Plant traits explain variation in symbiotic nitrogen fixation responses to global nitrogen enrichment: a meta-analysis

Abstract

Anthropogenic nitrogen enrichment is widely expected to suppress symbiotic nitrogen fixation in terrestrial ecosystems. Nevertheless, observed symbiotic nitrogen fixation responses remain incompletely explained by exogenous nitrogen inputs, climate, and edaphic factors. In this meta-analysis, we integrate 908 globally distributed field measurements to identify the key predictors that improve simulation of symbiotic nitrogen fixation responses to nitrogen enrichment. On average, symbiotic nitrogen fixation declines by 33.0% upon nitrogen enrichment, with the reduction being more pronounced in non-croplands than croplands. Models considering only environmental factors overestimate symbiotic nitrogen fixation decline relative to observations. The better performance of plant traits like plant growth and biomass allocation (shoot:root ratio) partially buffer symbiotic nitrogen fixation suppression under nitrogen enrichment. Integrating both environmental factors and plant performance traits improves predictive accuracy of symbiotic nitrogen fixation responses by 42.7% and brings the simulated symbiotic nitrogen fixation reductions into close agreement with observations. The alterations in plant performance traits are thus critical for explaining variability in terrestrial symbiotic nitrogen fixation responses, and incorporating plant trait dynamics in Earth System Models can quantitatively partition the compensatory symbiotic nitrogen fixation supported by nitrogen-fixing plant growth from the direct negative impact of nitrogen inputs.

Anthropogenic nitrogen enrichment reduces symbiotic nitrogen fixation, but environmental factors alone cannot explain this response. This meta-analysis shows that integrating both environmental factors and plant performance traits improves the predictive accuracy of symbiotic nitrogen fixation responses by 42.7%.

Introduction

Symbiotic nitrogen (N) fixation (SNF), carried out by plant-microbe symbionts (e.g., rhizobia and Frankia), sustains critical terrestrial ecosystem functions, including biogeochemical N cycling and gross primary production^1,2. Globally, SNF replenishes 50-70 Tg N yr⁻¹ in croplands^3,4 and 31-66 Tg N yr⁻¹ in non-croplands^1,4. Anthropogenic N enrichment, especially via atmospheric deposition and synthetic fertilizer inputs, has markedly altered SNF dynamics^5,6. Consistent with carbon-economic theory, the declines in SNF under exogenous N sufficiency are confirmed, particularly for facultative fixers^7,8. Yet the observed magnitude of N-induced SNF suppression remains poorly predicted by environmental factors alone⁹. In particular, substantial taxon-specific variation persists under comparable conditions^10–12, implicating unaccounted plant traits (e.g., performance traits) as key regulators. However, the role of plant performance traits in predicting SNF responses to N enrichment remains largely overlooked, particularly at large spatial scales. A comprehensive understanding of SNF responses and their predictors is essential to project terrestrial N budgets accurately, thus informing ecosystem functionality assessments under anthropogenic change.

Environmental factors, including exogenous N inputs, climate, and edaphic conditions, have been extensively investigated for their roles in regulationg SNF responses^9,13. Carbon-economic theory posits that exogenous N inputs (e.g., nitrate and ammonium enrichment) suppress legume SNF, as direct N uptake is energetically cheaper than maintaining N-fixing symbionts^14,15. Accordingly, legumes in N-rich conditions preferentially assimilate soil bioavailable N while downregulating SNF^16,17, thereby weakening plant-microbe mutualism under N enrichment¹⁸. This suppression operates mechanistically through reduced rhizobia colonization, diminished rhizobia populations, and curtailed nodule biomass^17,19,20. Climatic and edaphic factors can also modulate SNF sensitivity to N enrichment. For instance, SNF declines upon N enrichment were less pronounced at high mean annual temperatures⁹, while elevated soil carbon: N ratios and pH buffer diazotrophs against N inhibition¹³ and consequently attenuate SNF declines. Despite the importance of environmental factors, these factors explain only 2-33% of observed variation in SNF responses upon N enrichment^9,13, leaving a large proportion of variability attributable to unidentified predictors.

Plant performance traits emerge as critical predictors of SNF responses to environmental change. Biomass allocation patterns substantially contribute to observed SNF variability²¹, e.g., in Glycine max, SNF increases by 35% as shoot biomass rises from 842.9 g m⁻² to 1200.2 g m⁻²²², while root biomass consistently correlates with SNF magnitude²¹. Higher shoot:root ratios often correlate with enhanced SNF, potentially through improved light interception, as light availability is a key factor regulating the symbiotic efficiency of legumes^23,24. Given that N enrichment frequently enhances biomass accumulation and shoot:root ratios in N-fixing plants^25–27, these trait adjustments may partially alleviate SNF declines by augmenting photosynthetic capacity and carbon supply to fuel energetically costly N fixation. Accordingly, we hypothesized that plant trait plasticity may explain a significant fraction of the residual heterogeneity in SNF responses to N enrichment (Hypothesis 1). Critically, the quantitative global-scale contribution of plant performance traits to SNF responses, beyond environmental predictors, has yet to be assessed.

