A total of 1,410 soil samples was collected across different land-use types in French Guiana between 2016 and 2023 as part of various research initiatives and soil sampling campaigns (Fig. 1A, Suppl. material 1: table SS1). Forest habitats included terra firme, seasonally flooded, white-sand, and cloud forests (ONF 2015). Sampling in forests followed a standardized design of 1-ha plots at each site, with 16 sampling points arranged on a 20 × 20 m grid. At each point, composite samples were collected, each consisting of five soil cores (Fig. 1B). The sampling design differed slightly for the Montagne Kotika site (10 plots), where only eight samples were collected within a 0.5-ha plot due to the particular topographical features of the area. In coastal savannas and pastures, the sampling design consisted of smaller plots (0.25 ha), each yielding five composite samples. Urban areas were sampled through a citizen science initiative involving school students, using simple non-composite samples collected in private backyards, schoolyards, parks, or other urban or semi-urban soils.

Soil cores were collected with a trowel cleaned with distilled water before each use. After removing surface litter, the upper 10 cm of soil was sampled either from a single point for simple samples or from five points for composite samples. For composite samples, the five subsamples were sampled as shown in Fig. 1B and then pooled and homogenized. A total of 15 g of each soil sample was transferred into sterile bottles containing 40 g of silica gel for rapid desiccation and storage prior to DNA extraction. The remaining soil was used to measure soil physicochemical properties.

DNA was extracted from 15 g of desiccated soil using a phosphate-based buffer and processed following the protocol of Taberlet et al. (2012, 2018). This approach was chosen over commercial kits due to its reduced cost, ease of implementation, and ability to work on larger soil volumes while preserving the overall signal retrieved (Zinger et al. 2016). In summary, after mixing the soil with 15 mL of phosphate buffer (Na₂HPO₄; 0.12 M; pH 8), 2 mL of the suspension was loaded onto NucleoSpin Soil columns (Macherey-Nagel kit, 740780.50) and processed following the manufacturer’s protocol. Negative extraction controls, consisting of 15 mL of phosphate buffer only, were included in each DNA extraction batch.

Polymerase chain reaction (PCR) amplifications were performed for four DNA markers targeting bacteria (16S rRNA gene), eukaryotes (18S rRNA gene), fungi (ITS1 region), and insects (mitochondrial 16S rRNA gene) (Table 1). Each PCR reaction was carried out in a total volume of 20 μL, containing 10 μL of AmpliTaq Gold Master Mix (Life Technologies, Carlsbad, CA, USA), 5.84 μL of nuclease-free water (Thermo Fisher Scientific, Waltham, MA, USA), 0.25 μM of each primer, 3.2 μg of BSA (Roche Diagnostics, Basel, Switzerland), and 2 μL of DNA template, which was diluted 10-fold prior to PCR to reduce the effects of inhibitors. Thermocycling conditions for each DNA marker are described in Table 1.

For forest sites, one negative extraction control per site and one negative PCR control per plate were amplified and sequenced alongside the samples. For urban areas, savannas, and pastures, negative controls, including extraction and PCR controls, were randomly assigned on the plates. Positive controls consisted of mock communities of plant and fungal DNA; no mock communities were included for bacteria or eukaryotes, and these controls were used to guide data curation. Two PCR replicates were performed for each forest sample and control, whereas only one replicate was performed for soils collected in urban areas, pastures, and savannas. Consequently, cautious cross-habitat comparisons using this dataset are recommended, and future studies aiming to harmonize replication across these habitats should add technical replicates or internal standards. Nevertheless, the consistently robust PCR yields obtained across these nonstandardized samples serve as reliable proxies for both high initial DNA copy numbers and strong overall sample amplifiability, minimizing the likelihood of widespread false negatives.

