One-way analysis of variance (ANOVA) was performed to assess the effects of fertilization treatments on soil properties, the diversity and abundances of soil bacterial and fungal communities, and microbial carbon metabolism and qCO2 using Bonferroni’s post hoc test in SPSS 20.0 software (SPSS, Chicago, IL, USA). The significance testing in one-way ANOVA was based on the data from which the samples originated both followed a normal distribution and had the same variances [78 ]. Principal coordinate analysis (PCoA) was used to evaluate the biochar impacts on the Bray-Curtis distances of the bacterial and fungal community compositions [79 (link)]. We conducted the “capscale” function of the R package vegan (version 3.1.2) to calculate the Bray-Curtis dissimilarities for principal coordinate analysis and “permutest” permutation-based testing for the calculation of the significance values [80 ]. The compositional similarity was calculated as 1 minus the Bray-Curtis dissimilarity.
The non-amended (CK and CF) and biochar-amended (LB, MB, and HB) samples were separately examined for biochar effects on soil bacterial and fungal networks. The OTUs presented either in all non-amended or in all biochar-amended samples were kept for the subsequent network constructions, respectively. The co-occurrence patterns of the bacterial and fungal communities were constructed by calculating multiple correlations and similarities with co-occurrence network (CoNet) inference [81 (link)]. We used an ensemble approach based on the four measurements, including Pearson and Spearman correlations and Bray-Curtis and Kullback-Leibler dissimilarities between pairwise OTUs. A valid co-occurrence was considered a statistically robust correlation between taxa when the correlation threshold was above 0.7 and the P value was below 0.01. The P values were merged using Brown’s method for the four measurements [82 (link)] and then adjusted using the Benjamini-Hochberg procedure to reduce the chances of obtaining false-positive results [83 ]. Network visualization was conducted using Gephi software [84 ]. Nodes indicated individual microbial taxa (OTUs) in the microbiome network [26 (link)]. Network edges represented the pairwise correlations between nodes, suggesting a biologically or biochemically meaningful interactions [12 (link)]. The modules were the clusters of closely interconnected nodes (i.e., groups of co-existing or co-evolving microbes) [26 (link)]. The microbial networks were searched to identify highly associated nodes (clique-like structures) using Molecular Complex Detection (MCODE) introduced for the Cytoscape platform [85 (link)]. The algorithm identifies seeded nodes for expansion by computing a score of local density for each node in the graph. Over 90% accuracy of MCODE predictions yielded, when an overlap score was above 0.2 threshold. The calculated topological characteristics of the bacterial and fungal networks included the following: the numbers of positive and negative correlations, average path length, graph density, network diameter, average clustering coefficient, average connectivity, and modularity. The roles of individual nodes were estimated by two topological parameters: the within-module connectivity Z, which quantified to what extent a node connected to other nodes in its own module, and the among-module connectivity P, which quantified how well the node connected to different modules [86 (link)]. The nodes with either a high value of Z or P were defined as keystone taxa, including module hubs (Z > 0.25, P ≤ 0.62; critical to its own module coherence), connectors (Z ≤ 0.25, P > 0.62; connect modules together and important to network coherence), and network hubs (Z > 0.25, P > 0.62; vital to both the network and its own module coherence) [87 (link)]. For network modules, the module eigengene could summarize the closely connected members within a module [88 (link)]. The singular value decomposition of the module expression matrix was used to represent the module eigengene networks [89 (link)]. The module eigengene of a module was defined as the first principal component of the standardized module expression data [90 (link)]. Then, the relationships between soil properties, microbial diversity, network module eigengenes, and SOC mineralization (microbial carbon metabolism and qCO2) were evaluated using Spearman’s rank correlation test.
Random forest modeling was used to quantitatively assess the important predictors of carbohydrate catabolism and qCO2 involving soil properties and the microbial community. Soil properties included soil pH, SMC, SOC, total nitrogen, total phosphorus, total potassium, available nitrogen, available phosphorus, available potassium, and cation exchange capacity, while the microbial community included the biomass, diversity, composition, and network of soil bacterial and fungal communities. The bacterial and fungal biomass were characterized by bacterial and fungal PLFAs. The bacterial and fungal diversities were represented by the Shannon index based on the rarified same sequencing depth. The compositions of soil bacterial and fungal communities were indicated by the first principal coordinates (PCoA1). The bacterial and fungal networks were represented by the module eigengenes that were significantly related to diversity and carbohydrate metabolism. The importance of each factor was evaluated by the increase in the mean square error between the observed and predicted values when the predictor was randomly permuted [91 (link)]. This accuracy of importance was measured for each tree and was averaged across the forest. Accuracy of importance was estimated for each observation using the left-out individual predictions (called “out-of-bag” estimation) and then averaged over all observations [92 (link)]. These analyses were conducted using the randomForest package [93 ], and the significance of the model and predictor importance was determined using the A3 and rfPermute packages, respectively [94 , 95 ]. Structural equation modeling (SEM) was applied to determine the direct and indirect contributions of soil properties and microbial community to carbohydrate catabolism and qCO2 [96 (link)]. SEM analysis was conducted via the robust maximum likelihood evaluation method using AMOS 20.0 (AMOS IBM, USA). The SEM fitness was examined on the basis of a non-significant chi-square test (P > 0.05), the goodness-of-fit index (GFI), and the root mean square error of approximation (RMSEA) [97 ].