Divergent baseline N conditions between non-croplands and croplands due to long-term anthropogenic management can fundamentally alter SNF responses to experimental N enrichment. Chronic fertilization in croplands elevates soil N availability while simultaneously disrupting legume-rhizobia mutualism and reshaping microbial communities^28,29. Consequently, legumes in croplands may acclimatize to high exogenous N inputs and exhibit attenuated SNF sensitivity relative to non-croplands^18,30,31. In contrast, N-fixing plants in non-croplands experience more marked declines in nodule biomass and SNF following N inputs due to limited microbial adaptation, abrupt symbiosis breakdown, and competitive exclusion pressure from non-fixing species^7,32,33. Accordingly, we propose that N enrichment suppresses SNF more markedly in non-croplands than croplands (Hypothesis 2).

This study focuses on accurate SNF predictability rather than mechanistic causation, to identify plant predictors explaining heterogeneity in SNF responses to N enrichment across sites and taxa. Leveraging a global synthesis based on 908 field observations, we determine the dominant predictors of SNF responses to N enrichment. Two specific questions were addressed: (a) What proportion of variability in SNF responses is explained by environmental factors (exogenous N inputs, climate, and edaphic properties) versus plant performance traits? (b) What are the differences in SNF responses between croplands and non-croplands?

Results

Environmental forces are insufficient to explain SNF responses upon N enrichment

N enrichment reduced SNF (8.49 g N m⁻² yr⁻¹ of SNF upon N enrichment versus 9.51 g N m⁻² yr⁻¹ upon ambient conditions), partially attributable to decreased SNF per unit biomass (Fig. S1). On average, the weighted effect size of SNF was −0.40 (95% CI: –0.58 ~ −0.22), corresponding to a 33.0% global decline relative to controls (Fig. 1a). This suppression in SNF intensified poleward (slope = −0.01, p < 0.05) (Fig. S2), and SNF decline was aggravated with high levels of N enrichment rate (a slope of −0.16, p < 0.01) (Fig. 1b).

Fig. 1. Changes in symbiotic nitrogen fixation under nitrogen enrichment and environmental drivers.

Changes of symbiotic nitrogen fixation upon nitrogen enrichment (a), the relationship between the effect size of symbiotic nitrogen fixation and nitrogen enrichment rate (b), the relative contributions (c) and ranking of relative importance (d) of environmental drivers to the variability of symbiotic nitrogen fixation responses in terrestrial ecosystems. Sample size (n) refers to the number of independent comparisons between nitrogen-enriched and control treatments derived from different studies or sites; technical replicates within individual studies were not treated as independent observations. In a, each blue dot stands for the individual effect size based on comparisons between N-enriched and control plots (n = 606 independent observations from 67 studies). The central red dot indicates the mean effect size, with error bars representing the 95% confidence interval. The shaded density shows the distribution of effect sizes. In (b), each dot represents an independent comparison (n = 563 from 58 studies). Solid lines show fitted relationships (representing the mean regression line) and shaded bands indicate the 95% confidence intervals. The relationship was tested using a two-sided linear mixed-effects model with study as a random effect. Full statistical results for the linear mixed-effects model, including test statistics, degrees of freedom, effect sizes and confidence intervals, are provided in the Source Data file. The abbreviations MAT, N enrichment rate, soil C:N, AP, MBC and MBN are mean annual temperature, the magnitude of nitrogen application, carbon:nitrogen ratio of soil, available phosphorus, the content of carbon and nitrogen in microbial biomass, respectively. Percentages of relative importance denote the proportional contribution of each environmental predictor to the total explained variance in symbiotic nitrogen fixation responses. Source data are provided as a Source data file.

Environmental factors partially accounted for the global variation in SNF responses to N enrichment (Fig. S3). SNF decline attenuated in conditions with higher mean annual temperature (slope = 0.42, p < 0.05), while the effects of precipitation on SNF response were insignificant (p = 0.50; Fig. S3a, b). Edaphically, SNF decline amplified under lower soil pH (slope = 0.21, p < 0.05), greater carbon:N ratio (slope = −0.53, p < 0.05), and elevated available phosphorus (slope = −0.32, p < 0.05). Ambient soil ammonium (p = 0.17) and nitrate (p = 0.66) exerted negligible influences (Fig. S3c–g). Moreover, SNF decline was more pronounced in soils with higher microbial biomass carbon (slope = −0.33, p < 0.05) and microbial biomass N (slope = −0.41, p < 0.05) (Fig. S3h–i). Notably, N enrichment rate (relative importance = 37.4%) and soil carbon:N ratio (relative importance = 31.4%) emerged as dominant predictors of the variability of SNF responses at a global scale (Fig. 1). The combined factors (i.e., mean annual temperature, N enrichment rate, pH, soil carbon:N ratio, available phosphorus, the carbon and N contents in microbial biomass) explained approximately one third of the variability of SNF responses at a global scale (34.2%) (Fig. 1c).

Enhancement of plant performance traits ameliorated the SNF decline upon N enrichment

SNF response to N enrichment also diverged across plant taxa (ranging from −1.29 to 0.44; Fig. 2a). Plant trait plasticity buffers SNF declines upon N enrichment. The enhanced N-fixing plant biomass or shoot:root ratios upon N enrichment attenuated SNF decline, while changes in plant nutrient contents had little impact on SNF responses (Fig. 2). Specifically, SNF decline trended to neutralize with the enhancement of plant performance traits, with a slope of 0.47 for N-fixing plant biomass (p < 0.001) and 2.75 for the plant shoot:root ratios (p < 0.01), respectively (Fig. 2).