Amplifications employed a double-indexing strategy using a set of 32 × 36 octamer tags, with at least five differences between tags, attached to the 5’ ends of each primer (Coissac et al. 2012). This allowed each PCR product to be labeled with a unique combination of forward and reverse tags for sample identification. A total of 10 wells per PCR plate were left empty, serving as sequencing controls and unused tag combinations for downstream data curation.

PCR products were pooled, and sequencing libraries were prepared using the Illumina TruSeq Nano PCR-Free kit according to the manufacturer’s instructions, except that the ligation product was not PCR-amplified to reduce tag-jump biases (Schnell et al. 2015). Libraries were sequenced on different Illumina platforms depending on the marker: HiSeq for insects and MiSeq paired-end for the other markers (Genoscope, Toulouse, France, and San Diego, CA, USA).

Raw reads from each library were processed individually using a common bioinformatics pipeline. First, demultiplexing was performed with cutadapt v4.4 (Martin 2011). To allow sorting of slightly truncated reads, an overlap (-o, --overlap) of two nucleotides shorter than the length of the shortest tag-primer combination was allowed, and the error rate (-e, --error-rate) was set to 0.1. Reads were retained without truncation (--action retain). Because raw reads were in mixed orientation within the FASTQ files, demultiplexing was performed twice, and reads were later concatenated into per-sample FASTQ files.

Subsequent steps were carried out using the APSCALE pipeline v1.7.1 (Buchner et al. 2022), which incorporates VSEARCH (Rognes et al. 2016), cutadapt (Martin 2011), and LULU (Frøslev et al. 2017). Paired-end reads were assembled with a minimum overlap of 10 nucleotides and a maximum read difference of 25% over 199 nucleotides. Adapters were trimmed, and reads were filtered based on expected minimum and maximum lengths (±10 nucleotides), with a maximum error rate of 1. Reads were then dereplicated, and unique sequences with fewer than four reads were excluded. The remaining unique sequences were clustered at 97% identity. The LULU algorithm was applied to detect and exclude erroneous OTU occurrences using the default parameters, i.e., a minimum similarity of 84% and a minimum relative occurrence of 95%.

Final metabarcoding data quality control and additional curation were performed with the metabaR package (Zinger et al. 2021) to exclude contamination or degraded OTUs, flag failed PCRs, and reduce tag-jump bias. Metrics regarding the number of samples and controls, and read counts at the initial and final stages of the bioinformatics pipeline, are summarized in Suppl. material 1: table S2.

Read standardization was performed using the scaling with ranked subsampling (SRS) method from the R package SRS (Beule and Karlovsky 2020) to correct for differences in sequencing depth among samples. SRS operates directly on read counts and aims to equalize sampling effort while preserving the original community structure. It combines scaling to a common minimum sequencing depth (Cmin) with ranked fractional subsampling (Cfrac), thereby reducing the loss of information and compositional distortion associated with classical rarefaction approaches. The cleaning and standardization steps led to the loss of some samples for each marker, such as those containing a high proportion of contamination or an insufficient number of reads.

To combine data across different sequencing runs, OTU representative sequences were reclustered using SUMACLUST (Mercier et al. 2013) at 97% similarity (Bonin et al. 2023). Taxonomic assignment was performed on representative OTU sequences using marker-specific approaches (Table 1). For ribosomal universal markers (Bacteria and Eukaryota), the sequences were compared against SILVA version 138.1 with the SILVAngs pipeline (Quast et al. 2013). Insect OTUs were assigned using the ecotag program from OBITools (Boyer et al. 2016) with marker-specific reference databases built using the ecoPCR program (Ficetola et al. 2010) from the EMBL database v136 and a local database that included published records (Kocher et al. 2017; Talaga et al. 2017; Rongier et al. 2023) and newly generated sequences for the remaining taxa (currently in preparation). Fungal OTUs were assigned using DNAbarcoder (Vu et al. 2019) with a curated UNITE ITS1 database and a local database (Jaouen et al. 2019). When multiple databases were used for a given marker, the final assignment was selected using the taxodecider function from the metabaR R package. All OTUs corresponding to taxa not targeted by each marker were removed from the final dataset. Additionally, other groups, such as marine or aquatic taxa, were removed, as these assignments may result from taxonomic misassignment or incomplete reference databases, as well as occasional non-soil DNA inputs (e.g., transported DNA from animal excrements or watercourses).