The non-amended (CK and CF) and biochar-amended (LB, MB, and HB) samples were separately examined for biochar effects on soil bacterial and fungal networks. The OTUs presented either in all non-amended or in all biochar-amended samples were kept for the subsequent network constructions, respectively. The co-occurrence patterns of the bacterial and fungal communities were constructed by calculating multiple correlations and similarities with co-occurrence network (CoNet) inference [81 (link)]. We used an ensemble approach based on the four measurements, including Pearson and Spearman correlations and Bray-Curtis and Kullback-Leibler dissimilarities between pairwise OTUs. A valid co-occurrence was considered a statistically robust correlation between taxa when the correlation threshold was above 0.7 and the P value was below 0.01. The P values were merged using Brown’s method for the four measurements [82 (link)] and then adjusted using the Benjamini-Hochberg procedure to reduce the chances of obtaining false-positive results [83 ]. Network visualization was conducted using Gephi software [84 ]. Nodes indicated individual microbial taxa (OTUs) in the microbiome network [26 (link)]. Network edges represented the pairwise correlations between nodes, suggesting a biologically or biochemically meaningful interactions [12 (link)]. The modules were the clusters of closely interconnected nodes (i.e., groups of co-existing or co-evolving microbes) [26 (link)]. The microbial networks were searched to identify highly associated nodes (clique-like structures) using Molecular Complex Detection (MCODE) introduced for the Cytoscape platform [85 (link)]. The algorithm identifies seeded nodes for expansion by computing a score of local density for each node in the graph. Over 90% accuracy of MCODE predictions yielded, when an overlap score was above 0.2 threshold. The calculated topological characteristics of the bacterial and fungal networks included the following: the numbers of positive and negative correlations, average path length, graph density, network diameter, average clustering coefficient, average connectivity, and modularity. The roles of individual nodes were estimated by two topological parameters: the within-module connectivity Z, which quantified to what extent a node connected to other nodes in its own module, and the among-module connectivity P, which quantified how well the node connected to different modules [86 (link)]. The nodes with either a high value of Z or P were defined as keystone taxa, including module hubs (Z > 0.25, P ≤ 0.62; critical to its own module coherence), connectors (Z ≤ 0.25, P > 0.62; connect modules together and important to network coherence), and network hubs (Z > 0.25, P > 0.62; vital to both the network and its own module coherence) [87 (link)]. For network modules, the module eigengene could summarize the closely connected members within a module [88 (link)]. The singular value decomposition of the module expression matrix was used to represent the module eigengene networks [89 (link)]. The module eigengene of a module was defined as the first principal component of the standardized module expression data [90 (link)]. Then, the relationships between soil properties, microbial diversity, network module eigengenes, and SOC mineralization (microbial carbon metabolism and qCO2) were evaluated using Spearman’s rank correlation test.
Random forest modeling was used to quantitatively assess the important predictors of carbohydrate catabolism and qCO2 involving soil properties and the microbial community. Soil properties included soil pH, SMC, SOC, total nitrogen, total phosphorus, total potassium, available nitrogen, available phosphorus, available potassium, and cation exchange capacity, while the microbial community included the biomass, diversity, composition, and network of soil bacterial and fungal communities. The bacterial and fungal biomass were characterized by bacterial and fungal PLFAs. The bacterial and fungal diversities were represented by the Shannon index based on the rarified same sequencing depth. The compositions of soil bacterial and fungal communities were indicated by the first principal coordinates (PCoA1). The bacterial and fungal networks were represented by the module eigengenes that were significantly related to diversity and carbohydrate metabolism. The importance of each factor was evaluated by the increase in the mean square error between the observed and predicted values when the predictor was randomly permuted [91 (link)]. This accuracy of importance was measured for each tree and was averaged across the forest. Accuracy of importance was estimated for each observation using the left-out individual predictions (called “out-of-bag” estimation) and then averaged over all observations [92 (link)]. These analyses were conducted using the randomForest package [93 ], and the significance of the model and predictor importance was determined using the A3 and rfPermute packages, respectively [94 , 95 ]. Structural equation modeling (SEM) was applied to determine the direct and indirect contributions of soil properties and microbial community to carbohydrate catabolism and qCO2 [96 (link)]. SEM analysis was conducted via the robust maximum likelihood evaluation method using AMOS 20.0 (AMOS IBM, USA). The SEM fitness was examined on the basis of a non-significant chi-square test (P > 0.05), the goodness-of-fit index (GFI), and the root mean square error of approximation (RMSEA) [97 ].
Free full text: Click here
