Trait-based models are an emerging tool in ecology with the potential to link community dynamics, environmental responses and ecosystem processes. These models represent complex communities by defining taxa with trait combinations derived from prior distributions that may be constrained by trade-offs. Herein I develop a model that links microbial community composition with physiological and enzymatic traits to predict litter decomposition rates. This approach allows for trade-offs among traits that represent alternative microbial strategies for resource acquisition. The model predicts that optimal strategies depend on the level of enzyme production in the whole community, which determines resource availability and decomposition rates. There is also evidence for facilitation and competition among microbial taxa that co-occur on decomposing litter. These interactions vary with community investment in extracellular enzyme production and the magnitude of trade-offs affecting enzyme biochemical traits. The model accounted for 69% of the variation in decomposition rates of 15 Hawaiian litter types and up to 26% of the variation in enzyme activities. By explicitly representing diversity, trait-based models can predict ecosystem processes based on functional trait distributions in a community. The model developed herein illustrates that traits influencing microbial enzyme production are some of the key controls on litter decomposition rates.
A major goal in ecology is to link the structure of biological communities with processes at the ecosystem level (Zak et al. 2003). Making this link has been a challenge due to the diversity and complexity of most communities, which may contain hundreds or thousands of members contributing to a given process (Loreau et al. 2001). Furthermore, the potential for specific organisms to contribute to ecosystem processes is often unknown or poorly characterised in most biological communities.
Theoretical approaches and models are important tools for linking community structure and ecosystem processes (Loreau 2000). Models can bridge gaps in empirical knowledge to predict how biological communities respond to environmental change at ecosystem to global scales (Bardgett et al. 2008). In addition, these approaches are potentially useful for integrating new sources of data, such as the genomic and metagenomic data sets that are now rapidly generated for many types of biological communities (Tringe et al. 2005).
Although models have an important role to play in the integration of community ecology with research on global environmental change, there are few mathematical tools available for representing biological diversity at the ecosystem to global scale. In particular, few large-scale models of ecosystem function incorporate community interactions, such as complementary resource use, facilitation and competition among species. However, ecosystem processes depend on community attributes and the biological diversity present in many systems (Loreau et al. 2001; Reich et al. 2001). Models that do represent biological diversity generally take a functional group approach, whereby organisms are assumed to belong to a small number (1–15) of functionally distinct units (Allison & Martiny 2008). Organisms within a functional group are considered to be effectively the same in the model. This approach has been used in many models of the global ocean (Moore et al. 2000; Le Quéré et al. 2005), global vegetation (Foley et al. 1996; Moorcroft et al. 2001) and litter decomposition (Moorhead & Sinsabaugh 2006).
Recent advances in trait theory may provide an alternative approach for representing the link between biological diversity and ecosystem processes (Litchman & Klausmeier 2008). Trait-based models can predict rates of ecosystem processes across environments based on the relationship between response and effect traits (Webb et al. 2010). Response traits determine how organisms respond to changes in the environment (e.g. drought tolerance), whereas effect traits determine the effect of organisms on ecosystem processes (e.g. photosynthetic capacity); some traits, particularly those involved in resource acquisition, may influence both responses and effects (Lavorel & Garnier 2002). For example, plants with high water-use efficiencies might dominate desert environments where water is scarce. However, these plants are at a disadvantage in mesic environments because high water-use efficiency trade-offs against low photosynthetic capacity (Wright et al. 2001). Thus, the environment can influence ecosystem processes (such as primary production) via correlations between response and effect traits of organisms.
As an example of this approach, Follows et al. (2007) used a trait-based model to simulate the distributions of phytoplankton taxa and patterns in oceanic primary productivity. Each taxon was assigned a set of trait values for nitrogen (N), phosphorus (P) and light acquisition as well as temperature tolerance. These trait values were drawn at random from empirically derived distributions and constrained by known trade-offs among traits. For example, larger cells were not only assumed to have higher intrinsic growth rates but also higher half-saturation constants for nutrient uptake, meaning that they were less effective at drawing down nutrients. Taxa defined in the model competed along resource gradients driven by a physical ocean model. Approximately 20 of 78 taxa initially present reached high abundances in the global ocean with different taxa dominating in different regions. Taxa with low growth rates and efficient nutrient uptake dominated the oligotrophic tropical and subtropical oceans. The physiological traits and distributions of these hypothetical taxa corresponded well with ecotypes of Prochlorococcus, which dominates phytoplankton communities in oligotrophic ocean waters at low latitudes. Furthermore, the simulated distribution of net primary productivity was qualitatively similar to satellite observations.
The objective of this article is to build on Follows et al. (2007) to develop a trait-based model of plant litter decomposition, which has been recognised for decades as an important ecosystem process (Tenney & Waksman 1929; Olson 1963). Most large-scale models represent decomposition as an exponential decay process outside the direct control of soil organisms (Friedlingstein et al. 2006; Thornton et al. 2007; Todd-Brown et al. 2012). However, the diversity, traits and biological interactions present in decomposer communities may influence litter decay (Gessner et al. 2010). Therefore, developing models that represent the biological diversity of decomposer communities could improve predictions of decomposition rates.
