This study was conducted on Hainan Island (18∘10′–20∘10′ N and 108∘37′–111∘03′ E), in southern China. The total area of Hainan Island is about 34 000 km2 (Lopez et al., 2009). It is the largest island within the Indo-Burma Biodiversity Hotspot of tropical Asia (Francisco-Ortega et al., 2010) and has a tropical monsoon climate. Hainan Island has a warm and humid climate year-round, with an average annual temperature of 22–26 ∘C. Rainfall is abundant, ranging from 1000 to 2600 mm yearly, with an average annual precipitation of 1639 mm. The rainy season occurs from May to October, with a total precipitation of about 1500 mm, accounting for 70 %–90 % of the total annual precipitation. Only 10 %–30 % of the total annual precipitation falls within the dry season, from November to April. The central part of Hainan Island is mountainous and contains old-growth tropical rainforests and monsoon forests. Rubber plantations are found on the plateaus surrounding the central mountainous zone.

The tropical rainforests of Hainan are mainly distributed in Bangwangling, Diaoloshan, Wuzhishan, Yinggeling and Jianfengling national forest parks. Therefore, we selected a tropical rainforest at each of these sites for our study objectives. There are about 530 000 ha of rubber plantations on Hainan Island. Five rubber plantations were selected, one in Wanning, Ledong, Danzhou, Haikou and Qiongzhong, respectively. These five rubber plantations are located in the east, south, west, north and middle of Hainan, respectively (Fig. S1 in the Supplement). We selected 25- to 30-year-old rubber trees (i.e., mature rubber plantations) for our study objectives. Latex harvest and fertilizer application are common management practices in rubber plantations. Usually, compound fertilizer (1–1.5 kg per tree) and organic fertilizers (20–25 kg per tree) were applied once or twice a year. More information on the study sites is provided in Table S1 in the Supplement. Before soil sampling, we sterilized the soil drill with 75 % alcohol. After the removal of the litter layer using a 5 cm diameter steel drill, topsoil (0 to 20 cm) was collected, homogenized and passed through a 2 mm mesh sieve. For each site, 13 soil samples were collected; thus, a total of 130 samples were collected between the rubber plantation and tropical rainforest sites. Soil sampling was performed twice in 2018, once in January (dry season) and once in July (rainy season), resulting in a total of 260 soil samples (130 per forest type). Soil samples were divided into two parts. One was used to analyze soil water content, soil pH, total nitrogen (TN), total phosphorus (TP), total potassium (TK), nitrate nitrogen (NN), ammonium nitrogen (AN), available phosphorus (AP) and potassium (AK). Soil water content (%) was measured gravimetrically. Soil pH was measured in a soil : water suspension (1 : 2.5, w/w) using a pH meter. TN was determined using micro-Kjeldahl digestion followed by steam distillation. TP and TK were measured following digestion with NaOH. NN and AN were determined by steam distillation and indophenol-blue colorimetry, respectively. Soil samples were extracted with NaHCO3, and the extract was then used to measure available soil phosphorus AP via molybdate-blue colorimetry. To measure soil AK, soil was extracted with ammonium acetate, and the extract was then loaded onto an atomic absorption spectrometer with ascorbic acid as a reductant (Chen et al., 2019). Soil properties of the rubber plantation and rainforest sites are shown in Table S2. The other part of each soil sample was stored in an ultralow-temperature (-80 ∘C) refrigerator for later DNA extraction.

Microbial DNA was extracted from 0.5 g of soil using E.Z.N.A.^® soil DNA kits (Omega Bio-tek, Norcross, GA, USA) following the manufacturer's protocol. The fungal ITS1 hypervariable region was amplified using the PCR primers ITS1F (5′-CTTGGTCATTTAGAGGAAGTAA-3′) and ITS2R (5′-GCTGCGTTCTTCATCGATGC-3′) (Adams et al., 2013). For bacteria and archaea, the V4 hypervariable region of the bacterial 16S rRNA gene was amplified using the PCR primers 515FmodF (5′-GTGYCAGCMGCCGCGGTAA-3′) and 806RmodR (5′-GGACTACNVGGGTWTCTAAT-3′) (Walters et al., 2016; Sampson et al., 2016). The PCR reactions were conducted using the following approach: an initial 3 min denaturation at 95 ∘C, 27 cycles of 30 s at 95 ∘C, 30 s of annealing at 55 ∘C, 45 s of elongation at 72 ∘C, and a 10 min final extension at 72 ∘ C.

