Abstract
Microbial interaction networks support many ecosystem services, including the regulation of crop diseases. One of the current challenges is automatically reconstructing these networks from metabarcoding data and monitoring their responses to environmental change. Here, we evaluated the ability of network inference methods to detect changes in crop-associated microbial networks. We used grapevine as a model plant system and assessed the impact of vineyard management (conventional versus organic) on the alpha- and beta-diversity of the fungal communities of grapevine leaves. We also inferred replicated networks of fungal associations, to compare network alpha- and beta-properties between management systems and to generate hypotheses concerning fungus-fungus interactions. We found that the richness, diversity and evenness of fungal communities were significantly higher in organic plots, and that community composition differed between management systems. Erysiphe necator, the causal agent of grapevine powdery mildew, was significantly more abundant in conventional plots, consistent with visual records of disease symptoms, whereas several yeast species were significantly more abundant in organic plots. Vineyard management also had a significant impact on the beta-properties of fungal association networks, but the high turnover of associations between plots precluded the generation of robust hypotheses concerning interactions between fungal taxa, casting doubts on the relationship between microbial association networks and plant health. Network inference methods therefore require improvement and validation before use in the next-generation biomonitoring of disease control services provided by the crop microbiota. As things stand, community-level data appear to be a more reliable and statistically powerful option than network-level data for monitoring the ecosystem services provided by the plant microbiota.
1 Introduction
Interactions between organisms and between organisms and their abiotic environment regulate the ecological processes underlying ecosystem services (Mace et al., 2012). Ecological interactions (e.g. predation, mutualism, parasitism) at a single point in space and time are usually represented as a network, with the organisms as nodes and the interactions as links (Pocock et al., 2012). Current challenges focus on understanding how and why these networks vary in space and time (Pilosof et al., 2017; Pellissier et al., 2018), and which network properties should be conserved or enhanced to sustain ecosystem services (Tylianakis et al., 2010; Montoya et al., 2012; Raimundo et al., 2018).
Global environmental changes due to human activities alter ecological networks by shifting species distributions and disrupting the interactions between species (Scheffers et al., 2016; Pecl et al., 2017). Agriculture, a key driver of global change (Tilman et al., 2002), has for instance an impact on both ecological communities (Tuck et al., 2014; Seufert and Ramankutty 2017) and ecological networks (Tylianakis et al., 2007; Macfadyen et al., 2009; Ma et al., 2019). Theoretical frameworks have been developed for the comparison of ecological networks between contrasting environmental conditions or along environmental gradients (Poisot et al., 2012; Tylianakis and Morris 2017; Pellissier et al., 2018; Delmas et al., 2019). By analogy with the α- and β-diversity of ecological communities, these frameworks define α- and β-properties for ecological networks, corresponding to whole-network metrics (e.g. connectance) and dissimilarities between pairs of networks, respectively (Pellissier et al., 2018).
Network ecology originates from the study of trophic links between macro-organisms (Ings et al., 2009) and initially ignored smaller organisms (Lafferty et al., 2006), but the importance of microorganisms is now recognized, with microbial interactions supplying ecosystem services such as biogeochemical cycling, provisioning services or disease regulation (Falkowski et al., 2008; van der Heijden et al., 2008; Berendsen et al., 2012). Microbial networks are affected by environmental changes (Creamer et al., 2016; Morriën et al., 2017), and it has been suggested that their properties could be used as potential bioindicators of environmental quality and ecosystem functioning (Karimi et al., 2017). The regulation of disease development in holobionts, defined as individual animals or plants together with all their associated microorganisms (Zilber-Rosenberg and Rosenberg 2008), is mediated by microbial interactions, including direct antagonistic interactions between the microbiota and pathogen species (Arnold et al., 2003; Koch and Schmid-Hempel, 2011; Kamada et al., 2013; Kemen 2014; Laur et al., 2018) and indirect interactions involving activation of the host immune system by the microbiota (Perazzolli et al., 2012; Kamada et al., 2013; Ritpitakphong et al., 2016; Vogel et al., 2016; Hacquard et al., 2017). The subset of a host-associated microbial network consisting of a pathogen and its interacting partners is known as the pathobiome (Vayssier-Taussat et al., 2014; Brader et al., 2017). The biomonitoring of holobiont health requires elucidation of the microbial interactions forming pathobiomes (Durán et al., 2018), identification of the network-level indicators of host disease susceptibility (Agler et al., 2016, Poudel et al., 2016) and monitoring of the impact of anthropogenic environmental change on these indicators. A combination of next-generation sequencing and machine-learning is required to achieve these objectives (Bohan et al., 2017; Derocles et al., 2018) and this study aimed to evaluate some of the methods developed to date.