The model developed herein is based on the premise that litter decomposition is a microbial process dependent on extracellular enzyme production (Schimel & Weintraub 2003; Fontaine & Barot 2005; Moorhead & Sinsabaugh 2006; Allison et al. 2010). The microbial traits involved in enzyme production define a resource assimilation strategy to obtain energy, carbon (C) and nutrients from complex chemical compounds (Vetter et al. 1998). This strategy has costs and benefits that may vary with environmental conditions, relate to organismal fitness and thus be subject to natural selection. I developed the model to assess which microbial traits and trait combinations might be selected to enhance litter exploitation and hence decomposition under varying resource conditions. I also asked how interactions among microbial taxa respond to different trait values and trade-offs among a set of traits. To validate the model, I compared model predictions of litter decay rates and enzyme activities to empirical data from a field study with 15 litter types decomposing under fertilised and unfertilised conditions.
Model structure and pools
To integrate a mechanistic description of microbial decomposition with the trait distributions of diverse microbial communities, I developed the Decomposition Model of Enzymatic Traits (DEMENT). This model is spatially explicit and simulates litter decomposition driven by diverse communities of microbes producing extracellular enzymes (Table 1; Fig. 1). The user can specify distributions for traits such as enzyme production rates, enzyme kinetic parameters and C use efficiencies. DEMENT runs on a discrete daily time step and is coded in the R programming language (R Development Core Team 2011). Spatial structure is implemented by allowing microbial cells to interact on a grid with dimensions x × y where the edges contact each other, forming a torus. Spatial structure is essential for realistic representation of microbial interactions and diversity on a solid litter substrate (Allison 2005). The grid is analogous to the surface of a decomposing leaf, and multiple microbial cells may occupy the same grid box.
Table 1. Ranges and units for model parameters
Number of iterations
Number of enzymes in community
Number of substrates
Number of uptake transporters
Number of taxa
Activation energy for enzymes
Activation energy for uptake
1 × 10−6–5 × 10−6
mg enzyme day cm−3
Slope for Km–A0E relationship
Intercept for enzyme Km−A0E relationship
1 × 10−7–5 × 10−7
mg biomass day cm−3
Slope for Km−A0U relationship
Intercept for uptake Km−A0U relationship
1 × 107
mg substrate mg−1 enzyme day−1
Minimum Arrhenius constant for enzymes
1 × 108
mg substrate mg−1 enzyme day−1
Maximum Arrhenius constant for enzymes
1 × 107
mg substrate mg−1 biomass day−1
Minimum Arrhenius constant for uptake
1 × 108
mg substrate mg−1 biomass day−1
Maximum Arrhenius constant for uptake
Fractional change in cellulose decay per unit lignocellulose index
Minimum number of enzymes capable of degrading each substrate
Minimum number of uptake transporters capable of taking up each monomer
Maximum number of enzymes a taxon may produce
Coefficient determining strength of specificity-efficiency trade-off
Intercept for C use efficiency trade-off
Slope for C use efficiency trade-off
Minimum per enzyme C cost as a fraction of uptake
Maximum per enzyme C cost as a fraction of uptake
mg mg−1 day−1
Minimum per enzyme C cost as a fraction of biomass
1.33 × 10−5–28 × 10−5
mg mg−1 day−1
Maximum per enzyme C cost as a fraction of biomass
Per enzyme N cost as a fraction of C cost
Per enzyme P cost as a fraction of C cost
Per enzyme biosynthetic cost as a fraction of C cost
mg mg−1 day−1
Per transporter uptake C cost
mg cm−3 day−1
Input rate of substrate
mg cm−3 day−1
Input rate for monomers
mg cm−3 day−1
Input rate for NH4+ + NO3−
mg cm−3 day−1
Input rate for PO43−
Enzyme turnover rate
Biomass turnover rate
Initial monomer present as a fraction of initial substrate
Initial NH4+ + NO3− concentration
Initial PO43− concentration
Initial enzyme concentration
Initial microbial biomass concentration per taxon per grid box
Initial microbial N fraction
Initial microbial P fraction
Minimum fraction of biomass as C
Minimum fraction of biomass as N
Minimum fraction of biomass as P
Threshold C concentration for cell death
Threshold N concentration for cell death
Threshold P concentration for cell death
Maximum cell C concentration before reproduction occurs
Probability of a taxon occupying a grid box
Maximum dispersal distance
In the community, there are NB microbial taxa and NE types of extracellular enzymes that degrade polymeric substrates to produce organic monomers. NS is the number of polymeric substrates, which in the current model version includes 10 compounds often found in plant litter and 2 additional substrates representing inactive enzymes and dead microbial biomass (see Table S1 in Supporting Information). Each substrate is associated with one organic monomer, and there are two additional monomer-like pools representing inorganic N (NH4+ + NO3−) and P (PO43−). Polymeric substrates, monomers and inorganic nutrients enter the grid at rates IS, IM, IN and IP, although IS and IM were set to zero for all runs reported herein. Initial C and nutrient concentrations are given in Table S1, and their associated initial monomer concentrations are represented as a fraction FMS of the initial substrate concentration. These values are based on the average chemistry of leaf litter from the validation data set (see below). A fraction L of the mineral nutrient pools are leached from the grid at every time step.
