Microbial Functional Gene Abundance‐Integrated Modeling of Global Methane Sinks in Upland Soils Under Future Climate Change

ABSTRACT

Methanotrophs are key microbial regulators of soil methane (CH₄) sinks, but the global impact of their functional gene abundance on CH₄ oxidation remains unquantified. This gap limits the integration of key functional genes abundance parameters (e.g., pmoA) into soil CH₄ sink model. We integrated meta‐analysis, machine learning, and process‐based modeling to assess the relationship between pmoA gene abundance and soil CH₄ uptake. Our developed Functional Gene Abundance‐Based Methanotrophy Model (FGA‐MeMo) incorporates pmoA as a proxy for CH₄ oxidation capacity, significantly improving model simulations. FGA‐MeMo estimates global upland soil CH₄ uptake at 45.74 ± 0.26 Tg year⁻¹, which is 56%–58% higher than MeMo model. Under SSP5‐8.5 scenario, this increases to 64.68 ± 0.35 Tg year⁻¹ by 2100, with mid‐ and high‐latitude regions showing enhanced CH₄ oxidation due to greater pmoA abundance. These findings highlight the importance of integrating microbial functional genes into Earth system models for improved CH₄ cycle predictions.

This study integrates machine learning, microbial functional gene data, and process‐based modeling to advance the prediction of global upland soil methane (CH₄) sinks under changing climates. (a) Workflow for predicting pmoA gene abundance using environmental covariates; (b) Global map of pmoA abundance showing methanotrophic hotspots; (c) Spatial pattern of soil CH₄ sink simulated by the Functional Gene Abundance–Based Methanotrophy Model linking pmoA to CH₄ oxidation rates. This framework bridges genomic and ecosystem scales to quantify terrestrial CH₄ sinks under climate change.

1. Introduction

Atmospheric methane (CH₄), a potent greenhouse gas, exhibits a global warming potential (GWP) over a 100‐year timescale that is approximately 27.0 to 29.8 times that of carbon dioxide (CO₂), even though its atmospheric concentration is far lower than that of CO₂ (IPCC ²⁰²³; Saunois et al. ²⁰²⁵). Over the past century, the rise in atmospheric CH₄ has driven a global average temperature increase of roughly 0.5°C (with a range of 0.3°C–0.8°C), accounting for approximately 33% of total global warming (IPCC ²⁰²³). As a key component of Earth's biosphere, upland soils, including those in forests and grasslands, play a critical role in the global CH₄ cycle. Current models estimate that the annual CH₄ uptake by global upland soils is approximately equivalent to 86%–167% of the annual net increase in atmospheric CH₄ (Saunois et al. 2025). However, this estimate remains highly uncertain, largely because existing models fail to account for key functional gene abundance (e.g., the absolute abundance of pmoA), which limits the accuracy of global soil CH₄ sink assessments (Wang et al. 2022; Hedenec et al. ²⁰²⁴; Ou et al. ²⁰²⁴).

Net soil CH₄ flux is governed by the balance between aerobic CH₄ oxidation and anaerobic CH₄ production. In upland ecosystems, where soil aeration is favorable, CH₄ oxidation rates exceed production rates, resulting in these soils acting as a net CH₄ sink (Hedenec et al. 2024; Ou et al. ²⁰²⁴). The process of CH₄ oxidation is mediated by a specialized group of functional microorganisms known as aerobic methanotrophs. Generally, aerobic methanotrophs can be classified into high‐affinity and low‐affinity types based on their affinity for CH₄ (Kolb 2009). Among them, low‐affinity methanotrophs are capable of oxidizing CH₄ at elevated concentrations, whereas high‐affinity methanotrophs can effectively oxidize CH₄ even at ambient (or trace) levels (Kolb et al. 2005; Kolb ²⁰⁰⁹; Pratscher et al. ²⁰¹¹; Schmider et al. ²⁰²⁴). High‐affinity methanotrophs may constitute more than 90% of the total methanotroph population in some forest ecosystems (Kolb et al. 2005; Pratscher et al. ²⁰¹¹). During CH₄ oxidation, these microbes use particulate methane monooxygenase (pMMO), an enzyme encoded by the pmoA gene, to oxidize CH₄. This step is the rate‐limiting stage of the entire metabolic pathway (Lüke and Frenzel 2011; Hedenec et al. ²⁰²⁴; Ou et al. ²⁰²⁴). Due to its pivotal role, the absolute abundance of the pmoA gene is widely regarded as reflecting the potential capacity for aerobic CH₄ oxidation and is used as a molecular proxy for methanotroph abundance (Ahirwar et al. 2018; Bhardwaj and Dubey ²⁰²⁰; Li et al. ²⁰²⁰; Zhang et al. ²⁰²¹). However, the abundance of methanotrophs exhibits significant spatial heterogeneity across different ecosystem (Zhang et al. 2021; Hedenec et al. ²⁰²⁴; Ou et al. ²⁰²⁴). This variability is driven by environmental factors such as climate (temperature and precipitation), soil pH, soil organic carbon, nitrogen deposition, and land‐use types, thereby modulating the soil CH₄ sink capacity (Zhang et al. 2021; Hedenec et al. ²⁰²⁴). Nevertheless, current global soil CH₄ sink models rely primarily on the physical diffusion of atmospheric CH₄ into soils and typically incorporate only environmental factors (e.g., temperature, moisture, and soil pore structure), while overlooking key methanotrophy genes, including absolute abundance of pmoA (Kolb 2009; Curry ²⁰⁰⁷; Zhuang et al. ²⁰¹³; Murguia‐Flores et al. ²⁰¹⁸; Lafuente et al. ²⁰¹⁹). Consequently, this structure limits the model's capacity to capture the spatiotemporal variation of soil CH₄ sinks at the regional scale. Thus, integrating key methanotrophy genes into existing soil CH₄ sink models is essential to better characterize regional variations in CH₄ oxidation capacity, thereby improving the accuracy of predictions regarding global soil CH₄ uptake and its future dynamics.