Fig. 2. Taxonomic variation in symbiotic nitrogen fixation responses and relationships with plant performance traits.

The variations in symbiotic nitrogen fixation response among plant species (a), and bivariate relationships between response of symbiotic nitrogen fixation and the change in plant performance traits and nutrient contents upon nitrogen enrichment (b–e). In a, the colour indicates the families of plants. In b–e each point represents an independent comparison (sample sizes are shown in the panels). Solid lines show fitted relationships (representing the mean regression line) and shaded bands indicate the 95% confidence intervals. Statistical significance was assessed using linear mixed-effects models with two-sided tests. Exact p values are shown in the panels; full statistical results (including test statistics, degrees of freedom, effect sizes and confidence intervals) are provided in the Source Data file. Source data are provided as a Source data file.

Integrating plant performance traits, soil properties, microbial traits, N enrichment rate, and climatic factors explained 48.8% of the variability in SNF responses to N enrichment (Fig. 3). This represents a substantial improvement of 42.7% (a relative increase in explanatory power) over the prediction considering merely environmental factors (34.2% explained variance in Fig. 1c), which was consistently confirmed by both linear mixed-effects model and structural equation model (SEM). In the SEM, soil carbon:N ratio was the crucial driver of the variability in SNF responses to N enrichment (relative importance = 35.8%), with a standardized coefficient of −0.43 (p < 0.001) (Fig. 3). Moreover, plant performance traits (relative importance = 24.6%) were also important for modulating SNF responses to N enrichment, with the standardized coefficient being 0.29 (p < 0.001) (Fig. 3). The N enrichment rate explained 11.7% of the relative importance in SNF responses with a standardized coefficient of −0.30 (p < 0.001) (Fig. 3). The remaining variables (i.e., mean annual temperature and microbial traits) collectively accounted for 24.6% of the relative importance in SNF responses (Fig. 3). Moreover, environmental factors indirectly mediated SNF responses upon N enrichment through plant performance trait modification, with a standardized coefficient of −0.56 for mean annual temperature (p < 0.001), 0.11 for N enrichment rate (p < 0.05), 0.41 for soil carbon:N ratio (p < 0.001), and 0.60 for microbial traits (p < 0.001), respectively (Fig. 3b).

Fig. 3. Environmental and plant controls of symbiotic nitrogen fixation responses at the global scale.

The relative contributions (a) and multiple influences (b) of the environmental factors and plant performance traits for the variability of responses in symbiotic nitrogen fixation upon nitrogen enrichment. In a, the bar is the relative importance of the variability of responses in symbiotic nitrogen fixation, and the dots with error bars are the standardized regression coefficients (representing the mean effect), with 95% confidence intervals (n = 606 independent observations from 67 studies). Statistical significance was assessed using linear mixed-effects models with two-sided tests. In b, numbers on arrows are standardized path coefficients. The solid and dashed arrows are significant and insignificant relationships, respectively (two-sided tests; p ≤ 0.05). Overall structural equation model fit is indicated by Fisher’s C test (p = 0.409). All analyses in this figure that involve frequentist statistical inference are reported using standardized effect sizes and 95% confidence intervals. Full model outputs and fit statistics are provided in the Source Data file. The abbreviations MAT, N enrichment rate, soil C:N, AP, MBC and MBN are mean annual temperature, the magnitude of nitrogen application, carbon:nitrogen ratio of soil, available phosphorus, the carbon and nitrogen content in soil microbial biomass, respectively. Source data are provided as a Source data file.

Comparison of SNF response to N enrichment between ecosystem types

SNF decline was more pronounced in non-croplands (−0.52; 95% CI: −0.77 to −0.28) than croplands (−0.17; 95% CI: −0.31 to −0.04) (p < 0.001) (Fig. 4a). Environmental factors alone poorly explained the variability of SNF responses in croplands (37.0%) and non-croplands (36.0%) (Fig. 4b, c). Incorporating plant performance traits considerably improved predictive accuracy, enhancing the explanatory power by 63.5% in croplands and 43.3% in non-croplands (Fig. 4d, e). Final models explained 60.5% and 51.6% of total SNF response variability in croplands and non-croplands, respectively (Fig. 4d, e).

Fig. 4. Ecosystem differences in symbiotic nitrogen fixation responses and their predictors.

The comparison of symbiotic nitrogen fixation upon nitrogen enrichment (a), the explanations for variability in symbiotic nitrogen fixation responses from models considering only environmental drivers (b, c), and the importance of environmental drivers and plant performance traits (d, e) between croplands and non-croplands. In a, each point represents one independent comparison (croplands: n = 198; non-croplands: n = 408). The central red point indicates the mean effect size, with error bars representing the 95% confidence interval; shaded distributions show the distribution of effect sizes. In b–e the percentages indicate the proportion of variance explained by the models. Percentages of relative importance denote the proportional contribution of each predictor to the total explained variance in symbiotic nitrogen fixation responses. The abbreviations MAT, N enrichment rate, soil C:N, AP, MBC and MBN are mean annual temperature, the magnitude of nitrogen application, carbon:nitrogen ratio of soil, available phosphorus, and the contents of carbon and nitrogen in soil microbial biomass, respectively. Source data are provided as a Source data file.