Stoichiometry and turnover
All chemical litter compounds contain a defined combination of C, N and P with stoichiometric transfers between pools of different compounds. The stoichiometry of substrates and monomers is fixed at the initial values derived from Table S1. Monomers are assumed to reflect the stoichiometry of their associated polymers, which represents a simplification for P-containing compounds since phosphatases can release inorganic P independent of polymer hydrolysis (McGill & Cole 1981). The initial stoichiometry of microbial biomass is specified based on the initial biomass C concentration B0 and the N and P fractions BN and BP. Microbial biomass stoichiometry can then vary freely so long as biomass C, N and P fractions remain above FC, FN and FP. If the biomass fractions fall below these thresholds, excess elements are eliminated through mineralisation of CO2 or inorganic nutrients. This formulation prevents microbes in the model from developing unrealistic stoichiometric ratios while allowing for some flexibility in elemental composition. A substrate pool representing dead microbial material has variable stoichiometry dependent on the dying biomass. The stoichiometry of all enzymes is fixed by the proportions ZEN and ZEP for N and P. A substrate pool of inactive enzymes retains this stoichiometry, which is a simplification because some elements, such as P, are lost faster than others. Constant stoichiometry across all enzymes is an additional simplification because microbial proteins involved in the assimilation of a given element may contain significantly lower concentrations of that element; however, the average effect on element concentration is only 5–10% (Baudouin-Cornu et al. 2001). Enzyme inactivation and microbial biomass turnover are first order processes with rate constants τE and τB. Microbes with biomass C or nutrient contents below Cmin, Nmin or Pmin are assumed to be too small to survive and are transferred to the dead microbial pool.
Substrates are converted to monomers through the activity of extracellular enzymes. Each substrate is assumed to be degraded by at least ES randomly chosen extracellular enzymes. Conversely, each enzyme must degrade at least one substrate to prevent microbes from producing enzymes with no function. These rules mean that in situations where NE > NS, some (or all) of the substrates are degraded by more than one enzyme, and the choice of redundant enzymes is random. After meeting the criteria that all enzymes degrade at least one substrate and all substrates be degraded by at least one enzyme, enzymes are randomly assigned to substrates with replacement, so a single enzyme may act on multiple substrates. Some (or all) enzymes must degrade multiple substrates if NE < NS. For the parameters used in this study, only a minority of the enzymes acted on one or occasionally more additional substrates.
Substrate decay rate Dij is a function of enzyme and substrate concentrations as specified by the Michaelis-Menten relationship:
where i = [1… NE], j = [1… NS ], (Vmax)ij is a matrix of maximum catalytic rates, Ei is a vector of enzyme concentrations, Qij is a binary matrix specifying which enzymes degrade which substrates (0 = no degradation, 1 = degradation), Sj is a vector of substrate concentrations and (Km)ij is a matrix of half-saturation constants. Since some substrates may be degraded by multiple enzymes, the elements of Dij are summed across enzymes to determine a vector DNs of total decay rates for each substrate. For simplicity, the model does not constrain the maximum rate of substrate decay, even though binding sites for enzymes could become limiting at high enzyme concentrations (Schimel & Weintraub 2003). However, decay cannot exceed the amount of substrate present in a given grid box, regardless of enzyme concentration. As in the Guild Decomposition Model (GDM; Moorhead & Sinsabaugh 2006), DEMENT assumes that lignin shields cellulose such that cellulose decomposition depends linearly on the lignocellulose index (LCI) according to the relationship: , where λSlope represents the fractional decline in decay rate per unit LCI.
The (Vmax)ij values for substrate degradation are based on the temperature-dependent Arrhenius equation:
where R is the ideal gas constant in kJ mol−1 K−1, T is the temperature in Kelvin, (EaE)j is a vector of activation energies for each substrate (Table S1), and (A0E)ij is a matrix of values in units of mg product mg−1 enzyme day−1 chosen at random from a uniform distribution with range A0Emin to A0Emax. For a given temperature and activation energy, the values of (A0E)ij are directly proportional to the enzyme catalytic constant, or number of reactions catalysed per unit time by each active site. The range for A0Emin to A0Emax given in Table 1 corresponds to the low end of reported catalytic constants for extracellular enzymes (Scheer et al. 2011). Activation energies for soil hydrolytic enzymes vary from 13 to 94 kJ mol−1 with most values in the range of 20–50 kJ mol−1 (McClaugherty & Linkins 1990; Trasar-Cepeda et al. 2007).
Monomer uptake requires transporters that are associated with microbial biomass, and there are NU transporters in the model (10 for this study). The algorithm for assignment of transporters to monomers (including inorganic N and P) is similar to the assignment of enzymes to substrates. Analogous to the extracellular enzymes, the assignment algorithm ensures that each transporter binds to at least one monomer, and each monomer corresponds to at least one transporter. Likewise if NU > NS + 2, some (or all) of the monomers are taken up by more than one transporter, and if NU < NS+ 2, some (or all) transporters must bind to multiple monomers (the number of monomers = NS + the two inorganic nutrient pools).
