Bayesian Pairwise Compositional Lotka–Volterra Modeling Infers Potential Rhizosphere Microbial Suppressors of Ralstonia pseudosolanacearum
Article information
Abstract
The Ralstonia solanacearum species complex (RSSC) is a major soil-borne pathogen of solanaceous crops. During a field experiment originally designed to monitor rhizosphere and episphere microbiomes in two pepper cultivars, a naturally emerging and asymptomatic Ralstonia dominance event was detected in the rhizosphere without visible wilt symptoms. This unexpected occurrence provided an opportunity to characterize asymptomatic RSSC dynamics and their microbial interactions under field conditions. Full-length 16S rRNA amplicon sequencing showed that one ASV (Sq_1) was nearly absent from the episphere but increased sharply in the rhizosphere from week 3 onward, dominating 20–80% of samples during weeks 7–10. Phylogenetic comparison with 93 historical Korean RSSC isolates placed Sq_1 within a 16S-defined lineage corresponding to pepper-associated R. pseudosolanacearum biovars 3 and 4. Sq_1 abundance accounted for a large portion of β-diversity turnover in the rhizosphere. After within-plot correlations were meta-analyzed, selected taxa were evaluated using a Bayesian pairwise compositional Lotka–Volterra (pcLV) model, which identified three taxa (Sq_272, TRA3-20; Sq_178, Bradyrhizobium; and Sq_124, Bryobacter) that consistently exerted inhibitory effects on Sq_1 per-interval growth. Supported by the longitudinal design and the high accuracy of PacBio full-length 16S sequencing, these findings highlight potential microbial suppressors of RSSC and demonstrate the utility of pcLV modeling for resolving directional interactions at the ASV level.
Plant-associated microbiomes assemble in a structured manner that profoundly influences plant health and ecosystem function. These microbiomes comprise highly diverse communities that are not distributed randomly, but instead segregate into distinct aboveground and belowground compartments. Root exudates create nutrient-enriched, chemically buffered microhabitats, known as the rhizosphere. These exudate-driven niches recruit specific taxa, ultimately leading to the formation of characteristic root-associated microbial communities (Berendsen et al., 2012; Bulgarelli et al., 2013). In contrast, the episphere, which comprises plant surfaces, exposes microbial colonizers to completely different challenges, such as nutrient deprivation, desiccation, temperature fluctuations, and direct sunlight. These factors serve as strict filters for colonization, allowing only a subset of stress-tolerant microorganisms to survive (Vorholt, 2012). Evidence from both model plants and field-grown trees has consistently indicated that these environmental contrasts generate clear compartmental differentiation between root- and leaf-associated microbiomes (Bodenhausen et al., 2013; Trivedi et al., 2020). This differentiation reflects the interplay of fundamental ecological filters with host-driven recruitment processes, underlining the structured nature of plant-microbiome assemblies.
Within this ecological context, Ralstonia solanacearum species complex (RSSC) are an important group of soil-borne pathogens that cause bacterial wilt in solanaceous crops such as banana and tobacco. It is well known that this wilt is caused by penetration from the soil, colonizing the xylem and blocking water flow, and this is attributed to manipulation of the host response by the pathogen’s Type III secretion system and Type III effectors. Furthermore, comparative genome analyses demonstrated that the RSSC pan-genome contains a diverse pool of over 100 type III effector genes, with most strains harboring approximately 70 effectors. This diversity facilitates the infection of diverse hosts, and plants co-evolve with the pathogen to detect effectors, thereby preventing invasion (Genin and Denny, 2012; Peeters et al., 2013). Therefore, resistant or tolerant cultivars may harbor symptomless colonization, complicating the detection and management of agriculture (Du et al., 2017; Lebeau et al., 2011). Regulatory and diagnostic frameworks recognize these challenges and emphasize multi-target or confirmatory assays for accurate identification (EFSA PLH Panel et al., 2019; EPPO, 2022; Garcia et al., 2019).