Consistently, the change in N-fixing plant biomass upon N enrichment were the important predictor for SNF response variability in both ecosystem types (Fig. 4d, e), while the influence of other factors on SNF response differed between ecosystem types. Soil carbon:N ratio accounted for 43.2% of relative importance for SNF responses in non-croplands, while it explained only a little in croplands (relative importance = 0.3%) (Fig. 4d, e). The relative importance of N enrichment rate for SNF responses was 22.7% in croplands, while it played a less important role in non-croplands (relative importance = 8.9%) (Fig. 4d, e).

Integrating plant performance traits improved the predictability of SNF response

The simulated SNF decline (−37.8%, 95% CI: −44.8% to −30.9%) exceeded the observed SNF decline (−33.0%, 95% CI: −42.4% to −23.6%) (Table 1) when predictions were conducted using environmental variables alone (extracted from the full model). Similarly, the predicted SNF decline across ecosystem types based on environmental variables alone (−22.3% and −43.0% for croplands and non-croplands, respectively) was substantially greater than the observed declines (−15.9% and −40.7% for croplands and non-croplands, respectively) (Table 1). Crucially, models incorporating both environmental factors and plant performance traits achieved strong concordance with observed SNF declines, e.g., −17.0% (95% CI: −20.8% to −12.6%) for croplands, and −39.7% (95% CI: −49.1% to −30.2%) for non-croplands.

Table 1.

Comparisons of the predicted responses of symbiotic nitrogen fixation against the observed counterparts by the linear mixed-effects model

Ecosystem types	Observed SNF decline	SNF decline predicted with environmental factors	SNF decline predicted with both environmental factors and plant performance traits
Croplands	−15.91% (−23.12% to −1.31%)	−22.26% (−26.03% to −18.49%)	−16.96% (−20.79% to −12.59%)
Non-croplands	−40.69% (−54.11% to −27.27%)	−42.95% (−53.95% to −31.95%)	−39.69% (−49.14% to −30.24%)
Overall	−33.02% (−42.42%to −23.62%)	−37.81% (−44.76% to −30.86%)	−32.63% (−38.96% to −26.30%)

The values in parentheses are 95% confidence intervals for the responses of symbiotic nitrogen fixation. Here, percentage values are the SNF decline under N enrichment compared with that in ambient conditions (means and 95% confidence intervals).

Discussion

Understanding how and to what extent N enrichment influences the magnitude of SNF is an important issue. Preceding meta-analyses had revealed that the magnitude of SNF was reduced by exogenous N input, albeit based on limited datasets^9,13. The SNF declined by 33.0% under N enrichment based on a large dataset (Table 1 and Fig. 1), which was compatible with preceding estimates^9,13, while revealing substantial response heterogeneity. Crucially, climate and edaphic factors jointly explain minor fractions of the variability in SNF responses at large spatial scales^9,13, leaving a massive amount of unexplained variance. Here, we demonstrate that the residual variation can be largely attributed to plant performance traits, which modulate the SNF responses. Models incorporating these plant performance traits improved predictive accuracy of SNF responses by 42.7%, a relative improvement over the prediction considering environmental factors alone at a global scale (Figs. 1 and 3), underscoring the essential role of plant traits in forecasting SNF upon anthropogenic perturbation.

Alterations in plant performance traits are vital for understanding SNF responses to N enrichment

Despite a declining trend in SNF upon N enrichment, SNF responses exhibited considerable variability (Fig. 1a), implicating that factors beyond exogenous N inputs may also be important to explain SNF responses (Fig. 1b). The stronger N-fixing plant growth and increased shoot:root ratios consistently mitigated the negative effects of N enrichment on SNF, indicating that greater carbon supply to symbionts and expanded rhizosphere habitats partially compensate for the decline in SNF (Figs. 2–3). In the absence of direct physiological proxies, these widely reported plant performance traits can serve as scalable proxies for photosynthetic capacity, root system, and carbohydrate availability. A recent global meta-analysis showed that N-fixing plant biomass is increased by 56% following the N enrichment²⁷. Accordingly, plants with larger shoot biomass may possess a more powerful capacity in photosynthesis³⁴, which potentially supplies additional photosynthates for SNF (1 g N fixed at the expense of 2.8-4.8 g carbon from photosynthates)^35–37. Similarly, plants with a flourishing root system generally extend rhizosphere habitat, which can furnish more loci for N-fixing bacterial symbionts and secrete more semiochemicals (e.g., flavonoids) to recruit N-fixing bacteria towards forming intracellular nodules^38–40. Accordingly, SNF rises from 3.8 g N m⁻² yr⁻¹ to 5.9 g N m⁻² yr⁻¹ when the nodules' biomass increases from 17.7 g N m⁻² to 27.3 g N m⁻²⁴¹. Furthermore, N enrichment tends to increase shoot:root ratios by 37%²⁷, which may enhance light interception and carbon assimilation and thus favour SNF. As an example, SNF in Trifolium repens is increased by 26% with shoot:root ratio rising from 3.2 to 3.7⁴². Collectively, these results suggest that SNF decline under N enrichment is not a simple unidirectional suppression but a balance between negative environmental forcing and positive plant-mediated compensation. The enhancement of N-fixer biomass and carbon allocation capacity^43–45 provides a mechanistic explanation for this compensation, supported by consistent global patterns (Figs. 2–4). Occasionally, the sharp increase in the competitive ability of non-fixing species following external N addition can cause a pronounced reduction in N-fixer abundance and biomass at low soil N availability³², masking this SNF compensation. Despite these exceptions, recent global syntheses align with our findings that positive growth is a pervasive attribute of N-fixers under N enrichment^27,46. This work has scaled up the mechanistic understanding of the SNF compensation supported by plant growth under N enrichment, which is a critical step toward explicitly projecting SNF with models.