Monomer uptake rate Uik is also expressed as a Michaelis–Menten function:
where i = [1… NB], j = [1… NU], k = [1… NS + 2], Bi is a vector of microbial biomass for each taxon, GUij is a matrix of transporters possessed by each taxon, QUjk is a binary matrix specifying which transporters take up which monomers, and Mk is a vector of monomer concentrations. (VmaxU)jk is a matrix calculated similar to Eq. (2) except with a constant activation energy of EaU = 35 kJ mol−1, and (KmU)jk is a matrix of half-saturation constants for uptake. The elements of Bi (taxon biomasses) are multiplied by the elements of GUij (transporters possessed by each taxon) to generate a matrix containing the mass of every transporter in every taxon. The quotient in Eq. (3) generates a matrix containing the potential uptake of every monomer by every transporter (per unit transporter mass). These two matrices are then multiplied to generate the Uik matrix of actual uptake rates of each monomer by each taxon, summing over all j transporters. These calculations mean that monomers are taken up in proportion to taxon biomass, and a given taxon cannot take up a monomer unless it expresses the associated transporter(s). All taxa are assumed to possess the transporters for inorganic N and P. If the monomer pool size is limiting, the available monomer is partitioned among taxa in proportion to their calculated uptake rates.
Microbial growth is represented implicitly as the difference between uptake and loss processes. Microbes within a given grid box have the first opportunity to take up monomers, which gives an advantage to microbial biomass located in grid boxes with high enzyme activity. Any excess monomer is then assumed to instantaneously diffuse and equilibrate across the entire grid, since diffusion is rapid relative to the daily time step of the model. Immediately upon uptake, microbes respire a fraction of monomer C equal to 1 – ε, where ε represents C use efficiency. This fraction represents the energy expended during monomer metabolism and for simplicity is assumed to be constant across all monomers. Each microbial taxon has the capacity to produce n extracellular enzymes, where n is drawn from a uniform distribution in [0, Emax]. Emax represents the maximum number of enzymes a taxon may produce (which may be less than the total number in the community, NE). The number of substrates accessible to a taxon may exceed n because some enzymes may target multiple substrates. Microbial taxa produce enzymes by allocating a fraction of biomass and monomer uptake to enzyme production. The biomass fraction represents constitutive enzyme production. For each enzyme produced, the biomass and uptake fractions are chosen at random from uniform distributions with range βECmin to βECmax or ZECmin to ZECmax, respectively, thus allowing taxa to vary in their rates of enzyme production. The amount of N and P allocated to enzymes is determined by enzyme stoichiometry (ZEN, ZEP), and microbes respire additional C to supply energy for enzyme biosynthesis in proportion to ZEB.
Taxa that produce an extracellular enzyme are always assumed to be able to take up the monomer produced by the enzyme (Tettelin et al. 2001). If a taxon produces no enzymes, it is assigned one organic monomer transporter at random so that it can acquire C; thus all taxa can take up C, N and P. Taxa are also assigned a random number (including zero) of additional transporters beyond those linked to the taxon's enzymes. These rules allow for a range of “cheater” strategies whereby microbes benefit from the enzymes of other taxa (Allison 2005). The daily maintenance cost for each uptake transporter is ZUC, expressed as a fraction of microbial biomass C.
When the biomass of a given taxon in a grid box reaches the threshold Bmax, it splits into half and disperses to a grid box in any direction up to dispersal distance δ. The original grid box (δ = 0) may be chosen, but the biomass will then be subject to division again during the next time step. Dispersal is independent of whether or not the destination grid box is occupied.
Specificity-efficiency – If the specificity-efficiency trade-off is implemented, the Vmax of each enzyme activity from Eq. (2) is recalculated as:
where θ is a coefficient that determines the slope of the trade-off and SE is the total number of substrates degraded by the enzyme. The rationale behind this trade-off is that greater specialisation on a given chemical substrate in the enzyme active site would reduce activity on other substrates, whereas active sites of more promiscuous enzymes should accommodate a broader range of substrates with lower efficiency (McLoughlin & Copley 2008). Km values are not adjusted for the change in Vmax, which reduces the Vmax/Km ratio.
Vmax−Km—There is some evidence that Vmax correlates negatively with binding affinity (i.e. positively with Km), but the form of this relationship is uncertain (Somero 1978; Stone et al. 2012). If the Vmax−Km trade-off is implemented, Km values (for enzymes and/or uptake transporters) are calculated as a linear function of (A0)ij with intercept KmInt and slope KmSlope. Vmax is proportional to (A0)ij, thus introducing a trade-off between enzyme velocity and binding affinity.
Carbon use efficiency-enzyme—If this trade-off is implemented, microbes that produce more extracellular enzymes are assumed to have higher ε according to the function:
This relationship represents the possibility that microbes investing in enzyme production may be selected to use resources more efficiently. Due to associated energetic costs, increasing levels of enzyme production in the model will reduce ε within a taxon (del Giorgio & Cole 1998). However, Eq. (5) assumes that evolution has selected for taxa that compensate for these energetic costs by increasing assimilation efficiency, an adaptation consistent with the low growth rates of some enzyme-producing taxa (Schmidt & Konopka 2009).
The grid was initialised with a homogeneous distribution of each substrate and monomer. Each grid box was colonised at random by each taxon with probability pT = 0.01, meaning that each taxon occupied ~ 100 grid boxes on the 100 × 100 grid. Each taxon in an occupied grid box was initialised with biomass C concentration B0 = 1 mg cm−3, such that the average biomass density across the grid was ~ 1 mg cm−3 with 100 taxa and pT = 0.01.