This study seeks to develop a more mechanistically robust soil CH₄ sink model by incorporating key methanotrophy gene (pmoA gene) into existing modeling frameworks. Given the limited availability of global data on pmoA gene, we use machine learning to predict a spatially explicit global map of the absolute abundance of the pmoA gene in upland soils. Based on this map, we then develop a functional gene‐integrated soil CH₄ sink model, that is, FGA‐MeMo. Using this new model, we provide a more precise assessment of the CH₄ sink capacity of global upland soils and project how this sink may change under future environmental conditions. This work advances our mechanistic understanding of the soil CH₄ sink and strengthens the scientific foundation for predicting biosphere–climate interactions, ultimately supporting the development of climate change mitigation policies.

2. Materials and Methods

2.1. Development of a Global Dataset of pmoA Gene Absolute Abundance in Upland Soils

This study focuses on upland ecosystems, where methane oxidation is the dominant process and net fluxes function as a sink for atmospheric CH₄. This study characterizes the CH₄ oxidation capacity in upland soils using the pmoA gene as a marker, quantified by quantitative polymerase chain reaction (qPCR). This method enables simultaneous DNA amplification and quantification and has been widely used to determine the absolute abundance of the pmoA gene in soil (Kolb et al. 2003; Lafuente et al. ²⁰¹⁹; Hedenec et al. ²⁰²⁴). To develop a global dataset of the absolute abundance of the pmoA gene in upland soils, we conducted a literature search in the Web of Science (https://www.webofscience.com) database using the following keywords: [(forest) or (dryland) or (grassland) or (farmland) or (agricultural) or (upland) or (city) or (urban) or (park)] and (pmoA) and (soil). We screened the studies using the following criteria: (1) Soil samples: The experimental soil samples used in the studies were all collected from field sites, rather than derived from laboratory incubations or artificial growth systems. This ensures that the measured properties are representative of natural soil environments; (2) Selection of ecosystem types: All soil samples were collected from upland areas. This study focused on natural ecosystems where aerobic processes predominate, as the activity and distribution of methanotrophs in these environments differ significantly from those in wetlands or rice paddies with higher moisture levels. Therefore, we excluded anthropogenic or anaerobic environments such as landfills and paddy fields to ensure that the results accurately represent the characteristics of typical upland ecosystems; (3) Soil depth: The studies focused on the surface soil layer (0–20 cm) in upland ecosystems, which is the most active zone for methanotroph activity and experiences the most frequent gas exchange with the atmosphere; (4) Quantification method: qPCR was used in the included studies to quantify the absolute abundance of the pmoA gene. The primer sets selected in this study were based on a published meta‐analysis (Hedenec et al. 2024), specifically including four primer systems: A189/A661, A189/A601, A189/A682, and A189/A650.

For controlled experiments involving manipulations such as warming and altered precipitation regimes, only data from the control groups were retained for analysis. For seasonal measurements, we used the seasonal means. After removing duplicates within the meta‐analysis article, the remaining literature was included in subsequent quantitative analyzes. Through systematic retrieval, we identified 72 original peer‐reviewed research articles, including a meta‐analysis study (Hedenec et al. 2024). We then standardized all soil pmoA gene absolute abundance data to the common units of copies per gram of dry soil, hereafter abbreviated as copies g⁻¹ soil. We also collected the latitude and longitude of all plots, pmoA gene absolute abundances, and CH₄ flux data from each article.

2.2. Global Mapping of pmoA Gene Abundance and Its Drivers in Upland Soils

Understanding the relationship between environmental factors and the absolute abundance of the pmoA gene is crucial for predicting its global distribution in soils. We first gathered data on 27 environmental variables, including 3 related to human activities, 2 to climate, 2 to biological factors, and 20 to soil properties. Detailed information on data sources and variable definitions is provided in Table S1. All environmental predictor variables were resampled to a spatial resolution of 1°. We extracted environmental data for all sampling locations included in the pmoA gene dataset. To reduce the influence of outliers and facilitate comparison, all environmental variables were standardized (z‐scores). We then examined the relationship between the absolute abundance of the pmoA gene and these environmental variables. First, we calculated the mutual information (MI) between each environmental variable and the absolute abundance of the pmoA gene. MI quantifies the dependency between two variables and can detect both linear and nonlinear relationships; statistical significance was assessed using permutation tests. Next, to evaluate multicollinearity, we computed the Pearson correlation coefficients among the environmental variables.

Based on the results of environmental driver analysis, we selected key environmental variables to predict the global distribution of pmoA gene abundance in soils. First, we excluded environmental variables that showed no significant mutual information with the pmoA gene abundance. Next, for pairs of variables with absolute correlation coefficients exceeding 0.6, we removed the one with the lower MI value to minimize multicollinearity (Dormann et al. 2013). Feature selection aims to retain information‐rich and less redundant variables, thereby improving model stability and explanatory power. To optimize predictions of the global distribution of pmoA gene abundance, we evaluated five machine learning algorithms: Random Forest (RF), XGBoost, K‐Nearest Neighbors (KNN), Support Vector Machine (SVM), and Elastic Net Regression (ENR). We randomly assigned 80% of the data to the training set and reserved the remaining 20% as an independent test set. Hyperparameter tuning was performed using grid search with 5‐fold cross‐validation on the training set to optimize model performance. Finally, we evaluated the performance of each model on the test set using multiple metrics, including the Pearson correlation coefficient (r), coefficient of determination (R ²), mean absolute error (MAE), and root mean squared error (RMSE). The best‐performing model was selected based on these metrics. We then predicted the global distribution of pmoA gene abundance in upland soils at a 1° × 1° resolution using global environmental data. However, due to the absence of desert samples in the training data, predictions for desert regions were excluded from the final map to avoid unreliable extrapolation.

