice of the model is rooted in the theory of reactivity continuum (Boudreau and Ruddick, 1991), which emerges as a realistic model of organic matter decay (Vähätalo et al., 2010; Koehler and Tranvik, 2015; Mostovaya et al., 2016) and assumes that each individual reactive type within bulk DOM degrades following the single-exponential kinetics. For the purpose of model fitting, only formulae that were detected on at least 8 days of the 12-point time series were considered. The model fitting was conducted using generalized least squares modeling, and the goodness of model fit was described with normalized root-mean-square error (NRMSE) (Mostovaya et al., 2017). The time series containing a lot of noise were not considered in the subsequent analysis. The threshold for noise was established at NRMSE = 35.8%–a 95% quantile of the NRMSE distribution calculated from a series of repeated measurements of DOM NEqPIW standard (Green et al., 2014) of steady concentration (15 mg C L⁻¹). Between 89.7 and 97.3% of the formulae in the experimental datasets met this criterion. Since the context of our study was DOM decomposition, we were interested in the molecular formulae that exhibited net decay, and, hence, only used the formulae with positive k_exp in the subsequent analysis. It should be remembered that the apparent exponential decay coefficient k_exp represents the net of simultaneous loss and formation of the corresponding formula. The formulae demonstrating the net decay constituted the majority in our datasets (76.1−98.1% of all formulae with accepted NRMSE). The formulae demonstrating the net accumulation (Figure S2) are not discussed further.

The relationship of k_exp with molecular characteristics, including the H/C and O/C ratios, DBE, AI_mod, NOSC, and molecular mass were approximated using generalized additive models (GAM) based on cubic splines (R function stat_smooth in package ggplot2). The model fitting, and all calculations were performed using R (version 3.3.2, R Development Core team, 2016).

We focused on two groups of formulae from the comparison between the non-manipulated and UV-manipulated experimental treatments: (1) shared formulae, found in both treatments and (2) the formulae found only in UV treatment, i.e., corresponding to photoproduced molecules. For the shared formulae, apparent decay coefficients k_exp were compared between the treatments. To explore possible similarities in chemical characteristics and reactivity of photoproduced molecules, the corresponding formulae were plotted in van Krevelen space with apparent decay coefficients k_exp as a third, color-coded dimension.

Between 1,308 and 1,820 formulae could be assigned an apparent decay coefficient k_exp in each treatment (see Formula-Specific Apparent Decay Coefficient k_exp in Methods). To compare the distribution of k_exp we considered the following groups of compounds: aliphatics, highly unsaturated and phenolic compounds, plant polyphenols, and polycondensed aromatics. The values of k_exp varied from 4.3 × 10⁻⁸ to 0.02 day⁻¹. While most of k_exp in this range were found within each group of compounds, the exact distribution of k_exp, i.e., the prevalence of lower or higher k_exp, differed between the groups (Figures 1, 2, Figures S3, S4).

The lowest k_exp were mostly associated with highly unsaturated and phenolic compounds (Figures 1, 2, Figures S3, S4). This group is ambiguous with respect to aromaticity–molecules may contain aromatic rings, or may simply have a high density of double bonds in a different configuration (Koch and Dittmar, 2006). Given that the DOM of boreal lakes originates largely from terrestrial vascular plants (Thurman, 1985; Hessen and Tranvik, 1998), a reasonable assumption can be made that most of the highly unsaturated and phenolic compounds found in our experiment are lignin-derived.

In non-manipulated treatments, representing the natural lake water, the share of slowly decaying formulae was also substantial within the groups of polyphenols and polycondensed aromatics (Figures 1A, 2A,B, Figures S3, S4). In one of the lakes (Figure 2C), and in UV-manipulated treatment (Figures 2E,F) these groups were almost as reactive as aliphatics.

Relatively low reactivity of vascular plant-derived components of DOM has been reported previously (Benner et al., 1987; Moran and Hodson, 1994; Benner and Kaiser, 2011). However, we do not suggest that the low reactivity of lignin-derived and polyphenolic compounds is ubiquitous. Our results showed the presence of high apparent decay rates within each group of compounds (Figure 2, Figure S4). Accordingly, fast decomposition of lignin phenols was recently documented for Amazon River DOM (Ward et al., 2013). The fact that some fraction of phenolic compounds is likely to readily participate in decomposition at the timescales of several months was also previously reported (Mostovaya et al., 2016).

