Bioinformatic Approaches Reveal Metagenomic Characterization of Soil Microbial Community

As is well known, soil is a complex ecosystem harboring the most prokaryotic biodiversity on the Earth. In recent years, the advent of high-throughput sequencing techniques has greatly facilitated the progress of soil ecological studies. However, how to effectively understand the underlying biological features of large-scale sequencing data is a new challenge. In the present study, we used 33 publicly available metagenomes from diverse soil sites (i.e. grassland, forest soil, desert, Arctic soil, and mangrove sediment) and integrated some state-of-the-art computational tools to explore the phylogenetic and functional characterizations of the microbial communities in soil. Microbial composition and metabolic potential in soils were comprehensively illustrated at the metagenomic level. A spectrum of metagenomic biomarkers containing 46 taxa and 33 metabolic modules were detected to be significantly differential that could be used as indicators to distinguish at least one of five soil communities. The co-occurrence associations between complex microbial compositions and functions were inferred by network-based approaches. Our results together with the established bioinformatic pipelines should provide a foundation for future research into the relation between soil biodiversity and ecosystem function.


Introduction
Soil is considered to be the most diverse natural environment on the Earth [1,2]. The soil microbial communities harbor thousands of different prokaryotic organisms that contain a substantial number of genetic information, ranging from 2,000 to 18,000 different genomes estimated in one gram of soil [3]. One of the most important issues in the field of soil ecology is to uncover the complex relationships between microbial compositions and functional diversity in soil.
Based on traditional approaches for cultivating and isolating soil microorganisms, early studies have focused on culturable bacteria which only account for less than 1% of soil microbial populations [4]. These studies have already discovered many novel genes encoding interesting enzymes and antimicrobials in soils via functional screens and clone-based Sanger sequencing [1,5,6]. Due to the recent advent of High-Throughput Sequencing (HTS) technologies, metagenomic sequencing approaches have been applied to investigate characterizations of diverse soil microbial communities, including target sequencing of the phylogenetic marker gene encoding 16S rRNA [7,8] and whole-metagenome shotgun sequencing [9][10][11][12][13]. However, the majority of 16S rRNA gene-based studies are committed to the interpretation of community composition but poorly focus on the functional and metabolic properties in a microbial community [14]. In addition, integrated bioinformatic analyses for microbial community-level taxonomic affiliation, metabolic reconstruction, and interaction network, seems to be less studied for the highly diverse soil ecosystems. Currently, MG-RAST [15], IMG/M [16], and CAMERA [17] are the major databases that can support deposition and analysis of metagenomic datasets. Uploading large sequencing data and the subsequent analysis jobs on these web servers sometimes take long waiting time and even weeks. The computational pipelines implemented by these prominent platforms are capable of processing many analysis tasks, but some approaches for special biological inference and graphical visualization still need to be complemented [18].
Recently, together with the rapid development of the Human Microbiome Project, numerous computational tools and methodologies have been developed for effective interpretation and visualization of taxonomic and metabolic profiling of complex microbial communities [19,20] and could be applied to the analysis of the soil microbiota. Particularly, some outstanding computational techniques that could better explain the complexity and heterogeneity of microbial communities are still less applied in the study of the soil microbiota, e.g. prediction of metagenomic biomarkers and network-based correlation analyses [21,22]. In this study, we aim to explore the characterizations of the soil microbiota through integrating the current state-of-the-art bioinformatics tools. A collection dataset of 33 publicly available soil metagenomes was investigated in a custom metagenomic data mining pipeline for explaining and visualizing microbial compositions and metabolic potential. A full spectrum of metagenomic biomarkers and a network of taxon co-occurrence patterns were inferred to hopefully provide some new insights into the underlying mechanisms of complex ecological relationships in the soil microbial community.

Ethics statement
No specific permissions were required for the described field studies. The study locations are not privately owned and the field studies did not involve endangered or protected species.