To clarify the terminology used in this study, the wilt-causing strains historically assigned to Pseudomonas were transferred to the new genus Ralstonia in 1995 based on polyphasic evidence (Yabuuchi et al., 1995). Subsequent multilocus and population genetic studies introduced the phylotype scheme (I–IV), which tracks broad biogeography (Fegan and Prior, 2005). Comparative genomics has shown that phylotype IV strains (R. syzygii and the banana blood disease bacterium) form a single genomic species that is distinct from other RSSC members (Remenant et al., 2011). A formal revision proposed a widely adopted split into three species: R. solanacearum sensu stricto (phylotype II), R. pseudosolanacearum (phylotypes I and III), and R. syzygii (phylotype IV with subspecies syzygii, indonesiensis, celebesensis) (Safni et al., 2014).
Many microbiome studies have identified co-occurring taxa using correlation or network analysis; however, such associations conflate direct interactions with shared environmental responses and lack mechanistic interpretability (Röttjers and Faust, 2018; Weiss et al., 2016). To overcome these limitations, dynamic modeling frameworks, such as the generalized Lotka–Volterra (gLV) model (MacArthur, 1970), are increasingly employed to estimate interaction signs and stability properties from time-series data (Bucci et al., 2016; Stein et al., 2013). However, simultaneous fitting of a gLV system across many taxa is often impractical under sparse and irregular sampling conditions, leading to parameter explosion and unstable estimates.
This study aimed to characterize the interactions shaping the rhizosphere and episphere microbiomes of pepper (Capsicum annuum). During the field experiment, we unexpectedly observed a naturally occurring, asymptomatic episode of Ralstonia dominance in the rhizosphere, providing a unique opportunity to investigate RSSC dynamics under field conditions without wilt symptoms. Because such events are rarely documented with high-resolution sequencing, we sought to identify microbial taxa that modulate, suppress, or accompany the rise of this lineage. To achieve this, we analyzed the time-series community data using a Bayesian pairwise compositional Lotka–Volterra (pcLV) modeling framework, an extension of the compositional Lotka–Volterra model (Joseph et al., 2020). This approach enabled the inference of potential facilitative or inhibitory effects on the Ralstonia lineage based on directional, lagged relationships rather than non-temporal correlations.
Materials and Methods
Field experiment and sample collection
Two pepper cultivars, Kaltanpass and Criollo de Morelos 334 (CM334), were cultivated in a 9 × 17 m field (35°49′33.3″N, 127°02′38.5″E) from May 11 to August 21, 2023. Kaltanpass (Nongwoobio, Suwon, Korea) and CM334 cultivars (Mexico) were used in this study. The field was divided into ten plots (five per cultivar), each comprising two parallel raised-bed rows. The samples were collected weekly for ten weeks from each plot. Samples from Kaltanpass were labeled with prefix A, and CM334 with prefix B.
Rhizosphere and leaf episphere samples were collected every Monday at 9:00 AM for ten weeks starting on June 19. One plant of each cultivar assigned to each plot was uprooted. After shaking off soil, roots and leaves were soaked in Phosphate-buffered saline (PBS, pH = 7.0, NaCl 8.0 g, KCl 0.2 g, Na2HPO4 1.44 g, KH2PO4 0.24 g per 1 L) and sonicated in an ice-cooled water bath (JAC-2010, Kodo Technical Research, Korea) for 20 min. The PBS volume was adjusted according to plant growth to ensure complete submersion. After sonication, plant tissues were discarded, and microbial suspensions were pelleted by centrifugation at 10,000 ×g for 20 min (2236R; LaboGene, Allerød, Denmark) and stored at −80°C.
In total, 100 rhizosphere and 60 episphere samples were collected (10 plots × 10 weeks for the rhizosphere and 10 plots × 6 weeks for the episphere). Episphere sampling began at week 5 when sufficient leaf biomass was available, whereas rhizosphere samples were collected throughout the entire period.
DNA extraction and amplicon library preparation
DNA was extracted from rhizosphere (500 mg) and episphere (10–20 mg) pellets using a FastDNA Spin Kit for Soil (MP Biomedicals, Solon, OH, USA). Samples were lysed in sodium phosphate and Matrix Tube buffers using a FastPrep-24 homogenizer at 6 m/s for 40 s, followed by the addition of Protein Precipitation Solution, centrifugation, and DNA binding to the matrix solution. Bound DNA was washed and eluted according to the manufacturer’s instructions.
