The composition and structure of endophytic microbial communities, as well as their interactions with host plants, are strongly influenced by local environmental conditions [26]. Consequently, all samples of P. ginseng, P. quinquefolium, and P. notoginseng were sourced from their respective core production areas, which are the regions where plant-endophyte co-evolution occurs, as detailed described in a previous study [27]. Simply, sampling was conducted in mid-September 2019 in Baishan City, Jilin Province, for P. ginseng and P. quinquefolius, and in late October 2019 in Wenshan Prefecture, Yunnan Province, for P. notoginseng. All plants were sampled at the late root thickening stage. To capture the growth year-dependent microbial variations, three independent fields, each planted with 2-year-old (2y), 3-year-old (3y), or 4-year-old (4y) Panax plants, were selected per species. From each field, nine samples (biological replicates) were collected, and each sample consisted of 10 healthy plants. A total of 81 plant samples (3 Panax species × 3 growth year × 9 replicates) were obtained. After collection, plant leaf and root samples were rinsed, surface-disinfected, pulverized in liquid nitrogen to a fine powder, and subsequently archived at −80°C, respectively.

Genomic DNA was extracted from leaf and root tissues (the FastDNA SPIN Kit for soil (MoBio Laboratories, Inc., Carlsbad, CA, USA). The fungal internal transcribed spacer (ITS) region was amplified using the host-blocking primer pair ITS1F/ITS2R and sequenced on an Illumina MiSeq platform (Shanghai Biozeron Co., Ltd., Shanghai, China) [28]. After quality filtering and chimera removal, operational taxonomic units (OTUs) were clustered at 97% similarity and taxonomically assigned against the UNITE database [29,30,31]. Ecological networks for leaf and root communities were constructed separately based on robust Spearman correlations (|ρ| > 0.7, FDR < 0.001) among OTUs with relative abundance >0.01% [32,33]. Keystone taxa within each network were defined as nodes exhibiting both high connectivity (degree > 20) and high centrality (betweenness centrality >5000) [9,34].

Since amplicon analysis identified both Plec and Clad in all three Panax species, we only selected P. notoginseng (a representative species) for subsequent strain isolation experiments, with specific sample collection years detailed in Section 2.1. Based on the amplicon results, P. notoginseng was selected for fungal isolation [28]. Endophytic fungi were cultured from surface-sterilized tissues on potato dextrose agar (PDA) medium and purified by repeated sub-culturing. Pure strains were identified by ITS sequencing. Strains with the highest sequence identity to the keystone genera Plectosphaerella (Plec) and Cladosporium (Clad) were selected for functional assays.

The selected strains were co-cultured with a mixed saponin substrate (containing ginsenosides Rb1, Rb2, Rc, Rd, Rg1, Re, and notoginsenoside R1; Yuanye Bio-Technology Co., Ltd., Shanghai, China) in potato dextrose broth (PDB) medium for 7 days. Culture extracts were analyzed by High-Performance Liquid Chromatography (HPLC; Agilent Technologies Inc., Santa Clara, CA, USA) to quantify saponin conversion, using an uninoculated substrate as control [28].

The β-glucosidase activity of the two keystone strains was measured using p-nitrophenyl-β-D-glucopyranoside (pNPG; Yuanye Bio-Technology Co., Ltd., Shanghai, China) as substrate. Culture supernatants harvested at different time points were incubated with pNPG, and the reaction was terminated with sodium carbonate. Enzyme activity was determined by measuring the absorbance at 405 nm [35].

Genomic DNA of the two keystone strains was extracted using the E.Z.NA™ HP Fungal DNA Extraction Kit (E.Z.NA HP; OMEGA Bio-tek Inc., Norcross, GA, USA) and sequenced on both Illumina and PacBio platforms (Shanghai Biozeron Co., Ltd., Shanghai, China). A hybrid assembly strategy was employed to generate high-quality genome assemblies [36,37,38,39,40]. For transcriptome analysis, strains were cultured in yeast extract-peptone-dextrose (YPD) medium and cellobiose-modified YPD (Cel-YPD) medium. Total RNA was extracted using the Fungal Total RNA Isolation Kit (Sangon Biotech (Shanghai) Co., Ltd., Shanghai, China) and subjected to RNA sequencing (RNA-seq) on Illumina and PacBio platforms (Shanghai Biozeron Co., Ltd., Shanghai, China). Additionally, a pooled sample from multiple media was used for full-length transcriptome sequencing (PacBio Iso-Seq; Shanghai Biozeron Co., Ltd., Shanghai, China) to aid genome annotation.