Collection and quality control of metagenomic datasets
Thirty-three metagenomes sampled from five natural soil environments were publicly available and collected in the present study: 14 from grassland, seven from forest soil, nine from desert, two from Arctic soil, and one from mangrove sediment. The metagenomic datasets used can be downloadable according to the list of sequence accession numbers or web links shown in Table S1. All datasets have been produced by whole-metagenome shotgun sequencing using the Roche 454 or Illumina platforms. More reference information about these chosen metagenomes was listed in Table 1. For the datasets of FASTQ formatted sequence reads without quality control, we performed a quality check of bases by using the package Biopieces (http://www.biopieces.org). Low quality ends per read were trimmed by trim_seq. Trimming progressed until all bases in a 3-bp stretch with minimum quality score of 20. High quality reads were retained if satisfying the following criteria: minimum average quality score of 15 in a sliding window of 20 bp; minimum read length of 50 bp.

Estimation of microbial composition
MetaPhlAn v1.7 [24] and BLAST v2.2.22 [25] were employed for profiling the taxonomic clades in the metagenomic datasets. Briefly, metagenomic reads were firstly mapped to the MetaPhlAn reference database composed of unique clade-specific marker genes using BLASTN. The non-default parameters used for BLASTN sequencing similarity searching were as follows: E-value cutoff of 1e-10, word size of 12, and minimum alignment length of 75 nt. Relative abundance scores at all taxonomic levels from the domain level to the species level were then estimated by MetaPhlAn. In the text, mean values of abundances were shown for the mentioned taxon. To assess the compositional similarity among soil samples from different microbial communities, the Bray-Curtis measure of beta diversity [26] was employed to compare all pairwise taxonomic abundances between each sample-pair using a R function vegdist in the package vegan [27]. The permutation-based multivariate analysis of variance (PER-MANOVA) and 2D stress value were then estimated. Based on the resultant Bray-Curtis similarity distance matrix, non-metric multidimensional scaling (NMDS) was adopted to visualize the dispersion of community structure. Multivariate analysis was carried out using vegan [27] and R (http://www.R-project.org) [28].

Metabolic reconstruction of metagenomes
Metabolic reconstruction was carried out using the HUMAnN methodology designed for the functional analyses of meta'omics [29]. High quality reads were initially mapped to the characterized protein functional database KEGG Orthology v54 [30] using the accelerated translated BLAST program USEARCH v6.0.307 [31]. The cutoff E-value was set to 1e-6 and best hits were then used to estimate relative abundances of KEGG orthologous (KO) gene families by HUMAnN v0.98. Base on the resulting KO information, MinPath was used to calculate the coverage and relative abundances of KEGG modules that are manually defined functional units [32]. Circular cladograms representing microbial taxonomic compositions and metabolic modules were implement-ed by using a standalone graphical tool GraPhlAn v0.9.5 (http:// huttenhower.sph.harvard.edu/GraPhlAn).

Detection of metagenomic biomarkers
In order to further test whether some taxa/metabolic modules are significantly overrepresented in the individual soil habitat, statistical analyses were performed according to the inferred relative abundances. Differentially abundant features were identified by the approach of the linear discriminant analysis (LDA) effect size (LEfSe) and could be used as metagenomic biomarkers [21]. As the sample size is not very large in this test, the significance threshold of the alpha parameter for the Krushkal-Wallis (KW) test among classes was set to 0.01 and the cut-off logarithmic LDA score was 2.0. These analyses were performed through the Galaxy server [33]. Additionally, a non-parametric test of Spearman rank correlation between the relative abundances of each KO entry and taxonomic unit was employed to estimate co-variation of community composition and functional features using the R function cor.test.

Detection of microbial interactions
A recently developed computational methodology was used to investigate microbial co-occurrence and co-exclusion relationships within and between soil sites [22]. The microbial network of significant co-occurrence and co-exclusion interactions was built by a Cytoscape plugin CoNet 1.0b2 (http://psbweb05.psb.ugent. be/conet/). The taxonomic abundances estimated by MetaPhlAn were used to prepare an input matrix consisting of data from three sites (grasslands, deserts, and forest soils). The analysis was carried out with the non-default parameters listed below: 50 initial top and bottom edges; four similarity measures (Spearman, Pearson, Kullbackleibler, and Bray Curtis); edgeScores for the randomization routine; 1000 permutations and bootstraps. The resulting networks were merged based on the Simes method [34] and Benjamini-Hochberg false discovery rate (FDR) correction [35]. The FDR cutoff was set to 0.05. The ensemble co-occurrence network was visualized by Cytoscape 2.8 [36].

