Machine learning and omics based systems and methods for assessing microbial functional composition

Abstract

Systems and methods for determining information representing an amount of an indicator of a soil trait based on a machine learning engine, and systems and methods for training the machine learning engine are described. An example method includes obtaining values of a plurality of environmental factors relating to a location, and determining the information representing the amount of the indicator by inputting the values of the plurality of environmental factors into a machine learning engine. The machine learning engine is trained using sample datasets each of which includes values of at least a portion of the plurality of environmental factors of a sample site and information representing an amount of the indicator of the sample site.

Description

MACHINE LEARNING AND OMICS BASED SYSTEMS AND METHODS FOR ASSESSING MICROBIAL FUNCTIONAL COMPOSITIONSTATEMENT REGARDING FEDERALLY FUNDED RESEARCH

[0001] This invention was made with government support under Grant Nos. 2022055 awarded by NSF, and 2022-67020-36123 awarded by USDA / NIFA. The government has certain rights in the invention.CROSS-REFERENCE TO RELATED APPLICATION S)

[0002] This application claims priority to the provisional application with serial number 63 / 520,572, titled “MACHINE LEARNING AND OMICS BASED SYSTEMS AND METHODS FOR IDENTIFYING MICROBIAL FUNCTIONAL COMPOSITION,” filed August 18, 2023. The entire contents of the above noted provisional application are incorporated by reference as part of the disclosure of this document.TECHNICAL FIELD

[0003] The disclosed technology relates to machine learning based systems and methods for assessing microbial functional composition using omics and environmental data.BACKGROUND

[0004] Soil microbial communities synthesize and release diverse extracellular enzymes for catalyzing soil carbon and nutrient cycling and thereby regulate soil carbon sequestration and nutrient availability of the ecosystem. These extracellular enzymes are involved in decomposing various soil organic matter (SOM) components. The amounts and / or compositions of these enzymes may vary across different environments.SUMMARY

[0005] An aspect of the present document relates to a system and a method for determining information representing an amount of an indicator of a soil trait. Another aspect of the present document relates to a system and method for training a machine learning engine configured to determine information representing an amount of an indicator of a soil trait. In a further aspect of the present document, the above-described systems or methods are embodied in the form of processor-executable code and stored in one or more non-transitory computer-readable storage media. The code included in the computer readable storage media when executed by one or moreprocessors, causes the one or more processors to implement the methods described in the present document.BRIEF DESCRIPTION OF DRAWINGS

[0006] Figure 1 illustrates information representing the amounts of enzyme functional classes (EFCs) in soil samples from various regions across the United States.

[0007] Figure 2 illustrates Taylor diagrams that compare ensemble models’ predictions of information representing the amounts of various EFC ensembles or EFCs against observed data in accordance with example embodiments of the present document.

[0008] Figure 3 illustrates the relevance (quantified using Gini value) of environmental factors for predicting information representing the amounts of various EFC ensembles in accordance with example embodiments of the present document.

[0009] Figure 4 illustrates predicted information representing the amounts of EFCs involved in soil organic matter (SOM) decomposition using environment-informed machine learning models in accordance with example embodiments of the present document.

[0010] Figure 5 illustrates predicted information representing the amounts of various EFC ensembles at multiple levels of environmental conditions.

[0011] Figure 6 is a flowchart of a process for determining information representing the amount of an indicator of a soil trait in accordance with example embodiments of the present document.

[0012] Figure 7 is a flowchart of a process for training a machine learning engine configured to determine information representing the amount of an indicator of a soil trait in accordance with example embodiments of the present document.

[0013] Figure 8 is a flowchart of a machine leaning pipeline for training, validating, and / or predicting information representing the amount of an EFC in accordance with example embodiments of the present document.

[0014] Figure 9 illustrates results of sensitivity analysis on model performance metrics with various split ratios of training and validating data in accordance with example embodiments of the present document.

[0015] Figure 10 illustrates relevance (quantified using Gini value) of environmental factors for predicting information representing the amounts of various EFC ensembles in accordance with example embodiments of the present document.

[0016] Figure 11 illustrates a principal component analysis (PCA) of relevant environmental factors in accordance with example embodiments of the present document.

[0017] Figure 12 illustrates predicted information representing the amounts of various EFCs using environment-informed machine learning models in accordance with example embodiments of the present document.

[0018] Figure 13 illustrates predicted information representing the amounts of various EFCs involved in lignocellulose decomposition at multiple levels of environmental conditions in accordance with example embodiments of the present document.

[0019] Figure 14 illustrates predicted information representing the amounts of various EFCs for N-rich soil components decomposition, mineralization, and transformation at multiple levels of environmental conditions in accordance with example embodiments of the present document.

[0020] Figure 15 illustrates predicted information representing the amounts of various EFCs for P-rich soil components decomposition, mineralization, and transformation at multiple levels of environmental conditions in accordance with example embodiments of the present document.

[0021] Figure 16 illustrates soil C / N ratios of various land covers of multiple sites across the United States.DETAILED DESCRIPTION1 Introduction

[0022] The disclosed embodiments relate to methods, devices and systems that among other features and benefits predict soil enzyme functional composition across diverse environments.

[0023] Microbial enzyme-mediated soil organic matter (SOM) decomposition regulates many key ecosystem functions such as elemental cycling, soil carbon sequestration, and soil fertility. However, representing microbial processes in Earth system models (ESMs) is still a challenge due to the lack of a clear understanding regarding the spatial patterns of functional diversity within microbial communities and how diverse environments regulate these communities. We synthesized published metagenomics data across the US to identify microbial genes for general enzyme functional classes (EFCs) involved in soil carbon (C), nitrogen (N), and phosphorus (P) cycles. EFCs including groups of gene-encoded enzymes involving decomposition pathways from a specific chemical compound, e.g., carbohydrates and lignin for C, proteins and chitin for N, and nucleic acids, phospholipids, and inositol for P. We integrated environmental data to trainand test a machine learning (ML) pipeline for predicting the potential EFC composition. Eight EFCs for soil C, N, and P cycling were identified across the continental U.S., and we found their relative abundances varied with spatial differences in edaphic, climate, and vegetation conditions. Soil water regulated the abundance of total extracellular enzymes by decreasing enzyme investment for C and nutrient acquisition from complex soil substrates under waterlimited environments. Soil C, C / N ratio, and soil texture regulated the relative abundances of different enzymes for decomposing diverse soil substrates. Our findings highlighted that environmental factors could be incorporated into the developed ML model to predict soil enzyme functional composition across diverse environments.

[0024] Microbial-mediated soil organic matter decomposition regulates soil nutrient cycling, carbon sequestration, and soil fertility. Elucidating the environmental regulation on the spatial patterns of functional diversity within microbial communities may allow better assessment of the implication of microbial-mediated soil decomposition for ecosystem functions. We synthesized published molecular biological data of soil microbial communities across the US to identify general microbial enzyme functional classes involved in soil carbon, nitrogen, and phosphorus cycles. An enzyme functional class refers to a group of microbial functions for decomposing a group of similar soil substrates. We trained and tested a machine learning pipeline to learn how environmental factors regulate the composition of microbial functional classes. Merely for illustration purposes and not intended to be limiting, our results indicated that eight microbial functional classes for soil carbon, nitrogen, and phosphorus cycling were identified across the continental U.S. Their relative composition varied with spatial differences in edaphic, climate, and vegetation conditions. Soil microbial communities decreased their functions for decomposing soil organic matter under water-limited environments. Soil texture, soil carbon content, and its ratio to soil nitrogen regulated the relative composition of microbial functions for the decomposition of diverse soil substrates. Our finding provides a machine learning tool for quantifying the functional composition of soil microbial communities under diverse environments.

[0025] Soil microbial communities synthesize and release diverse extracellular enzymes for catalyzing soil carbon and nutrient cycling and thereby regulate soil carbon sequestration and nutrient availability of the ecosystem. To understand the mechanisms that regulate key ecosystem functions, such as nutrient cycling, organic matter decomposition, soil health, andplant productivity, it is beneficial to elucidate the role of microbially enzyme-mediated soil biogeochemical cycling. Improved understanding of these mechanisms can advance the predictive power of the land carbon-climate feedback, which is a major source of uncertainty in the projected future climate.

[0026] To date, interpreting and predicting the role of microbial enzyme-mediated soil biogeochemical cycling remains a challenge. Although soil biogeochemical models have been advanced to explicitly represent microbial functions in soil organic matter (SOM) decomposition and variation with soil temperature, moisture, and chemical substrates, many models still exhibit large uncertainties in reproducing the observed response of microbial carbon emission to environmental change. To overcome this deficiency in models, a more detailed understanding of the taxonomic and functional composition of soil microbial communities and their variability across diverse environments is needed.

[0027] High-throughput omics technologies, such as metagenomics, metatranscriptomics, and metaproteomics (“omics”) can identify both the taxonomic composition and structure of microbial communities. For example, metagenomics may be utilized to identify genes to produce extracellular enzymes that enable lignocellulose degradation, to study their implications for soil biogeochemistry in a tropical peat swamp forest, and / or to quantify microbially-mediated hydrocarbon degradation rates and bioremediation treatments. As another example, a multi-site metagenome dataset may be used to identify genes encoding for the enzyme polyethylene terephthalate hydrolase and / or estimate its global distribution. As a further example, metagenomic datasets may be used to identify the global biogeography of microbial functional traits involved in inorganic N transformation processes, such as nitrification, denitrification, ammonia assimilation, etc. Previous efforts have focused on gene-encoded enzyme functions involved in specific metabolic processes. Fewer studies have examined enzymes responsible for the decomposition of soil organic carbon (C), nitrogen(N), and phosphorus (P).

[0028] Applying omics to identify community-level microbial functions involved in SOM decomposition may be complex and redundant, with multiple soil organic substrates relying on common enzymes and many enzymes catalyzing similar substrates. In a site-level study, we utilized metagenomics and metaproteomics to identify 118 genes responsible for producing soil enzymes involved in soil C, N, and P cycling. For illustration purposes and not intended to be limiting, we grouped these gene-encoded enzymes into eight enzyme functional classes (EFCs).An EFC may include a group of enzymes that catalyze the decomposition or mineralization of similar chemical substrates (e.g., carbohydrates). The EFC approach may reduce the complexity and redundancy of omics information and / or enable linkages with site-level soil biogeochemical prediction. Applying the EFC approach to a soil biogeochemical model at a tropical site may improve predictions of soil C emission under nutrient-limited conditions. However, questions remained as to whether the identified 118 genes for enzyme production for soil C, N, and P cycling and the corresponding eight EFCs can be found across diverse environments and how the relative abundances of these EFCs vary with environmental conditions.