Functional assignments were performed separately for each marker using different tools, databases, literature sources, and expert knowledge. To facilitate the functional classification of taxonomically annotated OTUs, the Soil Food Web Ontology (SFWO) (Le Guillarme et al. 2023) was used as a global framework for reconciling taxonomic and semantic heterogeneities among data sources.

For bacteria, the retrieved taxa were assigned a functional guild primarily using the Bacteria and Archaea Traits database (Madin et al. 2020), as well as FAPROTAX (Louca et al. 2016). Additionally, some generalizations were applied: all unclassified Cyanobacteriota, Chloroflexota, and Chlorobiota were grouped as phototrophs, while all ChlamydiaeOTUs were classified as zooparasites unless otherwise specified. Potential pathogens, including phytoparasites, zooparasites, and microbial parasites, were identified at the genus level using the MBPD (Yang et al. 2023) and EID2 (Wardeh et al. 2015) databases. Genera listed as potential zoopathogens in EID2 but absent from MBPD were individually verified, and only those known to cause diseases in humans or animals, excluding commensals or rare taxa, were retained. Phytopathogens were identified using MBPD, and microbial parasites using EID2. The main functional guilds identified were phototrophs, chemoheterotrophs, microbial parasites, phytoparasites, zooparasites, and taxa involved in the iron (Fe), manganese (Mn), nitrogen (N), and sulfur (S) cycles. Guilds related to biogeochemical cycles included taxa specifically associated with each cycle. OTUs that could not be classified into these categories were further assigned as copiotrophs or oligotrophs following Stone et al. (2023) when the information was available.

For eukaryotes, functional assignments were performed for fungi, animals, and protists using taxa retrieved from the different DNA markers (Table 1). Animals, excluding insects, and protists were annotated from the eukaryote marker (Euka02; Table 1), whereas fungi and insects were annotated from their respective group-specific markers (Fung01 and Inse01), which provide higher taxonomic resolution. Consequently, no functional annotations were performed for fungal and insect taxa retrieved from the eukaryotic marker. Plant sequences were not functionally classified due to insufficient taxonomic resolution.

Protists were defined as all eukaryotic taxa not belonging to the Animalia (Metazoa), Plantae (Spermatophyta), or Fungi kingdoms. Functional annotations were based on trophic information from Adl et al. (2019) and Singer et al. (2021). Protists were assigned to the following functional groups: microbivores, osmotrophs, phytoparasites, phototrophs, saprotrophs, zooparasites, and zoosymbionts. When possible, microbivores were further divided into bacterivores and eukaryvores.

For animals, functional information, i.e., trophic preferences, was derived primarily from Potapov et al. (2022) and Brown et al. (2024) at the phylum, order, or class level. Animal taxa were assigned to the following functional groups: bacterivore, detritivore, fungivore, microbivore, omnivore, phytophage, predator, and zooparasite.

For fungal taxa, a list of ectomycorrhizal fungal genera from the Neotropics curated by specialists was used to assign OTUs to the ectomycorrhizal guild (Romero-Gutiérrez et al. n.d.). The remaining OTUs annotated at the genus level were assigned to guilds using the FungalTraits database (Põlme et al. 2020), based on their primary lifestyle. However, when the primary lifestyle did not correspond to ecological roles from the soil compartment (e.g., epiphytes living on leaf surfaces), but the secondary lifestyle did, the secondary lifestyle was used for classification. OTUs identified at higher taxonomic ranks were classified using FUNGuild (Nguyen et al. 2016), retaining only guilds corresponding to the categories above. All Glomeromycetes were classified as arbuscular mycorrhizal fungi. Fungal taxa were assigned to the following functional groups: saprotrophs (soil, wood, litter, or other), mycorrhizal (arbuscular or ectomycorrhizal), plant pathogens, mycoparasites, animal parasites, lichen parasites, lichenized fungi, endophytes, and epiphytes. Other guilds, which were represented by fewer than five OTUs, were grouped into an “Other” category, including unspecified pathotrophs, unspecified symbiotrophs, moss symbionts, and algal parasites.