One set of simulations examined the model sensitivity to each trade-off by varying the strength of the trade-off and comparing the results to simulations with no trade-offs (Table S2). All other simulations were run by default with “weak tradeoffs” where all trade-offs were implemented at their lowest non-zero values. To test for the effect of substrate quality on decomposition, simulations were run with doubled N concentrations or doubled lignin concentrations in litter (Table S1). Concentrations of other substrates were reduced proportionately to keep the total litter C concentration constant. Another set of runs varied the maximum enzyme production rate from 0.0005 to 0.0105 mg enzyme C mg−1 uptake C. Maximum constitutive enzyme production was varied in proportion to these values (Table 1). To test whether enzyme or taxon diversity was limiting decomposition, the number of enzymes in the community was increased to 50 for one set of runs, and the number of taxa was increased to 200 in a second set of runs. For the second set of runs, the colonisation probability of each taxon was reduced to 0.005 to keep overall density constant. Increased enzyme diversity and the Vmax–Km trade-off for uptake had no effect on decomposition, so these results are shown in Table S3 and are not discussed further.
Although most simulations were run for 500 days, runs with weak trade-offs were continued for 9 more 500-day pulses of litter input and decomposition to represent density-dependent microbial colonisation of new litter. The period of 500 days is somewhat arbitrary but corresponds to the time at which mass loss approaches completion. Each pulse started by setting substrate, monomer and enzyme concentrations back to their initial values. The average initial microbial biomass density across the grid was reset to ~ 1 mg cm−3, and the colonisation probability for each taxon was proportional to its maximum density in the previous litter pulse. Because colonisation is a discrete random process, taxa present at very low densities were generally lost from the community during subsequent litter pulses. These taxa were not permitted to re-invade the microbial community after being lost.
DEMENT was validated by comparing predictions of litter decomposition rates and enzyme activities to data on 15 fertilised and unfertilised litter types from the Hawaiian Islands (Allison & Vitousek 2004). The model was initialised with weak trade-offs, and substrate chemistry was specified using ash-free C fractions and nutrient concentrations from the empirical study (Appendix S1). Parameters for uptake-induced and constitutive enzyme production were selected to generate a decay coefficient matching the average value for 11 unfertilised leaf litter types in the validation data set. This parameter optimisation was performed with the average substrate chemistry of these 11 litter types. Using the initial substrate conditions, simulations were run to predict decomposition rate constants and enzyme activities. Predicted enzyme activities were calculated as the sum of the average concentration of each enzyme contributing to substrate degradation, multiplied by its Vmax. This metric was chosen to be comparable to the empirical measurements of enzyme activity, which used high concentrations of soluble substrate analogues and therefore approximate enzyme concentrations rather than in-situ activities. The model substrates targeted by each enzyme were defined as cellulose for cellobiohydrolase, chitin for N-acetyl-glucosaminidase, lignin + dead microbial biomass for polyphenol oxidase, and phospholipids and nucleic acids for acid phosphatase. Predicted enzyme activities were calculated for each of three sampling dates in the empirical study. Since the empirical study also included an N + P fertilisation treatment, the simulations were repeated with initial inorganic pools of 0.5 mg N cm−3 and 0.5 mg P cm−3, and inputs of 0.00274 mg cm−3 day−1 for both N and P. These values were calculated to match the nutrient addition rates in the field study.
All model scenarios and validation runs were replicated at least eight times, and the decomposition rate was determined based on the loss of C from all initial substrate and monomer pools, not including microbial by-products (i.e. C in dead biomass and inactive enzymes). Decay coefficients were calculated as the slope of a line fit to ln(mass remaining) vs. time by least-squares linear regression with the intercept forced through ln(initial mass). Since decomposition sometimes levelled off before 500 days, the regression was only fit to time points with < 90% of the mass loss achieved at 500 days. Relationships among variables and pairwise spatial associations of taxa across the grid were tested using the Pearson correlation coefficient. I also calculated a weighted average spatial association for each taxon with all other taxa, where the weights were proportional to the number of mutually occupied grid boxes. Tests of spatial association were conducted at the time when total enzyme biomass reached its peak on the grid because enzyme biomass is a good index of integrated microbial activity and interaction. The effect of model scenario on decomposition rates and correlation coefficients was tested using one-way analysis of variance. Model predictions were compared to validation data using least-squares linear regression, and all statistical tests were run in R.
Patterns in litter decomposition
With default (weak trade-offs) parameter values, the model predicted plausible patterns in decomposition and microbial community dynamics. Total litter mass and the mass of litter chemical components declined in an exponential pattern over time (Fig. 2). Most model runs also predicted the accumulation of dead microbial biomass at later stages of decomposition, consistent with the formation of microbial by-products that may contribute to soil organic matter formation. Most taxa increased in abundance during the model runs, but the timing of peak abundance differed across taxa (Fig. 3a). Taxa with more enzymes tended to have greater maximum densities that were reached later during decomposition (Fig. 3b,c). Microbial biomass stoichiometry was reasonable, with predicted C:N ratios of 5–9 and N : P ratios of 6–11, consistent with C:N ratios of 8.6 and N : P ratios of 7 observed in a recent meta-analysis (Cleveland & Liptzen 2007). However, N:P ratios were low (~ 2) in some validation model runs with N + P fertilisation, reflecting luxury uptake and storage of P. Doubling litter N concentration had no effect on decomposition, whereas doubling litter lignin concentration significantly reduced decomposition rates from 2.80 ± 0.35 year−1 to 1.39 ± 0.13 year−1 (Table 2; Table S3).