Ecological networks have long been built on the basis of observations of ecological interactions. Meta-barcoding approaches are now providing additional information (Evans et al., 2016), and making it possible to generate hypotheses about interactions between organisms that are difficult to observe, such as microorganisms (Faust and Raes 2012; Berry and Widder 2014). Hypotheses concerning microbial interactions can be formulated on the basis of co-abundance data derived from environmental DNA metabarcoding data (i.e. the sample × taxa matrix) via statistical or machine-learning approaches (Vacher et al., 2016). The representation of all positive and negative statistical associations between abundances for microbial taxa generally gives rise to a network resembling a hairball (Röttjers and Faust 2018). The challenge is to go beyond networks of this type, by removing spurious associations not due to microbial interactions and linking the remaining positive and negative associations to possible interaction mechanisms (Derocles et al., 2018). In a metabarcoding dataset, the total number of sequences per sample is arbitrary, imposed by the sequencer. Sequence counts contain only relative abundance information for species, and comparisons not taking this feature, known as compositionality, into account can result in the identification of artifactual associations (Gloor et al., 2017). Early methods of microbial network inference, such as SparCC (Friedman and Alm 2012), made use of log ratios of counts to overcome this bias. The second cause of spurious associations is the joint response of microbial taxa to abiotic or biotic factors, creating indirect associations reflecting taxon requirements, particularly for samples collected in heterogeneous environments (Röttjers and Faust 2018). Early studies dealt with this issue by first using regression to eliminate the environmental factors from the sequence counts and then inferring networks from the residuals (Jaskuschkin et al., 2016; Biswas et al., 2016). Several methods of network inference integrating environmental covariates, such as HMSC (Ovaskainen et al., 2017), PLN (Chiquet et al., 2018), FlashWeave (Tackmann et al., 2018) and MAGMA (Cougoul et al., 2018), have since been developed. Last, but not least, the taxonomic resolution of the nodes must be fine enough to capture variations in ecological interactions between microbial strains (Röttjers and Faust 2018). Novel bioinformatics approaches fully exploiting the resolution of molecular barcodes, such as DADA2 (Callahan et al., 2017), therefore seem highly appropriate. However, despite these advances, the inference of microbial networks from metabarcoding data is still in its infancy (Layeghifard et al., 2017), and inferred interactions should be interpreted with caution (Weiss et al., 2016; Röttjers and Faust 2018; Freilich et al., 2018; Barner et al., 2018; Zurell et al., 2018), because only a few validation experiments to date have been successful (Das et al., 2018; Wang et al., 2017).
In this study, we investigated whether changes in crop-associated microbial networks could be detected by combining current metabarcoding and network inference approaches. The monitoring of these changes is important, because these microbial networks possess intrinsic properties that hinder their invasion by pathogens (Murall et al., 2017) and form a barrier to disease development (Brader et al., 2017; Hacquard et al., 2017). Our objectives were (1) to assess the influence of crop management (conventional versus organic agriculture) on network α- and β-properties, (2) to investigate whether network-level properties are more sensitive to change than community-level properties and (3) to determine whether microbial networks can generate robust hypotheses concerning microbial interactions. We inferred microbial networks from multiple samples collected in homogeneous, replicated agricultural plots, using DADA2 (Callahan et al., 2016) and SparCC (Friedman and Alm 2012). We used grapevine as a model plant species and focused on the fungal component of its foliar microbiota, which contains several major pathogens.