Environmental factors also modulate SNF responses to N enrichment. SNF decline upon N enrichment attenuated with increasing temperature (Fig. S3), implying that warming can partly counteract the decline of SNF, likely through enhanced plant growth and resultant carbon supply^21,45. A latitudinal pattern of SNF responses (Fig. S2) further supports the role of temperature in dictating SNF under N enrichment. Mean annual temperature was used as a predictor representing coarse-grained climate conditions, soil temperature was rarely reported across studies, even as microbial activity and rhizobia-legume symbiosis are closer to soil temperature. Mean annual precipitation showed no clear relationship with SNF responses, presumably because irrigation in croplands confounded the signal; the lack of irrigation data across studies precluded explicit assessment. Soil properties (soil carbon:N ratio and pH) also affect SNF responses to N enrichment (Fig. 3a). In N-deficient environments, e.g., those with higher carbon:N ratio, legumes are more reliant on SNF^47,48, and SNF responses to abrupt N enrichment may be more dramatic. Meanwhile, rhizobia prefer neutral soils⁴⁹ over acidic soils^50,51. The soils with higher initial pH levels can effectively buffer against acidification caused by exogenous N inputs, and eventually alleviate SNF decline^51,52. Consequently, the decline of SNF in alkaline soils was lower than that in acidic soils (Fig. S3). Moreover, climatic and edaphic factors regulate SNF responses, both directly and indirectly via plant performance traits (Fig. 3b). For instance, warmer climates can enhance plant growth and shift carbon allocations⁵³, which in turn modulate SNF efficiency⁵⁴. Embedding both direct and indirect effects in models enables a mechanistic scaling-up of SNF predictions that accounts for both environmental constraints and plant-mediated compensation.

Different magnitudes of SNF response to N enrichment between ecosystems

N enrichment led to a more pronounced SNF decline in non-croplands than croplands (Fig. 4a). The divergence arises from several reasons, e.g., legumes in croplands acquire a smaller fraction of N from SNF under exogenous N inputs compared with non-croplands⁵⁵. Legumes in croplands are not expected to substantially decrease SNF because SNF baseline (7.28 g N m⁻² yr⁻¹) is relatively lower than that in non-croplands (9.24 g N m⁻² yr⁻¹)²¹. Non-croplands are commonly phosphorus-limited^56,57; N enrichment can exacerbate this deficiency, tightening phosphorus constraints on the N fixation process and ultimately leading to more pronounced SNF suppression (e.g., phosphorus content is a key driver of SNF variability in Fig. 4c)⁵⁸. Moreover, under sustained exogenous N inputs, leguminous crops may undergo adaptive domestication that prioritizes soil N acquisition and yield over SNF^30,31. Such long-term adaptations lessen crop dependence on SNF, thereby buffering SNF decline even under N enrichment^30,59. In contrast, non-croplands are typically dominated by wild or less-domesticated N-fixing plants that exhibit greater sensitivity to abrupt N increases due to limited evolutionary adaptation to chronic N enrichment. For instance, wild Trifolium populations experience nodule biomass reduction and competitive exclusion upon N enrichment in non-croplands, driving pronounced SNF suppression^30,32. Given that the global SNF in croplands now approximates that in non-croplands⁴ and that the SNF in these two ecosystem types exhibits a different sensitivity to N enrichment (Fig. 4 and Table 1), Earth System Models should explicitly address the contrasting SNF dynamics in croplands and non-croplands under anthropogenic perturbation.

Critical predictors of SNF responses, with the exception of plant performance traits, differed between ecosystems, e.g., N enrichment rate in croplands versus soil carbon:N ratio in non-croplands (Fig. 4). Regardless of SNF decline magnitude, legumes in croplands exhibited stronger responses to N enrichment rate than to edaphic factors, reflecting long-term fertilization legacies where soils approach N saturation⁶⁰, yet plant-microbe interactions remain responsive to incremental N inputs within constrained SNF ranges. As for non-croplands, the exogenous N inputs derive mainly from atmospheric N deposition, and N deposition magnitude in non-croplands is far less than the amount of N fertilization in croplands. As a consequence, bioavailable N is relatively scarce in non-croplands (e.g., soil carbon:N ratio is 13.3 for grasslands, and 19.9 for temperate coniferous forests) compared with that in croplands (soil carbon:N ratio = 12.5)⁶¹. Under N scarcity, legume species are more reliant on SNF, and thus their sensitivity to exogenous N is amplified. Rhizobia inhabiting soils with high carbon:N ratios are more likely to adapt to low-N environments and maintain close symbioses with leguminous plants. The mutualistic associations are more easily disrupted by external N inputs, leading to stronger SNF declines under similar N enrichment^13,18,30. Therefore, SNF responses to N enrichment in non-croplands depends on the initial soil carbon:N ratio. Recognizing the differences in the importance of predictors in different ecosystems is an essential step towards refining N fixation parameterizations in Earth System models. Accordingly, these models should adopt ecosystem-specific schemes⁶², assigning driver weights (e.g., N enrichment rate, soil carbon:N ratio and available phosphorus) that reflect the biogeochemical context of croplands versus non-croplands.