Table 2. Analysis of variance statistics and response direction (+ or −) for tests of model sensitivity to trade-offs or different parameter scenarios
The model performed reasonably well in predicting empirically determined decomposition rate constants with an overall R2 of 0.69 and a slope of 1.55 ± 0.19, meaning the model under-predicted decomposition rates (Table 3, Fig. 4a). For unfertilised litter, the R2 improved to 0.80 and the slope declined to 1.35 ± 0.18, whereas predictive power was lower for fertilised litter (Table 3). Similar to the empirical study (Allison & Vitousek 2004), the model predicted a nearly significant increase in decomposition rate with nutrient fertilisation across all litter types (F1,210 = 3.85, P =0.051, anova). However, the model outputs also showed a significant fertilisation×litter type interaction (F14,210 = 2.17, P <0.01) whereby slow-decaying litter responded positively to fertilisation and fast-decaying litter responded negatively. This interaction was not significant in the empirical study, but there was a tendency for litter types with low and intermediate decay rates to show stronger positive responses to fertilisation.
Table 3. Regression statistics for empirically measured vs. model-predicted decomposition rate constants for 15 unfertilised and fertilised litter types from Hawaii
Intercept ± SE
Slope ± SE
−0.98 ± 0.53
1.55 ± 0.19
−0.80 ± 0.52
1.35 ± 0.18
−1.55 ± 0.99
1.93 ± 0.38
Model performance in predicting enzyme activities varied by enzyme and was weaker than for decomposition rates (Table 4). For the cellulose-degrading enzyme cellobiohydrolase, the overall R2 was 0.17 and slightly higher (0.22) for unfertilised litter (Fig. 4b). R2-values for the other enzymes ranged from 0.04 to 0.26, with greater predictive power in unfertilised than in fertilised litter (Table 4). There was no obvious variation in model performance by sampling date, meaning the model predicted enzyme activities with roughly equal skill throughout decomposition.
Table 4. Regression statistics for empirically measured vs. model-predicted extracellular enzyme activities for 2–3 time points in 15 unfertilised and fertilised litter types from Hawaii
Allocation to extracellular enzyme production
Varying the maximum rate of enzyme production had a significant impact on decomposition rates and patterns in microbial abundance (Table 2). Decay coefficients increased nearly linearly with increasing enzyme production before levelling off at ~ 8 year−1 (Fig. 5a). When the maximum rate of enzyme production was < 0.005 mg C mg−1 uptake C, taxa with a greater number of enzymes (and greater enzyme production rates) dominated the community (positive correlation between maximum taxon density and number of enzymes, Fig. 5b). Above this level, taxa with fewer enzymes were more abundant in the community. Thus, high individual rates of enzyme production are advantageous when enzyme concentrations are low, but disadvantageous when enzyme levels are high.
Prediction of community interactions
The densities of taxa in competition should correlate negatively where they overlap, whereas the reverse should be true for taxa that facilitate one another. The taxa represented in DEMENT often showed evidence of facilitation, with the average spatial association between one taxon and all others being positive or zero (Fig. 6, black squares). However, the strength and direction of pairwise interactions among taxa varied widely in the simulations. At very low rates of enzyme production, most interactions were weak because overall densities were low. Only abundant taxa that produced many enzymes showed positive associations (Fig. 6a). At slightly higher rates of enzyme production, microbial densities and interaction strength increased, such that all taxa showed positive associations on average. This change was largely driven by a disproportionate increase in positive interactions with less-abundant cheater taxa (Fig. 6b). As enzyme production increased further, the relative abundance of enzyme producers declined, and they showed strong positive associations with other producer taxa and neutral or negative associations with cheater taxa (Fig. 6c). Thus, producer taxa not only facilitate each other but also compete with abundant cheaters for reaction products. At very high levels of enzyme production, producers decline in density due to high enzyme costs (Fig. 6d). Taxa with intermediate levels of enzyme production show strong positive interactions with each other. Under these conditions, cheaters show weak associations with other taxa, likely because monomer availability is high across the grid and they need not overlap with enzyme producers.
Responses to traits and diversity
Model predictions were sensitive to the Vmax−Km and enzyme specificity-efficiency trade-offs but not the ε-enzyme trade-off (Table 2). Increasing the Vmax−Km slope by a factor of five had a significant negative effect on taxa producing more enzymes, resulting in a 70% decline in decomposition rate (Table S3). Strengthening this trade-off reduced overall monomer production, which tended to reduce cheater growth and interaction with other taxa (Fig. S1b). Increasing the specificity-efficiency slope by a factor of 10 resulted in a 44% decline in decomposition rate (Table S3) and also reduced monomer availability and cheater interactions (Fig. S1c).