2.3. A Functional Gene Abundance–Based Methanotrophy Model and Its Future Projections

The Soil Methane Oxidation Model (MeMo) is a process‐based model of soil CH₄ consumption that simulates CH₄ diffusion and oxidation at the soil–atmosphere interface. It has been extensively validated in previous studies across a range of ecosystems (Zhou et al. 2024). In the model, the CH₄ oxidation rate constant is defined as follows: (Equation (1), (2), (3))

$$J_{{\mathit{CH}}_4}=-D_{{\mathit{CH}}_4}\left(-A\sqrt{\frac{k_d}{D_{{\mathit{CH}}_4}}}+B\sqrt{\frac{k_d}{D_{{\mathit{CH}}_4}}}\right)$$ (1)

$$\begin{array}{c}{D}_{{\mathit{CH}}_4}=D_{0{\mathit{CH}}_4}\times G_T\times G_{\mathit{\text{soil}}}\end{array}$$ (2)

$$\begin{array}{c}{k}_d=k_0\times r_T\times r_{\mathit{sm}}\times r_N\end{array}$$ (3)

where $J_{{\mathit{CH}}_4}$ denotes the soil CH₄ flux; $D_{{\mathit{CH}}_4}$ represents the CH₄ diffusion coefficient in soil; k_d denotes the CH₄ oxidation rate constant; A and B serving as integration constants; $D_{0{\mathit{CH}}_4}$ indicates the reference diffusion coefficient at 0°C and 1 atm(0.196 cm² s⁻¹); Factors $G_T$ and $G_{\mathit{\text{soil}}}$ account for the influences of soil temperature and pore structure on CH₄ diffusivity; k₀ is the base oxidation rate, assigned specific values for different ecosystem types; and r_T, r_sm, and r_N represent the response functions for soil temperature, soil moisture, and nitrogen deposition, respectively (Table S2).

In the newly developed FGA‐MeMo model, we incorporate the absolute abundance of the soil methane monooxygenase gene (pmoA), a molecular proxy for CH₄ oxidation capacity, as a key input to scale the CH₄ oxidation rate constant (k_d) (Equation 4). Given the strong positive relationship between the absolute abundance of the pmoA gene and CH₄ uptake observed across sites. The quantitative relationship between pmoA gene abundance and soil CH₄ uptake was evaluated using log‐linear, log‐power function, and exponential transformation models (the latter being mathematically equivalent to a standard power law). The exponential transformation model yielded the optimal fit (R ² = 0.520, p < 0.001; Figure S1), Consequently, an exponential transformation function (r_mt) was introduced to modulate the base oxidation rate (k₀) by pmoA gene abundance, thereby determining the CH₄ oxidation rate constant (k_d) (Equation 4–5):

$$\begin{array}{c}{k}_d=k_0\times r_{\mathit{mt}}\times r_T\times r_{\mathit{sm}}\times r_N\end{array}$$ (4)

$$\begin{array}{c}{r}_{\mathit{mt}}=a\times \mathit{exp}\left(\beta \times {\mathit{log}}_{10}\left(\text{Absolute abundance of}\mathit{\text{pmoA}}\right)\right)\end{array}$$ (5)

where α and β are parameters optimized via Bayesian optimization using the mlrMBO package in R (Bischl et al. 2017). The search space was set to α ∈ [0, 5] and β ∈ [0, 1], with the augmented expected improvement (AEI) acquisition function. Model calibration for FGA‐MeMo was performed by minimizing the root mean square error (RMSE) between simulated and observed CH₄ efflux data. During calibration, the key parameters in the temperature and moisture response functions were kept constant. After 1000 iterations, the 10 best‐performing parameter sets were selected to form an ensemble. The final prediction was taken as the mean of the model outputs from these sets, providing a robust estimate that accounts for parameter uncertainty (see Table S2 for detailed results).

The training and validation datasets for the FGA‐MeMo model were derived from a publicly available dataset (Cheng et al. 2026). The dataset was randomly partitioned into training and validation sets using a 7:3 ratio, with stratification by ecosystem type to ensure representative sampling. The global distribution map of the pmoA gene for upland soils was generated based on predictions from the machine learning model that achieved the highest validation performance (lowest RMSE).

We then used FGA‐MeMo models to project future changes in the global soil CH₄ sink under two Shared Socioeconomic Pathway (SSP) scenarios: SSP1‐2.6 (representing a low‐emission pathway with sustainable development and efficient resource use) and SSP5‐8.5 (representing rapid economic growth with limited climate change mitigation) at a spatial resolution of 1° × 1°. First, due to the current lack of comprehensive future projections of environmental variables (Table S1) under the SSP1‐2.6 and SSP5‐8.5 scenarios, we held these environmental factors constant at their present‐day values. We then applied the previously trained random forest model to project future annual absolute abundances of soil pmoA gene, allowing only mean annual air temperature (MAT) and mean annual precipitation (MAP) to vary dynamically according to scenario‐specific climate projections over time. Environmental drivers, including MAP, MAT, soil temperature, soil moisture, nitrogen deposition, and atmospheric CH₄ concentration, were obtained from CMIP6 and ISIMIP3b simulations under the SSP1‐2.6 and SSP5‐8.5 scenarios (Coupled Model Intercomparison Project Phase 6: https://aims2.llnl.gov/search/cmip6/; Inter‐Sectoral Impact Model Intercomparison Project 3b: https://data.isimip.org/search/query/nhx/tree/ISIMIP3b/) (For detailed data sources, see the Data Availability section). To quantify the impact of uncertainty in soil pmoA gene abundance predictions on global soil CH₄ sink estimates, we employed a Monte Carlo simulation approach. Specifically, we assumed that the predicted pmoA values for each grid cell followed a normal distribution, with the mean and standard deviation (SD) defined by the Random Forest model's predictions and their corresponding prediction errors, respectively. Based on this distribution, we performed 100 stochastic iterations. In each iteration, we generated a random realization field of global pmoA abundance and input it into the FGA‐MeMo model to recalculate the global soil CH₄ sink. Finally, the uncertainty in the global soil CH₄ sink attributable to pmoA prediction errors was characterized by the standard deviation of the 100 simulation outcomes. Subsequently, we analyzed the spatial distribution of differences in their predictions to identify regions with significant divergence. An area‐weighted average of the CH₄ sink was computed for each grid cell to estimate the global soil CH₄ sink. The effective land area for each grid cell was derived from the ESA WorldCover 2021 10‐m resolution land cover dataset (https://doi.org/10.5281/zenodo.7254221), excluding non‐upland environments such as wetlands, water bodies, deserts, and glaciers. We also analyzed changes in the global soil CH₄ sink and the contribution of different latitudinal bands under future scenarios. Finally, we compared our results with simulations from the MeMo model.