Genomic DNA was sent to Macrogen (Seoul, Korea) for full-length 16S rRNA (V1–V9) library preparation and sequencing using the PacBio platform. Amplicons were generated using barcoded 27F (5′-AGRGTTYGATYMTGGCTCAG-3′) and 1492R (5′-RGYTACCTTGTTACGACTT-3′) primers under the following PCR conditions: 95°C for 3 min; 25 cycles of 95°C for 30 s, 57°C for 30 s, and 72°C for 60 s; final extension at 72°C for 5 min. PCR products (approximately 1.5 kb) were pooled, converted into SMRTbell libraries, and sequenced. Circular consensus sequence reads were generated and demultiplexed using SMRT Link.
Sequence processing and taxonomic assignment
CCS reads were processed using DADA2 (v1.32.0) in R (v4.5.1). Primers were removed allowing up to two mismatches, and reads were quality-filtered using full-length 16S thresholds (1,400–1,500 bp, minimum quality score ≥2, no ambiguous bases), dereplicated, processed with PacBio-specific error-learning, and denoised into amplicon sequence variants (ASVs). As characteristic of PacBio CCS HiFi data, the retained reads exhibited very high accuracy, with mean Phred scores typically exceeding Q80. Chimeric sequences were removed using the pooled chimera-detection procedure, and taxonomic classification was performed using DECIPHER (v3.2.0) against SILVA (v138.2), followed by species-level assignment. Nonbacterial sequences (chloroplasts and mitochondria) were removed. The sequencing depth and ASV richness per sample were evaluated using rarefaction curves (Supplementary Fig. 1).
Bioinformatic and statistical analyses
β-diversity was assessed using Bray–Curtis dissimilarity and visualized by Principal Coordinates Analysis (PCoA). The statistical significance of community differences was tested using PERMANOVA with 9,999 permutations using selected ASVs as explanatory variables. ASVs of interest were aligned with 16S rRNA reference sequences using ssu-align (v0.1.1), and maximum-likelihood phylogenies were inferred using IQ-TREE 3 (v3.0.1).
To identify candidate taxa associated with Sq_1, within-plot correlations were computed using additive log-ratio (ALR)-transformed abundance. Temporal autocorrelation was accounted for by fitting an autoregressive error structure of order 1 (AR(1)), which reduces the inflation of correlation estimates due to time-series dependence, thereby partially limiting confounding from shared temporal trends. Correlations were subsequently aggregated across plots utilizing a random-effects meta-analysis, specifically the DerSimonian–Laird method. Taxa were included in the analysis if they met the following criteria: they were detected in five or more plots, co-occurred with Sq_1 in at least 25% of samples per plot, exhibited less than 40% co-zeros, and adhered to a false discovery rate threshold (q ≤ 0.10, weighted q-values).
The gLV model (MacArthur, 1970), which describes how the population dynamics of one taxon are influenced by its own growth and interactions with other taxa, was adopted to infer microbial interactions with Ralstonia ASV Sq_1. In this framework, termed Bayesian pcLV, taxon i was treated as the target (response) and taxon j as the source (predictor). The models were fitted in a pairwise manner, considering only two taxa at a time (one as the target and the other as the source) rather than attempting to model all taxa simultaneously. This strategy substantially improved the stability and convergence, which are often problematic in community-wide gLV models when the time series are sparse. Here, t denotes the sampling week and Δt = 1 week (the sampling interval). The continuous-time ALR ordinary differential equation formulation, termed the compositional Lotka–Volterra model by Joseph et al. (2020), underlies our pcLV framework, which is employed in a discretized form for sparse weekly sampling.
where ALRi and ALRj denote the ALR-transformed relative abundances of taxa i and j, respectively, assuming that the total microbial biomass is approximately constant (or changes slowly) between adjacent time points. r0 represents the intrinsic growth rate of the target taxon, ãii and ãij denote the self- and cross-interaction coefficients, and et is the residual process noise, modeled as an Ornstein–Uhlenbeck process with optional Student’s t-distributed errors. The models were derived from the original gLV equations and reformulated for compositional data. The detailed mathematical derivation of this ALR-based formulation and the rationale for adopting a pairwise modeling strategy rather than fitting a full model including all taxa simultaneously are provided in the Supplementary Methods section.