General characterization of soil community composition
To explore comprehensive characterizations of taxonomic compositions in the soil microbiota, 33 metagenomes sampled from five soil habitats (i.e. grassland, forest soil, desert, Arctic soil, and mangrove sediment) were included in this analysis (Table 1). Based on the assessment by MetaPhlAn, a total of 63 clades (11 phyla and 53 genera) were identified at $0.5% abundance in at least one sample (Table S2). Proteobacteria was the most dominated phylum in the microbial community of soil, $70% abundance detected in all soil sites except for the microbiota in the desert samples ( Figure 1A). In desert, both phyla Proteobacteria and Actinobacteria exhibit almost identical abundance: 30% for Proteobacteria and 29% for Actinobacteria. In addition, Firmicutes and Bacteroidetes, which are the two major phyla dominating the human microbiome [37,38], were not frequently present in the soil microbial communities. Particularly, bacterial species within the Firmicutes rarely occurred in soil. As the taxonomic distribution of environmental metagenomic sequences are greatly affected by distinct reference databases [23], the 16S amplicon approach should provide more accurate taxonomic profiling than metagenome shotgun sequencing [12]. Previous amplicon surveys of 16S rRNA gene have pointed out that bacterial phyla Acidobacteria, Actinobacteria, Bacteroidetes, Proteobacteria, and Verrucomicorbia are often abundant and ubiquitous in soil [12,39]. Although the clade-specific marker gene database in MetaPhlAn has successfully validated the composition of the human microbiome [24], it still needs to be updated with more genomes sequenced recently from various environments. At the family level, several families were observed to be evidently more prevalent in and specific to one soil habitat or closely related habitats ( Figure 1B). For instance, the family Methylocystaceae was dominated and almost equally present in both sites of forest soil (16.4% abundance) and grassland (17.1%) comparing with the other three soil habitats. Additionally, among five soil habitats, both families Ectothiorhodospiraceae (37.0%) and Desulfobulbaceae (30.8%) were found to be extremely abundant in the microbiota of mangrove sediment. The enrichment of these families could be reasonably explained by the selective pressures acting on certain ecological sites. For example, organisms within Methylocystaceae are usually methanotrophs that can metabolize methane as their only carbon source and involved in methane oxidation [40,41]. The DNA-level evidence identified herein may support the oxidation of methane observed in forest soil and grasslands [42,43]. In addition, the microbiota of mangrove sediments is known to be sampled from anaerobic and hyperhaline seawater [23]. The corresponding environmental features should be beneficial for the dominance of Ectothiorhodospiraceae and Desulfobulbaceae in this particular habitat. The former comprises the most halophilic eubacteria [44] and bacteria in the latter family are strictly anaerobe sulphate reducers [45]. However, it is worth mentioning that taxonomic profiling of individual metagenome is visually distinguished from those of the other metagenomes within the same soil habitat ( Figure S1). This is probably because publicly available soil metagenomes were generated by different research groups and varied in sampling strategies as well as sequencing methods. Thus, analysis of more soil metagenomes newly sequenced or coming soon is still required to statistically support the findings of soil biodiversity in the present study.