Implications and limitations

While SNF and exogenous N inputs collectively supply bioavailable N in terrestrial ecosystems^55,63, their combined effects on plant productivity and carbon sequestration^64,65 exhibit critical non-additivity due to N-induced SNF suppression. This suppression implies that anthropogenic N enrichment may yield diminishing returns in bioavailable N provisioning, potentially constraining plant productivity gains and altering ecosystem functionality. Crucially, the sensitivity of SNF to N enrichment is more pronounced in non-croplands than in croplands. Given the vast extent and integral ecosystem functions of non-croplands (e.g., forests and grasslands), such heightened vulnerability in SNF could cascade into altered plant community dynamics and compromised long-term ecosystem productivity.

Accurately representing SNF under elevated N is therefore essential to avoid systematic biases in projections of nutrient cycling, productivity, and carbon–climate feedbacks. However, effectively incorporating the mechanistic understanding in Earth System Models remains a fundamental challenge, since environmental heterogeneity complicates the direct extrapolation to macroscales. Global analyses, such as those presented here, allow scaling up the mechanistic understanding of the compensatory SNF supported by plant growth under N enrichment by accounting for both direct and indirect effects in Earth System Models. Specifically, the empirical relationships derived here (Eqs. 8 and 9) can be implemented within N fixation subroutines of Earth System Models, where environmental drivers define the baseline limitation of SNF, and plant performance traits act as dynamic modifiers that enable models to simulate how shifts in plant growth or carbon allocation mediate SNF responses. The observed SNF compensation, as supported by plant growth under N enrichment, provides the quantitative relationships and driver sensitivities needed to mechanistically parameterize the dynamic representation of SNF under N enrichment in current models^62,66,67. In that way, this study provides Earth System Models with the quantitative basis (Eqs. 8 and 9) to explicitly distinguish between SNF compensation generated by N-fixer growth and direct SNF suppression induced by N enrichment.

Some limitations are inescapable in the analyses. First, although legumes universally exhibit SNF suppression under N enrichment, the extent of SNF decline varies across species with different SNF strategies^7,68. For instance, some plant species with a facultative strategy can partially downregulate SNF by increasing soil N uptake when inorganic N is abundant, whereas species with an obligate strategy typically sustain SNF regardless of external N levels¹². Yet, some actinorhizal species (traditionally classified as obligate N-fixers) exhibited significant SNF suppression under N enrichment (Fig. 2a), suggesting broader plasticity than previously theorized^68,69. Nevertheless, due to data scarcity of plant N fixation strategies in current studies, systematic clarification was precluded. Second, predictors critical to SNF regulation, such as micronutrients (Mo and V), soil moisture, light availability, plant ontogeny, and finer-resolution physiological traits (e.g., photosynthesis and internal N status), could not be integrated due to data paucity. Additionally, areal SNF rates (e.g., g N m⁻² yr⁻¹) integrate both physiological efficiency and diazotroph biomass. Though SNF per unit biomass declined in available subsets (Fig. S1), future work is needed to systematically partition these components.

This work provides a more comprehensive understanding of the variability in SNF responses to N enrichment by integrating both environmental factors and plant performance traits. Our results demonstrate that SNF suppression under N enrichment is globally consistent (-33.0%), yet disproportionately stronger in non-croplands than croplands. This difference is attributable to agricultural legumes’ lower baseline SNF and their evolved nutrient acquisition strategies. Plant trait plasticity, e.g., plant growth and shoot:root ratios, accounts for significant heterogeneity in SNF responses, where enhanced trait performance correlates with attenuated SNF suppression. Integrating plant performance traits with environmental factors improves the predictive accuracy of SNF responses by 42.7%. These findings advance the fundamental understanding of terrestrial N cycle plasticity and offer a practical pathway to refine next-generation Earth System Models. Incorporating trait-mediated SNF responses will enhance projections of ecosystem functionality under global change scenarios.

Methods

Data compilation

Data on SNF responses upon N enrichment were assembled from peer-reviewed papers. Peer-reviewed papers published from 1990 to 2022 were searched using the keyword string (“nitrogen fixation” OR “N₂ fixation” OR “dinitrogen fixation”) AND (“nitrogen addition” OR “nitrogen fertilization” OR “nitrogen deposition” OR “nitrogen enrichment” OR “nitrogen application”), by means of Web of Science and China National Knowledge Infrastructure Database. The criteria used to screen the peer-reviewed papers were: 1. Experiments with N enrichment were conducted in the field; 2. The magnitude of SNF was reported in pairwise treatments, e.g., a treatment without N enrichment vs. a treatment with N enrichment, or the response of SNF to a gradient of N enrichment levels was provided; 3. The SNF was measured using the acetylene reduction assay or ¹⁵N isotope methods, applying the same method to pairwise treatments to minimize systematic errors. In total, 67 peer-reviewed papers matching the above criteria were used to assemble the dataset (Fig. S4 and Supplementary Data 1), where each unique comparison between a treatment and its corresponding control—defined by distinct species, N levels, or experimental units—constituted an independent observation.