2 Materials and methods
2.1 Study site and sampling design
Samples were collected in 2015, from an experimental vineyard (Fig. 1) located near Bordeaux (INRA, Villenave d’Ornon, France; 44°47’32.2”N 0°34’36.9”W). The experimental vineyard was planted in 2011 and was designed to compare three cropping systems: sustainable conventional agriculture (CONV), organic farming (ORGA) and pesticide-free farming (RESI) (Delière et al., 2014). The latter used a disease-resistant cultivar and was therefore not included in this study. The experiment had a randomized block design (Schielzeth and Nakagawa 2013) consisting of three blocks, each composed of three plots, one for each of the cropping systems tested. Each plot covered an area of 2100 m2 and was composed of 20 rows of 68 vines each, with a 1.60 m between rows and 0.95 m between vines in a single row.
The same cultivar – Vitis vinifera L. cv. Merlot noir grafted onto a 3309 C rootstock – was used in both the CONV and ORGA cropping systems. CONV plots were managed according to the general principles of integrated pest management (IPM), as listed in Appendix III of the 2009/128/EC Directive (European Commission 2009). ORGA plots were managed according to European Council Regulation (EC) No 834/2007 (European Commission 2007). ORGA plots were treated with copper and sulfur-based products, whereas additional phytosanitary products were allowed in CONV plots (Table S1). The cropping systems differed in terms of the types of pesticides applied and the timing of applications, but not in terms of doses (Table S1). All products and active ingredients were applied between the end of April and mid-August of 2015. Grapes were harvested on September 10, 2015. The incidence and severity of disease at harvest were higher in CONV plots than in ORGA plots for both powdery mildew (caused by the fungal pathogen Erysiphe necator) and black rot (caused by the fungal pathogen Guignardia bidwellii). Downy mildew symptoms (caused by the oomycete pathogen Plasmopara viticola) did not differ significantly between cropping systems (Table S2).
Grapevine leaves were collected a couple hours before grape harvest, from 20 vines per plot in the CONV and ORGA plots. Edge effects were avoided by selecting the 20 vines from the center of each plot. The third leaf above the grapes was collected from each vine, placed in an individual bag and immediately transported to the laboratory. In total, 120 leaves, corresponding to 1 leaf × 20 vines × 3 plots × 2 cropping systems, were therefore collected. Leaves were processed on the day of collection, with sterilized tools in the sterile field of a MICROBIO electric burner (MSEI, France). Three contiguous discs, each 6 mm in diameter, were cut from the center of each leaf, approximately 2 cm from the midrib. They were placed in the well of a sterile DNA extraction plate. The leaf disks were then freeze-dried overnight (Alpha 1-4 DA Plus, Bioblock Scientific).
2.2 DNA extraction and sequencing
Leaf disks (Fig. 1) were ground with a single-glass ball mill (TissueLyser II, Qiagen) and DNA was then extracted with a CTAB chloroform/isoamyl alcohol (24:1) protocol. A dozen “empty” wells (i.e. containing nothing but extraction reagents) were included on each plate as negative control samples for DNA extraction. Three of these negative control samples were randomly selected and pooled before sequencing. Three replicates of a fungal mock community, each consisting of an equimolar pool of DNA from 189 pure fungal strains, were also included as positive control samples (Pauvert et al., 2019).
The nuclear ribosomal internal transcribed spacer (ITS) region, which is considered to be the universal barcode region for fungi (Schoch et al., 2012), was then amplified with the ITS1F (5’-CTTGGTCATTTAGAGGAAGTAA-3’, Gardes and Bruns 1993) and ITS2 (5’-GCTGCGTTCTTCATCGATGC-3’, White et al., 1990) primer pair, which targets the ITS1 region. PCR was performed in an Eppendorf thermocycler (Eppendorf), with a reaction mixture (25 μl final volume) consisting of 0.04 U Taq polymerase (SilverStar DNA polymerase, Eurogentec), 1X buffer, 2 mM MgCl2, 200 μM of each dNTP, 0.2 μM of each primer, 1 ng.μl-1 bovine serum albumin (New England BioLabs) and 2 μl DNA template. A pseudo-nested PCR protocol was used, with the following cycling parameters: enzyme activation at 95°C for 2 min; 20 (1st PCR with regular primers; Table S3) and then 15 (2nd nested PCR with pre-tagged primers; Table S3) cycles of denaturation at 95°C for 30 s, 53°C for 30 s, 72°C for 45 s; and a final extension phase at 72°C for 10 min. “Empty” wells (i.e. containing nothing but PCR reagents) were included on each plate as a negative control for PCR. Three negative control samples were randomly selected and pooled before sequencing. In addition, the six PCR products corresponding to the 24th leaf of the six plots under study (three CONV plots, three ORGA plots) were split in two, with each half of the sample sequenced independently to serve as technical replicates for sequencing.