[0029] Previous studies have shown a relationship between microbial taxonomic composition and a range of abiotic and biotic factors at various scales, such as soil C, land cover type, soil N content, soil P content, soil C / N ratio, latitude, elevation, soil pH, precipitation, temperature, agricultural management, and soil texture. By incorporating these evident environmental regulators and microbial taxonomic composition to train machine learning (ML) models, recent efforts have enabled the assessment of the distribution of microbial taxonomic composition, structures, and specific functions, as well as their relationships with environmental factors. One study integrated ultraviolet light, vegetation conditions, soil C, and pH with a random forest model to predict worldwide soil bacterial alpha diversity. One study utilized a neural network to predict the relationships between microbial taxonomic composition and environmental factors at the global scale. An effort used random forest and a neural network to select relevant microbial taxa for predicting plant litter decomposition. These studies demonstrated the effectiveness of ML in interpreting the environmental regulations of soil microbial community diversity and leveraging these relationships to predict microbial diversity. Building upon these studies, it seems possible to use omics and environmental data-trained ML models to predict and assess the spatial pattern of soil EFC composition for soil C, N, and P cycling at large spatial scales.

[0030] The regulation of environmental factors on soil enzyme composition has been proposed using the resource optimization theory, which assumed that microbial communities may optimize enzyme synthesis to mitigate resource limitations. Observations at the site scale demonstrated this theory under nutrient-limited or water-limited conditions. However, it’s still unclear whether resource optimization strategies of microbial communities widely exist across heterogeneous environments. Here we hypothesized that: (i) the general gene-encoded potential enzymes regulating soil C, N, and P cycling can be identified across diverse environments and groupedinto EFCs; (ii) EFC abundances can be predicted by ML models using omics-informed EFC composition and the corresponding environmental factors ; (iii) according to resource optimization theory, the relative abundances of EFCs vary across diverse environments depending on spatial variability in climate, edaphic, and aboveground vegetation conditions. This study tests these hypotheses by integrating published soil metagenomics datasets and the corresponding edaphic, climatic, and vegetation information in diverse environments (Section 2.1) and applying them to develop and train an ML pipeline (Section 2.2). Section 3.1 addresses hypothesis (i) by identifying EFC compositions using collected metagenomics datasets across 20 representative sites in the continental U.S. Section 3.2 addresses hypothesis (ii) by training, validating, and testing ML models for predicting EFC abundances. Section 3.3 discusses the relevant environmental factors for predicting the abundances of eight EFCs and addressed hypothesis (iii). Section 3.4 applies the trained ML pipeline to generate a dataset to represent the EFC composition of soil microbial communities involved in soil C, N, and P cycling across the continental U.S.2 Example Materials and Methods2.1 Metagenomics and environmental data collection and integration

[0031] We used metagenomics to identify potential soil enzyme functional composition and integrated this information with the corresponding environmental data. Firstly, we extracted the genes related to SOM decomposition and inorganic N and P transformation from each metagenomic dataset and estimated their abundances by recording gene counts (section 2.1.1). Then the extracted genes and their abundances were grouped into eight EFCs (section 2.1.2). Finally, we collected environmental data related to each metagenomics dataset and paired it with the corresponding EFC abundances (section 2.1.3). This methodology enabled us to utilize omics to inform the microbial functional composition and link their spatial distribution to diverse environmental conditions.2.1.1 Metagenomics-informed potential enzyme functional composition

[0032] To identify microbial enzyme functional composition across the U.S. domain, we collected the existing metagenomics data from the Integrated Microbial Genomes and Microbiomes (IMG / M) system of the Department of Energy (DOE)'s Joint Genome Institute(JGI). Here we targeted the shotgun metagenomic sequencing data from soil samples in the environmental manipulation and gradients experiments. We collected 277 metagenomic datasets from 20 sites that represented diverse environments across the U.S. (Figure 8 of Appendix A). The gene functional composition of each metagenomic dataset was identified using JGI IMG / M genome set function profile identification tool. Here we identified genes regulating the production of specific enzymes, which we referred to as gene-encoded enzymes. The function of each gene-encoded potential enzyme was identified using the assigned enzyme commission (EC) number by the JGI function profile identification tool. The abundance of each gene-encoded enzyme within a dataset was quantified using reported gene counts assigned to each EC number.2.1.2 Identifying enzyme functional classes for catalyzing soil C, N, and P cyclingTable 1. Definition of enzyme functional classes (EFC)

[0033] Following our previous studies that identified 118 genes for enzyme production related to soil C, N, and P cycling, we examined the existence of these genes across diverse environments by searching them from each metagenomics dataset. To reduce the complexity and redundancy of omics-informed enzyme functional information, we employed our enzyme functional classification approach to classify 118 enzymes into three ensembles of EFCs (Table 1). The firstensemble represented EFCs for lignocellulose decomposition, including an EFC for carbohydrate decomposition (EFCcarbohydrate) and an EFC for lignin decomposition (EFCtignin). The second ensemble represented EFCs for N-rich soil components decomposition, mineralization, and transformation, including one EFC involved in organic N decomposition (EFCorganicx), one EFC involved in monomer N mineralization (EFCMonomerN), and one EFC for inorganic N transformation (EFCinorganicN). The third ensemble referred to EFCs for P-rich soil decomposition, mineralization, and immobilization, including one EFC for organic P decomposition (EFCorganicp), one EFC for monophosphorus mineralization (EFC ionomerp), and one EFC for inorganic P immobilization (EFCinorganicp). The detailed definition of each EFC was listed in Table 1 and included enzymes within each EFC can be found in supporting data set SI of Appendix A.

[0034] The abundance of an EFC was estimated by summarizing gene counts assigned to all enzymes within this EFC. As multiple metagenomic datasets from a site may have different total gene numbers, the raw abundance of an EFC was then normalized to the relative abundance by dividing by the total gene counts of the corresponding soil sample. The relative abundances of each EFC within an ensemble were then summed to represent the abundance of the corresponding EFC ensemble.2.1.3 Example Environmental data

[0035] To investigate the effect of environmental regulation on gene-encoded potential enzyme composition within each dataset, we collected the corresponding environmental metadata and experimental information for each dataset from the corresponding publications and metadata reported in the JGI IMG / M database. Nineteen biotic and abiotic factors that may affect SOM decomposition (Table SI of Appendix A were collected and grouped into five categories: climate drivers, soil physical properties, soil chemical properties, aboveground vegetation, land cover information, and experimental treatments (added N and P fertilizer amounts). When site-specific environmental data was unavailable, the corresponding data was extracted from global or regional databases (Table SI of Appendix A) using the site’s longitude and latitude. These environmental data were paired with the corresponding omics-informed EFCs abundance data to train and test the ML model for predicting EFC composition.2.2 Developing a machine learning pipeline for predicting environment-informedEFC composition

[0036] We built an ML pipeline to integrate environmental information for predicting gene- encoded enzyme functional composition across diverse environments (Figure 8 of Appendix A). Our ML pipeline was composed of (1) a data preprocess for preparing ML training, testing, and validation datasets, (2) a data augmentation scheme that increased training, testing, and validation data when omics data was limited for ML training; and (3) an ensemble learning framework that automatically aggregated prediction powers of multiple ensemble ML models.2.2.1 Example Data preprocessing

[0037] To make the data suitable for training our ML pipeline and increase prediction efficiency and accuracy, we performed the data preprocessing by factors pre-selection, categorical data conversion, and data standardization. Here factors data referred to collected biotic and abiotic environmental data corresponding to each omics dataset, including climate, soil physical, soil chemical, and vegetation factors (Table SI of Appendix A). Each variable represented one dimension in the factor data. We reduced collinearity among environmental factors by excluding soil N and clay content as they can be derived from other environmental factors, such as soil C / N ratio and soil silt and sand fraction. Finally, the relative abundances of three EFC ensembles and each EFC belonging to an ensemble (Table 1) within each dataset were labeled by the corresponding factor data. This preprocess generated a dataset of 277 instances for training and validating the ML pipeline. Each instance was composed of relative abundances of three EFC ensembles and eight specific EFCs, and the corresponding selected 17 environmental factors.2.2.2 Example Data augmentation

[0038] Our omics dataset was composed of 277 samples from 20 sites, which may limit the ability of the ML models to learn the relationships between environmental factors and EFC relative abundances. To overcome this challenge and increase the prediction power of our ML pipeline, we used a data augmentation technique to increase the amount of training, testing, and validation data. Data augmentation is the process of inflating limited data by either data warping or oversampling and has been used in ML-based image classification and identification. For example, the random erasing data augmentation increased image data by identifying an area inan image and randomly assigning each pixel in this area a random value in the range of RGB color scheme (0-255). Following the same data inflating strategy, we reproduced the data by randomly removing the enzymes in each EFC from the original data. We identified an enzyme within a randomly selected EFC in a soil sample based on given rules. The selected enzyme was analogized to a “pixel” in an image. Instead of assigning a random value, we set the gene abundance of the selected enzyme to zero. To keep enzyme composition within an EFC still reflecting the real EFC composition and therefore can be labeled by the corresponding environmental factor, we selected the enzyme following three rules: a) The abundance of the erased enzyme within an EFC was smaller than the standard deviation of EFC relative abundance; b) The abundance of the erased enzyme accounted for less than 20% of its affiliated EFC abundance; c) Removing this enzyme didn’t change the ranking of the relative abundances of eight EFCs. We performed this data augmentation process for 227 samples for ten iterations and inflated training and testing data to 3,047, including original and reproduced ones.2.2.3 Example ensemble machine learning framework

[0039] To increase the prediction power of the EFCs abundances under diverse environmental conditions, we developed an ensemble ML framework that first integrated environmental-labeled EFC composition data to train and test four ensemble ML models, including bagging, random forest, adaptive and gradient boosting, and then aggregated prediction outcomes of four ensembles models using stacked generalization ensemble (Figure 8 of Appendix A).