Amplicons were extracted from 2 % agarose gels, purified using the AxyPrep DNA gel extraction kit (Axygen Biosciences, Union City, CA, USA) and quantified using a QuantiFluor™ single-tube fluorometer (Promega, USA). Purified amplicons were pooled in an equimolar solution and then sequenced (paired-end, 2×250 bp) on an Illumina MiSeq platform according to standard protocols.

Shotgun metagenomic sequencing libraries were prepared and then sequenced by Majorbio, Inc. (Shanghai, China) using the Illumina HiSeq 2000 platform. The NRGene gene catalog was aligned against the Kyoto Encyclopedia of Genes and Genomes (KEGG) database using BLAST (Basic Local Alignment Search Tool, version 2.2.28+) (Qin et al., 2010).

Raw FASTQ files were demultiplexed and quality filtered using QIIME (Quantitative Insights Into Microbial Ecology, version 1.17; Caporaso et al., 2010). Operational taxonomic units (OTUs) were clustered with a 97 % similarity cutoff using UPARSE (Edgar, 2013), and chimeric sequences were identified and removed using UCHIME. Using the RDP (Ribosomal Database Project) classifier, the phylogenetic affiliation of each 16S rRNA gene and internal transcribed spacer (ITS) gene sequence was determined using a confidence threshold of 70 % with the SILVA 16S rRNA database and UNITE database, respectively (Amato et al., 2013). For each site, the relative abundance of different taxa (Good, 1953) and the Shannon diversity index were calculated based on resampled sequence data using mothur (Schloss et al., 2009). The raw reads were deposited into the National Center for Biotechnology Information (NCBI) Sequence Read Archive (SRA) database (Accession Number: SRP108394, SRP278296, SRP278319).

To make the co-occurrence network analyses simple, we only used the top 300 most abundant bacterial and fungal OTUs to analyze the network structure of the two forest types. These OTUs roughly match the OTUs with a relative abundance greater than 0.05 % (Jiao et al., 2016). Four networks, representing each habitat during each sampling period (rainforest and rubber plantations in the dry and rainy seasons) were constructed with 65 samples each. We also performed the bacterial–fungal community network analysis to investigate the soil microbial network complexity of each of the four. Interactions consisted of Spearman's rank correlations. Co-occurrence networks were constructed using only significant correlations of ρ>0.6 (P<0.01) (Barberan et al., 2012), as this cutoff includes a range of interaction strengths (De Vries et al., 2018). The networks were then visualized in R using the “igraph” package. To reveal the distribution pattern of correlation coefficients, the frequency of the coefficients of ρ>0.3 (P<0.01) were plotted. The “NetworkAnalyzer” tool in Cytoscape (version 3.4.0) was used to calculate network topology parameters including number of nodes, edges, degree, betweenness and closeness. The number of shared edges and unique edges as well as keystone OTUs were calculated to evaluate whether the network structure changed. We defined a shared network edge as the link (edge) between two species (species A and species B) that not only exist in the rubber plantation network but also exist in the rainforest network. In contrast, a unique edge was defined as a link that only exists in either the rubber plantations or rainforests. The numbers of shared edges and unique edges were calculated using igraph and were visualized using Venn diagrams. Keystone OTUs are known to be important for ecosystem structure and function and were selected on the basis of high network degree, high closeness centrality, and low betweenness centrality, as defined by Berry and Widder (2014). To evaluate the proportional influence of each phylum on bacterial and fungal network structure, node degrees of each phylum were calculated and bar plots were created. Correlation coefficients between species and functions were calculated based on metagenomics data. Here, we used the top 50 most abundant species and the top 50 KEGG functions (pathway level 3). The species and function correlation network was then constructed on the Majorbio cloud platform (https://cloud.majorbio.com/, last access: 23 January 2022). To reveal the relationship between microbial taxa and environmental variables, two-way correlation networks were also constructed on the Majorbio cloud platform. The topological role of each node in a network was assessed by the Zi and Pi values, where Zi represents the connectivity of node i within a module, and Pi measures the degree of the connection of node i with other modules (Roger and Amaral, 2005). All species can be divided into four groups according to simplified criteria (Olesen et al., 2007), namely peripherals (Zi<2.5 and Pi<0.62), connectors (Pi>0.62), module hubs (Zi>2.5) and network hubs (Zi>2.5 and Pi>0.62). The Zi and Pi values were calculated using the “GIANT” package in Cytoscape. The Zi–Pi plot was created with “ggplot2” in R.