We checked the quality of all the PCR products by electrophoresis in 2% agarose gels. PCR products were purified (CleanPCR, MokaScience), multiplex identifiers and sequencing adapters were added, and library sequencing on an Illumina MiSeq platform (v3 chemistry, 2×250 bp) and sequence demultiplexing (with exact index search) were performed at the Get-PlaGe sequencing facility (Toulouse, France).
2.3 Bioinformatic analysis
Based on the mock community included in the sequencing run, we found that analyzing single forward (R1) sequences with DADA2 (Callahan et al., 2016) was a good option for fungal community characterization (Pauvert et al., 2019). Using DADA2 v1.6, we retained only R1 reads with less than one expected error (based on quality scores; Edgar and Flyvbjerg, 2015) that were longer than 100 bp, and we then inferred amplicon sequence variants (ASV) for each sample. Chimeric sequences were identified by the consensus method of the removeBimeras function. Taxonomic assignments were performed with RDP classifier (Wang et al., 2007), implemented in DADA2 and trained with the UNITE database v. 7.2 (UNITE Community 2017). Only ASVs assigned to a fungal phylum were retained. The ASV table was then filtered as described by Galan et al., (2016) with a homemade script (https://gist.github.com/cpauvert/1ba6a97b01ea6cde4398a8d531fa62f9). The script removed ASVs from all samples for which the number of sequences was below the cross-contamination threshold, defined as their maximum number in negative control samples (Galan et al., 2016). Finally, we checked the compositional similarity of the technical replicates (Fig. S1), retaining the sample with the largest number of sequences for each replicate. The final ASV table contained 1116 ASVs, 112 samples and 4,760,068 high-quality sequences (Table S4).
2.4 Statistical analyses
Statistical analyses were performed with R software v3.4.1 (R Core Team 2017), with the packages lme4 (Bates et al., 2015), vegan (Oksanen et al., 2018), permute (Simpson 2016), phyloseq (McMurdie and Holmes 2013) including the DESeq2 extension (Love et al., 2014), and igraph (Csardi and Nepusz 2006). Data were manipulated and plots were created with reshape2, plyr and ggplot2 (Wickham 2007; 2011; 2016), cowplot (Wilke 2018), ggraph (Pedersen 2018) and VennDiagram (Chen 2018).
2.4.1 Effect of cropping system on community α-diversity
Generalized linear mixed models (GLMMs) were used to test the effect of cropping system on the richness, diversity and evenness of fungal communities. The models included the cropping system as a fixed treatment effect, the block and its interaction with the cropping system as random factors, and the sampling depth (defined as the total number of raw sequences per sample) as an offset (Bálint et al., 2015; McMurdie and Holmes 2014). Community richness was defined as the number of ASVs per sample. We used a logarithmic link function to model these count data, assuming a negative binomial distribution to deal with overdispersion (Zuur et al., 2009). Community diversity was measured with the Inverse Simpson index (Simpson 1949) and modeled with a Gaussian distribution and the logarithmic link function. Evenness was estimated with Pielou’s index (Pielou 1966) and modeled with a Gaussian distribution and the logarithmic link function. The offset was transformed according to the link function. The significance of the fixed treatment effect was finally assessed with the Wald χ2 test (Bolker et al., 2009).
2.4.2 Effect of cropping system on community β-diversity
Permutational analyses of variance (PERMANOVAs; Anderson 2001) were used to evaluate the effect of cropping system on compositional dissimilarities between fungal communities detected with the quantitative and binary versions of the Jaccard dissimilarity index (Chao et al., 2006; Jaccard 1900). The models included cropping system, sampling depth (log-transformed) and their interaction as fixed effects. Permutations (n = 999) were constrained within blocks. ASVs differing in abundance between cropping systems were identified with DESeq2 (Love et al., 2014), by calculating the likelihood ratio between a full model including block and cropping system as fixed effects and a simplified model including only the block factor. The estimated fold-changes in abundance were considered significant if the p-value was below 0.05 after Benjamini and Hochberg adjustment.