A custom Bayesian implementation was implemented using the R package cmdstanr (v0.9.0) with cmdstan (v2.36.0) using four Markov chain Monte Carlo (MCMC) chains. Each chain consisted of 1,000 warmups and 2,000 sampling iterations. Informed prior probability distributions were applied to the interaction coefficient, intrinsic growth rate (r0), and process noise parameters to improve convergence. The absolute magnitude of the interaction coefficient (ãij) was not emphasized; instead, we focused mainly on the signs (positive versus negative). To replace non-comparable raw coefficients with common weighting scheme, multiple candidate edges targeting the same taxon were adjusted using Bayesian stacking (Yao et al., 2018), which assigns weights based on cross-validated predictive performance. Stacking weights were calculated from cross-validated prediction performance, and they were maintained only for those passing Bayes False Discovery Rate (FDR) <0.05 and stacking weight >0.01. The model was included only when it met the moderate convergence thresholds (R̂ < 1.05, bulk/tail ESS > 400, ≤8 divergences, ≤80 max treedepth hits, and E-BFMI ≥ 0.30).
Results
Descriptive profiles of rhizosphere and episphere communities
To examine the temporal dynamics associated with pepper plants, bacterial communities from both the rhizosphere and leaf episphere of two distinct cultivars were systematically profiled over a period of ten weeks. Episphere communities were analyzed from weeks 5 to 10, when sufficient biomass was available, whereas rhizosphere samples were analyzed across all ten weeks (Fig. 1). Across all the samples, the ten most abundant genera were Pseudomonas, Ralstonia, Enterobacter, Paenibacillus, Methylobacterium, Sphingomonas, Chryseobacterium, Priestia, Acinetobacter, and Pantoea (Fig. 1). In the rhizosphere, strong Ralstonia-dominated profiles appeared sporadically in weeks 3–4 and became common from weeks 7 to 10, ranging from approximately 20–80% of the relative abundance across samples. Where Ralstonia was not dominant, Pseudomonas, Enterobacter, and Priestia were prevalent, whereas Acinetobacter showed a sharp increase at week six. In the episphere, Ralstonia was rarely detected; instead, Pseudomonas, Enterobacter, Paenibacillus, Methylobacterium, Sphingomonas, and Pantoea were consistently observed.
Relative abundance of the top 10 genera. The numbers at the top of each panel indicate the sampling week. The upper row of panels shows the relative abundance in the Episphere, while the lower row shows the relative abundance in the Rhizosphere. The top 10 genera are selected based on their average relative abundance across both compartments.
Phylogenetic identity of Sq_1
Despite the absence of wilt symptoms, rhizosphere dominance of Ralstonia suggests an association with R. pseudosolanacearum. Six Ralstonia-assigned ASVs (Sq_1, Sq_1233, Sq_2708, Sq_4852, Sq_6340, and Sq_8416) were compared with 93 full-length 16S rRNA sequences of R. pseudosolanacearum reported between 1998 and 2003 from fields across Jeju-do, Gyeongsangnam-do, Jeollanam-do, Jeollabuk-do, Gyeongsangbuk-do, and Chungcheongbuk-do (Cho et al., 2018), encompassing biovars 2, 3, and 4, which include pepper-pathogenic lineages. Phylogenetic analysis placed Sq_1 within clades containing biovars 3 and 4 from pepper fields, and its full-length 16S rRNA sequence was identical to that of multiple strains in these clades (Fig. 2). The remaining five Ralstonia ASVs did not cluster into the pepper-pathogenic biovar clade. Given the placement of Sq_1 within pepper-associated R. pseudosolanacearum lineages and its rise to dominance late in the time series, substantially exceeding that of other Ralstonia ASVs (Supplementary Fig. 2), associations with community structure were examined.
Full-length 16S rRNA phylogeny of Ralstonia ASVs and previously reported R. pseudosolanacearum isolates. Among the ASVs, Sq_1, Sq_1233, Sq_2708, Sq_4852, Sq_6340, and Sq_8416 are identified as Ralstonia. These sequences are aligned with the 16S rRNA sequences reported by Cho et al. (2018) using ssu-align, and a phylogeny is inferred with IQ-TREE 3. Biovar classification and host pathogenicity (potato, tomato, eggplant, and pepper) of the reported isolates, obtained from Cho et al. (2018), are visualized alongside the tree.