Our results showed that the bacterial–fungal community network was more complex in rubber plantations (4284 edges in the dry season, 7257 in the rainy season) than in primary forests (3650 in the dry season, 6507 in the rainy season) and that it was more complex during the rainy season than during the dry season (Fig. 1). Most nodes of bacterial networks (Fig. 2) and fungal networks (Fig. 3) varied with forest type in both the dry and rainy seasons. For bacterial networks, there were 2559 and 2501 edges in the forests and rubber plantations, respectively, in the dry season, but these two networks only shared 262 edges, accounting for only about 10 % the total edges (Fig. 4a, b). Similarly, these networks only shared 519 edges during the rainy season. For fungal networks, there were only four and five shared edges (i.e., no more than 5 % of the total edges) in the dry and rainy seasons, respectively (Fig. 4c, d).

The number of edges of bacterial and fungal networks were almost equivalent during the dry season. However, in the rainy season, there were more edges in the rainforest bacterial network than in the rubber plantation network (Table 1, Fig. 4b). For the network structure of the fungal community, more edges were observed in rubber plantations in the rainy season (Table 1, Fig. 4d). Similarly, there were no significant differences in the degree of the bacterial or fungal networks between the tropical rainforest sites and rubber plantations in the dry season (Fig. 5a, c). In the rainy season, rainforest sites had a higher bacterial network degree, while rubber plantations showed a higher fungal network degree (Fig. 5b, d). For bacterial networks, more nodes (OTUs) with a high degree (rubber plantation had two nodes with a degree greater than 75, rainforest had eight such nodes) were observed in the rainforest during the rainy season (Fig. S2B). For fungal networks, 15 nodes of a higher degree (degree greater than 25) were observed in rubber plantations; however, only 7 such nodes existed for rainforest sites (Fig. S2D). These results indicate that the fungal network structure in the rubber plantations was more complex than for tropical rainforests during the rainy season but that the reverse was true for bacteria.

When considering the ratio of positive to negative correlation coefficients, the bacterial–fungal community network revealed that rubber plantations (844 edges in the dry season, 1744 edges in the rainy season) have more negative links than rainforests (149 edges in the dry season, 489 edges in the rainy season), indicating that the rubber plantation network was more stable than that of the rainforests (Fig. 1). More correlations (greater than 0.3, P<0.01) were also calculated to investigate the positive to negative correlation ratio, and the results showed that negative correlations between bacterial and fungal OTUs in the rubber plantations were consistently stronger than in the tropical rainforests in both the dry season and the rainy season (Fig. 6).

Module hubs and network hubs have been commonly identified as keystone taxa in networks. However, for both the bacterial and fungal communities, neither tropical rainforest nor rubber plantation networks possessed module hubs and network hubs (Figs. S3, S4). The keystone taxa of bacteria were very different between rubber plantations and tropical rainforest sites in both the dry and rainy seasons. For bacteria, there were more keystone taxa at tropical rainforest sites than in the rubber plantations in both the dry and rainy seasons, indicating that the tropical rainforest networks had higher complexity. We found that some groups of Acidobacteria are keystone taxa at tropical rainforest sites but disappeared after forest conversion. There were more Actinobacteria in rubber plantations than at tropical rainforest sites (Table S3).

For fungi, more keystone taxa were observed in rubber plantations than at tropical rainforest sites during both the dry and rainy seasons, indicating that the rubber plantation networks were more complex. Most keystone taxa belong to Ascomycota, suggesting that members of this group are very important for network structure. In addition to forest conversion, seasonal changes also affect the keystone taxa of the fungal community network. There were more Basidiomycota OTUs in the dry season but more Ascomycota in the rainy season (Table S4).

Two-way correlation network analysis of the 300 most abundant OTUs and environmental variables revealed that elevation, AK and soil pH were the main drivers of bacterial community composition and had 106, 104 and 86 degrees, respectively (Fig. S5A). For fungi, elevation had the highest degree (61), followed by AK (51) (Fig. S5B). The relationship between OTU abundance and soil pH revealed that the soil pH was negatively correlated with the abundance of some species of Acidobacteria but was positively correlated with the abundance of some species of Chloroflexi and some species of Ascomycota (Fig. S6). AK was positively correlated with the abundance of Planctomycetes, Verrucomicrobia and Basidiomycota species but was negatively correlated with the abundance of Chloroflexi and Ascomycota. In other words, AK and soil pH were important drivers of microbial composition.

The total degree of each phylum was used to reveal the influence of each phylum on network structure (Fig. 7). For bacteria, Proteobacteria, Actinobacteria and Acidobacteria had a large influence on network structure. Acidobacteria and Planctomycetes contributed more to rainforest networks than to rubber plantation networks. However, Actinobacteria and Chloroflexi showed the opposite pattern. For fungi, Ascomycota and Basidiomycota had large influences on network structure. Both Ascomycota and Basidiomycota had stronger influences on rainforest networks than rubber plantation networks. The influence of Ascomycota was stronger during the rainy season than during the dry season, indicating that seasonal change also had an impact on fungal community networks. The degree of each phylum of bacterial and fungal network also revealed that seasonality had effects on the network degree (Fig. S7).

