Due to the high diversity and turnover in Amazonia we could expected a neutral community assembly due to dispersion limitation [72, 73] and a high redundancy of functional traits in the microbial community , which would result in less connected co-occurrence networks. However, here we show that Amazonian micro-organism communities form highly interconnected co-occurrence and co-exclusion networks when compared to those of other regions [75, 12]. For instance, global scale samples of the Tara Ocean project, with 313 samples and more than 1 million OTUs, including viruses, showed a network composed of 29,912 OTUs and 127,000 edges with an average degree (number of edges per node) of 4.26 . Our results, with a smaller sampling of 39 Amazonian plots (78 samples, 25% of the number of Tara Ocean samples) still resulted in a network of 72,577 edges involving 2,350 OTUs, with an average degree of 30.88. In a study surveying soil bacteria in France, which included 30 soil samples for different land use categories (forest, grassland, crop system, and vineyards), researchers found a range from 16,430 edges involving 1,083 nodes in forests to 2,046 edges involving 1,342 nodes in vineyards . Our study comprised 24 samples for terra-firme, 24 for igapós, 18 for campinas, and 12 for várzeas, and ranged from 6,223 edges involving 1,047 nodes in terrra-firme to 321 edges involving 273 nodes in várzeas. Although these studies use different methods to calculate the networks, our results demonstrate the relatively high complexity of Amazonian networks and highlight the importance of biotic and/or abiotic interactions in this region.
H1: The Amazonian-wide network will be mostly composed by organisms associated with organic decomposition: As expected in our H1, considering bacterial metabolism, there is a combination of bacteria that rely on respiration (e.g., Acetobacteraceae and Planctomycetaceae) and those that rely on fermentation (e.g., Fusobacteriales and Lactobacillales). These two groups complement each other in ecological functionality and indicate a high turnover of organic material underpinned by both metabolisms. Some groups present in the networks are active in degrading cellulose (e.g., Acidothermaceae and Polyangiaceae), which is in agreement with our litter samples that had a high amount of plant debris. There are also a few nitrogen-fixating groups, but they do not seem to be very common. Yet, one third of the bacteria recorded in our networks are not assigned to any functional group or preference, stressing the need for further investigation of Amazonian bacterial communities. Furthermore, the majority of fungi found here were saprotrophic, that is, decomposers of organic matter and, therefore, important agents in carbon cycling . Unfortunately, archaea were heavily underrepresented in the results (just 131 OTUs) of our sequencing approach. We do not consider this a biologically meaningful result but rather an artefact of the amplification process, since archaea are commonly found in soil samples, where they contribute to, e.g., ammonia oxidation [78, 79]. We therefore decided to exclude archaea from our analyses and refrain from making any statements about this group of organisms.
Deforestation alters patterns of co-occurrence, impacting ecosystem functions . The edges of Amazonian forest fragments have an increase of soil organic carbon probably due to high tree mortality , but potentially also due to an alteration of the microbial community , which may have an impact on the rate of organic decomposition and the ecosystem resilience capability. With the increasing anthropogenic pressures in Amazonia , it is crucial to understand how biological interactions are linked to decomposition in Amazonia and the potential effects of microbial community alteration in the world’s most diverse forest.
H2: Environmental soil properties, especially pH and organic carbon (Ritter et al. 2018), will be the most important factors to explain co-occurrence in all networks: Our hypothesis H2 was partially supported. The organic carbon and base saturation indices were key nodes in explaining co-occurrence in the general network, whereas several other soil properties were key nodes for the co-exclusion, such as soil texture, exchangeable bases, aluminum, and pH. The composition of micro-organisms in the soil is usually determined by physico-chemical conditions such as phosphorous [83, 84, 85], pH [86, 87], and soil moisture . Using the same data, Ritter et al.  showed that pH was the most important factor explaining the community turnover, and organic carbon the OTU richness . This is the reason why we expected the pH to be important for the co-occurrence network, but instead we found it to be important just for co-exclusion.
Our results showed a predominance of bacteria that grow in low pH regimes, especially Acidobacteria , which are very abundant in soil. Amazonian soils are acidic, with pH varying from 3.65 to 5.14 , which explains the large presence of acidophilic organisms in both co-occurrence and co-exclusion networks. However, in less acidic soils, other groups could co-occur with acidophilic organisms, but the acidity tends to lead to increased competition. Phyla associated with low pH, such as Acidobacteria and Proteobacteria [85, 90], may be better competitors in more acidic soils. Furthermore, temperature appears to select for groups that are tolerant to high temperatures up to 40 °C (Fig. S8), which can occur in Amazonia all year. The soil properties were less important in the habitat-specific networks. As these soil properties are associated with the habitat types, and the same habitat presents similar soil physico-chemical conditions , the network analysis of habitats may already be controlled by the variation in soil properties.
H3: The habitat-specific networks will show different network structures, with the presence of flood pulse (the periodic inundation of floodplains along certain rivers) as the main factor explaining network properties: Our hypotheses H3 was also partially supported. We expected that the high environmental stress in igapós and várzeas due to the seasonal flooding would act as an abiotic filter, keeping the same set of specialized organisms that co-occur for long periods of time, which would result in a dense and highly connected network. However, both habitat-specific networks were sparsely connected considering analyses within localities (Fig. S9-S12). Both habitats remain submerged during most of the year, up to 240 days . In our data, the seasonally flooded forests – the várzeas and igapós – were more similar to each other in terms of community composition , which could be related to similar environmental filters linked to stress by flooding [92, 93], although the transitivity, that may indicate niche filtering, of the habitat networks was similar. Other factors, such as the random colonization due to smaller area size than in terra-firme, fragmented habitat distribution [94, 95], or the seasonally introduced species with the flood pulse that brings organism from river curse, could to some extent randomize the presence of OTUs. Várzeas differ from igapós by being a more fertile habitat, as their waters carry sediments from the Andes . Therefore, a higher rate of colonization by micro-organisms is to be expected in várzeas than in igapós, which are bathed by acidic, low-fertile waters [96, 97]. Várzeas may therefore have more micro-organisms that could survive this particular stress condition, not just specialists. Indeed, várzeas had fewer specialist OTUs than igapós . The difference between the sampling size of igapós (24 replicates) and that of várzeas (12 replicates), though, makes a more general comparison between these habitats difficult.