Structure similarity and taxonomic biomarkers of soil microbial communities
For a glimpse of structural similarity of soil microbial communities, ecological dissimilarity indices Bray-Curtis similarity scores were inferred and summarized in Table 2. The PERMA-NOVA test demonstrated that taxonomic compositions of microbial communities were significantly varied among soil habitats (p = 0.001). Meanwhile, the NMDS plot in Figure 2 further illustrated the compositional similarity among 33 samples from five soil sites. These results demonstrated that the microbiota from the same soil habitat should be more similar to each other. The community structure similarity is also influenced by varied geographical locations. E.g., the soil samples from grasslands were intensively clustered together and the corresponding similarity score (Bray-Curtis index 0.8060.07) is indeed the highest among all inter-or intra-group comparisons ( Table 2). On the contrary, the Bray-Curtis similarity score between nine desert samples is the lowest (0.5860.16) among all intra-group comparisons, perhaps due to their sampling environments: three samples from hot deserts but the remaining ones from cold deserts [12]. Likewise, A2 sampled from the edge of the Arctic Circle is distant from A1 from high Arctic soil (Table 1). In addition, it was observed that the distances of most samples between forest soil and grasslands were closely clustered ( Figure 2) and the Bray-Curtis similarity score was consistently high (0.7660.07). Whereas, the microbiota from two extreme conditions, desert and mangrove soil, respectively, exhibited the greatest compositional dissimilarity (0.3760.07) among all inter-group comparisons.
To further investigate the taxonomic distribution and differentially abundant clades of diverse soil ecosystems, we compared the abundances of microbial compositions at each taxonomic level. Figure 3A shows a cladogram visualizing all detected microbial compositions ($0.5% abundance) from domain to species, respectively. Based on the inferred taxonomic profiling of all samples, a statistical strategy for discovering metagenomic biomarkers was carried out by LEfSe and determined 46 differentially abundant taxa (Table S2). Among these differentially abundant taxa, 10 and 12 were found to be family-and genus-level biomarkers, respectively ( Figure 3A). These detected taxonomic biomarkers could be used as candidate indicators to distinguish at least one microbial community of five individual soil habitats. E.g., two families Beijerinckiaceae and Methylocystaceae that consist of methanotrophic taxa [46] were detected to be family-level biomarkers (P value ,0.01) that were most abundant in the forest soil and grassland, respectively. The abundances of both families were found to be significantly decreased in desert and mangrove sediment ( Figure 3B). The abundance differences of these methanotrophs might be positively associated with the expected capability of methane oxidation among distinct soil ecosystems. Although the organisms within the Alphaproteobacteria class were most differentially abundant in the grassland community, a genuslevel biomarker within Alphaproteobacteria was specially enriched in the communities of forest soil and Arctic soil, respectively ( Figure 3A). Intriguingly, the desert community had two phylum-level markers, Cyanobacteria and Chloroflexi, both of which showed the highest abundance in deserts comparing with other soil sites ( Figure 3A). Bacteria in both phyla can produce their energy through photosynthesis [47]. It was worth noting that the family Oscillatoriaceae within Cyanobacteria was significantly enriched in the desert microbial community ( Figure 3B). The enrichment of these bacterial groups should be consistent with the following environmental features of deserts: extreme arid, strong light, and poor nutrient conditions. In addition, six species were found to be differentially abundant, some of which were uniquely present in the individual soil habitat. E.g., Rubrobacter xylanophilus only occurs in the microbiota of deserts. R. xylanophilus is the most thermophilic actinobacterium known and bears extreme tolerance to desiccation [48]. Bradyrhizobium japonicum, an agriculturally important species of legume-root nodulating [49], was found to be the most abundant in grassland and second in forest soil.