Two-way network analysis of the 50 most abundant species (all belonging to bacteria groups based on metagenomic data) and the 50 most abundant KEGG functions revealed the interaction between microbial composition and functions. The results showed that soil network structure at rainforests sites (460 edges in the dry season, 1750 in the rainy season) was more complex than that of rubber plantations (223 in the dry season, 451 in the rainy season) (Fig. 8), indicating that more links between microbial species and functions were observed in the soils of rainforest sites. Both rubber plantation and rainforest networks were more complex in the rainy season than in the dry season. We also found that metabolism was the most important function in soil microbial networks. Surprisingly, species of Actinobacteria were negatively correlated with function in rubber plantations.

Land-use changes increasingly threaten biodiversity, particularly in tropical forests (Gibson et al., 2011). However, we still have little understanding of how soil networks respond to forest conversion, such as when rainforests are converted to rubber plantations. Our results showed that forest conversion had large effects on both soil bacterial and fungal networks. Tropical rainforest bacterial networks had more edges (Table 1) and a higher degree (Fig. 4) than rubber plantation networks, especially during the rainy season, indicating that the rainforest bacterial network was more complex than the rubber plantation network. This finding is consistent with previous observations that soil bacterial networks were more complex in natural systems than in crop soil (Karimi et al., 2019). Further study showed that soil networks become more connected through the process of ecological restoration (Morriën et al., 2017). The observed decreases in network complexity and cohesion support the hypothesis that cropping may enhance the isolation of bacterial taxa (Karimi et al., 2019), which results in lower numbers of connections in the network. In addition, at the microscale, the structure of tilled soils is more homogeneous, and the soil pores are less connected than in soils that are not tilled (Pagliai et al., 2004), such as rainforest soil. In nature, soil ecosystems are highly heterogeneous, as hot spots of soil microbial biodiversity can form spatial and temporally within soil aggregates (Bach et al., 2018). This spatial heterogeneity likely plays an important role in the interactions among microbes and the mechanisms by which more complex and diverse communities drive various nutrient-cycling processes on small spatial scales (Wagg et al., 2019).

A large number of studies employing microbial network analysis have enriched our understanding of microbial co-occurrence patterns in various soil ecosystems. However, very little is known about whether differences in the structure of microbial networks have consequences for microbiome functioning (Wagg et al., 2019). Our results demonstrated that more species were correlated with metabolism in a natural system than in an agricultural system, especially during the rainy season. This is in line with a previous study conducted in Sumatra, Indonesia, which found that the transformation of forest to rubber plantation results in a 10 %–16 % decrease in community metabolism (Barnes et al., 2014). Fewer interactions between microbial species (most of them are bacteria) and functions in rubber plantations demonstrated that forest conversion reduced the soil bacterial network complexity.

Surprisingly, rainforest bacterial networks were characterized by fewer edges (Table 1) and a lower degree (Fig. 4) than the rainforest bacterial networks, meaning that rubber plantation bacterial networks were more complex than those of native forests. Our results were not consistent with a study conducted in Xishuangbanna (Song et al., 2019) which showed that tropical forest conversion reduced fungal network complexity, but out work is consistent with previous observations which found that fungal community networks were more organized in disturbed forest compared with primary forest (Chen et al., 2019). The observations of Banerjee et al. (2019) showed that organic agricultural fields harbored much more complex fungal networks with many more keystone taxa than conventionally managed fields. Forest conversion resulted in shifts in fungal composition from Basidiomycota to Ascomycota (Fig. S8), as seen in previous investigations (Lan et al., 2021, 2020a). Previous work showed that Basidiomycota species display higher drought sensitivity than Ascomycota species (Taniguchi et al., 2018); this would result in a shift in the richness and abundance of Basidiomycota species (Fig. S8). Many Basidiomycota species are capable of long-distance dispersal (Egidi et al., 2019; Geml et al., 2012), which may result in a decrease in fungal network complexity. This possibly explained why Ascomycota OTUs contribute more to the network structure than Basidiomycota (Fig. 6). Overall, reductions in the abundance and richness of Basidiomycota species led to an increase in fungal links in rubber plantations.