2.4. Statistical Analysis

Use one‐way ANOVA to examine differences in the absolute abundance of the pmoA gene across different land‐use types. First, we performed Levene's test to assess homogeneity of variances within groups and test for normality. Then we proceeded to multiple comparisons: if variances are homogeneous, we used Tukey HSD; if variances are not homogeneous, we used the Kruskal‐Wallis test. We used Locally Estimated Scatterplot Smoothing (LOESS) to explore nonlinear trends between the absolute abundance of the pmoA gene and key environmental drivers. LOESS does not require specifying a global functional form, allowing it to flexibly capture local patterns in data with complex structures (Jacoby 2000). All statistical analyzes described in the article were conducted using R (v4.3.2).

3. Result

3.1. Drivers of Absolute Abundance and Distribution of pmoA Gene in Global Upland Soils

This study synthesized data from 72 previously published studies that quantified the absolute abundance of the pmoA gene in soils, yielding a final dataset of 698 records. The dataset encompasses a variety of ecosystem types, with relatively undisturbed natural systems (e.g., forests, shrublands, grasslands) accounting for 80% and human‐impacted environments (e.g., croplands and urban areas) making up the remaining 20%.

Across the entire dataset, the mean pmoA gene abundance is 6.48 ± 1.50 log₁₀ (gene copies g⁻¹ soil) (n = 698) (Figure 1b). Notably, pmoA gene abundance varies substantially across different soil ecosystems: forest soils exhibit the highest abundance (7.43 ± 0.12 log₁₀ (gene copies g⁻¹ soil), n = 75), whereas urban soils show the lowest (4.82 ± 0.06 log₁₀ (gene copies g⁻¹ soil), n = 75). Statistical analysis confirmed that these differences are significant (p < 0.05) (Figure 1c). Furthermore, pmoA gene abundance shows a significant positive correlation with soil CH₄ uptake rates (Figure S1, p < 0.01). The observed differences in pmoA gene abundance across ecosystem types are highly consistent with current ecological understanding, suggesting that the globally compiled pmoA gene abundance data from upland soils in this meta‐analysis are unlikely to be strongly influenced by methodological biases (Lafuente et al. 2019; Hedenec et al. ²⁰²⁴). This dataset therefore provides a robust basis for predicting the global distribution of pmoA genes in upland soils.

FIGURE 1.

Distribution of sampling site from the literature (a), histogram of the global absolute abundance of the pmoA gene in soils (b), and pmoA gene abundance across land‐use types (c). Different lowercase letters indicate significant differences (p < 0.05) based on one‐way ANOVA followed by Tukey's post hoc test.

The mutual information (MI) between each of the 27 environmental variables and pmoA gene absolute abundance exceeded 0.2 (p < 0.05; Figure 2a), indicating that all these variables exert varying degrees of influence on pmoA gene abundance. Among them, the strongest associations with pmoA gene abundance were observed for soil calcium carbonate (CaCO₃), pH, soil extractable magnesium (Mg²⁺), and soil organic carbon (SOC). Locally weighted scatterplot smoothing (LOWESS) trends revealed that pmoA gene abundance changes with soil calcium carbonate content: it remains relatively stable at low concentrations, increases markedly at intermediate levels, and eventually reaches a saturation state at high concentrations. When calcium carbonate content exceeds 4%, the absolute abundance of the pmoA gene plateaus, with no further significant increases (Figure 2b).

FIGURE 2.

Mutual information (MI) between 27 environmental factors and absolute abundance of the pmoA gene in soils (a). The four environmental variables with the strongest associations with soil pmoA gene abundance are soil calcium carbonate (b), pH (c), exchangeable magnesium (d), and soil organic carbon (e). Each panel represents a scatter plot of pmoA gene abundance versus the corresponding environmental factor; the green line represents the locally weighted scatterplot smoothing (LOESS) trend, and the shaded area denotes the 95% prediction interval.

A unimodal response of pmoA gene abundance to soil pH was observed, with peak abundance occurring within the optimal pH range of 7.2–8.0 (Figure 2c). The relationship between pmoA gene abundance and extractable magnesium content followed a pattern similar to that of calcium carbonate: when extractable Mg exceeds 0.02 cmol kg⁻¹, pmoA gene abundance plateaus (Figure 2d). Meanwhile, the association between pmoA gene abundance and soil organic carbon (SOC) is nonlinear, with a threshold at approximately 1.5% SOC. Below this threshold, pmoA gene abundance decreases as SOC increases; above it, abundance increases with rising SOC (Figure 2e).

Following feature screening, nine environmental variables were excluded from further analysis: base saturation (BS), cation exchange capacity (CEC), electrical conductivity (EC), exchangeable sodium percentage (ESP), exchangeable calcium (Ca²⁺), exchangeable potassium (K⁺), gross domestic product (GDP), heterotrophic respiration (Rh), and total nitrogen (TN) (Figure 2a and Figure S2). The remaining 18 environmental variables were selected as input variables for subsequent modeling (Figure 3a).

FIGURE 3.

Workflow for the machine learning‐based prediction of the absolute abundance of the pmoA gene in global upland soil (a), performance of each model on the test dataset (b), and comparison of model evaluation metrics (c). The metrics include Pearson correlation coefficient (r), coefficient of determination (R²), mean absolute error (MAE), and root mean square error (RMSE). The Pearson correlation coefficient measures the linear association between predicted and observed values but does not reflect the magnitude of absolute prediction errors. The coefficient of determination quantifies the proportion of variance in observations explained by the model but is sensitive to bias. The mean absolute error represents the average absolute difference between predictions and observations, providing an intuitive measure of average error magnitude with lower sensitivity to outliers. The root mean square error assesses predictive accuracy as the square root of the average of squared errors, making it more sensitive to large deviations and outliers.

Among the five candidate models, the Random Forest model outperformed the other four on the test set, achieving a Pearson's correlation coefficient (r) of 0.83, a coefficient of determination (R ²) of 0.68, a mean absolute error (MAE) of 0.53 log₁₀ (gene copies g⁻¹ soil), and a root mean square error (RMSE) of 0.82 log₁₀ (gene copies g⁻¹ soil) (Figure 3). Compared with the K‐nearest neighbors (KNN), support vector machine (SVM), and elastic net regression (ENR) models, its RMSE was 10.30%–29.49% lower; it also slightly outperformed the extreme gradient boosting (XGBoost) model. Thus, the Random Forest model was selected as the optimal model for subsequent analyzes.