[0040] We selected and incorporated these algorithms to utilize their diverse prediction characteristics. Bagging fitted multiple decision trees on different subsets of a training dataset, then combined the predictions from all models. This method was highlighted by its ability to increase prediction accuracy compared with a single predictor trained on the whole training set. The random forest extended the bagging algorithm by randomly sampling subsets of factors for predicting each data sample. This algorithm resulted in greater model diversity and usually allowed searching for a desirable factor for performing predictions and increasing prediction accuracy. Unlike bagging and random forest, boosting combined several weak learners (base models) and boosted them into a strong learner. Here we applied the two popular boosting methods: adaptive boosting and gradient boosting. Adaptive boosting increased the weight of data that had bad performance previously in the current training process. Instead, gradientboosting sequentially added predictors to a training process and allowed for correcting its predecessor by minimizing the residual errors that the predecessor made. Overall, boosting focused more on under-fitted data points by directly increasing their weights (adaptive boosting) or indirectly leveraging residual error (gradient boosting). Therefore, boosting efficiently improved microbial functional composition predictions of regions with fewer samples (‘underfitted data points’), such as the southeast with only 14 samples. Finally, the predictions of bagging, random forest, adaptive boosting, and gradient boosting were integrated using stacking, a method to minimize the ensemble prediction error generated from multiple base estimators. Instead of simply using base models’ average as the ensemble prediction, stacking performed this aggregation by fitting a model (e.g., random forest) to the target variables with all predictions of base estimators as the input factors. Therefore, using stacking as the last step of this ML framework allowed us to utilize the prediction power of four ensembles and generate the final prediction.

[0041] To reduce prediction uncertainty and better assess the overall prediction ability, we integrated Monte Carlo cross-validation to split 3047 data into training, testing, and validation data randomly and repeat the training, validating, and testing of each ensemble model for 1000 iterations. During each iteration, the training, validation, and testing data sizes accounted for 36%, 24%, and 40% of the total data, respectively. Our sensitivity analysis of model performances to different training data sizes indicated that using 36% of the total data as the training data may achieve a desirable model performance (Figure 9 of Appendix A).

[0042] To train the ML models, we performed the following ML training, validation, and testing processes for predicting the abundances of each EFC and each EFC ensemble, respectively. First, bagging, random forest, adaptive boosting, and gradient boosting were trained using the environmental factors-labeled EFC abundance data from the training dataset, respectively (Figure 8 of Appendix A). Environmental factors from the validation dataset were then incorporated into each trained ensemble model to predict the abundances of the corresponding EFC or EFC ensemble. The resultant predictions of each model were then labeled with the real abundances of the corresponding EFC or EFC ensembles from the validation dataset and stacked into a new dataset to train a gradient boosting ensemble to make the stacking prediction of the abundances of EFC or EFC ensembles. Key hyperparameters (Table S2 of Appendix A) used in each model were optimized using the grid search approach. Finally, we applied the testingdataset to the trained four first layers of ensemble models and the stacking model to assess our ML pipeline's ability to predict the abundances of eight EFCs and their EFC ensembles (Figure 8). The detailed processes of training, validating, and testing each ML model were described in the Text SI of Appendix A.

[0043] Each model's ability to predict the relative abundance of an EFC or EFC ensemble was quantified by calculating four metrics: (1) the coefficient of determination (R2), (2) the Pearson correlation coefficient (R), (3) the root-mean-square error (RMSE) between observed and predicted EFC or EFC ensemble abundances and (4) standard deviation (SD) of predicted EFC or EFC ensemble abundances. The overall prediction performance of this ML pipeline was quantified as the Mean ± SD of R2 values of the final stacking model based on 1000 iterations. A higher R2 value indicated better agreement between observed and predicted relative abundances of an EFC. A lower SD of R2 values based on 1000 prediction interactions suggested the convergence of model prediction agreement.2.3 Interpreting ML-predicted environment-EFC composition relationship

[0044] We performed factor relevance analysis and Principal Component Analysis (PCA) to interpret the ML-predicted environments-EFC composition relationship. First, we assessed the relative importance of each environmental factor for predicting EFC composition via Gini importance analysis, which measured the average gain of purity by splitting with a given factor variable in the decision tree. Each decision tree divided the data within the training dataset into subsets with a similar abundance of an EFC or EFC ensemble based on a specific environmental factor. The change of data variance within each subset resulting from the split of this environmental factor was calculated as the Gini value or impurity. A high Gini value means low similarity in a group. A significant decrease in Gini value before and after splitting with an environmental condition indicated this factor was effective in classifying similar samples and predicting EFC composition. We calculated how each environmental factor, on average, decreased the Gini value for all nodes in all decision trees. The average decreases by preprocessed 17 environmental factors over all trees in an ensemble were standardized to represent their relative importance to EFC abundance predictions.

[0045] We further performed PCA using the Scikit-learn PCA package to reduce the dimension of relevant environmental factors and identify principal environmental factor components thatmay regulate the relative abundance of each EFC and EFC ensemble across diverse environments. The relevance or importance of each principal environmental component was measured by its accounted variance of relevant environmental factors. The projection of an environmental factor on a principal component was used to assess its relative contribution to the corresponding principal component. In our study, each sample’s environmental conditions were visualized by the top three principal components and their factor contributions.2.4 Predicting composition of EFCs involved in soil C, N, and P cycling across the diverse U.S. environments

[0046] To identify the spatial variability of the composition of EFCs involved in the soil C, N, and P cycle across diverse U.S. environments, we incorporated environmental factors from 1278 sites into our ML pipeline. Here 1278 sites were independent of sites used in training, validating, and testing the ML pipeline and from the global soil organic C and N dataset, which reported soil C and N contents for more than 3500 soil profiles and standardized at a depth of one meter. Building upon this dataset, we extracted other environmental factors for the ML pipeline from multiple regional-scale datasets (Table SI of Appendix A). In detail, climate variables were collected from the PRISM. Soil texture and elevation data were from the Vegetation / Ecosystem Modeling and Analysis Project (VEMAP) dataset and USGS National Elevation Datasets (USGS, 2000), respectively. Vegetation ND VI and NPP data were from prior products derived from Landsat 30m and MODIS 250m by others. Land cover data were from National Land Cover Dataset (NLCD). Ecoregion and soil order were from Gridded Soil Survey Geographic Database (gSSURGO) (Soil Survey Staff, 2021). Finally, this environmental metadata for EFC predictions included ML pipeline-relevant environmental factors for 1278 sites across the U.S. We assessed the differences in ML -predicted EFC composition across the gradients of the relevant environmental factors to interpret the regulation of principle environmental factors on the spatial distribution of EFC composition.3. Representative Results3.1 Example Gene-encoded potential enzyme functional classes for soil C, N, and P cycling in the collected metagenomics datasets

[0047] The 118 general gene-encoded enzymes involved in SOM decomposition, mineralization, or inorganic N and P transformation were identified from all 277 soil samples across 20 sites. These enzymes accounted for about 5% of total gene abundances within each soil sample.

[0048] Although general EFC composition was identified across 20 sites, the relative abundances of each EFC showed large variability among different sites depending on their environmental heterogeneity. By grouping EFCs within various sites into five geographical regions, we found that the relative abundances of EFCs varied with regions across the U.S.

[0049] Figure 1 illustrates information representing the amounts of enzyme functional classes (EFCs) in soil samples from various regions across the United States. Specifically, Figure 1 shows relative abundances of (a) total EFCs for soil C, N, and P cycling, (b) EFCs for lignocellulose decomposition (EFCcarbohydrate and EFCrignin), (c) EFCs for N-rich soil components decomposition, mineralization, and transformation (EFCorganicN, EFCMonomerN, EFCinorganicN), and (d) EFCs for P-rich soil decomposition, mineralization, and immobilization (EFCorganicP, EFCwonomerP, EFCinorganicp) of collected soil samples across different regions in U.S. The error bars represented the standard deviation of the corresponding EFC abundances from samples within the same region. The range of each region can be found in Figure 8 of Appendix A.Letters represent the results of Tukey’s test, with different letters representative of statistically significant differences between the two corresponding regions (p<0.05).

[0050] In general, the abundances of EFCs ensemble for lignocellulose decomposition were the lowest in the Southwest sites (1.50%±0.2%) but the highest in the Midwest sites (2.60%±0.5%) (Figure la). The relative abundance of EFCs specifically for carbohydrate decomposition (EFCcarbohydrate) was the lowest in the Southwest sites (1.24%±0.25%) but the highest in the Midwest sites (2.36%±0.49%) (Figure lb). In contrast, the relative abundance of EFCs for lignin decomposition (EFCrignin) was significantly higher in the Southwest sites than that in the Midwest sites (p value<0.05). Among 20 sites within five regions, the abundance of EFCrignin was the highest in the Southeast sites (0.28%+0.02%) (Figure lb).

[0051] The Midwest sites had statistically lowest abundances (1.40%±0.44%) of the EFC ensemble involved in the decomposition, mineralization, and transformation of N-rich SOM (EFCN) among five regions (p value<0.05) (Figure la). The lower ensemble EFCN abundances were mainly attributed to the lower abundances of specific EFCs for N-rich SOM decomposition (EFCorganicN) and inorganic N transformation (EFCinorganicN) (Figure 1c). Among the five regions, the Southwest had the lowest abundance of EFC for monomer N mineralization (EFCMonomerN) (0.04%±0.01%) and also had statistically lower abundances of EFCorganicN (0.88%±0.18%). These regional differences in the abundances within the EFCN ensemble indicated the distinct preferences of N acquisition sources among different regions by the microbial communities.

[0052] The Midwest sites had 0.50%±0.09% of the EFC ensemble involved in the decomposition, mineralization, and transformation of P-rich SOM (EFCp), which was statistically lower than the other sites (p value<0.05) (Figure la). The relative abundances of specific EFCs for the P-rich SOM decomposition (EFCorganicp) in the West and the Midwest sites were statistically lower than that in the Northwest and the Southeast sites. The relative abundances of EFCMonomerP were around 0.15% at most sites, except in the Midwest sites, which had the statistically lowest abundances of 0.10%+ 0.05% (Figure Id). In contrast, the abundance of EFC for P immobilization (EFCinorganicp) did not show significant differences among 20 sites.