Metabolic potential and functional biomarkers of soil microbial communities
Besides microbial composition, metabolic potential of soil microbial communities was also investigated. In this study, we focused on the KEGG modules that are tight functional units composed of approximately 5 to 20 genes and beneficial for biological interpretation of metagenomes [29,30,38]. To further enhance the performance of statistical inference on the functional analysis, two soil sites with limited samples (two samples from Arctic soil and one sample from mangrove sediment) were excluded. After translated BLAST searching against the database of KO gene families, we found an average of ,33.6% of reads mapped to at least one KO entry (Table S3). Based on the metabolic reconstruction of 30 metagenomic datasets using HUMAnN, Figure 4 shows 119 functional modules detected in the microbial communities of grasslands, forest soils, and deserts (Table S4). Of these functional modules, we found 20 core metabolic modules that were almost entirely present at .90% coverage in all soil metagenomes tested ( Table 3). Some of these core modules were essential for basic life activities of prokaryotic cells in soil, such as central carbon metabolism (M00002-3 In addition, all the remaining core modules were involved in certain transport systems, three (M00207, M00222, M00239) of which are also detected in the core modules of the human microbiome [29]. On the other hand, three functional modules (M00026, M00133, M00319) were differentially covered among three soil sites (Figure 4; Table S5). It was worth noting that structural complex module manganese/zinc/iron transport system (M00319) was completely present only in the deserts but appeared to be absent in both grasslands and forest soils. It indicates that deserts microbiota is well-equipped with metal acquisition systems that play potential roles in the maintenance of metal homeostasis [50].
Furthermore, 33 functional modules were detected to be differentially abundant in at least one of three soil sites ( Figure 4 and Table S4). Interestingly, two thirds of these modules were significantly enriched in the microbiota of deserts in comparison to the microbiota of grasslands and forest soils. Of them, three metabolic modules (M00165-167) are involved in the reductive pentose phosphate cycle (Benson-Calvin cycle), which is the main pathway for the conversion of atmospheric CO 2 to organic compounds [51]. It was worth noting that these overrepresented modules involved in carbon fixation might be consistent with high abundance of photosynthetic organisms Cyanobacteria present in the microbiota of desert. Additionally, eight structural complex modules detected to be functional biomarkers in deserts are responsible for the transport of metallic cation (M00317, M00319), mineral and organic ion (M00321, M00299), saccharide and polyol (M00201, M00199), glutamate (M00233), and urea (M00323). On the other hand, we found that two metabolic modules (M00022: Shikimate pathway and M00237: Branched chain amino acid transport system) were significantly overrepresented in grasslands and forest soils comparing with plant-free deserts ( Figure 4 and Table S4). Both modules are associated with plant-derived metabolites [52,53]. These results showed that some modular metabolic activities are likely to be associated with the individual soil ecosystem. However, more metagenome samples from different sites are needed for accurately statistical validation of these characterized modules as promising biomarkers for diverse soil communities.

Correlation between microbial compositions and functions
Similar to the approach presented by Segata et al. [38], we assessed the correlations between microbial compositional and functional enrichment. The results showed that some significant associations between taxonomic clades and functional gene families were detected in the soil microbial communities (Spearman non-parametric test; Benjamini-Hochberg corrected p-value ,0.01) ( Figure S2). Notably, several taxonomic biomarkers possessed by individual microbial community mentioned above were further confirmed by the related strong associations between gene families and taxonomic clades. E.g. the gene petA (K02634) encoding apocytochrome f protein involved in photosynthesis was positively associated with the members of Cyanobacteria (Spearman test; q-value ,0.001), one of the earliest prokaryotic organisms which can carry out oxygenic photosynthesis on Earth [47]. In addition, a significantly positive correlation (Spearman test; q-value ,0.001) between methanotrophs Methylocystaceae and the gene mcl (K08691) coding for malyl-CoA lyase was observed in the microbial community of grassland. The enrichment of protein Mcl involved in both pathways of methane metabolism and carbon fixation, should be consistent with the featured metabolic activities of these methanotrophs.