Based on the Random Forest model, the estimated pmoA gene abundance in soils ranges from 2.43 to 9.30 log₁₀ (gene copies g⁻¹ soil) with a mean of 6.74 ± 0.95 log₁₀ (gene copies g⁻¹ soil). A clear latitudinal gradient is evident: pmoA gene abundance increases toward mid‐latitudes and declines thereafter. The model predicts the highest pmoA gene abundance in northern Central Asia and the lowest in Japan (Figure 4).

FIGURE 4.

Random forest predictions of the absolute abundance of the pmoA gene in global upland soil (a), frequency distribution of predictions (b), and latitudinal distribution of predictions (c). Light green: Standard deviation of predictions; orange dots: Observed pmoA values grouped by latitude.

To assess prediction uncertainty, the standard deviation of predictions across the 1000 regression trees in the Random Forest model was used as a metric. Results showed that uncertainty is lower in mid‐latitude regions, indicating greater model confidence in these areas. In contrast, predictions for low‐latitude and high‐latitude regions exhibit higher uncertainty (Figure S3).

3.2. Development and Validation of a Soil CH₄ Sink Model Incorporating Functional Gene Abundance

In this study, the absolute abundance of the pmoA gene, used as a proxy for CH₄ oxidation capacity, was integrated to constrain the soil CH₄ oxidation rate constant (k_d). This integration led to the development of a new model, that is, FGA‐MeMo, which was subsequently evaluated against the original MeMo model. Detailed parameter values for both models are provided in Table S2.

Model evaluation results demonstrate that FGA‐MeMo outperforms MeMo in predicting the global spatial patterns of soil CH₄ sinks. Performance was assessed on the test set using four metrics: Pearson's correlation coefficient (r), coefficient of determination (R ²), mean absolute error (MAE), and root mean square error (RMSE). On the test set, FGA‐MeMo achieved higher fitting precision than the MeMo model, evidenced by a substantial 62.98% improvement in R ² (Figure 5 and Figure S4). In contrast, the decrease in RMSE was limited to 4.30% (Figure 5). FGA‐MeMo exhibited markedly enhanced predictive performance specifically in mid‐to‐high latitudes (30°–90°) (Figure S4).

FIGURE 5.

Comparison of global soil CH₄ sink simulated by the FGA‐MeMo and MeMo models against observational data from the test dataset (a). Differences between predicted and observed values were calculated and model performance was evaluated using four metrics: Pearson correlation coefficient (b), coefficient of determination (R²) (c), mean absolute error (MAE) (d), and root mean square error (RMSE) (e).

According to FGA‐MeMo simulations, the global mean soil CH₄ sink during 2015 was 503.38 ± 227.61 mg CH₄ m⁻² year⁻¹, corresponding to an annual total sink of 46.91 ± 0.61 Tg CH₄ year⁻¹ (± indicates the interannual standard deviation for 2015–2019 under the SSP1‐2.6 scenario, Figure 6). According to FGA‐MeMo, the soil CH₄ sink exhibits a distinct latitudinal pattern, peaking at 10°–20° N and 30°–40° S in the Northern and Southern Hemispheres, respectively, before tapering off. In contrast to MeMo, FGA‐MeMo reproduces the observed latitudinal gradient with greater fidelity (Figure 6c).

FIGURE 6.

Spatial pattern of global soil CH₄ sink simulated by FGA‐MeMo (a), frequency distribution of simulated sink values (b), latitudinal profile of the simulated sink (c), and projected changes in the global soil CH₄ sink under future climate scenarios (d).

At the regional scale, FGA‐MeMo predictions indicate enhanced soil CH₄ sinks in the Mediterranean, the Middle East, northern Central Asia, Africa, the US‐Mexico border region, central South America, and Australia (Figure 6). Conversely, lower sink magnitudes are predicted for near‐equatorial regions and high‐latitude areas of the Northern Hemisphere. Notably, compared to the MeMo model, FGA‐MeMo estimates substantially higher soil CH₄ uptake in northern Central Asia, central Africa, and central South America (Figure S5).

3.3. Response of Global Upland Soil CH₄ Sink Under Future Climate Scenarios

Under the SSP1‐2.6 scenario, a low‐emission pathway centered on sustainable development, the FGA‐MeMo model projects a gradual decline in the global upland soil CH₄ sink, with the sink decreasing to 27.19 ± 0.15 Tg CH₄ year⁻¹ by 2100. Even amid this downward trend, FGA‐MeMo's estimates of the CH₄ sink remain consistently higher than those of the MeMo model throughout the 21st century (Figure S6).

In contrast, under the SSP5‐8.5 scenario, a high‐emission pathway defined by rapid economic growth and limited climate mitigation measures, the model projects a different trajectory: the global upland soil CH₄ sink first increases, peaks at 71.27 ± 0.38 Tg CH₄ year⁻¹ around 2075, and then declines slightly, reaching 64.68 ± 0.35 Tg CH₄ year⁻¹ by 2100 (Figure S5). Analysis of latitudinal patterns further reveals regional disparities in soil CH₄ sink responses: the relative contribution of low‐latitude regions (< 40°) to the global soil CH₄ sink shows a declining trend, whereas that of high‐latitude regions (> 40°) exhibits an increasing trend (Figure S7).

4. Discussion

4.1. Advancing Global CH₄ Sink Modeling by Integrating Functional Gene Abundance

In upland soils, methanogenesis is negligible due to well‐aerated conditions—an environment unsuitable for this anaerobic process. As a result, these soils function predominantly as a net sink for atmospheric CH₄, with their CH₄ flux dynamics primarily regulated by aerobic methanotrophs (Kolb 2009; Zhang et al. ²⁰²¹; Saunois et al. ²⁰²⁵). However, current global estimates of the upland soil CH₄ sink rely heavily on simplified mechanistic models that overlook spatial variability in CH₄ oxidation capacity. This oversight limits the accuracy of CH₄ uptake simulations across complex environmental gradients (Kolb 2009; Curry ²⁰⁰⁷; Zhuang et al. ²⁰¹³; Murguia‐Flores et al. ²⁰¹⁸; Lafuente et al. ²⁰¹⁹).