[0053] Besides differences in EFC abundances among sites within the collected metagenomics dataset, there was also a consistent trend of EFC abundances across these sites. The abundances of EFCc were usually higher than EFCN, and EFCc and EFCN were usually 3 times more than EFCP (Figure la). In detail, the abundances of EFC carbohydrate were around 5-10 times the abundance of EFCLignin (Figure lb). The abundances of EFCorganicN were around 10 to 20 times more than EFCMonomerN, and 20% to 50% more than EFCinorganicN (Figure 1c). The abundances of EFCorganicp were around 2-3 times the abundances of EFCMonomerP, which were usually 20% more than EFCinorganicp (Figure Id). Overall, there were high abundances of EFCs for the decomposition of carbohydrates, organic N, and organic P, and low abundances of EFCs for the decomposition of lignin, monomer N, and monomer P.3.2 Accuracy of Example ML model for soil EFC abundances predictions

[0054] We assessed the developed ML pipeline for predicting EFC abundances using four metrics. Metrics R2, R, and RMSE of predicted EFC abundances were used to evaluate theagreement between omics-informed EFC abundance and the prediction by each ML model within the ML pipeline. In addition, we assessed the ML model's ability to capture the variation of observed EFC abundance by comparing the SD values of omics-informed EFC abundances with the SD values of the predicted abundances by each ML model. All reported metrics were the mean values of 1,000 iterations of the ML predictions. The reported mean values of R, RMSE of predicted EFC abundances, and the SD values of both omics-informed and predicted EFC abundances were shown using the Taylor diagrams (Figure 2). The reported mean R2and its variation among 1000 iterations of the ML predictions were shown in Table S3 of Appendix A.3.2.1 Example ML model performance for the training dataset

[0055] All five example ensemble models used in our ML pipeline showed agreements between omics-informed abundances of EFCs and EFC ensembles and the corresponding predictions using training data, with the averaged R2 values above 0.87 (Table S3 of Appendix A) and R values above 0.9 (Figure 2). The modeled RMSE values were usually less than 30% of SD values of the omics-informed abundances of an EFC or EFC ensembles within the training dataset (Figure 2), indicating that prediction error was far smaller than the spatial variability of omics-informed abundances of the corresponding EFC or EFC ensembles. The SD values of R2 of all EFCs and EFC ensembles in over 1000 iterations were less than 0.02 (Table S3 of Appendix A), indicating that the ML pipeline was stable in predicting EFC composition using the training dataset. The consistency between modeled and observed SD values for EFCC, EFCN, EFCP, EFCcarbohydratc, EFCorganicN, EFClnorganicN, and EFCMonomcrP (Figure 2a-C, d, f, h, j) indicated that the ML models also captured the spatial variability of abundances of these EFCs and EFC ensembles within the training dataset. However, the ML models slightly underestimated the variability of abundances of EFCLignin, EFCMonomerN, EFCorganicP, and EFCinorganicP within the training dataset, as shown by the lower modeled SD values compared to the corresponding measured values (Figure 2e, g, i, k).

[0056] Figure 2 illustrates Taylor diagrams that compare ensemble models’ predictions of information representing the amounts of various EFC ensembles or EFCs against observed data in accordance with example embodiments of the present document. Specifically, Figure 2 shows Taylor diagrams to compare five ensemble models’ predictions of abundances of (a-c) three EFC ensembles and (d-k) eight specific EFCs, with respect to the observed data in terms of root meansquare error (RMSE), correlation coefficient (R), and standard deviation (SD), in a polar coordinate system. Here RMSE, R, and SD values were the average of 1000 iterations of predictions. Circle and triangle represented simulation with training and testing datasets, respectively. The five ensemble models used here include a gradient boosting model, an adaptive boosting model, a bagging model, a random forest model, and a stacking model.

[0057] There was no significant difference in the ability of the five ensemble models to predict the abundances of most EFCs and EFC ensembles within the training dataset, except for EFCorganicP and EFCinorganicP. EFCorganicP and EFCinorganicP were better predicted by bagging, random forest, and final stacking models. The R values of the bagging, random forest, and stacking predictions were 0.93 (Figure 2i) for EFCorganicP and 0.90 (Figure 2k) for EFCinorganicP respectively, which were almost 3%-8% higher than the predicted R-value by the adaptive boosting and gradient boosting. Also, bagging, random forest, and stacking slightly better captured the variability of abundances of EFCorganicP and EFCinorganicP within the training dataset, with the predicted SD values closer to the measured SD values than the adaptive and gradient boosting (Figure 2k and i).3.2.2 Example ML model performance for the testing dataset

[0058] All five trained example ensemble models within the ML pipeline were able to capture the abundances of all three EFC ensembles and all eight EFCs within the testing dataset, with the averaged R2 values around 0.84 (Table S3 of Appendix A) and most of R values above 0.9 (Figure 2), and RMSE values less than 50% of observed SD values (Figure 2). Most of the SD values of R2 for EFC and EFC ensembles over 1000 iterations of predictions were lower than 0.05, and EFCinorganicP had SD values of more than 0.1 in bagging, random forest, and stacking (Table S3 of Appendix A). This result indicated that the ML pipeline exhibited stable performance in predicting most EFCs and EFC ensembles within the testing dataset. In addition, the trained ML models were able to capture the variability of abundances of most EFC and EFC ensembles within the testing dataset, except for EFCtignin, EFCMonomerN, EFCorganicP, and EFCinorganicP. The abundance variabilities of these EFCs within the testing dataset were underestimated, as indicated by lower predicted SD values than the corresponding observed SD values (Figure 2e, g, i, k).

[0059] Five ensemble models also didn’t show significant differences in their prediction power for most EFCs in the testing dataset, except for the prediction of EFCinorganicP. Compared to the equivalent prediction power of bagging, random forest, and gradient boosting for EFCinorganicP abundances using the training dataset, stacking and adaptive boosting appeared to better predict EFCinorganicP abundance. The R values of EFCinorganicP prediction by stacking and adaptive boosting were around 0.78, which were around 3% higher than the R values of the bagging, random forest, and gradient boosting (Figure 2k).

[0060] Our results indicated that the developed ML pipeline could integrate soil omics and the corresponding environmental data to train ML models to predict the EFC compositions of soil microbial communities. ML model performance in predicting the abundances of EFCinorganicP was usually not as good as for other EFC abundances (Figure 2). This model deficiency may result from the lack of inorganic P and soil C / P ratios in our environmental datasets. Although five ensemble models within the ML pipeline showed similar prediction power of the abundances of EFC ensembles and EFCs, the final stacking model improved the prediction of EFCinorganicP abundances despite the lack of inorganic P and soil C / P information in our training and testing dataset.3.3 Examples of relevant environmental factors for predicting EFC composition

[0061] To interpret the relevance of environmental factors with EFC abundances, we ranked the relevancies of environmental factors applied in adaptive, gradient boosting, random forest, and bagging using the Gini factor importance analysis. As the resultant predictions of these four ML models were then applied in the training stacking model, the identified relevant environmental factors for these four ML models were then integrated to identify relevant environmental factors for the final stacking prediction in our ML pipeline. The larger Gini value indicated that selecting the corresponding environmental factor was more relevant and effective for predicting EFC abundances. To summarize the main environmental factors that regulate EFC abundances, we performed PCA analysis to decrease the dimension of multiple relevant environmental factors.

[0062] Soil C / N ratio, land cover type, ND VI, soil C content, elevation, and latitude were the most relevant factors for predicting EFCc abundances by four ensemble models (Figure 3a). Specifically, these environmental factors played equivalent roles in predicting EFCcarbohydrate abundances by random forest, bagging, and gradient boosting (Figure 10a of Appendix A).However, the land cover type was the most relevant factor for predicting EFCcarbohydrate by the adaptive boosting (Gini value=27%) (Figure 10a of Appendix A). In contrast, soil C / N ratio and annual accumulated precipitation were the dominant factors for predicting EFCLignin, with accumulated importance of more than 50% by the random forest, gradient boosting and bagging (Figure 10b of Appendix A).

[0063] Soil C / N ratio, soil C content, land cover type, latitude, ecoregion, silt content, elevation, ND VI, and sampling depth were relevant environmental factors for predicting EFCN abundances by four ensemble models (Figure 3b). Among these factors, the soil C / N ratio was the most relevant factor, with the Gini value around 20 - 25% for the predictions by four ensemble models (Figure 3b). Soil C content was the second relevant factor for predicting EFCN by random forest (Gini value=20%), bagging (Gini value=20%), and gradient boosting (Gini value=16%), while soil sampling depth was the second most relevant factor for predicting EFCN by adaptive boosting, with the Gini value of 20% (Figure 3b).

[0064] Figure 3 illustrates the relevance (quantified using Gini value) of environmental factors for predicting information representing the amounts of various EFC ensembles in accordance with example embodiments of the present document. Specifically, Figure 3 shows relevance (quantified using Gini value) of environmental factors for predicting the abundances of three EFC ensembles: (a) EFC ensemble for lignocellulose decomposition (EFCc); (b) EFC ensemble for N-rich soil components decomposition, mineralization, and transformation (EFCN). (C) EFC ensemble for P-rich soil decomposition, mineralization, and immobilization (EFCp). The relevance value (in terms of the Gini value) was averaged over 1000 iterations of predictions. Here we only showed the Gini values for the top 10 most relevant environmental factors identified in our studies.

[0065] For three EFCs within the EFCN ensemble, the top three relevant factors for predicting EFCinorganicN abundance by random forest, gradient boosting, and bagging models were land cover type (Gini value=29%), latitude (Gini value =28%), and soil C / N ratio (Gini value =20%). In contrast, soil C / N ratio was the most dominant factor (Gini value =45%) in adaptive boosting prediction, followed by soil C content (Figure lOe of Appendix A). Silt content was dominant for predicting EFCorganicN abundance by random forest, gradient boosting, and bagging with Gini values of 62% (Figure 10c of Appendix A). Besides silt content, soil C content was also relevant for predicting EFCorganicN abundance by adaptive boosting. Silt and soil C contents had Ginivalues of more than 20%, compared with lower than 10% of Gini values of the remaining factors (Figure 10c of Appendix A). Four ML models had similar factor importance for predicting EFCMonomerN abundances, with elevation being dominant (Gini value =45%), followed by soil sampling depth and latitude with Gini values of 16% (Figure lOd of Appendix A).