2.4.3 Effect of cropping system on network α-properties
Fungal association networks were inferred at plot level (Fig. 1) with the SparCC algorithm (Friedman and Alm 2012) implemented in FastSpar (Watts et al., 2019) with default SparCC values. Ten networks per plot were constructed by varying the percentage P of ASVs included in the network (with P ranging from 10% to 100% of the most abundant ASVs in the plot). Networks had ASVs as nodes and a positive or negative link between ASVs in cases of significant associations between abundance. Six α- properties were calculated for all networks: number of links, network density, number of connected components, diameter of the largest component, mean node degree and proportion of negative links (Table S5). The effect of cropping system on these properties was investigated by performing Wilcoxon rank-sum tests for every value of P. The Benjamini-Hochberg procedure was used to correct p-values for multiple testing.
2.4.4 Effect of cropping system on network β-properties
The topological distance between networks was calculated for all pairs of networks, with the D index defined by Schieber et al., (2017) for binary networks. The dissimilarity of associations between networks, βWN according to the framework described by Poisot et al., (2012), was then calculated for all pairs of networks with the binary Jaccard dissimilarity index. βWN was then partitioned into two components (Poisot et al., 2012): the dissimilarity of associations between ASVs common to both networks (βOS) and the dissimilarity of associations due to species turnover (βST). PERMANOVA was used to evaluate the effect of cropping system on the topological distance between networks (D) and the dissimilarity of associations between networks (βWN, βST and βOS). The models included cropping system, P and their interactions as fixed effects. The permutations (n=999) were constrained within blocks. Consensus networks were built to identify robust associations that could indicate biotic interactions between fungal strains.
3 Results
The foliar fungal communities were dominated by Ascomycota in both ORGA (87.2% of sequences) and CONV (96.8%) plots. They were largely colonized by an ASV assigned to the Aureobasidium genus. More than half the sequences belonged to this ASV (Table 1), whatever the cropping system. The causal agent of powdery mildew, Erysiphe necator, was among the 10 most abundant fungal species. The proportion of sequences assigned to this pathogen species was higher in CONV than in ORGA plots (Table 1), consistent with the visual records of disease symptoms (Table S2).
3.1 Effect of cropping system on community α- and β-diversity
Fungal community richness, diversity and evenness were significantly higher in ORGA than CONV plots (Wald χ2 = 4.74, p = 0.029; Wald χ2 = 8.28, p = 0.004; Wald χ2 = 12.88,p < 0.001, respectively; Fig. 2). The composition of foliar fungal communities differed significantly between cropping systems (Table 2), in terms of both the relative abundance (Fig. 3A) and presence-absence of ASVs (Fig. 3B). DESeq2 analysis revealed that four ASVs, including the fungal pathogen Erysiphe necator, were significantly more abundant in CONV plots, whereas 10 other ASVs, including several yeast species (from the genera Vishniacozyma, Sporobolomyces and Filobasidium), were significantly more abundant in ORGA plots (Fig. 3C). Principal coordinate analysis (PCoA) with the Jaccard quantitative index (Fig. 3A) revealed large differences in the relative abundances of ASVs between samples within a cropping system. The first axis of the PCoA accounted for 32.5% of the variance in community composition but did not discriminate between cropping systems. It was significantly correlated with the relative abundance of the dominant ASV (assigned to the Aureobasidium genus) (Spearman ρ = 0.96; p < 0.001).