The dataset encompasses variation in site conditions (e.g., climatic and edaphic factors), species identity, N treatment rate, and plant performance traits (plant biomass, shoot:root ratios, N and phosphorus contents). The data capture key sources of heterogeneity across studies, although some influential variables, such as micronutrients (Mo and V), soil moisture, light availability, plant ontogeny, and finer-resolution physiological traits, were rarely reported across studies. Here, plant performance traits (plant biomass, shoot:root ratios, plant N and phosphorus contents) refer exclusively to N-fixing species rather than community-level vegetation, ensuring SNF responses reflect diazotroph physiology. Information on each experimental site was extracted from the peer-reviewed papers, including geographical location (latitude and longitude), climatic factors (mean annual temperature and precipitation), plant performance traits (plant biomass, shoot:root ratios, plant N and phosphorus contents), N enrichment rate (i.e., the magnitude of N application), soil properties (pH, soil carbon:N ratio, ammonium and nitrate concentration) and ecosystem type. Complementary microbial traits (carbon, N and carbon:N ratio in soil microbial biomass) were obtained from a soil microbes dataset⁷⁰ and matched by geographical coordinates. Forests and grasslands were grouped as ‘non-croplands’ in this analysis, based on shared characteristics, e.g., minimal management and absence of long-term fertilization, which functionally differentiate them from croplands. Finally, the dataset comprised 908 field observations, encompassing 305 observations from croplands and 603 observations from non-croplands (350 observations from forests and 253 observations from grasslands) worldwide (Figure S5).

Calculation of SNF response to N enrichment

The natural logarithm response ratio (lnRR) can portray the change in SNF and plant performance traits upon N enrichment. A positive value of lnRR indicates that the variable is increased upon N enrichment, and vice versa. The lnRR and corresponding variance ($v_i$) were calculated via Eqs. (1) and (2), respectively.

$$\ln \mathrm{RR}=\ln \left({\mathrm{X}}_t/X_c\right)$$ 1

$$v_i=S_t^2/\left(N_t{X}_t^2\right)+S_c^2/\left(N_c{X}_c^2\right)$$ 2

where ${\mathrm{X}}_t$ and ${\mathrm{X}}_c$ are the means of SNF or plant performance traits in a treatment with N enrichment and a treatment without N enrichment, respectively; ${\mathrm{S}}_t$ and ${\mathrm{S}}_c$ are the standard deviations of the treatment with N enrichment and the treatment without N enrichment, respectively; ${\mathrm{N}}_t$ and ${\mathrm{N}}_c$ are the sample sizes of the treatment with N enrichment and the treatment without N enrichment, respectively.

To address the issue of statistical non-independence among multiple treatment groups sharing the same control group within a single study, a variance-covariance matrix (V) was constructed. The diagonal elements of variance-covariance matrix represent the individual variances (vi), calculated from Eq. (2), whereas off-diagonal elements of variance-covariance matrix represent covariances between lnRRs sharing the same control group. The covariance between two lnRRs from treatment groups i and j, sharing the same control group, was estimated using:

$$Cov\left({lnRR}_i,{lnRR}_j\right)=S_c^2/\left(N_c{X}_c^2\right)$$ 3

The variance-covariance matrix ($V$) was incorporated into the subsequent meta-analysis to accurately reflect the dependency structure of the data. After constructing the variance-covariance matrix ($V$), we then calculated the weighted response ratios (${RR}_{++}$) using a mixed-effect model. In the mixed-effect model, there are two sources of variances: (1) within-study variance, represented by the variance-covariance matrix ($V$), and (2) between-study variance (${\tau }^2$). To integrate both sources of variance into the meta-analysis, we defined a weight matrix ($W$) as:

$$W={\left(V+{\tau }^{2}I\right)}^{-1}$$ 4

where $V$ is the constructed variance–covariance matrix; ${\tau }^2$ represents the between-study variance which was obtained by the restricted maximum likelihood (REML) method; $I$ is an identity matrix of the same dimensions as $V$, with diagonal elements equal to 1 and off-diagonal elements equal to 0. Subsequently, the ${RR}_{++}$, and standard error ($Se$) were calculated one after another by the following equations:

$${RR}_{++}=\left(1^{T}WlnRR\right)/\left(1^{T}W1\right)$$ 5

$$se=\sqrt{1/\left(1^{T}W1\right)}$$ 6

where $W$ is the weight matrix that incorporates the variance–covariance matrix ($V$) and the between-study variance component (${\tau }^2$); $\ln RR$ represents the vector of lnRR values calculated for each treatment group; $1$ denotes a vector of ones with the same length as $\ln RR$; and $1^T$ indicates the transpose of the vector $1$. The significance level was set at p < 0.05. Weighted response ratios (${RR}_{++}$) of SNF were determined separately for the entire dataset and for croplands and non-croplands separately. The funnel plots and Egger’s regression tests were used to test the publication biases of variables in this study. Our result showed that no significant publication bias was detected for SNF responses in either the globes or acorss ecosystem types (Table S1 and Fig. S6). Moreover, the independent-samples t-tests were performed to compare the lnRR of SNF between two ecosystem types, and the raw SNF values between N enrichment and control conditions. The geographical variation of lnRR of SNF with latitude was tested using a linear mixed-effect model.