The reactivity of aliphatic compounds was predominantly high (Figures 1, 2, Figures S3, S4), which agrees with existing concepts of OM lability (Sun et al., 1997; D'Andrilli et al., 2015). Studies have shown that labile fractions of DOM are often of algal origin (e.g., Guillemette et al., 2013), and there is evidence that algal DOM is to a large extent aliphatic (H/C > 1.5) (Mangal et al., 2016). Yet, we do not imply that aliphatic compounds in our study are exclusively of algal origin, because they may also be components of fresh terrestrial DOM of increased lability. In the study evaluating the persistence of DOM in boreal lakes Kellerman et al. (2015) found that aliphatic compounds were abundant in lakes with higher water residence time. While, at first, this seems to contradict our results, it should be considered that persistence is not necessarily caused by the resistance to degradation and sedimentation, but can be also a result of the constant replenishment. This might be especially relevant for algal DOM, as the conditions for photosynthesis improve in more transparent waters of the lakes with higher water residence time.

It should be reemphasized that k_exp formed a continuous distribution within each group of compounds, with both high and low k_exp present (Figure 2, Figure S4). The average behavior of each compound group was, hence, determined by the character of the distribution and whether it was shifted toward lower or higher values of k_exp. This gives further support to the idea that overall DOM decomposition kinetics emerges from the decay kinetics of multitude of individual constituents (Mostovaya et al., 2017). The presence of a wide range of reactivities within each compound group, however, signifies that molecular formulae with different bulk properties can be equally reactive. This likely reflects the complexity of structural constraints on reactivity and suggests that higher structural resolution is needed to further ascertain the reasons behind DOM reactivity in lake water.

Our experiment characterized the aerobic decomposition of DOM typically present in boreal humic lakes, not limited by nutrients. While it is not possible to say whether the observed trends in compound reactivity would be retrievable from the other aquatic settings, they might be characteristic of other lakes with high terrestrial DOM input. At the same time, it has been shown before that the freshness of terrestrial DOM can influence the rates of microbial DOM utilization (Berggren et al., 2010a,b). This exemplifies that the exact chemical composition might be a more universal predictor of decay patterns than the degree of allochthony. With the increasing number of studies using the ultrahigh resolution mass spectrometry to understand the molecular drivers of DOM reactivity in a variety of systems (Gonsior et al., 2013; Sleighter et al., 2014; D'Andrilli et al., 2015; Kellerman et al., 2015; Hawkes et al., 2016b; Riedel et al., 2016; Kamjunke et al., 2017) a better understanding of the intrinsic controls of aquatic DOM reactivity becomes increasingly possible to achieve.

We examined the relationship of k_exp with compounds' saturation (H/C), modified aromaticity index (AI_mod), double bond equivalency (DBE), molecular mass, oxygen content (O/C ratio), and nominal oxidation state of carbon (NOSC). Despite the large variability in the data, fitting the GAM models allowed us to identify general trends (Figures 1, 3, Figures S3, S5).

The reactivity of aromatic compounds (H/C~below 1.1) was on average higher than reactivity in the “highly unsaturated” region (H/C~between 1.1 and 1.5) (Figure 1, Figure S3), and AI_mod showed curved relationship with k_exp: the k_exp first decreased with increasing AI_mod, but at AI_mod between 0.25 and 0.33 began to increase again (Figures 3A,B, Figure S5). As was pointed out above, highly unsaturated compounds may contain aromatic structures but their content is assumed to be lower than in the groups of plant polyphenols and polycondensed aromatics. A high contribution of aromatic structures is typical for the colored DOM of humic lakes, and traditionally considered to be associated with low reactivity (Tranvik, 1998). Hence, at the first sight our findings seem surprising. However, substantial evidence has accumulated that this material is not as persistent as often assumed (Tranvik, 1992), and in boreal lakes colored DOM disappears often faster than bulk DOM (Koehler et al., 2012; Weyhenmeyer et al., 2012). Likewise, FT-ICR-MS analysis of DOM from a large number of lakes (Kellerman et al., 2015) showed that the prevalence of aromatic compounds decreases with increasing water residence time of lakes, suggesting that they are selectively removed. The concept of universally low reactivity of aromatic DOM has been further challenged as higher-than-previously-thought lability of lignin-derived components was described in a variety of systems (Holmes et al., 2008; Spencer et al., 2008; Ward et al., 2013). The variability in reactivity observed for aromatic DOM reflects the fact that compounds that belong to a large common group (e.g., aromatics, lignin) can be extremely diverse with respect to exact structure and chemical properties.