Communities in campinas (Fig. 4A) resulted in a smaller network (fewer nodes and fewer edges) than terra-firme (Fig. 4C) and igapós (Fig. 4E, Table 2), but larger than várzeas (Fig. 4G). However, terra-firme and igapós had more replicates (24 each) than campinas (18 replicates). Regarding network properties, campinas and terra-firme are similar (Table 2). However, the habitat network of campinas shows two distinct modules (Fig. 4A) that either represent two different metabolic pathways, as the dominant bacterial phyla are different in each module, or indicate that there is a geographical clustering . The geographical clustering could be related to the natural fragmentation of campina habitats, which makes colonization more random than in more continuous habitats such as terra-firme [94, 95]. Yet, comparing the campinas networks within localities, the localities bathed by acidic, sediment-poor black waters (Jaú and Cuieras) had more modular networks (Fig. S9-S12, Table2). That is also true for terra-firme, that had the low modularity (0.47 in co-occurrence network; Table 2), in other words a more connected network in Caxiuanã, a locality bathed by rich sediments from a white-water river (Fig. S9-S12). Stressful localities may have a more specialized co-occurrence community due to the environmental filters such as low pH, while less stressful localities may allow more organisms to co-occur.
General networks composition: Overall, the key nodes of our networks are mainly affiliated to Proteobacteria (29 nodes, of which 17 belong to Alphaproteobacteria), Acidobacteria (19), Planctomycetes (12), and Actinobacteria (10, Table S3). These are common groups of bacteria in soils in general  and also in our samples [36, 37], being the most frequent groups in our networks. Alphaproteobacteria are a highly diverse clade of Gram-negative bacteria with several biological functions, including metabolizing C1 compounds , fixating nitrogen , endosymbiosis such as the widespread and important genus Wolbachia , and also intracellular pathogenicity . Actinobacteria are important for the decomposition of organic matter from soils , making a range of nutrients available to other organisms, which probably explains the importance of this group in the co-occurrence network. Planctomycetes, a phylum of mostly aquatic free-living bacteria, but also found in soils , were also rich in our samples, with their distribution linked to soil properties such as soil organic matter, Ca2+, and pH . Interestingly, even though Planctomycetes are associated with aquatic environments, they were not the richest in seasonally flooded forests (Fig. 4).
The co-occurrence network for all samples is densely connected, with low modularity and without differentiation explained by the habitat types, since most OTUs were classified as habitat generalists. If habitat type was the strongest variable to explain co-occurrence, the network should present modules, since groups belonging to a particular module should have similar environmental preferences [69, 104]. For instance, through analysis of deforestation along an Amazonian transect, it was possible to differentiate modules more associated with different degrees of deforestation . The lack of modularity in the network for all samples may be due to the majority of OTUs (> 98% of nodes in all our networks) being classified as habitat generalists (present in three or four habitat types). The specialist OTUs (present in just one habitat) were usually local specialists, i.e., recorded in one or very few plots (Fig. 2), making the co-occurrence and co-exclusion patterns for specialist organisms hard to detect. Finally, even though networks for habitats are different from each other, the resulting pattern when including all samples may not show any apparent modularity, as patterns for each habitat are overlaid, and OTUs may present additional connections with OTUs from other habitats, smoothing or hiding patterns from individual habitats.
The high number of local specialist OTUs is in agreement with general microbial patterns elsewhere [105, 106] and also with distribution patterns of Amazonian tree species . It is also in agreement with the competitive exclusion theory, which postulates that species with low competitive skills should be excluded from the community by highly competitive species . Yet, these “poor” competitors remain in the communities as rare species . In a global-scale study, rare species were found to have the highest positive association with each other, whereas common species had more negative associations . However, that study relied on the number of observations of macro-organism species (plants and animals) and defined rare species in terms of low abundance. Abundance is difficult to quantify in metabarcoding studies due to PCR biases, such as the false negatives . These biases will affect the detection of OTUs with low abundance beyond the intrinsic low detection probability and the stochastic distribution in their habitats due to their small abundance . Yet, we had a relatively highly populated co-exclusion networks, which may be related to a high competition in soil communities, and further supports the exclusion theory in Amazonian microbial soil communities.
As microbial metabolic functions can differ or overlap, indicating metabolic complementarity or redundancy, they can explain either co-occurrence due to mutualism or co-exclusion due to competition . For instance, several bacterial phyla are syntrophic, with complementary ecosystem functionality, such as microbial interactions for anaerobic methane oxidation, thermodynamic degradation, and nutritional exchange . These associations potentially reinforce co-occurrence patterns among these groups and can explain the bacterial dominance in the co-occurrence networks. On the other hand, bacteria are also dominant in the co-exclusion networks, probably due to competition between other bacteria and also with fungi and protists. Examples of antagonist relationships between bacteria and fungi are well established [112, 113]. Most protists recovered in the networks feed on bacteria and some on fungi (or both), with few eukaryotic parasites recorded. Although this pattern is very different from what was found in other Neotropical soils , our results can be explained due to the prevalence of bacteria in our networks.