[0066] Unlike the dependence of the predictions of EFCC and EFCN on multiple environmental factors, the prediction of EFCp abundances by four ensemble models was dominated by soil sampling depth, with the Gini value of 70%-80% (Figure 3c). Specifically, soil sampling depth was the most relevant factor, among those tested, for predicting EFCorganicP abundances with a Gini value of 58% (Figure lOf of Appendix A). Besides soil depth, soil sand content, soil C / N ratio, and temperature were also relevant for predicting EFCinorganicP abundances by random forest, gradient boosting, and bagging. These four factors had Gini values of 15% - 20%, followed by soil C content with Gini values around 8% (Figure lOh of Appendix A). In comparison to the strongest regulation of soil depth on the abundance predictions of EFCorganicP and EFCinorganicP, silt content was the most relevant factor for the prediction of EFCMonomerP abundances, with Gini values of around 20% for the prediction by random forest, bagging, and gradient boosting, and 36% for the prediction by adaptive boosting prediction (Figure 10g of Appendix A). In addition, annual accumulated precipitation, soil C / N ratio, and land cover type played equivalently relevant roles in predicting EFCMonomerP abundances by random forest, bagging, and gradient boosting with Gini values of 15%, followed by soil C content, sampling depth, and latitude with Gini values of 8%. (Figure 10g of Appendix A). In contrast, ND VI was the second relevant factor for predicting EFCMonomerP by adaptive boosting, with a Gini value of 20% (Figure 10g of Appendix A).

[0067] In summary, the four example ML models identified consistently essential environmental factors for predicting three EFC ensembles and their eight EFCs, including soil C / N ratio, land cover type, soil C content, latitude, elevation, ecoregion, soil sand and silt fraction, soil sampling depth, ND VI, and annual accumulated precipitation. To better interpret the co-regulation of these relevant environmental factors on EFC abundance predictions, we classified these relevant environmental factors into three principal components using PCA. Three principal components explained 44.1%, 21.6%, and 11.3% of the total variance of these relevant environmental factor data (Figure 11 of Appendix A). The first principal component was mainly composed of soil C content, soil C / N ratio, elevation, land cover type, and soil depth, representing the verticalvariability in chemical properties in the soil samples. The second principal component was mainly contributed by soil silt fraction, annual accumulated precipitation, ecoregion, and ND VI, illustrating vegetation factors associated with soil water availability. The third principal component included annual mean temperature, ND VI, elevation, latitude, and soil sand content, representing vegetation factors associated with temperature topography.3.4 Example ML pipeline prediction of the spatial variability of EEC composition across diverse U.S. environments

[0068] By incorporating available environmental factors data from the 1278 sites into the trained and tested ML pipeline, we predicted the abundances of all EFCs involved in soil C, N, and P cycling ranged from 3.5% to 4.8% across 1278 soil sites in the U.S. (Figure 4a). The spatial differences in abundances of total EFCs involved in soil C, N, and P cycling resulted from the regional variability in annual accumulated precipitation and soil C content. Sites with annual precipitation above 1000mm and soil C content above 1.0% usually held more than 4% of EFCs involved in C, N, and P cycling (Figure 4b, c).

[0069] Figure 4 illustrates predicted information representing the amounts of EFCs involved in soil organic matter (SOM) decomposition using environment-informed machine learning models in accordance with example embodiments of the present document. Specifically, Figure 4 shows predicted relative abundance of EFC involved in SOM decomposition using environment- informed ML pipeline, (a) Spatial distribution of predicted EFC relative abundances at 1278 sites, (b)-(c) Predicted EFC relative abundances at different soil C content (%) and mean annual accumulated precipitation, respectively. Here each box plot described the mean, 25% and 75% quantile, minimum and maximum values, and outliers of predicted EFC abundances at a level of a relevant environmental feature. Only the factors with the top two importance (quantified using the Gini value) were shown. Letters on each box represented the results of Tukey’s test, with different letters representative of statistically significant differences between two levels of an environmental feature (p<0.05).

[0070] Consistent with the prediction using training and validation datasets, our predictions showed that EFCc was the richest (1.27% - 2.93%), followed by EFCN (1.24%-2.06%) and then EFCP (0.51%-0.68%) across 1278 prediction sites (Figure 5a, d, g). Compared with other soils, soils with statistically more than 2.0% EFCC abundances usually had C / N ratios above 30, soil Cabove 1%, and / or were covered by crops and wetlands (Figure 5a-c). In contrast, soils with these properties usually showed a lower than 1.6% abundance of EFCN (Figure 5d-f). EFCp was relatively rich (> 0.58%) in the soils with silt fraction lower than 38%, C / N ratios below 10, and / or SOC below 0.2% (Figure 5g-i). These results indicated that the soils with richer carbon and relatively lower soil organic N had relatively higher EFCc abundances but lower EFCN and EFCP abundances.

[0071] Specifically, abundances of EFCcarbohydrate and EFCugnin showed agreeable spatial patterns but distinct magnitudes. EFCCarbohydrate abundances (1.0% to 3.0%) were around 10 times higher than EFCiagnin abundances (0.13% to 0.28%) across 1278 prediction sites (Figure 12a-b of Appendix A). EFCcarbohydrate abundances were usually more than 2% in the soils with more than 1% of soil C content (Figure 13b of Appendix A) and covered by crops and wetlands (Figure 13c of Appendix A). EFCiagnin abundances were usually more than 0.2% in the soil with annual precipitation above 1500 mm (Figure 13d of Appendix A) and / or soil C / N ratios above 30 (Figure 13e of Appendix A). These results suggested that C-abundant crops and wetland soils favored high abundances of EFCcarbohydrate, whereas high C / N ratios and moist soils were usually rich in EFCiagnin.

[0072] Figure 5 illustrates predicted information representing the amounts of various EFC ensembles at multiple levels of environmental conditions. Specifically, Figure 5 shows predicted abundances of three EFC ensembles at different levels of environmental conditions. Subplots on the top, middle and bottom rows referred to the predicted abundances of EFCc, EFCN, and EFCp at different levels of the top three relevant environmental predictors, respectively. At each row, the order of the subplot from left to right represented the 1st, 2nd, and 3rd relevant environmental factors for predicting the corresponding EFC abundance. Here EFCp variations with different levels of soil sampling depth were not plotted, as the depth of soil sampling was uniform in the prediction dataset. Here letter(s) on each box represent(s) the result(s) of multiple comparisons using Tukey’s test (p<0.050).

[0073] The abundances of EFCorganicN, EFCxionomerN, and EFCinorganicN varied from 0.71%- 1.3%, 0.03%-0.09%, and 0.45%-0.81% across 1278 prediction sites, respectively (Figure 12c-e of Appendix A). Sites had more than 0.9% of EFCorganicN abundances with lower than 15 soil C / N ratios, silt fraction below 25%, and sand fraction above 36% (Figure 14a-c of Appendix A). Most soil sites had EFCxionomcrN abundances above 0.05% and were mainly distributed in thesoils with elevations below 1000m (Figure 14d of Appendix A) and annual mean temperature above 10 °C (Figure 14e of Appendix A). More soil sites showed less than 0.69% EFCinorganicN abundances and were mainly distributed in the sites covered by herbaceous, crops, and wetlands soils (Figure 14f of Appendix A). The results implied that relative abundances of different forms of organic N mineralization depended on soil physicochemical properties, whereas relative abundances of EFCinorganicN depended on the land cover type.

[0074] Three EFCs within the EFCp ensemble had different regional patterns across 1278 prediction sites in the US. There was a trend that the sites with higher abundances of EFCinorganicP (>0.11%) usually showed lower abundances of EFCorganicP (<0.32%) and EFCMonomerP (<0.16%) (Figure 12f-h of Appendix A). This spatial pattern was correlated to spatial distribution in soil texture, soil C / N ratios, and annual accumulated precipitation. The soils with less than 0.32% EFCorganicP were mainly located in the soil with 26% < sandy fraction < 44% (Figure 15a of Appendix A). Less than 0.16% EFCMonomerP abundances were mainly distributed in the soils with silt fraction above 38% and / or annual precipitation above 1500 mm (Figure 15b-c of Appendix A). In contrast, EFCinorganicP abundance was usually higher than 0.11% in soils with C / N ratios lower than 40 or sand fraction below 26% (Figure 15d-e of Appendix A). These results suggested that EFCs for microbial inorganic P immobilization were relatively richer in fine texture and N-abundant soils, whereas EFCs for organic P decomposition and mineralization were relatively abundant in lower precipitation and loam texture soil environments.4. Discussion4.1 The geographical distribution of soil gene-encoded potential enzyme functional composition and its implication for predicting soil C, N, and P cycling

[0075] We identified genes encoding for enzyme production that are involved in the decomposition and mineralization of SOM and inorganic N and P transformations across diverse environments in the U.S. This is a regional -scale study that utilized shotgun metagenomics data to identify general soil enzyme functional composition involved in soil C, N, and P cycling. Although many previous studies have investigated the geographical composition and diversity of soil microbial communities at the taxonomic level using bacterial 16S rRNA or fungal ITS genes data, few efforts have been made to investigate the spatial distribution and diversity of microbial functions. Our study leveraged previous efforts that utilized metagenomic datasets to identify the global biogeography of microbial functional traits involved in inorganic N transformation. Herewe applied the novel EFC concept to identify the geographical distribution of microbial functional traits related to SOM decomposition, which involved decomposing many similar soil substrates by many gene-encoded enzymes from diverse microbial species. Our other site-level study has indicated the effectiveness of this EFC methodology for simplifying the functional complexity and redundancy of omics data. In this study, we generated a dataset that applied this EFC method to quantify the genetic control of the composition of general microbial functional traits for SOM decomposition across diverse environments.4.2 Chemical, physical, and biological control on gene-encoded enzyme functional composition of soil microbial communities

[0076] Our results integrated regional-scale omics and environmental data to demonstrate that the local edaphic, climate, and vegetation conditions are dominant predictors of the relative abundances of EFCs and EFC ensembles involved in soil C, N, and P cycles. Specifically, these environmental factors dominated the abundances of different potential soil enzyme functions across U.S.

[0077] Soil water regulated the abundance of extracellular degradative enzymes. Our results showed that the total abundances of EFCs involved in soil C, N, and P cycling were lower in the regions with annual precipitation below 1000 mm, suggesting the decreased relative abundances of gene-coded enzymes for C and nutrient acquisition from complex soil substrates under waterlimited environments. This finding supported the resource optimization strategy under waterlimited conditions, e.g., soil microbial communities might increase resource utilization to mitigate water stress while decreasing investment in microbial growth and production of extracellular enzymes. This water stress response may result from the lower return of the breakdown products due to diffusional constraints. To mitigate the effects of water stress, microbial communities may instead increase enzyme production for synthesizing cell osmolytes or microbial dormancy. Although the specific response mechanisms of soil microbial communities to water availability may vary with different frequencies and intensities of precipitation, our study systematically demonstrated the reduction in extracellular degradative enzymes in response to lower precipitation at a regional scale.