3.2 Effect of cropping system on network α- and β-properties
In total, we obtained 60 fungal association networks, corresponding to six plots (Fig. 1) per P category (Fig. 4A). None of the six network α-properties (Table S5) differed between cropping systems (Table S6), but all were significantly correlated with the percentage P of ASVs included in the network (Table S7). The total number of links and the mean degree increased with P, whereas the number of connected components, network diameter, connectance and the proportion of negative links decreased (Fig. 4B; Table S7). Similarly, the topological distance between networks did not differ between cropping systems, but was influenced by P (Table 3). By contrast, cropping system had a significant effect on the overall dissimilarity of associations (βWN) and the dissimilarity of associations between shared ASVs (βOS; Table 3 and Fig. 5). Fungal associations also varied considerably between plots within a cropping system (Fig. 4A and Fig. 6). When all ASVs were used for network construction (P=100%), only two associations were common to all three network replicates of the ORGA system (Fig. 6A; Fig. S2A) and only six were common to all three network replicates of the CONV system (Fig. 6B; Fig. S2B). The two associations replicated in the ORGA system were negative associations between the dominant ASV (assigned to the genus Aureobasidium) and Cladosporium ramotenellum, and between Vishniacosyma victoriae and Neofusicoccum parvum (Fig. S2A). Five of the six associations replicated in the CONV system were positive, the remaining negative association being that between the dominant ASV (assigned to the genus Aureobasidium) and Epicoccum nigrum (Fig. S2B). No association common to all six networks was identified (Fig. 6C).
4 Discussion
Plant-associated microbial interaction networks protect plants against disease (Kemen 2014; Hassani et al., 2018). There is, therefore, a crucial need to monitor their responses to drivers of environmental change, such as agricultural practices. Here, we assessed the effects of cropping system (organic versus conventional) on plant-associated microbial communities and networks, using grapevine as a model plant. We focused on the fungal component of the grapevine foliar microbiota because it contains several pathogen species. We analyzed both microbial community α- and β-diversities and microbial network α- and β-properties, to determine whether network-level metrics were better indicators of environmental change than community-level metrics (Karimi et al., 2017, Ma et al., 2018). We found that the α-properties of microbial networks (i.e. whole-network metrics; Pellissier et al., 2018) did not differ between cropping systems. Their variations were correlated with the percentages of the most abundant taxa included in the network, which is a network inference methodological parameter. By contrast, some β-properties of microbial networks differed significantly between cropping systems, revealing a difference in microbial associations between organic and conventional vineyards. This difference remained significant when network pairwise comparisons were based on shared taxa only, suggesting that the differences between organic and conventional networks were due to re-associations of fungal taxa rather than a turnover of taxa. As expected from previous results, microbial community α- and β-diversities differed significantly between cropping systems. Our results therefore indicate that both community-level and network-level metrics can be used to monitor changes in the crop microbiota. Using our findings, we tried to determine which metrics were most likely to contain information about plant health.
Community-level analyses showed that the richness, diversity and evenness of the grapevine foliar microbiota were significantly higher in organic than conventional vineyards, consistent with the recent findings of Kernaghan et al. (2017) (but see Castañeda et al., 2018). The cropping system also significantly affected the composition of grapevine foliar fungal communities, as reported in previous studies (Schmid et al., 2011; Pancher et al., 2012; Varanda et al., 2016; Kernaghan et al., 2017; Castaneda et al., 2018). Several fungal taxa differed significantly in abundance between cropping systems. For instance, Erysiphe necator, the causal agent of grapevine powdery mildew, was significantly more abundant in conventional than in organic plots. These results are consistent with visual assessments of disease symptoms, indicating that, despite their numerous biases, metabarcoding data do contain some quantitative information useful for monitoring plant disease development (Sapkota et al., 2015; Jakuschkin et al., 2016; Makiola et al., 2018). Several yeast strains, assigned to the genera Vishniacozyma, Sporobolomyces and Filobasidium, were significantly more abundant in organic plots. These yeast genera are frequently detected on leaf surfaces due to their tolerance of irradiation and they might influence plant growth by producing plant hormone-like metabolites (Kemler et al., 2017).