Sq_1-associated community dynamics
In the episphere (Fig. 3A), PCoA revealed scattered, irregular sample distributions with no clustering by week or Sq_1 abundance. PERMANOVA confirmed that the episphere composition was not described by Sq_1 abundance (R2 = 0.014, P = 0.878). In contrast, the rhizosphere samples (Fig. 3B) showed a clear pattern: higher-Sq_1 samples clustered toward the positive PCo1 axis, and Sq_1 explained a significant fraction of the rhizosphere compositional variation (R2 = 0.129, P < 0.001). These results indicate that Sq_1 was a major axis of β-diversity turnover in the rhizosphere during the late sampling period, whereas no comparable structuring was evident in the episphere.
Beta diversity revealed by PCoA. Bray–Curtis distance is calculated from the relative abundance of each sample, followed by PCoA. Panel A shows episphere samples, and Panel B shows rhizosphere samples. The color gradient of the points indicates the relative abundance (%) of Ralstonia ASV Sq_1. In the rhizosphere, Sq_1 reaches up to 80% dominance and significantly explains 12.9% of the variation in the overall PCoA distribution (PERMANOVA, R2 = 0.129, P ≤ 0.001).
To identify taxa whose temporal trajectories consistently covaried with Sq_1, pairwise correlations with Sq_1 were computed for 1,455 ASVs that reached a mean relative abundance of ≥0.01% within individual plots. Correlations were then meta-analyzed across the plots. This procedure yielded a distinct subset of taxa significantly associated with Sq_1 (Table 1): 14 showed positive associations (Sq_3, Sq_59, Sq_203, Sq_54, Sq_705, Sq_853, Sq_306, Sq_310, Sq_116, Sq_134, Sq_82, Sq_357, Sq_179 and Sq_63), and 11 showed negative associations (Sq_37, Sq_61, Sq_79, Sq_98, Sq_97, Sq_124, Sq_222, Sq_178, Sq_376, Sq_272 and Sq_395). These correlations delineated candidates for mechanistic follow-up but could not distinguish direct ecological interactions from indirect or environment-driven co-occurrences.
Taxa showing significant correlations with Ralstonia ASV Sq_1 based on meta-analyzed within-plot correlations in rhizosphere
In this study, pcLV models were applied to analyze the relationships between Sq_1 and each prescreened partner in order to estimate the directional influences on per-interval growth rates (ΔALRi/Δt). The resulting interaction network for the rhizosphere is illustrated in Fig. 4. The edges represent inferred effects, with their widths indicating the relative contributions to predictive performance. The nodes are identified by ASV IDs, alongside their corresponding taxonomic classifications. Modeling was performed for 26 ASVs (Sq_1 and 25 Sq_1-associated ASVs from Table 1). Among the models that passed all diagnostics and satisfied a Bayes FDR < 0.05, 33 significant edges were detected among the 24 nodes, comprising 22 positive and 11 negative effects. Notably, no effect of Sq_1 on the per-interval growth rate of other taxa surpassed these thresholds. Instead, Sq_1 had negative influences, consistent with inhibition or competition from the TRA3-20 (Sq_272), Bradyrhizobium (Sq_178), and Bryobacter (Sq_124).
Bayesian pcLV interaction network in the rhizosphere. Edges denote directional effects from the source taxon to the target taxon on the target’s per-interval growth (ΔALR/Δt) using lagged predictors. Color encodes the effect sign (blue, positive facilitation; red, negative inhibition/competition), and edge width is proportional to the relative ELPD based stacking weight, which sums to 1 per target. Only edges passing Bayes-FDR < 0.05 and stacking weight > 0.01 are shown. Each node is labeled with its ASV ID, and its taxonomic identifications are also provided. Model parameters are estimated using Bayesian inference via MCMC, and only models meeting moderate convergence thresholds (R̂ < 1.05, bulk/tail ESS > 400, ≤8 divergences, ≤80 max treedepth hits, and E-BFMI ≥ 0.30) are included.