Decay rate tended to be slightly higher for compounds with lowest DBE (Figures 3C,D, Figure S5), but generally DBE did not have a strong effect on k_exp.

We did not find strong indications of the effect of molecular mass on k_exp in the non-manipulated treatment (Figure 3E, Figure S5), and in the UV treatment reactivity tended to decrease with increasing mass in some cases (Figure 3F, Figure S5). It should be noted that our data covered the DOM mass range of ~150−600 Da. Other studies considered apparent masses up to 1−10 kDa (Tranvik, 1990; Amon and Benner, 1996), acquired using ultrafiltration techniques, and showed that bacteria preferentially utilize substrates of higher apparent molecular weight. A systematic comparison of apparent (i.e., assessed on the basis of separation techniques such as ultrafiltration) vs. actual (i.e., assessed based on high resolution methods targeting individual molecules) molecular weight across lakes showed that the patterns of the two are not systematically related (Kellerman, 2015). Our results indicate that in natural lake water actual molecular mass might not be important for reactivity, as opposed to other characteristics, such as compound class and, to a certain degree, nominal oxidation state of carbon (see below). The possible reasons for emerging relationship between molecular mass and reactivity in the UV treatment are discussed in section Treatment Effects.

Apparent decay coefficient k_exp was positively related to oxygen content, i.e., O/C ratio (Figure 1, Figure S3), and tended to be lowest at intermediate levels of NOSC (Figure 4, Figure S6). Figure 4 and Figure S6 demonstrate that exact relationships of k_exp with NOSC can differ between the groups of compounds. For all lakes, treatments, and replicates, reactivity of plant polyphenols increases nearly linearly with increasing NOSC (Figure 4, Figure S6). For other compound classes relationship of k_exp with NOSC was less straightforward, but demonstrated a similar tendency in many cases (Figure 4, Figure S6).

There is no absolute agreement in the literature on the exact relationship between oxygen content and reactivity of OM. Negative relationship between O/C and bioavailability of DOM (assessed as bacterial growth) has been reported by Sun et al. (1997). In contrast, based on the results of high resolution mass spectrometry analysis, Kim et al. (2006) concluded that biodegradation selectively removed compounds with higher O/C. However, there are also studies that did not find a strong connection between O/C and bioavailability of humic DOM (Hunt et al., 2000), or a direct correlation between NOSC and degradation rate coefficient of DOM compounds (Lechtenfeld et al., 2014).

There is evidence for a positive relationship between the degree of oxidation and apparent radiocarbon age of marine DOM (Lechtenfeld et al., 2014). It seems logical to assume that, unlike more reduced substrates, oxidized DOM has low energetic value for DOM decomposers (Perdue, 2009; Kattner et al., 2011). LaRowe and Van Cappellen (2011) challenge this concept and argue that removal of electrons from a substrate is thermodynamically more favorable at higher NOSC. Yet, the authors note that thermodynamic predictions may not match the observations due to complexity of factors affecting decomposition. Accordingly, in a conceptual model of DOM degradation in the oceans Williams (1999) proposes that there is a general mismatch between the susceptibility of DOM components toward microbial degradation and Gibb's free energy, i.e., apparently energetically favorable substrates appear to be recalcitrant, and vice versa.