Recent studies emphasize the need for community‐level functional attributes‐integrated biogeochemical models—models that link methanotroph community characteristics to ecosystem‐scale CH₄ uptake (Wang et al. 2022; Ou et al. ²⁰²⁴; Gao et al. ²⁰²⁰; Guo et al. ²⁰²⁰; Tao et al. ²⁰²⁴; Li et al. ²⁰²⁵; Zhou et al. ²⁰²⁵). Specifically, the functional gene abundance (quantified via the pmoA gene) more accurately reflects both the size of the functional microbial community and its metabolic potential. At regional scales, this metric also shows a stronger correlation with soil CH₄ oxidation rates (Zhang et al. 2021; Li et al. ²⁰²⁵; Zhou et al. ²⁰²⁵). Thus, incorporating the absolute abundance of the pmoA gene into global soil CH₄ sink models is critical for improving their performance.

This study quantifies the relationship between the absolute abundance of the pmoA gene and the soil CH₄ sink at the global scale (Figure S1; R ² = 0.52) and predicts the spatial distribution of this abundance in global upland soils (Figure 4). Based on these data, we developed a regulatory factor, r_mt, to quantify the influence of methanotrophic communities on the base oxidation rate (k₀). In the original MeMo model, methanotrophic activity is governed by uniform temperature and moisture response functions, with k₀ assigned fixed values ranging from 1.6 × 10⁻⁵ to 5 × 10⁻⁵ s⁻¹ depending on ecosystem type. Consequently, the CH₄ oxidation potential within each ecosystem is assumed to be homogeneous (Saunois et al. 2025; Murguia‐Flores et al. ²⁰¹⁸). In contrast, FGA‐MeMo treats r_mt as a continuous variable, allowing CH₄ oxidation potential to vary dynamically across grid cells in response to pmoA gene abundance, with values ranging from 0.26 to 7.62. This mechanistic representation, which integrates functional gene data with the spatial heterogeneity of soil CH₄ sinks, more realistically captures the spatial variability of microbial communities. This advancement is identified as the key driver behind the significantly improved accuracy of the FGA‐MeMo model (Figure 5) (Murguia‐Flores et al. ^{2018, 2021}).

Notably, while FGA‐MeMo demonstrates a significant improvement in explanatory power (R ²) compared to MeMo, the reduction in RMSE remains modest. This discrepancy likely stems from the fact that FGA‐MeMo retains the uniform soil temperature and moisture response functions inherited from the original MeMo model. In reality, however, methanotrophic community composition varies significantly across environmental gradients, which presumably leads to divergent responses to soil temperature and moisture (Kolb 2009; Jiang et al. ²⁰²⁵; Cheng et al. ²⁰²⁶). Currently, globally observational data on methanotrophic community composition remain extremely scarce. Moreover, recent studies have revealed that high‐affinity and low‐affinity methane oxidation capabilities can coexist within the same microbial taxa (Schmider et al. 2024). Constrained by both the paucity of comprehensive community‐level observations and an incomplete understanding of the functional complexity of methanotrophs, current models are unable to explicitly represent their functional diversity or the heterogeneity in their responses to environmental drivers. Consequently, although FGA‐MeMo successfully captures the spatial patterns (hotspots and coldspots) of CH₄ sinks via pmoA abundance, it struggles to precisely simulate the absolute flux magnitudes under diverse climatic conditions. This systematic bias not only limits the reduction of RMSE but also leads to an overestimation of the global soil CH₄ sink by 16–17 Tg CH₄ year⁻¹ relative to the MeMo model, slightly exceeding the mean estimates from most atmospheric inversion studies (Saunois et al. 2025; Lee et al. ²⁰²³). Future model development should therefore integrate metagenomic data to better resolve the functional differentiation among methanotrophic taxa and establish phylogenetically informed environmental response parameters. Such advances would enable a more mechanistic and biologically grounded linkage between microbial identity and the rates of methane oxidation they mediate.

4.2. Climatic Controls and Functional Gene Abundance Co‐Determine Global Patterns of Upland Soil CH₄ Sinks

FGA‐MeMo simulations reveal a non‐linear latitudinal pattern in the global upland soil CH₄ sink (Figure 6). This spatial pattern is primarily governed by climatic controls: in low‐latitude tropical regions, high precipitation reduces soil aeration, significantly suppressing methanotrophic metabolic activity; whereas in high‐latitude regions, low‐temperature constraints diminish methanotroph activity (Kolb 2009; Murguia‐Flores et al. ²⁰²¹). Notably, climate factors do not represent the sole deterministic mechanism. Methanotroph abundance also plays a critical role in regulating the spatial heterogeneity of the global CH₄ sink (Nazaries et al. 2018; Lafuente et al. ²⁰¹⁹; Hedenec et al. ²⁰²⁴).

The distribution of the pmoA gene reveals a pronounced mid‐ to high‐latitude enrichment pattern in methanotroph abundance, with hotspot regions concentrated in northern Europe and northern Central Asia. Correspondingly, soil CH₄ uptake is also projected to be elevated in these regions (Figure 4). This aligns with previous studies showing that higher methanotroph abundance in mid‐ to high‐latitude regions (e.g., northern Europe and northern Central Asia) significantly enhances soil CH₄ consumption capacity (Martins et al. 2017; Belova et al. ²⁰²⁰; Sabrekov et al. ²⁰²⁰).

Two key mechanisms could drive this pattern: First, nutrient‐rich forest soils in northern Europe provide favorable organic carbon conditions for methanotroph growth (Martins et al. 2017; Sabrekov et al. ²⁰²⁰; Lee et al. ²⁰²³; Tao et al. ²⁰²³). Second, in northern Central Asia, despite low precipitation and high evaporation, well‐aerated soils under low‐moisture conditions enhance gas diffusivity. This increases the availability of CH₄ as a substrate, thereby stimulating methanotroph metabolism and growth (Belova et al. 2020; Lee et al. ²⁰²³).

In contrast, in arid regions of central Africa adjacent to deserts, where pmoA gene abundance is low. This discrepancy primarily reflects prolonged drought stress, which limits soil water availability and directly suppresses methanotroph activity, and ultimately reduces methanotroph abundance (Angel and Conrad 2009). Consequently, compared to the MeMo model, FGA‐MeMo predicts a markedly smaller increase in soil methane uptake for this region relative to other areas (Figure S5).

In summary, integrating the absolute abundance of the pmoA gene into process‐based models enables a more realistic representation of how environmental changes (e.g., shifts in soil nutrients, drought conditions) affect the soil CH₄ sink, thereby improving the accuracy of global CH₄ uptake estimates. This study provides a feasible framework for linking community‐level functional attributes to ecosystem‐scale carbon cycle modeling and offers insights for developing more mechanistic and predictive terrestrial biogeochemical models.