Predicted protein sequences were functionally annotated against multiple public databases (Nr, Swiss-Prot, eggNOG, KEGG, GO) [39,40,41,42,43]. Glycoside hydrolase (GH) families GH1 and GH3 were identified using hidden Markov model (HMM) searches [44,45], and phylogenetic trees were constructed using FastTree 2 [46]. Secreted proteins were predicted using SignalP and DeepTMHMM [47,48]. Secondary metabolite gene clusters were predicted with antiSMASH [49]. To elucidate evolutionary relationships, we performed comparative genomic analyses with publicly available genomes from related species, including gene family clustering [50], phylogenetic tree construction, estimation of divergence times [51,52], and analysis of gene family expansion/contraction [53].

Detailed bioinformatic protocols are provided in the File S1.

Spearman correlations, respectively (Fig. 1a). The root network was dominated by Ascomycota taxa, while the leaf network mainly consisted of Ascomycota and Basidiomycota nodes. The leaf network had slightly more nodes and far more connections compared to the root network. The proportions of positive edges in the root and leaf networks were 95.54% and 91.30%, respectively, indicating potentially mutualistic interactions dominated within-tissue interactions (Fig. 1a). Multiple topological properties of the ecological network were further calculated. The leaf network exhibited higher average degree and density but lower clustering coefficient and modularity compared to the root network, indicating that the root fungal community may tend to interact more locally rather than globally compared with leaf mycobiome (Fig. 1b). Based on the node degree and betweenness centrality, 32 and 167 keystone OTUs were identified in root and leaf networks, respectively (Fig. 1c). Counting at the genus level showed that Plectosphaerella and Cladosporium contributed the highest proportions of keystone taxa in the root and leaf endophytic compartments, respectively.

Correlation analyses were performed between the relative abundances of Plectosphaerella in roots and Cladosporium in leaves and the contents of individual saponins in roots and leaves, respectively. In the roots of Panax plants, the relative abundance of Plectosphaerella was significantly and positively correlated with the contents of ginsenosides Rg1, Re, Rd, and NR1 (p < 0.05) (Fig. 1d). In the leaves of Panax plants, the relative abundance of Cladosporium was significantly and positively correlated with the abundances of ginsenosides Rg1, Re, Rd, Rb2, F1, and NR1 (p < 0.05). It is worth noting that both genera showed the highest correlation with the content of ginsenoside Rd in their respective compartments (Plectosphaerella: ρ = 0.62, p < 0.001; Cladosporium: ρ = 0.69, p < 0.001). These results indicate that the two keystone genus in root and leaf may be associated with the accumulation of specific ginsenosides in each compartment.

From the fresh root and leaf tissues, the Plectosphaerella (designated Plec) and Cladosporium (designated Clad) strains were obtained through isolation, culture, and purification, respectively. Based on colony morphology, microscopic characteristics, and a phylogenetic tree constructed using full-length ITS sequences, the two fungal strains were identified as Plectosphaerella niemeijerarum and Cladosporium crousii (Fig. 2a,b). Further, the full-length ITS sequences of Plec and Clad were blasted against the keystone OTUs of Plectosphaerella and Cladosporium identified in amplicon sequencing of roots and leaves. The lowest BLAST score of Plec against 26 Plectosphaerella-genus OTUs identified in roots was 270, with the highest E-value of 8 × 10⁻⁷⁶; the lowest BLAST score of Clad against 48 Cladosporium-genus keystone OTUs identified in leaves was 322, with the highest E-value of 5 × 10⁻⁹¹ (Table S1). These results indicate that the Plec and Clad strains isolated from the roots and leaves of P. notoginseng can well represent the keystone taxa obtained from ecological network analysis.

After co-culturing the two strains with mixed saponins, high-performance liquid chromatography was used to determine the changes in saponin content. Plec and Clad significantly reduced the content of Rb1 in the culture medium and increased the content of Rd (Fig. 2c). Furthermore, calculation of the Rb₁-to-Rd conversion efficiency of the two strains revealed that Clad exhibited a higher conversion efficiency than Plec (Fig. 2c), with detailed data provided in Table S2. Since Rd can be obtained by hydrolyzing a β-glucoside bond at the C20 position of Rb1, the above changes in ginsenoside content support the expectation that both fungi have β-glucosidase activity.