The ability of a microbial community to resist and recover from disturbances is defined as the stability of the community (De Vries and Shade, 2013). The ratio of positive to negative network links constitutes the balance between facilitative and inhibitive relationships within a network (Karimi et al., 2017). Theoretical studies, for example, predict that ecological networks that consist of weak interactions are more stable than those with strong interactions (Neutel et al., 2002; Coyte et al., 2015) and that compartmentalization and the presence of negative interactions increase the stability of networks under disturbances (Coyte et al., 2015; Rooney et al., 2006; Stouffer and Bascompte, 2011). In our study, more negative correlations were detected in the rubber plantation, indicating that the network structure of rubber plantation soils was more stable than that of rainforest soils (De Vries et al., 2018). Forest conversion from rainforest to rubber plantations resulted in shifts in bacterial composition from Proteobacteria to Chloroflexi and shifts in fungal composition from Basidiomycota to Ascomycota (Lan et al., 2021). Proteobacteria and Basidiomycota are copiotrophic microorganisms, whereas Chloroflexi and Ascomycota are oligotrophic (Yang et al., 2019; Lan et al., 2021). This indicated that forest conversion resulted in shifts in soil microbial composition from more k-strategists to more r-strategists (De Vries and Shade, 2013). r-strategists are quick-growing and less resistant but are more resilient to disturbances. However, growth is not the only mechanism that could maintain community stability, as the responses of microorganisms to disturbance are very complex.

Forest conversion results in a loss of plant diversity, a loss of plant biomass and an increase in soil pH (Lan et al., 2017a, b). Rubber plantations had a significantly higher pH, which explains the relative decrease in the abundance of Acidobacteria (Lan et al., 2017a). Our results demonstrate that forest conversion drives a shift in which soil microbes act as keystone species (Table 1). We found that many OTUs of Acidobacteria fit our criteria as keystone species for rainforest sites, which is consistent with previous findings (Banerjee et al., 2018). Unexpectedly, OTU11388 and OTU11373, both Acidobacteria, were observed in rainforest soils in both the dry and rainy seasons, indicating that Acidobacteria were very important for rainforest soil bacterial networks (Fig. 6, Table S3). Higher AK concentrations resulted in a higher abundance and more taxa of Actinobacteria (Fig. 9), which suggests that Actinobacteria contributed more in rubber plantation than in rainforest (Fig. 6). Indeed, forest conversion reduced the abundance of Actinobacteria OTUs (Fig. S9). Due to human disturbance in rubber plantations, the soil will inevitably be slightly polluted with herbicides and domestic garbage. A previous study observed a member of Actinobacteria in contaminated soil (Jiao et al., 2016).

Forest conversion also increases land-use intensity (Brinkmann et al., 2019), including the application of fertilizer and herbicide. The application of herbicides has also been linked to significant decreases in root colonization and spore biomass of arbuscular mycorrhizal fungi in tropical agriculture (Zaller et al., 2014). Soil nutrient concentrations show a decline around the roots of rubber plantations compared with those from rainforests (Sahner et al., 2015). Our study was no exception. For instance, the AK and TN concentrations were significantly lower in rubber plantations than in samples from the rainforest sites (Table S2). Higher concentrations of AK reasonably explained the higher contribution of Basidiomycota to the network structure (Fig. 8b) due to the positive association of AK with Basidiomycota.

Spatiotemporal heterogeneity can be a major driver of the abundance and distribution of keystone taxa in soil, which is a highly heterogeneous and multifaceted environment (Mills et al., 1993; Power et al., 1996; Mouquet et al., 2013). Seasonal variability determines the structural and compositional properties of microbiomes in an environment; as such, a keystone species might be present only during a specific season or time period (Banerjee et al., 2018). It was interesting that more bacterial OTUs were identified as connectors during the rainy season than during the dry season. Connectors have been characterized as generalists (Olesen et al., 2007), and generalists drive covariation among communities in a network (Chen et al., 2019). Previous observations have demonstrated that some keystone taxa that were found in the dry season disappeared during the rainy season (Lan et al., 2018). Seasonal changes possibly explain why some keystone taxa were observed in the rainy season but not in the dry season.

Forest conversion results in a decrease in the abundance of Proteobacteria and an increase in Actinobacteria. Most species of Proteobacteria were positively correlated with metabolic function, whereas most of Actinobacteria were negatively associated with metabolic function (Fig. 8). Therefore, the changes in the abundance of these two phyla result in a reduction in microbial community function after forest conversion. Because the metabolic function of a specific species is often affected by environmental conditions (Louca et al., 2018), some species are not correlated with any function in the dry season but are correlated with metabolic function in the rainy season, indicating that some microorganisms do not participate in the metabolic process during the dry season, especially in the rubber plantations. In conclusion, forest conversion as well as seasonal variation had a great impact on soil microbial community functions.