Our results are corroborated by the findings of Kellerman et al. (2015) in the study of molecular composition of freshwater DOM across 109 Swedish lakes. This study demonstrated that degradation processes preferentially remove aromatic oxidized terrestrial compounds in lakes with longer water residence times. Considering how different our approach is from Kellerman et al. (2015), the similarity in the findings is striking, further emphasizing that terrestrially derived, predominantly aromatic DOM is a dynamic component within the organic matter field. There is also a possibility that both in our experiments and in the study by Kellerman et al. (2015) formulae with higher NOSC were partially lost through non-biological processes (e.g., aggregation and flocculation). It has been shown that compounds with higher NOSC are preferentially precipitated by Fe and Al salts (Riedel et al., 2012; Lavonen, 2015). Studies on the flocculation of DOM in lakes suggest that allochthonous colored DOM is a dominant precursor of organic matter settling onto the sediment (von Wachenfeldt and Tranvik, 2008; von Wachenfeldt et al., 2009). Although we lack the data to adequately assess the importance of non-biological pathways for our experiment, relationship between NOSC and non-biological DOM removal could be an interesting topic for the future studies.

All trends of k_exp with molecular parameters were consistent between different lakes and replicates (Figures S3–S6 in the supplement, complementing the figures displayed in the main text). Hence, our results provide additional support to the idea that patterns of DOM utilization are in principle predictable from the intrinsic properties of DOM (Seitzinger et al., 2005; Kellerman et al., 2015), but we also acknowledge that there is high variability in reactivity which limits the potential to predict it from broad molecular properties, such as element ratios. The variability in our data could be due to the fact that each molecular formula describes a large number of potential isomers (Hertkorn et al., 2007; Zark et al., 2017), and the variations in specific structure likely translate into variations in decay behavior. Despite this limitation, our method was able to capture the general trends in reactivity of different DOM groups, including the effects of UV pretreatment on distribution of apparent decay coefficients k_exp.

Between 87.7 and 97.9% of formulae found in non-manipulated treatments were also detected in the UV-manipulated treatments. Interestingly, the reactivity (k_exp) of these shared compounds was often higher in the UV-manipulated treatment (Figure 5, Figure S7). Despite a general increase in reactivity, also indicated by the shift toward higher values of k_exp in all compound groups (Figure 2, Figure S4), the recalcitrant core in the region of highly unsaturated compounds persisted in the UV-manipulated treatment (Figures 1B, 2, Figures S3, S4).

A possible reason for higher DOM reactivity in the UV treatment is that the microbial community was stimulated by labile compounds produced as a result of DOM exposure to UV, and this triggered an ability to decompose formerly “recalcitrant” compounds. This phenomenon, known as the priming effect, is commonly described from terrestrial ecosystems (Fontaine et al., 2007; Kuzyakov, 2010). However, an increasing number of studies in aquatic systems have failed to corroborate its existence in aquatic ecosystems (Bengtsson et al., 2014; Catalán et al., 2015; Dorado-García et al., 2016). The main mechanisms of priming in soils are thought to involve nutrient limitations, co-metabolism, and higher enzyme production (Blagodatskaya and Kuzyakov, 2008). All our experiments were carried out with the addition of inorganic nutrients, therefore microbes should not have been limited by the availability of specific elements. Therefore, it seems unlikely that nutrient priming is responsible for increase in compounds decomposition in UV-manipulated treatment.

DOM degradation kinetics can also depend on the concentration of the individual DOM constituents (Dittmar, 2015). For example, it was proposed that the millennium-scale stability of DOM in the deep ocean is due to extreme dilution of the individual DOM compounds, and possibly not due to structural constraints (Arrieta et al., 2015). In our experiment, initial concentration of DOC was 30% lower in the UV treatment than in the non-manipulated treatment, yet decomposition in the UV treatment proceeded faster. Previous studies also showed that overall degradation and kinetics in freshwater systems are not constrained by DOC concentration within the concentration range studied here (Eiler et al., 2003; Koehler et al., 2012). Hence, changes in structural properties of DOM are most likely to be responsible for the differences in decay rates between the treatments.

We considered that UV treatment may lead to broad compositional or structural changes of the organic matter, such as loss of aromaticity (Stubbins et al., 2010) or loss of carboxylic acid functionality (Hawkes et al., 2016b) but these types of chemical change were not apparent from average H/C or O/C ratios (respectively), and loss of intensity of individual molecular masses did not show clear patterns in these respects (Figures S8, S9). It is still possible that individual isomers were affected but such changes, but not on a broad scale. As was mentioned above, each molecular formula encompasses a large number of structural isomers (Zark et al., 2017). Therefore, while the majority of formulae identified in the UV treatment were identical to those in the non-manipulated treatment, the isomeric mixture of structures underlying each molecular formulae could differ between the treatments.