Furthermore, the pNPG method was used to determine the relative β-glucosidase activity of the fungal culture broth at different culture times. From 72 h to 144 h, the catalytic capacity of the pNPG substrate in the culture broths of the two fungal strains gradually increased, and the relative catalytic activity of the Clad strain was higher than that of the Plec strain (Fig. 2d). The gap between the two strains increased with the extension of the culture time. Overall, these results indicate that the two strains isolated in this study have β-glucosidase activity and can convert ginsenoside Rb1 into Rd.

The genomes of Plec and Clad strains were sequenced and assembled using PacBio and Illumina sequencing platforms to gain mechanistic insights of the functional roles of keystone taxa. For the Plec and Clad strains, a total of 36,220,183 bp and 33,581,758 bp clean reads from Illumina sequencing were obtained for k-mer analysis and genome modification, respectively. GenomeScope was used to generate a histogram of sequencing depth distribution (k = 21), with the heterozygosity rate of 0.67% and 0.23% as well as the repeat rate of 3.70% and 3.48% for Plec and Clad, respectively (Fig. S1). The size of assembled genome (consisting of 36 scaffolds) for Plec is 36 Mb, with the longest scaffold, the N50 length, the N90 length, and the GC content being 185 kb, 9.7 kb, 7.2 kb, and 57.42%. For Clad, the assembled genome was 33.5 Mb in size, consisting of 51 scaffolds. The longest scaffold is 236 kb, with an N50 value of 11.3 kb and an N90 value of 9.3 kb. The GC content of Clad genome is 52.24% (Table 1). A total of 12,218 genes were identified in the Plec genome, with an average length of 1491 bp, while the Clad genome contained 10,583 genes with an average length of 1546 bp (Table 1). In addition, only less collinearity within the two genomes is identified (Fig. 3a). Based on searches against multiple databases including NR, GO, eggNOG, KEGG, and Swiss-Prot, a total of 92.9% and 90.8% of the genes in the Plec and Clad genome are functionally annotated, respectively (Fig. 3b, Table S3).

The single-copy orthology genes-based phylogenetic trees are constructed for Plec and Clad strains to explore their evolutionary positions (Fig. 3c and Table S4). The results indicate that the Plec strain is closely related to P. plurivora and P. sp. P0831, while the Clad strain has a close evolutionary relationship with other Cladosporium species, especially C. velox and C. colombiae (Fig. 3c).

To explore the molecular basis for the conversion of ginsenoside Rb1 to Rd by the two keystone fungi, we focused on β-glucosidases, particularly those from the Glycoside Hydrolase family 3 (GH3), which are known to hydrolyze the glucosyl moieties of Rb1. Phylogenetic analysis revealed evolutionary relationships between the GH3 sequences of both fungal strains and reference sequences from organisms reported to transform Rb1 (Fig. 4a). To identify candidate genes with potential activity, we examined the transcriptional response of these GH3 genes to cellobiose, a substrate for β-glucosidases. RNA-seq analysis showed that 11 GH3 genes in Plec and 5 in Clad were significantly upregulated under cellobiose induction (Fig. 4b). These upregulated GH3 genes represent strong candidates encoding the β-glucosidases responsible for the initial transformation of Rb1 to Rd.

Analysis of gene family evolution indicated significant expansion of specific families in both Plec and Clad strains compared to their respective taxonomic relatives (Fig. 5a). Gene Ontology (GO) enrichment analysis of these expanded families highlighted their strong association with transmembrane transport, response to various hormones, and cellular components such as mitochondria, membranes, and vacuoles (Fig. 5b). Additionally, hydrolase activity terms were significantly enriched in Clad. These results suggest that genomic expansions enhancing transmembrane transport and environmental responsiveness may underpin the adaptive strategies and keystone roles of these fungi.

Genomic analysis further revealed that both keystone strains possess a rich complement of biosynthetic gene clusters (BGCs) for secondary metabolism, with numbers typical within their respective genera (Fig. 6a), and these BGCs are unevenly distributed across their genomic scaffolds (Fig. 6b). Prediction of the secretome identified a substantial number of secreted proteins harboring signal peptides in both strains (Table S6). GO enrichment analysis of these secreted proteins showed significant overrepresentation of functions related to carbohydrate/polysaccharide metabolism, cell wall organization, and notably, glycoside hydrolase activity (Fig. 6c). These genomic features imply that secondary metabolites and the secreted proteome, particularly enzymes involved in carbohydrate modification, are crucial for the interaction of these keystone fungi with their plant host.