Increasing the number of taxa in the model to 200 did not affect decomposition rates (Table 2; Table S3), but did alter community interactions. The maximum density of enzyme producers declined slightly relative to cheaters (Table 2), but interactions between cheaters and other taxa weakened, similar to the Vmax–Km trade-off (Fig. S1d). Weak interactions are likely due to the 50% lower initial densities of all taxa in the 200 taxon simulation.
When the microbial community was sorted through 10 litter pulses, taxon richness declined by 20–30% and decomposition rates nearly doubled. The taxa lost from the community generally produced few enzymes. The remaining taxa (mainly enzyme producers) showed weak positive spatial associations (Fig. S2). Although there was no complement of enzymes that clearly conferred an advantage to the dominant taxa (Fig. S3), there was a weak but significant positive correlation between maximum taxon density and average enzyme Vmax during the first litter pulse (r =0.16, P <0.001; Table S3). This relationship disappeared by the tenth pulse, suggesting that the taxa with favourable average Vmax values had come to dominate the community.
DEMENT predicts that decomposition rates and community interactions depend strongly on microbial enzyme traits. Like earlier models (Schimel & Weintraub 2003), DEMENT assumes that decomposition requires microbial extracellular enzymes. Because they are involved in resource acquisition, enzymes represent both response and effect traits – they determine how microbes respond to different resource availabilities, and they affect rates of substrate degradation. Thus, any traits or trade-offs related to the costs or benefits of enzyme production should impact microbial decomposition. Except for C use efficiency, all the microbial traits and trade-offs examined did affect litter decomposition or community interactions (Table 2). In particular, manipulating the rate of enzyme production illustrates how decay rates and cheater-producer interactions can change dramatically with increasing allocation to enzymes (Figs 5 and 6).
The predictions from DEMENT can be used to evaluate which response traits may be selected for under different conditions. For example, producing a broad range of enzymes is advantageous if total enzyme production in the community is low (Fig. 6a). This strategy remains favourable even as cheater populations and interactions increase (Fig. 6b), and after 10 litter pulses, many cheaters disappear from the community (Fig. S2). In contrast, high rates of enzyme production in the community favour cheating and drive broad-spectrum enzyme producers to low densities (Fig. 6d). These community dynamics are reflected in the rate of litter decomposition, which increases linearly at low enzyme production rates and saturates at higher rates (Fig. 5a). Enzyme production rates are difficult to measure in the field, but laboratory studies suggest that DEMENT's parameters may be realistic. Protease production by Pseudomonas fluorescens bacteria consumes ~ 0.5% of uptake (S. Allison, unpublished data), and protease yields of 0.3–0.9% have been found for Bacillus growing on glucose in continuous culture (Christiansen & Nielsen 2002).
The impact of trait trade-offs can also be understood by examining enzyme costs and benefits. Strengthening the enzyme specificity-efficiency and Vmax–Km trade-offs had the predictable effect of reducing decomposition rates (Table 2) due to reductions in enzyme efficiency or binding affinity. Although enzyme producer densities declined, these reductions in enzyme effectiveness actually had a stronger negative impact on cheater taxa (Fig. S1). By making enzymes less effective, these trade-offs reduced overall monomer production in the system, leaving fewer resources available for cheaters to exploit.
Although trait-based models can be useful for predicting an optimal suite of traits in a given environment (Follows et al. 2007), DEMENT shows that traits with ecosystem effects can feed back on the resource environment and alter the predicted optimal responses. In DEMENT, enzymatic traits determine resource availability for the whole community, and facilitation is common, as the average interaction between any two taxa tends to be positive or neutral. However, both competitive (negative) and facilitative (positive) spatial associations contribute to these averages (Fig. 6). In a community context, the response of producing more enzyme only makes evolutionary sense if total enzyme production is low. If enzyme production is high, taxa with few or zero enzymes can persist and grow to high abundance.
On most leaf litter substrates, facilitation likely reduces selection for a particular suite of extracellular enzymes, since resources can be acquired by taking up monomers released by other taxa. However, there was evidence from the model validation runs that certain enzymes are associated with higher densities on stoichiometrically imbalanced litter substrates. For instance, on some fern stem litter that was very nutrient poor, only taxa that produced a small number of nutrient-acquiring enzymes reached high relative densities (e.g. correlation between maximum density and number of enzymes = −0.48 ± 0.04 for Dicranopteris stems; Table S3).
Although there is clearly room for improvement, the model validation suggests that ecosystem processes such as plant litter decomposition may be predictable based on a mechanistic representation of diverse microbial communities and traits. DEMENT explained 69% of the variation in litter decay rates across 15 Hawaiian litter types with a moderate under-prediction of the actual rates. This explanatory power is similar to the R2 value of 0.72 obtained from a multiple regression model based on substrate chemistry in the original empirical study (Allison & Vitousek 2004). DEMENT's agreement with measured enzyme activities was not as good, although it explained a significant amount of variation for most of the enzymes. Microbial community dynamics and the accumulation of microbial by-products were also reasonable.