[0078] Soil C and C / N ratios indicated the relative abundances of different soil enzymes involved in C cycling. Soils with above 1.0% soil C and covered by crops and wetlands were relatively rich in EFC carbohydrate (>2.0%) (Figure 13b). This may be because litterfall from fast-growing croplands added easily decomposable carbohydrates into soil compared with forests and thus promoted the microbial synthesis of EFCcarbohydrate for acquiring energy from carbohydrates. Rich EFCcarbohydrate in wetlands was possibly due to fluctuating water tables, which created alternating aerobic and anaerobic conditions and stimulated the organic matter decomposition. EFCrignin was rich in soils with C / N ratios above 30 (>0.2%) (Figure 13e of Appendix A). These soils were mainly covered by forests and shrubs (Figure 16), where the input of lignin-abundant litterfall from woody vegetation stimulated microbial communities to synthesize more EFC ignin for decomposing lignin. Our finding provided mechanistic explanations of the observations that show increased cellulase activities with the addition of cellulose to the wetland soils and decreased decomposition rates of litter lignin in N-abundant or N-added soils. These results indicated that enzyme production of microbial communities for energy acquisition could be inducible by the presence of the corresponding complex soil carbon substrates.

[0079] Soil physiochemical properties and land cover types regulated the spatial patterns of different enzymes involved in N cycling. More than 0.9% of EFCorganicN in the soils with the coarse texture (Figure 14a-b of Appendix A) and below 15 of C / N ratios (Figure 14c of Appendix A) may result from more available organic N substrates associated with organic N desorption from minerals in the coarse soils. The rich EFCMonomerN in warm low-elevation soils (Figure 14d of Appendix A) was because that warmer soil temperatures (20 - 45 °C) favored microbial N mineralization by higher abundance of EFCMonomerN. Relatively higher EFCinorganicN in urban soils was probably associated with the pollution-related N input, as inorganic N transformation is positively linked to the supply of substrates from organic N mineralization.

[0080] Similarly, soil physiochemical properties also regulated enzyme synthesis involved in P cycling. Lower than 0.11% EFCinorganicP abundances in loam soil and C / N ratios above 40 (Figure 15d-e of Appendix A) may be because the N limitation for microbial communities in high C / N ratio soils decreased microbial inorganic P assimilation to keep the homeostatic C / N / P ratios of microbial communities. Moreover, EFCMonomerP abundance was also low with high precipitation, possibly because of P leaching and subsequent losses of P induced by the larger subsurface runoff. These findings indicated that soil hydrothermal conditions could regulate soil enzyme composition by affecting enzyme-catalyzed decomposition or mineralization rates or substrate availability, therefore, adjusting the needs of microbial synthesis of different enzymes.4.3 Environment-informed ML pipeline for predicting gene-encoded enzyme functional classification involved in soil C, N, and P cycling

[0081] Our study provided an exemplary integrated ML pipeline that enabled 1) predictions of information representing the amount of an EFC (e.g., the relative abundances of gene-encoded EFCs) involved in soil C, N, and / or P cycling; 2) interpretation of the variation of soil EFC with spatial variability in climate, vegetation, and edaphic conditions using the resource allocation strategy of microbial communities; and 3) developing regional -scale information (e.g., an EFC abundance dataset) for parameterizing initial enzyme functional compositions in soil biogeochemical models. ML technology, such as the Hidden Markov model, has been used to identify gene function in the bioinformatics. Recently, there have been increasing applications of advanced ML technology in interpreting environmental regulations on the diversity of soil microbial communities and predicting taxonomic structures and specific functions. Our study trained and applied the ML model to predict the environmental regulation on omics-informed EFCs involved in soil C, N, and P cycling. We have paired regional-scale omics data with the corresponding environmental factors to train an ML pipeline for predicting the microbial functional composition involved in the whole process of the soil C, N, and / or P cycling.

[0082] The developed exemplary ML pipeline, including data pre-processing, data augmentation, and ensemble-based ML model, highlighted three factors. First, the scheme to augment gene functional information from a limited omics dataset under the same environmental condition increased the ML prediction power when trained omics data was limited. Our results indicated that the augmentation of the paired omics and environmental factors dataset from 277 to 3047 increased the ML metric R2 by about 20%. Second, the stacking ensemble of predictions by four ML algorithms allowed us to harness the power of four ML algorithms for predicting EFC compositions and therefore increased prediction accuracy and stability. This approach may be beneficial when predicting an EFC abundance using limited environmental factor data. For example, predictions of P-related EFCs by stacking were significantly improved compared with an individual ML algorithm. Finally, the developed exemplary ML pipeline here can be easily applied to predict the abundances of EFCs involved in other microbial functions (e g., water stress adaptation) by pairing the relative abundances of targeted gene functions with desired environmental factors.

[0083] Our results may indicate that the eight example EFCs involved in SOM decomposition can be identified across diverse environments and their relative abundances under diverse environmental conditions are predictable using omics and environmental data-trained ML pipeline.

[0084] First, metagenomics can inform potential microbial functions under a specific environment. Metagenomics, metatranscriptomics, and metaproteomics data may be integrated to analyze expressed enzyme activities and outcomes. Our method of applying the EFC concept to upscale gene functions to the ecosystem level can also be used for metatranscriptomics and metaproteomics. The developed ML pipeline in this study can also be applied for incorporating metatranscriptomics and metaproteomic data with the corresponding environmental data to train and predict the actual enzyme functional feedback to environmental change. We may apply multi-omics data, including metagenomics, metaproteomics, and metatranscriptomics, to quantify each omics data-informed EFCs involved in soil C, N, and P cycling.

[0085] Second, our example ML pipeline was trained using 277 paired metagenomic data and 17 environmental factor data from 20 sites across the U.S. The application of augmentation technology intensified the ML pipeline’s prediction power with limited training data. Model prediction was successfully applied to elucidate the biological, physical, and chemical regulation of soil enzyme functional composition across diverse environments. The example ML pipeline may be further trained / improved using diverse environment-microbial community feedback in regions beyond CONUS and / or more paired omics and environmental data.

[0086] Finally, the example ML pipeline utilized four tree-based ML algorithms to predict EFC compositions across diverse environments. A deeper layer to stack the predictions of four example ensembles using gradient boosting and / or another model structure (e.g., a more complex model structure like neutral networks) may increase the prediction power (Table S3 of Appendix A) and / or better capture the variability of EFC relative abundance across diverse environments.5. Further Remarks

[0087] Our study identified eight example soil EFCs for soil C, N, and P cycling across diverse environments by using published metagenomics data across the U.S. By labeling the omics- informed EFCs abundances with the corresponding edaphic, climate, and vegetation information, we developed and trained an example ML pipeline to predict the composition of gene-encoded potential soil enzyme functions. Our results showed that soil water regulated the abundance ofextracellular degradative enzymes, whereas SOC and soil C / N ratios can be indicators of the relative abundances of enzymes for carbohydrate and lignin decomposition. Soil physiochemical properties and land cover types were the main regulators for the spatial patterns of different enzymes involved in nutrient cycling. Our example ML pipeline and EFC abundances datasets may be used as a tool and / or benchmark data for parameterizing feedback between microbial communities and environmental conditions and the implications for coupling carbon and nutrient cycling.

[0088] Figure 6 shows a flowchart of a process for determining information representing an amount of an indicator of a soil trait according to some embodiments of the present document. A system including at least one processor and at least storage device may be configured to implement the process 600. The system may obtain values of a plurality of environmental factors relating to a location at 610; and determine information representing an amount of the indicator by inputting the values of the plurality of environmental factors into a machine learning engine at 620. The machine learning engine may be trained using sample datasets. In some embodiments, each of at least some of the sample datasets may include values of at least a portion of the plurality of environmental factors of a sample site and information representing an amount of the indicator of the sample site. The plurality of environmental factors may include at least one factor relating to climate, vegetation, or an edaphic condition of the location. Examples of environmental factors may be found elsewhere in the present document. See, e.g., Table SI of Appendix A.

[0089] In some embodiments, the system may select the plurality of environmental factors from a large group of candidate environmental factors to reduce the complexity of the training and application of the machine learning engine, identify principal environmental factors that may regulate the relative abundance of the indicator (e.g., EFC, EFC ensemble) across diverse environments and / or over time. The selection process may include determining a plurality of degrees of relevance each of which assesses relevance between the soil trait and one of a plurality of candidate environmental factors; ranking the degrees of relevance; and selecting the environmental factors based on the ranking. See also, e.g., section 3.3 of the present document.

[0090] The indicator may be an omics-level indicator. For example, the indicator may include an enzyme group including one or more enzymes relating to the soil trait. An enzyme group may include one or more enzymes that relate to the soil trait. An enzyme group may correspond anenzyme function class (EFC), or an EFC ensemble, as described elsewhere in the present document. As used herein, the amount of the indicator may also be referred to as abundance of the indicator (e.g., EFC or EFC ensemble). As used herein, an amount may be an absolute value or a normalized value. The information representing the amount of an indicator may include, relate to, or be in the form of an abundance or a relevant abundance of the indicator. The soil trait may include at least one of soil organic matter (SOM) decomposition, inorganic nitrogen transformation, or phosphorus transformation. The soil trait may include at least one of soil carbon cycling, soil nitrogen cycling, or soil phosphorus cycling. The soil trait may be used to estimate one or more environmental or soil related features including, e.g., carbon release from soil, soil heath, etc. In some embodiments, the system may perform process 600 by inputting values of the plurality of environmental factors from various sites to generate a map including information representing the amount of the indicator and / or a map of the soil trait. In some embodiments, the system may perform process 600 by inputting values of the plurality of environmental factors for a location or region over time to generate a temporal distribution (e.g., over different weeks, months, seasons) of information representing the amount of the indicator and / or of the soil trait. See also, e.g., sections 2.4, 3.4, and 4 of the present document.

[0091] The machine learning engine may include a plurality of ensemble models and a stacking model. The machine learning engine may be trained to determine information representing an amount of an indicator of a soil trait based on values of a plurality of environmental factors. More description regarding the machine learning engine may be found elsewhere in the present document. See, e.g., Figures 7 and 8, and the description thereof.