Nevertheless, a critical limitation of this study must be acknowledged. It is well established that high‐affinity methanotrophs (specifically USCα, USCγ, and certain pmoA2 lineages) are the primary drivers of the upland soil methane sink (Kolb et al. 2003; Täumer et al. ^{2021, 2022}). However, the vast majority of existing soil pmoA datasets were generated using broad‐range primers⁴, which fail to differentiate between high‐ and low‐affinity lineages. Although specific primers for USCα exist, there is currently no universal primer set capable of simultaneously targeting all major high‐affinity groups (e.g., both USCα and USCγ) (Kolb et al. 2003; Täumer et al. ^{2021, 2022}). Consequently, there is a paucity of data to support global‐scale mapping of atmospheric methanotrophs. Despite this constraint, our analysis reveals that total pmoA gene abundance still exhibits a significant positive correlation with soil CH₄ uptake (R ² = 0.52; Figure S1). This positive relationship has been corroborated by multiple studies (Zhang et al. 2021; Hedenec et al. ²⁰²⁴) and likely arises because high‐affinity methanotrophs dominate the methanotrophic community (> 90%) in certain upland ecosystems, such as temperate forests (Kolb et al. 2005; Pratscher et al. ²⁰¹¹). Therefore, non‐specific total pmoA abundance can serve as an effective proxy for methane oxidation potential. In this sense, our work constitutes a “proof of concept”, demonstrating that integrating even existing total pmoA data into models can significantly enhance the simulation performance of soil CH₄ sinks. Naturally, future work should also develop molecular tools capable of simultaneously targeting methanotrophs with high‐affinity CH₄ oxidation functionality, or integrate metagenomics, to directly and quantitatively resolve the contribution of different methanotrophic groups to the global soil CH₄ sink (Liu et al. 2026).

4.3. Future Projections From a Process‐Based Soil CH₄ Sink Model Incorporating Functional Gene Abundance

Relative to the MeMo model, FGA‐MeMo projects a substantially larger increase in the global upland soil CH₄ sink under high‐emission scenarios. Under SSP5‐8.5, the sink grows steadily: by 2075, an additional 25.01 Tg CH₄ year⁻¹ of uptake is projected compared to 2015, exceeding MeMo's projection of 14.82 Tg CH₄ year⁻¹ (Figure S5). This enhanced predictive capacity stems primarily from FGA‐MeMo's improved representation of how regional in the absolute abundance of the pmoA gene modulates the soil CH₄ sink.

Under SSP5‐8.5, high‐latitude warming, characterized by shorter winters, permafrost degradation, extended growing seasons, and improved soil temperature‐moisture regimes, stimulates methanotroph activity and boosts soil CH₄ consumption (Murguia‐Flores et al. ²⁰²¹; Zhou et al. ²⁰²⁵). In contrast, tropical and subtropical regions face amplified hydroclimatic extremes due to heightened rainfall variability: wet areas experience increased precipitation, while dry regions become drier. This intensifies soil moisture stress, reduces CH₄ oxidation rates, and diminishes local CH₄ sink capacity (Murguia‐Flores et al. ²⁰²¹; Zhang et al. ²⁰²¹) (Figure S6). Against this backdrop, FGA‐MeMo simulations reveal a spatial redistribution of the global soil CH₄ sink, characterized by an enhanced proportion in temperate regions and a diminished role in tropical and subtropical areas (Figure S7).

By 2100, the contribution of boreal regions (latitude > 60°) to the global soil CH₄ sink increases by 2.38 Tg CH₄ year⁻¹ relative to 2015, surpassing MeMo's projected 1.44 Tg CH₄ year⁻¹ increase (Figure S6). This contrast demonstrates that, accounting for its spatial heterogeneity is essential to accurately capture the projected rise in high‐latitude CH₄ uptake under climate change. Nevertheless, current models estimate future changes in pmoA gene abundance based exclusively on variations in temperature and precipitation, potentially constraining the accuracy of projected trajectories for the soil CH₄ sink.

Under SSP5‐8.5, sustained enhancement of CH₄ uptake could promote methanotroph growth and accumulation, further amplifying the soil CH₄ sink—suggesting current models may underestimate absorption potential in high‐emission futures. Conversely, under SSP1‐2.6, a declining soil CH₄ sink could reduce methanotroph abundance, potentially leading models to overestimate long‐term CH₄ consumption. Thus, future models that couple microbial population dynamics (e.g., growth‐mortality balance, community succession) could enhance the realism of long‐term projections for soil CH₄ sink evolution. Beyond natural climatic controls, anthropogenic disturbances significantly weaken the soil CH₄ sink (Murguia‐Flores et al. ²⁰¹⁸). Our analysis reveals markedly lower absolute abundance of the pmoA gene in urban and agricultural ecosystems (Figure 1)—driven by soil compaction (and associated impaired gas diffusion) from land conversion, as well as nitrogen toxicity resulting from intensive fertilization and atmospheric nitrogen deposition (Lafuente et al. 2019; Feng et al. ²⁰²⁰; Hedenec et al. ²⁰²⁴). As urbanization and agricultural intensification expand, these anthropogenic stressors may further suppress methanotroph populations. The omission of these processes in current models, however, risks overestimating future CH₄ sink capacity (Saunois et al. 2025).