Another possible explanation for different reactivity of the same compounds in the two treatments might be that specific molecules were to a different extent entangled in colloidal or supramolecular structures. For example, soil-derived humic DOM can exist in the form of supramolecular complexes where molecules are stabilized by weak hydrophobic interactions and hydrogen bonds (Conte and Piccolo, 1999; Piccolo, 2001). Multiple studies indicated the decrease in apparent molecular weight of DOM as a result of UV exposure (Helms et al., 2008, 2014), suggesting a disruption of multiple bonds within aggregated and colloidal DOM structures.

Electrospray ionization can disperse loosely, non-covalently bound structures (Yamashita and Fenn, 1984). Accordingly, the formulae identified by FT-ICR-MS would carry no information on whether corresponding compounds were previously part of a supramolecular complex. It is possible that UV light cleaved the larger complexes into more available fragments, no longer protected from decomposition by multiple weak bonds. Supramolecular complexes of high apparent molecular weight (measured with size exclusion chromatography) can be associated with plant polyphenols and polycondensed aromatics (Kellerman, 2015). In our study, these groups of compounds increased in reactivity in the UV-manipulated treatment, which suggests that supramolecular structure could have been at least partly responsible for lower reactivity of polyphenols in non-manipulated treatment.

In the UV treatment k_exp in some cases decreased with increasing molecular mass (Figure 3F, Figure S5)–the relationship that was not observed in non-manipulated treatment (Figures 3E, Figure S5). Pretreatment with UV light likely diminished the effects of steric hindrance as molecular aggregates were broken into smaller fragments, and relationship between structure of molecules and k_exp became more apparent. Notably, in the UV treatment relationship of k_exp with molecular characteristics had more pronounced character (steeper curves of the fitted GAMs) in many cases (Figures 1, 3, 4, Figures S3, S5, S6).

Photoproduced compounds, identified as formulae unique to the UV treatment, appeared among all compound groups and covered the whole range of apparent decay coefficients k_exp (Figure 6, Figure S10). This shows that products of photoreactions are not restricted to the aliphatic region where many described photoproducts belong (Moran and Zepp, 1997) and are not necessarily highly reactive. Instead, some photoproducts can persist and act as a slowly degrading substrate on a timescale of several months.

It should be mentioned that the most reactive compounds (such as low molecular weight carboxylic acids) were not addressed in our study, as they can disappear within a few hours. It should also be taken into account that many photoproducts tend to end up below the detection limit or are inefficiently ionized (Stubbins et al., 2010; Raeke et al., 2016) and therefore might fall outside our analytical window.

While phototransformations of lake DOM has been extensively studied before (e.g., Bertilsson and Tranvik, 2000; Moran et al., 2000; Stubbins et al., 2010; Gonsior et al., 2013), the exact contribution of photoreactions to lake carbon budgets has been rarely assessed. The existing studies on small boreal lakes suggest that sunlight-induced CO₂ emission constitutes <10% of the mean annual CO₂ emission from these lakes (Koehler et al., 2014; Groeneveld et al., 2015). This fact can be related to relatively low mean annual irradiance in boreal latitudes and strong light attenuation in brown humic waters. The analysis of hundreds of freshwater systems in boreal Canada, however, showed that the increased terrestrial influence increases photodegradability of DOM (Lapierre et al., 2013), hence the contribution of sunlight-induced CO₂ emissions from lakes might increase with changing climate and increased terrestrial runoff. The carbon budgets (e.g., Koehler et al., 2014; Groeneveld et al., 2015) do not account for the effects of phototransformations on DOM bioreactivity and proneness to coagulation, as those are especially difficult to evaluate quantitatively. We here show that a complete loss of 30% of DOM after the UV exposure coincided with phototransformations that result in a systematic increase in compound loss rates trough biological and, possibly, abiotic removal (see also Mostovaya et al., 2016).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.