[0092] Figure 7 shows a flowchart of a process for training a machine learning engine configured to determine information representing an amount of an indicator of a soil trait according to some embodiments of the present document. A system including at least one processor and at least storage device may be configured to implement the process 700. In some embodiments, the machine learning engine may include a plurality of ensemble models and a stacking model. The plurality of ensemble models may include, e.g., at least one of a random forest model, a bagging model, an adaptive boosting model, or a gradient boosting model.

[0093] The system may obtain sample datasets at 710, divide the sample datasets into training datasets and validation datasets at 720, generate the plurality of ensemble models by training using the training datasets at 730, and determine the stacking model using the validation datasetsand the trained ensemble models at 740. Each of at least some of the sample datasets may include values of the plurality of environmental factors of a sample site and information representing an amount of the indicator of the sample site. See also, e.g., section 2.2 of the present document. In some embodiments, the sample datasets that the system may obtain at 710 may be pre-grouped or identified as the training datasets and the validation datasets, and accordingly the system may omit 720.

[0094] In some embodiments, a sample dataset may be generated by a process including: extracting, from soil metagenomics information of a soil sample from a sample site, a gene group including one or more genes that correlate with information representing an amount of the indicator of the soil sample; estimating the information representing the amount of the indicator based on a gene count of each of the one or more genes of the gene group in the soil sample; and generating the sample dataset by pairing values of the plurality of environmental factors of the sample site with the estimated information representing the amount of the indicator. Merely by way of example, the system may extract the gene group using a Joint Genome Institute (JGI) function profile identification tool or one or more other alternative gene function identification tools. See also, e.g., sections 2.1.1 and 2.1.3 of the present document. In some embodiments, the sample datasets used by the system to train one or more of the ensemble models or the stacking model may be generated and / or preprocessed by the system, obtained by the system from another source (e.g., a third-party system), or a combination thereof.

[0095] In some embodiments, the system may estimate the information representing the amount of the indicator of a sample dataset by a process including determining a raw sum by summing the gene counts of the group of genes in a soil sample corresponding to the sample dataset; and determining the information by normalizing the raw sum based on a total gene count of genes in the soil sample. See also, e.g., section 2.1.2 of the present document.

[0096] In some embodiments, the system may generate a synthesized dataset based on a data augmentation algorithm. Examples of applicable data augmentation algorithms include data warping, oversampling, or random erasing or insertion. See also, e.g., section 2.2.2 of the present document.

[0097] In some embodiments, the system may generate a sample dataset by obtaining values of the plurality of environmental factors based on location information of the sample site; and pairing the values of the environmental factors with information representing the amount of theindicator of a sample soil corresponding to the sample dataset. Merely by way of example, the location information of the sample site may include longitude and / or latitude of the sample site; accordingly, the system may obtain the values of the environmental factors based on the location information of the sample site by retrieving the values from a global or regional database based on the longitude and / or latitude of the sample site. See also, e.g., section 2.1.3 of the present document.

[0098] The stacking model may be trained by a process including: obtaining stacking model training datasets; and determining the stacking model by training based on an algorithm (e.g., the gradient boosting algorithm) and the stacking model training datasets. In some embodiments, the stacking model training dataset may be generated by a process including: for each of the validation datasets, determining predicted information representing an amount of the indicator by inputting values of the plurality of environmental factors of a soil sample corresponding to the validation dataset into each of the plurality of ensemble models; and generating a stacking model training dataset by stacking the predicted information representing the amounts of the indicator with corresponding information (representing the amount of the indicator) retrieved from the validation dataset. In some embodiments, stacking model training datasets used by the system to train the stacking model may be generated by the system, obtained by the system from another source (e.g., a third-party system), or a combination thereof. The stacking model may be configured to determine a stacked prediction result from predictions of the trained ensemble models that are obtained based on a same input into the trained ensemble models.

[0099] In some embodiments, the system may further divide the sample datasets into testing datasets, in addition to the training datasets and validation datasets. Merely by way of example, the system may divide the sample datasets is performed based on a Monte Carlo cross-validation algorithm. The system may test the machine learning engine after training using the testing datasets at 750. See also, e.g., section 2.2.3 of the present document.

[0100] Various operations disclosed herein can be implemented using a processor / controller is configured to include, or be coupled to, a memory that stores processor executable code that causes the processor / controller carry out various computations and processing of information. The processor / controller can further generate and transmit / receive suitable information to / from the various system components, as well as suitable input / output (IO) capabilities (e.g., wired or wireless) to transmit and receive commands and / or data. The processor / controller may, forexample, provide signals to control the operation of various components such as lasers and detectors that are disclosed herein. The processor / controller may be further configured to perform various method steps and computations that are disclosed in this patent document.

[0101] Various information and data processing operations described herein may be implemented in one embodiment by a computer program product, embodied in a computer- readable medium, including computer-executable instructions, such as program code, executed by computers in networked environments. A computer-readable medium may include removable and non-removable storage devices including, but not limited to, Read Only Memory (ROM), Random Access Memory (RAM), compact discs (CDs), digital versatile discs (DVD), etc. Therefore, the computer-readable media that is described in the present application comprises non-transitory storage media. Generally, program modules may include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Computer-executable instructions, associated data structures, and program modules represent examples of program code for executing steps of the methods disclosed herein. The particular sequence of such executable instructions or associated data structures represents examples of corresponding acts for implementing the functions described in such steps or processes.

[0102] Only a few implementations and examples are described and other implementations, enhancements and variations can be made based on what is described and illustrated in this patent document.Appendix A

[0103] This appendix provides further explanations and examples, and includes (1) a detailed description of five machine learning (ML) algorithms applied in the developed ML pipeline (see Text SI); (2) Figure 8 that shows the flowchart of the ML pipeline for training, validating, and predicting enzyme functional classes (EFCs); (3) Figure 9 for indicating the results of sensitivity analysis on model performance metrics with the different split ratios of training and validating data; (4) Figure 10 for showing the importance (quantified using the Gini value) of environmental factors for predicting the abundances of eight EFC ensembles; (5) Figure 11 for showing the results of principle component analysis (PCA) of important environmental factors; (6) Figure 12 that indicates ML model-predicted relative abundances of eight EFCs at 1278 USsites; (7) Figures 13-16 illustrating predicted abundances of eight EFCs at different levels of important environmental conditions; (8) Table SI for describing the environmental data used in training, validating, and testing the ML pipeline; (9) Table S2 for listing key hyperparameters of the ML models employed in the pipeline; (10) Table S3 for listing the mean and standard deviation of R2 values of five ML models used for predicting the abundances of eight EFCs and three EFC ensembles within the training and testing datasets; and (11) A separate data set SI for listing the composition of gene-encoded enzymes within each EFC involved in soil Carbon, Nitrogen, and Phosphorus cycling.Text SI: Machine learning (ML) algorisms applied in the developed example ML pipeline.

[0104] Bagging and Random Forest: For illustration purposes, we used the BaggingRegressor and RandomF orest Regressor algorithm implemented by the Scikit-learn Python package to train ensembles of multiple Decision Trees (M = 200 and 100 for Bagging and Random Forest, respectively) based on the part of processed environmental feature data and corresponding EFC or EFC ensemble abundances. Each tree in the Bagging model was trained on 30% of the training instances randomly sampled from the training subset. The prediction outcomes of all Decision Trees were then averaged to generate the final prediction. Random Forest was performed using the same process as the Bagging algorithm. The only difference was that the node of each tree in the Random Forest regressor was split by searching for the best environmental feature among a random subset of a total of 17 environmental features.

[0105] Adaptive Boosting: For illustration purposes, we used the AdaBoostRegressor from Scikit-learn, which implemented the AdaBoost.R2 algorithm, to predict the abundances of eight EFCs and three EFC ensembles based on 17 processed environmental features in our datasets. Here the AdaBoost.R2 algorithm applied 50 (M=50) decision trees sequentially, with each trying to correct its predecessors by adjusting the sampling probability of each soil sample. Each decision tree (m) was fitted to the whole training data. Each sample (i) within the training data was randomly selected initially with the same probability (Pii=l / n), where n is the training data size. Then the training samples with larger prediction loss ( ljm) were more probably to be selected for training the next decision tree (Eq. SI).

[0106] Here wimis the weight of the training sample i used in the Decision Tree m. is the learning rate parameter and was calibrated as 0.5. ljmis the loss of decision tree, m, for predicting training sample i, and quantified using an exponential loss function (Eq. S2).yimand yoiare the decision tree (m)-predicted and omics-informed abundance of an EFC, respectively. The higher the limvalue, the larger the prediction loss of the training sample i is by the decision tree m. The overall prediction confidence of the decision tree m ( / 3m) was then measured as a function of the weighted average loss of all training samples (Lm) (Eq. S3).Eq. S3

[0107] The larger Lmleads to the higher f3m. Therefore, when a weaker Decision Tree with lower prediction confidence (3mestimated a higher prediction loss ljmfor a training sample i, the corresponding sample i was given more sampling probability for fitting the next decision tree according to Eq. SI. This procedure was repeated until reaching the desired 50 decision trees or finding the perfect prediction when Lm—< 0 . The final prediction was estimated as the weighted median of applied decision trees, where the weighted index of each decision tree prediction (Wm) was estimated as a function of / 3m(Eq. S4).

[0108] Gradient Boosting: For illustration purposes, we used Scikit-learn’s GradientBoostingRegressor to train 70 (M=70) decision trees sequentially and build the ensemble with a greedy structure, scaled by a learning rate T| . The first tree was fitted to the environmental features-labeled abundance of an EFC from training data. Then each tree wasfitted to the prediction residuals of the previous ensemble of trees. Each of these fittings aimed to minimize Lm, the sum of the loss function of samples (Eq. S5).L,. =Y'FFF)~yF Eq 85

[0109] Here Lmis a loss function of squared error. Fm(xj) is the ensemble of predictions of m number of trees. hm(x;) referred to the prediction of decision tree m that fitted to the residuals of the previous ensemble. The final prediction of the abundance of an EFC of sample i was the sum of the predictions of all trees in the ensemble.

[0110] Stacking: For illustration purposes, after training the four ensembles (bagging, random forest, adaptive boosting, and gradient boosting), we used the stacked generalization ensemble to integrate all predictions by four ensemble models. We assigned a new gradient boosting of 50 decision trees with a max depth of three to be the final predictor in this stacking pipeline. First, the four trained ensemble models were applied to predict the abundances of eight EFCs or three EFC ensembles using the environmental features-labeled validation dataset. The resultant predictions by each model were then labeled with the corresponding EFC or EFC ensemble abundance from the validation dataset and applied to train the final predictor to aggregate the four EFC predictions into the final one. Finally, we incorporated the testing dataset into the trained ML pipeline to evaluate the model performance.