Network-level comparisons revealed an unexpectedly high degree of variability between network replicates within each cropping system. Replicate microbial networks for the same cropping system had very few associations in common, and no association was consistent found to be common to both cropping systems. Only two associations were common to all three replicates of the organic system. Both were negative and involved two plant pathogens — Neofusicoccum parvum (Bruez et al., 2014) and Cladosporium ramotenellum (Swett et al., 2016) — and two other species known to have antagonistic effects on plant pathogens: Vishniacozyma victoriae and Aureobasidium sp (Pertot et al., 2017; Gramisci et al., 2018). These hypothetical fungus-fungus interactions appear to be relevant but require experimental validation, to demonstrate that networks do contain information about plant health-related mechanisms. Several methodological biases might account for the high variability of fungal associations within a cropping system (i.e. the low replicability of microbial association networks). First, each network was built from 20 samples, whereas a minimum of 25 samples is generally recommended to obtain reliable networks (Berry and Widder 2014). Second, network nodes (i.e. amplicon sequence variants) may group together fungal strains with different interaction traits (McLaren and Callahan, 2018), despite their fine-scale taxonomic resolution (Callahan et al., 2017). Third, microbial associations may be triggered by spatial variations in environmental conditions (i.e. microclimate or leaf traits) rather than by biotic interactions (Derocles et al., 2018; Freilich et al., 2018). However, this third hypothesis is not particularly likely in our experimental system because the vineyard plots were adjacent to each other and planted with grapevine clones. The sampling design also limited microclimate variations, because we collected all leaves in less than two hours and controlled for the position of the sampled leaf on the vine. Alternatively, the associations detected may reflect the real interactions between fungal taxa. The turnover of associations might, in that case, be due to the functional redundancy of fungal taxa. Taxonomically different communities, involving different interactions between members, may have similar functions (Louca et al., 2016) and protect plant health in a similar way.
Based on these findings, we cannot recommend the use of network properties for monitoring the disease regulation services provided by the microbiota. The replicated microbial association networks inferred from metabarcoding data were highly variable within each set of environmental conditions and generated few robust hypotheses concerning interactions between fungi, precluding their use to monitor the barrier effect of microbial interactions against pathogens. By contrast, community-level metrics revealed clear-cut changes in the plant microbiota in response to environmental change and reflected the disease status of the plant. Moreover, they were more statistically powerful than network- level metrics, because many samples were required to infer each microbial network replicate. Hence, as things stand, machine-learning approaches using community-level microbial metabarcoding data (Cordier et al., 2018) appear to be a more reliable option for the next-generation biomonitoring of ecosystem services than microbial network inference.
5 Conflict of Interest
The authors declare that there is no conflict of interest.
6 Authors’ Contributions
CP performed the bioinformatic and statistical analyses and wrote the first draft of the manuscript. JV and CV performed the sampling and processed the leaf samples. LD managed the sampling site and provided data on phytosanitary treatments and disease symptoms. JV performed the DNA extractions and amplifications. MB coordinated the creation of the fungal mock community and the sequencing of all samples. CV conceived the study, supervised the analyses and made a major contribution to the writing of the final draft. All authors discussed the preliminary version of the results and revised the manuscript.
7 Funding
We thank the INRA MEM metaprogram (Meta-Omics of Microbial Ecosystems) for financial and scientific support. Sequencing was funded by the INRA MEM MetaBAR project and bioinformatic and statistical analyses were performed as part of the INRA MEM Learn-biocontrol project. Additional funding was received from the LABEX COTE (ANR-10-LABX-45), the LABEX CEBA (ANR-10- LABX-25-01) and INRA EcoServ metaprogram on ecosystem services (IBISC project) and the Aquitaine Region (Athene project, n°2016-1R20301-00007218). CP’s PhD grant was funded by the INRA and Bordeaux Sciences Agro (BSA). The management of the experimental site was partly funded by the AFB (French Agency for Biodiversity) within the DEPHY network.
9 Data availability
The raw sequence data were deposited in Dataverse and are available in the FASTQ format at https://doi.org/10.15454/3DPFN. The code is available as an archive at https://doi.org/10.15454/NSHUAQ.
8 Acknowledgments
We thank Dave Bohan, Lucile Muneret, Andreas Makiola and all members of the ANR NGB Consortium (ANR-17-CE32-0011) for helpful scientific exchanges on next-generation biomonitoring and for their comments on the manuscript. We also thank Gregory Gambetta, Guilherme Martins, Frédéric Barraquand, Isabelle Lesur and Adrien Rush for useful comments on preliminary results. We thank the Genotoul sequencing facility (Get-PlaGe) for sequencing and the Genotoul bioinformatics facility (Bioinfo Genotoul) for providing computing and storage resources. We also thank Julie Sappa from Alex Edelman & Associates for English language revision.