Drivers of SNF response

First, we evaluated the influence of individual environmental factors (climatic factors, N enrichment rate, soil properties, and microbial traits) and the changes in plant performance traits (expressed as the lnRR of N-fixing plant biomass, shoot:root ratios, plant N and phosphorus contents) on the lnRR of SNF; linear mixed-effect models were conducted with the “lme4” package in R (version 4.2.0) to account for potential non-independence. The equation was:

$$\text{Y}={\beta }_0+{\beta }_1\times \mathrm{X}+{\pi }_{study}+\mathrm{\epsilon }$$ 7

where $\mathrm{Y}$ is the lnRR of SNF, and $\mathrm{X}$ is the value of an individual driver, respectively; ${\beta }_0$, ${\beta }_1$, ${\pi }_{study}$ and $\mathrm{\epsilon }$ are the intercept, slope coefficient, the random effect of study, and sampling error, respectively.

Next, drivers with a significant effect on the lnRR of SNF were combined in new linear mixed-effect models. In particular, linear mixed-effect models without (Eq. 8) and with changes in plant performance traits (N-fixing plant biomass and shoot:root ratio) (Eq. 9) were separately run for all land use types, croplands, and non-croplands to evaluate to what extent the alteration of plant performance traits impacted SNF responses:

$$\text{Y}={\sum }_{i=1}^n{\beta }_i\times X_i+{\beta }_0+{\pi }_{study}+\mathrm{\epsilon }$$ 8

$$\text{Y}={\sum }_{i=1}^n{\beta }_i\times X_i+{\sum }_{j=1}^m{\alpha }_j\times {PTA}_j+{\beta }_0+{\pi }_{study}+\mathrm{\epsilon }$$ 9

where $\mathrm{Y}$ is the lnRR of SNF; $n$ is the number of environmental factors; ${\beta }_i$ is the coefficient of the $ith$ environmental factor; $X_i$ is the $ith$ environmental factor; ${\beta }_0$ is the intercept; ${\pi }_{study}$ is the random effect that considered the specific condition for each experiment; $\epsilon$ is the sampling error; $m$ is the number of plant performance traits; ${\alpha }_j$ is the coefficient for the effect size of the $jth$ plant performance trait; and ${PTA}_j$ is the effect size of the $jth$ plant performance trait.

Moreover, a hierarchical partitioning analysis was employed to quantify the independent effects (%) of each driver on the lnRR of SNF based on Eqs. 8 and 9, after z-score standardizing all continuous drivers (including climatic and edaphic factors and plant performance traits). The relative importance of each driver was calculated by R package “glmm.hp”⁷¹.

To evaluate the relative predictive contributions of environmental factors versus plant performance traits, we constructed a full linear mixed-effects model incorporating both variable types (Eq. 9). The explanatory power was then quantified by the following approaches. Initially, we extracted coefficients for environmental factors alone, and then extracted coefficients for both environmental factors and plant performance traits. Subsequently, these coefficients were used to calculate predicted SNF response values, which were then compared with the observed effect sizes to evaluate the predictive power of the two predictor sets. This analytical framework was applied consistently across all ecosystems, including both croplands and non-croplands.

Moreover, a SEM model was constructed to evaluate how SNF responses were dictated by environmental factors and the changes in plant performance traits. All causal relationships were included in a priori model, including direct and indirect relationships between SNF responses and environmental factors and plant performance traits. The contents of carbon and N in microbial biomass were combined into the composite variable “microbial traits” using the parameter estimates from the linear mixed-effects model. To enable direct comparison of coefficients, the logarithmically transformed data were further normalized (Z-score normalization). In addition, “study” was included as a random variable in the SEM. The “piecewiseSEM” and “lme4” packages in R were used to test the priori models. The performance of SEM models was evaluated by the Chi-squared test, and the model is considered eligible if p  >  0.05. The SEM models were optimized by stepwise refinement. The criteria of optimal SEM models were the smallest Akaike information criterion (AIC) and the highest p value.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Z.L.L., X.P.C. and Y.Z.Y. conceived this research. Y.Z.Y. collected and analyzed the data. Y.Z.Y. and Z.L.L. wrote the draft. Z.L.L., P.M.v.B., X.P.C., and S.L.N. revised manuscript. B.B.H., X.Z.D., and Y.Y.Z. performed the literature search and screened studies.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The microbial trait data (soil microbial biomass carbon and nitrogen contents) used in this study are available in the Dryad database at [10.5061/dryad.xd2547dhk]⁷⁰. The meta-analysis dataset generated in this study has been deposited in the Figshare repository at [10.6084/m9.figshare.30479132]⁷² Source data are provided with this paper.

Code availability

The codes have been uploaded to the Figshare repository [10.6084/m9.figshare.30479132]⁷².

Competing interests

The authors declare no competing interests as defined by Nature Portfolio or other interests that might be perceived to influence the results and/or discussion reported in this paper.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-69876-1.