Functional assignments of insect taxa were based on several literature sources, online databases such as the Global Ants Database, and expert knowledge, prioritizing Neotropical information. OTUs annotated at the family level were assigned using trophic data from Rainford and Mayhew (2015). For OTUs assigned at finer taxonomic resolution (e.g., subfamily, genus, species), additional references were consulted—particularly for ants and termites, where Rainford and Mayhew (2015) provided limited information. The main trophic guilds identified for insects were detritivores, fungivores, nectarivores, omnivores, parasites, phytophages, predators, and zooparasites. Termites, classified as detritivores, were further divided into soil-feeding and wood-feeding groups when possible.

The dataset contains associated metadata consisting of soil properties, climate, and other topographic variables. Soil physicochemical properties were measured on soil samples at the plot level or at the sample level for urban soil samples. For urban soil samples, these physicochemical analyses were conducted only on a subset of samples (59/233 samples; Suppl. material 1: table SS1). The measured properties were granulometry, i.e., the proportion of sand, clay, and silt, pH, total nitrogen (TN; in mg/kg), organic carbon content (OC; in mg/kg), available phosphorus (Olsen P; in mg/kg), and the carbon-to-nitrogen ratio (C:N). These analyses were conducted at the PAPC platform (Toulouse, France; https://chemicalanalysis.cnrs.fr/) and at the Laboratoire d’Analyse des Sols (Arras, France; https://las.hautsdefrance.hub.inrae.fr/).

The remaining variables were extracted from environmental layers in free-access, remote-sensing-derived environmental databases that provide information on geology, topography, bioclimates, and vegetation in French Guiana. This information was retrieved at the plot level or at the sample level for urban soil samples. Elevation data were obtained from Shuttle Radar Topography Mission digital elevation models (Farr et al. 2007). The 30 m resolution tiles available at www.guyane-sig.fr were obtained and aggregated to cover French Guiana. Based on elevation data, slope, aspect, topographic wetness index, and topographic convergence index were calculated using QGIS (QGIS Development Team 2023).

Bioclimatic variables were obtained from CHELSA, which provides downscaled model output temperature and precipitation estimates at fine resolution (30 arc seconds) (Karger et al. 2017). The 19 bioclimatic variables derived from these estimates were extracted, representing annual trends, seasonality, and extremes of climatic variables (available at https://chelsa-climate.org/bioclim/).

Sampling completeness was evaluated using sample-based accumulation curves. For each DNA marker, OTU accumulation curves were computed from 200 random permutations of sample order, providing mean diversity estimates and 95% confidence intervals. To estimate regional diversity, curves were fitted with common nonlinear accumulation models, i.e., Monod and negative exponential models. Models were compared using Akaike’s Information Criterion (AIC), and the asymptote of the best-fitting model was used as the estimate of regional diversity.

To evaluate coverage across environmental gradients in French Guiana, a Multivariate Environmental Similarity Surface (MESS; Elith et al. 2010) was computed using the climatic and topographic layers described above with the R package dismo (Hijmans et al. 2010). The layers were used at 100 m resolution when available for the analysis. The MESS analysis quantifies, for each pixel in the study area, how similar its environmental conditions are to those observed at the sampled sites. Positive values indicate that the pixel falls within the environmental range represented in the samples, whereas negative values highlight areas where one or more variables exceed the sampled range, revealing regions of environmental extrapolation beyond the dataset.