[0111] Figure 8 shows a flowchart of the ML pipeline for training, validating, and predicting enzyme functional classes (EFCs).

[0112] Figure 9 shows results of a sensitivity analysis on model performance metrics with the different split ratios of training and validating data. The label "b" denotes data corresponding to the bagging algorithm; the label "gb" denotes data corresponding to the gradient boosting algorithm; the label "ab" denotes data corresponding to the adaptive boosting algorithm; the label "rf ' denotes data corresponding to the random forest algorithm. The notation "b (under rf)" indicates that a portion of the curve corresponding to the bagging algorithm is beneath a portion of the curve corresponding to the random forest algorithm over a specific test ratio range, as illustrated.

[0113] Figure 10 illustrates relevance (quantified using Gini value) of environmental factors for predicting information representing the amounts of various EFC ensembles in accordance with example embodiments of the present document. Specifically, Figure 10 shows relevance (quantified using Gini value) of environmental factors for predicting the abundances of eight EFC ensembles: (a). EFC for lignin decomposition (EFCugnin), (b). EFC for decomposition of N- rich SOM (EFCorganicN), (c). EFC for monomer N mineralization (EFCMonomerN), (d). EFC for inorganic N transformation (EFCinorganicN), (e). EFC for P-rich SOM, (EFCorganicp), (f). EFC for monomer P (EFCwonomerp), (g). EFC for P immobilization (EFCinorganicp). The relevance was averaged over 1000 iterations of predictions. Here only the top ten relevant environmental factors for predicting each EFC were shown.

[0114] Figure 11 illustrates a principal component analysis (PCA) of relevant environmental factors in accordance with example embodiments of the present document. Specifically, Figure 11 shows the PCA of relevant environmental factors.

[0115] Figure 12 illustrates predicted information representing the amounts of various EFCs using environment-informed machine learning models in accordance with example embodiments of the present document. Specifically, Figure 12 shows predicted relative abundances of eight EFCs at 1278 sites using environment-informed ML pipeline, (a)-(h) represented predicted relative abundances of EFCc arbohydrate, EFCrignin, EFCorganicN, EFCMonomerN, EFCinorganicN, EFCorganicp, EFCMonomerP, and EFCinorganicp at 1278 sites across the U.S.

[0116] Figure 13 illustrates predicted information representing the amounts of various EFCs involved in lignocellulose decomposition at multiple levels of environmental conditions in accordance with example embodiments of the present document. Specifically, Figure 13 shows predicted abundances of two EFCs involved in lignocellulose decomposition at different levels of environmental conditions. Subplots referred to the predicted abundances of EFCcarbohydrate and EFCrignin at different levels of the top three and two relevant environmental predictors. At each row, the order of the subplot from left to right represents the 1st, 2nd, and 3rd relevant environmental factors for predicting the corresponding EFC abundance. The letter(s) on each box represent(s) the result(s) of the multiple comparison tests using Tukey’s test (p<0.050)

[0117] Figure 14 illustrates predicted information representing the amounts of various EFCs for N-rich soil components decomposition, mineralization, and transformation at multiple levels of environmental conditions in accordance with example embodiments of the present document.Specifically, Figure 14 shows predicted abundances of EFCs for N-rich soil components decomposition, mineralization, and transformation at different levels of environmental conditions. Subplots referred to the predicted abundances of EFCorganicN, EFCMonomerN, and EFCinorganicN at different levels of the top three, two, and one relevant environmental predictors, respectively. Among the top three relevant environmental factors, only the environmental factors with a statistically significant level difference in the corresponding EFC abundances were plotted. Here EFCMonomerN variations with different levels of soil sampling depth were not considered, as the depth of soil sampling was uniform in the prediction dataset. Here the letter(s) on each box represent(s) the result(s) of the multiple comparison tests using Tukey’s test (p<0.050).

[0118] Figure 15 illustrates predicted information representing the amounts of various EFCs for P-rich soil components decomposition, mineralization, and transformation at multiple levels of environmental conditions in accordance with example embodiments of the present document. Specifically, Figure 15 shows predicted abundances of EFCs for P-rich soil components decomposition, mineralization, and transformation at different levels of environmental conditions. Subplots referred to the predicted abundances of EFCorganicP, EFCMonomerP, and EFCinorganicP at different levels of the top one, two, and two relevant environmental predictors, respectively. Among the top three relevant environmental factors, only the environmental factors with a statistically significant level difference in the corresponding EFC abundances were plotted. Here the letter(s) on each box represent(s) the result(s) of the multiple comparison tests using Tukey’s test (p<0.050)

[0119] Figure 16 shows soil C / N ratios of different land covers of 1278 sites across the U.S.Table S1. Description of environmental dataTable S2. Key hyperparameters of ML models used in the ML pipelineTable S3. The coefficient of determination (R2) of five ML models for predicting the abundances of eight EFCs and three EFC ensembles within the training and testing datasets, respectively. Here Each R2value is reported as the averaged R2 value over 1000 iterations of predictions and its standard deviation.Data Set S1. The composition of gene-encoded enzymes within each EFC involved in soil Carbon, Nitrogen, and Phosphorus cycling.SUM decomposition and mineralization

Claims

WHAT IS CLAIMED IS:

1. A method for determining information representing an amount of an indicator of a soil trait, comprising: obtaining values of a plurality of environmental factors relating to a location; and determining the information representing the amount of the indicator by inputting the values of the plurality of environmental factors into a machine learning engine, wherein the machine learning engine is trained using sample datasets each of which includes values of at least a portion of the plurality of environmental factors of a sample site and information representing an amount of the indicator of the sample site.

2. The method of claim 1, wherein the plurality of environmental factors comprise at least one factor relating to climate, vegetation, or an edaphic condition of the location.

3. The method of claim 1, wherein the indicator comprises an enzyme group including one or more enzymes relating to the soil trait.

4. The method of claim 1, wherein the soil trait comprises at least one of soil organic matter (SOM) decomposition, inorganic nitrogen transformation, or phosphorus transformation.

5. The method of claim 1, wherein the soil trait comprises at least one of soil carbon cycling, soil nitrogen cycling, or soil phosphorus cycling.

6. The method of claim 1, wherein each of the sample datasets is generated by a process comprising: extracting, from soil metagenomics information of a soil sample from a sample site, a gene group including one or more genes that correlate with information representing an amount of the indicator of the soil sample; estimating the information representing the amount of the indicator based on a gene count of each of the one or more genes of the gene group in the soil sample; andgenerating the sample dataset by pairing values of the plurality of environmental factors of the sample site with the estimated information representing the amount of the indicator.

7. The method of claim 6, wherein extracting the gene group comprises using a Joint Genome Institute (JGI) function profile identification tool or an alternative gene function identification tool.

8. The method of claim 6, wherein estimating the information representing the amount of the indicator of a sample dataset comprises: determining a raw sum by summing the gene counts of the group of genes in a soil sample corresponding to the sample dataset; and determining the information representing the amount of the indicator by normalizing the raw sum based on a total gene count of genes in the soil sample.

9. The method of claim 1, wherein the sample datasets comprise at least one synthesized dataset generated by a data augmentation algorithm.

10. The method of claim 9, wherein the data augmentation algorithm comprises data warping, oversampling, or random erasing or insertion.

11. The method of claim 1, wherein each of at least one of the sample datasets is obtained by: obtaining values of the plurality of environmental factors based on location information of the sample site; and pairing the values of the environmental factors with the information representing the amount of the indicator of a sample soil corresponding to the sample dataset.

12. The method of claim 11, wherein:The location information of the sample site comprises longitude and / or latitude of the sample site, andobtaining the values of the environmental factors based on the location information of the sample site comprises retrieving the values from a global or regional database based on the longitude and latitude of the sample site.

13. The method of claim 1, wherein the machine learning engine comprises a plurality of ensemble models and a stacking model.

14. The method of claim 13, wherein the machine learning engine is trained by a training process comprising: dividing the sample datasets into training datasets and validation datasets; generating the plurality of ensemble models by training using the training datasets; and determining the stacking model using the validation datasets and the trained ensemble models.

15. The method of claim 14, wherein: the stacking model comprises a gradient boosting algorithm, and determining the stacking model comprises: for each of the validation datasets, determining predicted information representing an amount of the indicator by inputting values of the plurality of environmental factors of a soil sample corresponding to the validation dataset into each of the plurality of ensemble models; and generating a stacking model training dataset by stacking the predicted information with information representing the amount of the indicator retrieved from the validation dataset; and determining the stacking model by training based on the gradient boosting algorithm and the stacking model training datasets.

16. The method of claim 14, wherein: the sample datasets are further divided into testing datasets, andthe training process further comprises testing the stacking model and the ensemble models using the testing datasets.

17. The method of claim 14, wherein dividing the sample datasets is performed based on a Monte Carlo cross-validation algorithm.

18. The method of claim 13, wherein the plurality of ensemble models comprise at least one of a random forest model, a bagging model, an adaptive boosting model, or a gradient boosting model.

19. The method of claim 1, wherein the plurality of environmental factors are selected by a selection process including: determining a plurality of degrees of relevance each of which assesses relevance between the soil trait and one of a plurality of candidate environmental factors; ranking the degrees of relevance; and selecting the environmental factors based on the ranking.

20. A system for determining information representing an amount of an indicator of a soil trait, comprising: at least one processor coupled to at least non-transitory, computer-readable storage medium with instructions stored thereon, the instructions upon execution by the at least one processor, causing the at least one processor to perform operations including: obtaining values of a plurality of environmental factors relating to a location; and determining the information representing the amount of the indicator by inputting the values of the plurality of environmental factors into a machine learning engine, wherein the machine learning engine is trained using sample datasets each of which includes values of the plurality of environmental factors of a sample site and information representing an amount of the indicator of the sample site.

21. A method of training a machine learning engine configured to determine information representing an amount of an indicator of a soil trait, comprising: obtaining sample datasets;dividing the sample datasets into training datasets and validation datasets; generating a plurality of ensemble models by training using the training datasets; and determining a stacking model using the validation datasets and the trained ensemble models, wherein the stacking model is configured to determine a stacked prediction result from predictions of the trained ensemble models that are obtained based on a same input into the trained ensemble models.

Images