Simulated decomposition rates were sensitive to nutrient availability and litter C chemistry. Fertilisation with N + P increased decomposition rates in the validation runs on Hawaiian litter. DEMENT also predicted an interaction whereby high-turnover litter types decayed more slowly under fertilisation due to increased dominance by cheater taxa. This result is surprising because N addition is usually thought to inhibit the decomposition of more recalcitrant litter types, particularly in later stages of decay (Berg 1986; Carreiro et al. 2000). There was no evidence for nutrient inhibition in the 15 Hawaiian litter types used for model validation (Allison & Vitousek 2004), although many litter types did not respond to the treatment. Given that the effect of N on litter decomposition is controversial and inconsistent (Fog 1988; Carreiro et al. 2000; Hobbie 2005; Knorr et al. 2005), DEMENT's unusual predictions warrant additional study.
Challenges and data needs
Trait-based microbial models offer great promise for predicting ecosystem processes, but they will benefit from more information on response and effect traits, as well as the biochemical properties of enzymes. DEMENT is very sensitive to the relative costs and benefits of enzyme production, and these parameters are poorly known in microbial communities. Although genomic tools can reveal the number of enzyme genes that different taxa possess (Hess et al. 2011), it is essential to quantify the cost of enzyme production in terms of biomass and energy expenditure. Whether enzyme production is best represented as a fraction of uptake, or biomass or both must also be confirmed. Fortunately, there are a number of existing approaches that could be used to measure these quantities. Protein isolation and purification techniques can provide an estimate of enzyme biomass that could be tracked at different points during growth of an enzyme-producing culture (Liao & McCallus 1998). A complementary approach is the use of enzyme knockout strains to assess differences in biomass, growth, and fitness with and without a particular enzyme (Worm et al. 2000). Advances in metaproteomics offer the potential to quantify the expression of enzyme genes by specific microbial taxa in environmental samples (Schneider et al. 2010), and estimates of enzyme turnover could be improved by adding small quantities of purified microbial enzymes to decomposing litter and monitoring the loss of excess activity over time (Allison 2006).
More information is also needed on extracellular enzyme traits, such as Vmax, Km and substrate specificity. Data on these traits would help constrain the slopes of the trade-off relationships represented in DEMENT. Ideally Vmax and Km could be determined using complex substrates that are more ecologically relevant than many of the substrate analogues currently used to assay enzyme activities in soil and litter (Wallenstein & Weintraub 2008). For example, fluorescence-based assays using polymers as substrates have been developed (Arnosti 2004). These assays would have to be applied to a range of purified enzymes in environmental samples. Varying the substrate type could then give insight into the degree of substrate specificity.
Although DEMENT focuses on enzyme traits, there are other microbial traits that could also be relevant for predicting decomposition. For example, many microbes produce antibiotics that could generate antagonistic interactions between taxa (Williams & Vickers 1986; Raaijmakers et al. 2002). Fukami et al. (2010) found that antagonistic interactions among fungi reduced wood decomposition rates, although they did not determine if antibiotics were directly involved. Taxonomic differences in stoichiometric flexibility could also influence microbial responses to environments with differing nutrient supply rates or ratios (Makino et al. 2003; Bragg & Wagner 2007; Franklin et al. 2011).
DEMENT represents only unicellular microbes, but organisms with other growth forms also play key roles in litter decomposition. For instance, invertebrate animals are known to facilitate decomposition through litter fragmentation and digestion (Hieber & Gessner 2002). Filamentous fungi are critical decomposers in most ecosystems and possess traits distinct from unicellular bacteria, including large cell sizes and the ability to translocate nutrients through hyphal networks. Some groups of bacteria such as actinomycetes may also translocate nutrients through filamentous structures. However, the importance of trait differences between unicellular bacteria and filamentous microbes for decomposition remains unresolved (Strickland & Rousk 2010).
Trait-based ecology is emerging as a powerful framework for linking community structure with ecosystem processes (McGill et al. 2006). Making this link has been a major but elusive goal in microbial ecology, especially with the advent of new molecular approaches for assessing microbial genetic and metabolic diversity (Green et al. 2008). This goal has been hard to achieve because theoretical approaches have been unable to represent the complexity of microbial community structure and function. The results in this article demonstrate how trait-based models could be useful in bridging the gap between community structure and the process of litter decomposition. In addition, they reveal how microbial strategies for resource acquisition depend on interactions with other taxa and qualitative differences in resource supply.
Predictions of trait distributions and ecosystem processes from DEMENT, which embraced the complexity of microbial decomposer communities, should be testable with emerging molecular, physiological and biochemical data from microbial studies. Although challenges remain in extrapolating to field conditions, classical biochemical and physiological studies are important for assessing the response and effect traits of different microbial taxa and evaluating the strength of trade-offs among traits such as enzyme Vmax, Km and microbial C-use efficiency. These trade-offs as well as others omitted here ultimately constrain the distribution of traits that determine community dynamics and related ecosystem processes. The ultimate test of trait-based models will be to match their predictions with a growing body of empirical measurements in a variety of field conditions (Dinsdale et al. 2008). By embracing the complexity of microbial communities, trait-based models now make this test possible.
I thank Kathleen Treseder, Henri Folse, Matt Wallenstein and three anonymous referees for comments that improved the manuscript. This work was funded by the Office of Science (BER), US Department of Energy and the NSF Advancing Theory in Biology program.
SDA developed the model, conducted the analyses and wrote the manuscript.