Discussion
Root exudates generate buffered nutrient-rich microhabitats that behave as abiotic filters for rhizosphere assembly (Berendsen et al., 2012; Bulgarelli et al., 2013). In contrast, the leaf episphere is oligotrophic and is exposed to desiccation and UV stress, imposing stringent abiotic and biotic colonization barriers (Vorholt, 2012). Consistent with these constraints, Sq_1 was abundant in the rhizosphere, but remained rare in the episphere (Figs. 1 and 3; Supplementary Fig. 2). Although RSSC members can spread systemically through vascular tissues (Planas-Marquès et al., 2020), the scarcity of Sq_1 in the episphere suggests that environmental filters, rather than vascular access alone, limit their establishment. In the rhizosphere, Sq_1 attained a very high relative abundance (Fig. 1 and Supplementary Fig. 2) without visible wilt symptoms in pepper plants. This rhizosphere dominance could be due to the pathogen being initially recruited for infection but failing to invade the roots because of plant resistance (Peeters et al., 2013) or prior recruitment to the rhizosphere for asymptomatic infection (Du et al., 2017; Lebeau et al., 2011).
Full-length 16S rRNA placement showed identity between Sq_1 and strains within the RSSC clades, including pepper-pathogenic R. pseudosolanacearum biovars 3 and 4 (Cho et al., 2018; Peeters et al., 2013; Fig. 2). Taken together, the precise 16S identity with pepper-associated clades and sustained, symptomless dominance in the rhizosphere support the inference that Sq_1 represents a plausible pepper-adapted pathogenic lineage under the field conditions studied. However, because full-length 16S rRNA provides limited strain-level resolution within the RSSC (EFSA PLH Panel et al., 2019; EPPO, 2022; Garcia et al., 2019), Sq_1 should be interpreted as a 16S-defined lineage that may encompass multiple closely related strains, rather than a single clonal genotype. Therefore, this inference should not be considered a formal diagnostic determination. The latent infection and progressive dominance of Sq_1 coincided with a pronounced restructuring of the rhizosphere bacterial composition. Rhizosphere samples with higher Sq_1 abundance clustered toward the positive axis of PCo1, whereas samples with lower abundance were dispersed across the negative PCo1 space. No comparable segregation was observed in the episphere (Fig. 3A). Consistently, PERMANOVA indicated that Sq_1 abundance portrayed a substantial fraction of rhizosphere variation (R2 = 0.129, P < 0.001; Fig. 3B), but not the episphere variation (R2 = 0.014, P = 0.878). These results identify Sq_1 as a major axis of β-diversity turnover in the rhizosphere during the late sampling period. This pattern is consistent with previous findings that RSSC invasion reduces non-pathogenic bacterial diversity and increases sample-to-sample dissimilarities in tomato rhizospheres (Wei et al., 2018) and in field soils (Wu et al., 2024). Moreover, this finding aligns with observations in tobacco, where Ralstonia wilt-induced compartment-specific recruitment of bacterial taxa and reconfiguration of root microbiomes have been reported (Tao et al., 2022).
Correlation analysis provided an initial map of taxa covarying with Sq_1 (Table 1); however, such methods cannot distinguish direct interactions from indirect responses to shared conditions and often yield spurious links (Weiss et al., 2016). To probe directionality, a pairwise compositional Lotka–Volterra (pcLV) formulation was adopted, extending the cLV framework (Joseph et al., 2020) and tailored to sparse amplicon time-series data. A discretized ΔALR/Δt regression in log-ratio coordinates was used to accommodate compositionality and weekly sampling, with the cross-interaction coefficients interpreted by sign as facilitative (positive) or inhibitory (negative), consistent with cautions that such coefficients should not be over-interpreted in magnitude (Momeni et al., 2017). To partially address this limitation, Bayesian stacking (Yao et al., 2018) was applied, with weights estimated from the expected log predictive density (ELPD) and normalized to the sum of one across the targets. Although the cross-interaction coefficients (ãij) were primarily interpreted by the sign, models in which the coefficients were more accurate and stable were expected to receive proportionally greater ELPD-based stacking weights, thereby allowing the degree of information influence to be reflected in the final network. The model outputs were filtered using Bayes-FDR and Bayesian stacking to control for false positives, and robustness was confirmed using convergence diagnostics. Among the 25 ASVs that showed significant negative correlations with Sq_1 (Table 1), only three (Sq_272, Sq_178, and Sq_124) were retained after Bayes-FDR filtering in the pcLV analysis (Fig. 4 and Supplementary Fig. 3). These taxa exerted direct inhibitory effects on Sq_1. In addition, two taxa that appeared to be negatively correlated with Sq_1 in Table 1, Sq_272 and Sq_37, were inferred to be in the pcLV framework to facilitate Sq_178 and Sq_124, respectively, thereby exerting an indirect inhibitory effect on Sq_1. These findings illustrate that pcLV provides a finer resolution of influence decomposition than correlation when analyzing inter-microbial interactions. This difference arises because the correlation reflects non-temporal abundance changes, whereas pcLV infers directional influences by considering time-lagged effects on growth rates. The continuous-time ALR ordinary differential equation underlying the pcLV formulation was originally introduced by Joseph et al. (2020) and adapted in a discretized form for sparse weekly sampling.