Despite its advancements, the FGA‐MeMo model has several limitations: First, the inability to simulate CH₄ production: The model does not integrate CH₄ production processes, rendering it inapplicable to ecosystems with periodic inundation or frequent wet‐dry cycles (e.g., wetlands, rice paddies, riparian zones, and tropical forests under heavy rainfall). This limits its utility in regions with complex hydrology (Feng et al. 2020; Gauci ²⁰²⁵). Future research could integrate methanogenesis‐associated genes (mcrA) and methanotrophy‐associated genes (pmoA) into a unified framework to concurrently represent CH₄ production and oxidation, thereby facilitating a more process‐based understanding of net CH₄ source‐sink dynamics of net soil CH₄ fluxes. Second, spatial variability in prediction accuracy: The projected accuracy of the absolute abundance of the pmoA gene varies spatially. Performance is stronger in mid‐latitude regions (due to denser data coverage), but uncertainty remains high in low‐latitude arid zones and high‐latitude permafrost regions. These areas require enhanced field observations to address extreme environmental conditions and data scarcity (Figure S3). Third, the omission of key physiological regulators: Critical physiological factors, such as copper and iron, which govern particulate CH₄ monooxygenase (encoded by the pmoA gene) synthesis and thus influence methanotroph activity and abundance, have not been incorporated into the current model framework. This may compromise the accuracy of global predictions. Fourth, the neglect of woody vegetation CH₄ fluxes: Recent studies suggest tree stems can act as both CH₄ sources and sinks, but the underlying mechanisms remain unclear (Gauci et al. 2024; Gauci ²⁰²⁵; Keppler et al. ²⁰⁰⁶). Future work should quantify CH₄ fluxes from woody vegetation at the ecosystem scale and integrate these processes into regional CH₄ models to improve assessments of the terrestrial CH₄ budget (Martins et al. 2017).

5. Conclusion

We developed the FGA‐MeMo—a terrestrial CH₄ sink model that integrates the absolute abundance of the pmoA gene. Compared to the original MeMo model, FGA‐MeMo substantially improves the accuracy of global upland soil CH₄ sink estimates, enabling more mechanistically robust simulations in key regions (e.g., boreal forests of Northern Europe, northern Central Asia, and arid areas of Africa). Relative to MeMo, FGA‐MeMo estimates an additional 16–17 Tg CH₄ year⁻¹ in the global soil methane sink. Future scenario simulations using FGA‐MeMo indicate that under the high‐emission pathway (SSP5‐8.5), global upland soils show a markedly enhanced potential for increased CH₄ uptake, reaching 64.68 ± 0.35 Tg CH₄ year⁻¹ by 2100. This upward trend is particularly pronounced in high‐latitude regions—where higher methanotroph abundance confers stronger CH₄ oxidation capacity. These results highlight the need to systematically integrate community‐level functional attributes into Earth system models and underscore the growing importance of upland soils as a CH₄ sink in the global CH₄ cycle.

Author Contributions

Wensheng Xiao: conceptualization, methodology, software, data curation, investigation, validation, formal analysis, visualization, writing – review and editing, writing – original draft, resources. Xiaoqi Zhou: conceptualization, methodology, funding acquisition, project administration, formal analysis, supervision, resources, writing – review and editing, validation. Paul L. E. Bodelier: writing – review and editing, methodology. Jizhong Zhou: writing – review and editing. Li Cheng: conceptualization, validation, writing – review and editing. Zhifeng Yang: writing – review and editing. Gangsheng Wang: writing – review and editing, methodology.

Funding

This work was supported by the National Natural Science Foundation of China, No. 32171635.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Fit between the absolute abundance of the pmoA gene and soil CH₄ sink.

Figure S2: Pearson correlation coefficients between 27 environmental factors and the absolute abundance of the pmoA gene in upland soil.

Figure S3: Spatial uncertainty of the random forest model predictions for the absolute abundance of the pmoA gene expressed as standard deviation (SD).

Figure S4: Performance of the FGA‐MeMo and MeMo models on the test dataset for low‐latitude (a) and mid‐to‐high‐latitude (b) regions.

Figure S5: Spatial difference in global soil CH₄ sink between FGA‐MeMo and MeMo simulations at the grid‐cell level.

Figure S6: Projected changes in future global soil CH₄ sink simulated by FGA‐MeMo and MeMo under two Shared Socioeconomic Pathways (SSPs). SSP1‐2.6 envisions a more sustainable world with strong mitigation policies to reduce greenhouse gas emissions and promote clean energy and advanced technologies. In this pathway, aggressive emission reductions are expected to limit global warming to below 2°C by 2100, helping to mitigate the adverse effects of climate change. SSP5‐8.5 represents a fossil‐fuel‐driven development pathway with high economic growth and continued reliance on conventional energy resources (e.g., coal, oil, and natural gas). Under this scenario, global mean temperature is projected to rise by approximately 3.2°C to 5.4°C by 2100 relative to pre‐industrial levels, leading to significant climate change impacts, including increased frequency of extreme weather events and sea‐level rise.

Figure S7: Changes in the proportion of soil CH₄ sink contributed by different global regions to the total global CH₄ sink under the SSP1‐2.6 (a) and SSP5‐8.5 (b) scenarios, as simulated by the FGA‐MeMo and MeMo models.

Table S1: Detailed information on the gridded global datasets utilized in this study, including units of measurement, data categories, spatial resolution, temporal coverage (years), and original sources.

Table S2: Comparison of model parameters between FGA‐MeMo and MeMo (parameters for FGA‐MeMo are averaged from 10 best‐performing parameter sets).

Data Availability Statement

The atmospheric temperature and precipitation data used in the Random Forest model to project future changes in pmoA gene abundance were derived from simulations of the BCC‐CSM2‐MR model within the Coupled Model Intercomparison Project Phase 6 (CMIP6) (https://aims2.llnl.gov/search/cmip6/). Bulk density and clay content were obtained from http://globalchange.bnu.edu.cn. Land cover dataset was obtained from https://doi.org/10.5281/zenodo.7254221. N fertilizer application was obtained from https://doi.pangaea.de/10.1594/PANGAEA.861203. For future scenario projections of soil CH₄ sink in FGA‐MeMo and MeMo, atmospheric CH₄ mole fraction, soil temperature, and soil moisture data were derived from CMIP6 simulations of the BCC‐CSM2‐MR model (https://aims2.llnl.gov/search/cmip6/). Atmospheric N deposition data for future scenarios were sourced from the Inter‐Sectoral Impact Model Intercomparison Project Phase 3b (ISIMIP3b): https://data.isimip.org/search/query/nhx/tree/ISIMIP3b/. The pmoA gene abundance data and soil CH₄ flux data from upland soils collected through literature synthesis in this study, along with the global pmoA gene abundance predictions from the random forest model and the projected changes in the global soil methane sink, have been uploaded to Figshare (https://doi.org/10.6084/m9.figshare.30609971). The R code used in this study for data analysis, machine learning, and the FGA‐MeMo model has been uploaded to Figshare (https://doi.org/10.6084/m9.figshare.30609971).