Soil microbial interaction network
To further decipher complex ecological relationships in the individual soil microbial community, microbial association networks were inferred based on the estimated taxonomic profiling. In this case, we intended to focus on the microbial associations within the single soil habitat, i.e. forest soil, grassland, and desert. The resultant metagenome-wide networks comprised 126 significant associations among 66 phylotypes at or above the genus level (Benjamini-Hochberg corrected p-value ,0.05) ( Figure 5). Of these significant phylotype correlations, 54% was detected to be co-present and the remaining was mutually excluded. Interestingly, we found that three quarters (,74%) of co-occurrence patterns were constituted by the taxa within the same phyla; whereas nearly all co-exclusion patterns (,90%) consisted of the taxa from the distinct phyla. The evidence presented herein can again support the previous notion that phylotypes with closely evolutionary relationships usually tend to co-occur [8]. E.g., three families (Bifidobacteriaceae, Mycobacteriaceae, and Frankiaceae) belonging to the same class Actinobacteria showed pairwise positive correlation in the microbiota of desert ( Figure 5). Similar taxon co-occurrence pattern was also found between Bifidobacteriaceae and Frankiaceae in the microbiota of grassland. Additionally, two genera within the family Bradyrhizobiaceae co-occurred in the grassland community: one is nitrogen-fixing bacteria Nitrobacter and the other is . Taxonomic composition of soil microbial community based on the metagenomes from five soil habitats. A) Taxonomic cladogram showing all detected taxa (relative abundance $0.5%) in at least one sample. Taxonomic clades with more than five samples $0.5% abundance were used as inputs for LEfSe. Seven rings of the cladogram stand for domain (innermost), phylum, class, order, family, genus, and species (outermost), respectively. Enlarged circles in color are the differentially abundant taxa identified to be metagenomic biomarkers and the circle color is corresponding to the individual soil habitat in which the taxon is the most abundant among 5 soil ecosystems (Green for forest soil, red for grassland, purple for Arctic soil, blue for mangrove sediment, and orange for desert). B) The histograms of relative abundances of family-level biomarkers in each sample. Bacterial families significantly differential among all pairwise comparisons were illustrated. The average abundance of each family in the individual soil habitat was denoted by the horizontal line. doi:10.1371/journal.pone.0093445.g003 phototrophic bacteria Rhodopseudomonas. On the other hand, those mutually excluded bacteria were found to be evolutionarily unrelated. E.g., Sphingobacteriaceae belonging to the Bacteroidetes were negatively associated with Desulfovibrionaceae from the Proteobacteria and Rubrobacteraceae from the Actinobacteria in the microbiota of desert ( Figure 5). Although most phylotype associations in the network lack empirical evidence to support their natural presence, it provides some promising targets at least to shed light on the complex cooperative or competitive mechanisms among soil microorganisms.

Conclusions
In this study, comparative metagenomic characterizations of divergent soil microbial communities were described in details by an integrated bioinformatics analysis pipeline. Complicated phylogenetic and metabolic networks with a spectrum of taxonomic and functional biomarkers were comprehensively illustrated at the metagenome level for soil. Cooperative or competitive associations among microbes from diverse soil ecosystems were also inferred to understand complex microbial interactions in the soil metagenome. This study provides new insights into the relation between soil biodiversity and ecosystem function, and provides applicable analysis and visualization approaches for studying soil microbial communities.   . Metagenome-level metabolic reconstruction of the soil microbial community. KEGG BRITE hierarchical structures that are illustrated by the innermost four rings were used to cluster metabolic modules. The outermost ring composed of circles denotes KEGG functional modules detected in at least one of 30 metagenomes from three soil sites. Differentially abundant modules were inferred by LEfSe and illustrated by the enlarged circles in distinct colors: green stands for the modules most abundant in the forest soil, red for the grassland, and orange for the desert. The outermost rectangles denote core and differentially covered modules among three soil sites: $90% coverage stands for presence and #10% coverage for absence. doi:10.1371/journal.pone.0093445.g004 Table 3. Core metabolic modules shared by grasslands, deserts, and forest soils.  Figure 5. A global microbial interaction network of the soil microbial community. The network captured all significant associations (multiple corrected p-value ,0.05) among the abundances of phylotypes at or above the genus level in the soil microbial community within and across the three soil sites. Phylotypes were illustrated by nodes (light blue for grasslands, blue for forest soils, and yellow for deserts) and edges denote significant correlations between phylotypes: positive correlation colored in green means co-occurrence whereas negative correlation in red means mutual exclusion. The border of nodes was colored according to taxonomic affiliations at the class level. doi:10.1371/journal.pone.0093445.g005 the highest LDA score among pairwise comparisons of all sites. According to the non-strict and strict statistical strategy, 46 taxonomic biomarkers were detected to be significantly differential in at least one of five soil habitats.