The pcLV model indicated that Sq_1 did not exert detectable effects on the growth of other taxa but was consistently subject to negative influences from several community members (Fig. 4 and Supplementary Fig. 3). Most prominently, the TRA3-20 (Sq_272), Bradyrhizobium (Sq_178), and Bryobacter (Sq_124) families showed inhibitory effects on Sq_1. The family TRA3-20 is a hypothetical taxon arbitrarily designated in the SILVA database and assigned to the order Burkholderiales. Members of this order are known to produce secondary metabolites that provide competitive advantages, including siderophores, which have iron-chelating activity, and cepacin, which has antibacterial activity (Foxfire et al., 2021; Mullins et al., 2019). Bradyrhizobium has been reported to utilize structurally diverse siderophores (Ong and O’Brian, 2023), and recent studies have suggested that it may also influence microbial competition through the efflux of ferric xenosiderophore complexes from the periplasm (Ong and O’Brian, 2024). Although Bryobacter has not been reported to produce secondary metabolites like the two previously mentioned microorganisms, it has been documented in a Pearson correlation–based co-occurrence network analysis, where Bryobacter and Bradyrhizobium were both identified as keystone taxa associated with soil metabolites (Liu et al., 2020). Despite these negative effects, no positive effects on Sq_1 were detected across the sampling intervals (Fig. 4); however, Sq_1 steadily increased to dominance (Fig. 3B and Supplementary Fig. 2) in the rhizosphere. This contrast suggests that the host-associated advantages of RSSC (Peeters et al., 2013), such as strong adaptation to the pepper rhizosphere, outweighed the competitive pressure from neighbors. Host genotype and species likely play a key role in rhizosphere assembly. Resistant tomato lines can heritably recruit protective rhizobacteria (Yin et al., 2022), and tomatoes and peppers have been shown to recruit distinct beneficial taxa under RSSC challenge (Yan et al., 2024). Accordingly, Sq_1 dominance likely reflects an adaptation to host-driven ecological contexts rather than a mere lack of competition. These findings show that host–pathogen interactions can surpass community-level antagonism, highlighting the need to consider host-mediated processes in microbial interaction networks within plant rhizospheres. Developing strategies to enhance competitive suppression, like maintaining high densities of antagonistic microbes, could offset the advantages of RSSC and promote microbial biocontrol.
In this study, a time-series microbiome analysis was performed based on 16S rRNA full-length amplicons. In this study, R. pseudosolanacearum (Sq_1) clustered with pepper pathogenic lineages in the rhizosphere and inferred microbial candidates that could negatively affect their growth. Although Sq_1 gradually prevailed in the rhizosphere without any noticeable symptoms during the cultivation period, pcLV modeling, which accounts for covariates and time-lagged effects, suggested that the family TRA3-20 (Sq_272), genus Bradyrhizobium (Sq_178), and genus Bryobacter (Sq_124) were candidate inhibitory or competitive bacteria. Although further research is required to isolate and validate these candidates, the pcLV modeling approach provides a valuable starting point for identifying microorganisms that may suppress soil-borne plant pathogens.
Notes
Conflicts of Interest
No potential conflict of interest relevant to this article was reported.
Acknowledgments
This study was supported by the Research Program for Agriculture Science and Technology Development (Project No. PJ01727501) of the Rural Development Administration, Republic of Korea.
Electronic Supplementary Material
Supplementary materials are available at The Plant Pathology Journal website (http://www.ppjonline.org/).