Plants provide a growth environment for microorganisms and interact with them to achieve better survival and evolution [54]. As a unique and critical group within microbial communities, keystone microbial taxa can exert a significant impact on the community structure and function of the microbiome in the corresponding niche, primarily by modifying the microenvironment of that niche [9]. In this study, representative keystone endophytes were identified from the root and leaf niches of Panax species. It was confirmed that all these strains can convert ginsenoside Rb1 to Rd and possess β-glucosidase activity. Furthermore, to verify that the two strains do not produce saponins by themselves, we determined the saponin content in the PDB culture medium, yet no peaks matching those of the known saponins were detected. This observation could be attributed to the high specificity of the ginsenoside biosynthetic pathway, which encompasses the mevalonate (MVA)-derived triterpene skeleton branching pathway. The key enzymes involved include dammarenediol synthase, cytochrome P450 monooxygenases of the CYP716A subfamily, and uridine diphosphate glycosyltransferases, all of which are considered to be unique to plants of the genus Panax to date. This process of altering the host’s secondary metabolites essentially constitutes a modification of the microenvironment, which may further reshape the microbial community structure of the corresponding niche. The antibacterial effect of ginsenosides has been widely verified [55]; however, ginsenosides with different structures exhibit varying antibacterial activities. For instance, in experiments investigating the growth-inhibitory effects of protopanaxatriol type (PPT-type) and protopanaxadiol type (PPD-type) saponins against Porphyromonas gingivalis and Fusobacterium nucleatum, PPD-type saponins were found to have stronger antibacterial effects than PPT-type saponins, with ginsenoside Rd showing the most prominent activity [56]. Furthermore, recent studies have revealed that Panax species possess a plant chemical defense mechanism mediated by a two-component system containing β-glucosidase [57]. Specifically, in response to pathogen infestation, plants synthesize a variety of secondary metabolites to resist pathogens. However, some of these compounds are toxic to the plants themselves. To address this, plants produce a class of glycosylated, inactive defensive compounds [58,59]. In addition to protecting plants from autotoxicity, glycosylation enhances the solubility of defensive compounds, thereby facilitating their storage (typically in vacuoles) [60]. Under physiological conditions, β-glucosidase and glycosylated defensive compounds are spatially separated. When pathogens invade, plant tissues are damaged, bringing β-glucosidase into contact with the defensive compounds. This interaction triggers the hydrolysis of glycosyl groups, ultimately achieving the effect of resisting pathogens [61]. Although plants inherently possess this defense mechanism, our study suggests that endophytic keystone fungi may secrete β-glucosidase during pathogen invasion to modify the microenvironment, assist plants in resisting pathogens, and thereby maintain the balance of the entire microbial community.

Based on the experimental results of saponin transformation by two strains, this study explored the β-glucosidase-encoding genes in their genomes; both strains contained 23 sequences encoding this enzyme, with generally complete structures. These sequences involved in encoding β-glucosidases may serve as downstream regulators adapted to invasion by different pathogens, and when different pathogens invade, the promoters upstream of β-glucosidase-encoding genes located at different positions in the keystone taxa may respond to different toxins released by the pathogens, thereby synthesizing β-glucosidases that further form a plant chemical defense mechanism mediated by a two-component system of β-glucosidases together with Panax plants to resist pathogens. Furthermore, in the gene family contraction and expansion analysis of comparative genomics, we also found that the Gene Ontology enrichment results of the significantly expanded gene families of the Clad strain relative to the family level showed functions related to glycoside hydrolase activity, while no such related functions were detected in the enrichment results of the significantly expanded gene families of the Plec genus relative to the family level; in the enrichment analysis of transmembrane secretory proteins with bioactive peptides in the two strains, both strains were enriched in multiple glycoside hydrolase functions, which further indicates that the secretion of glycoside hydrolases may be one of the reasons for their status as keystone taxa.

The current study has limitations in investigating the saponin-transforming function of the strains: while in vitro data are reliable, in vivo validation in host plants is lacking. Future research should re-inoculate the strains onto original host tissues to verify their saponin-transforming capacity in vivo, enhancing findings’ comprehensiveness and robustness. Additionally, analysis of secondary metabolite gene clusters in the two keystone strains yields testable hypotheses. Regarding host ginsenoside biosynthesis: Plec (4 acyltransferase domains) and Clad (3) may contribute to ginsenoside structural diversification (e.g., malonyl-ginsenosides) [62]; squalene synthase in both may promote synthesis via key intermediate supply [24]; Clad’s two HMGL-like domains (reverse of rate-limiting HMGR) may restrict synthesis under specific conditions [63]; CYP57 (Plec) and CYP58 (Clad) are candidates for oxidative modification (analogous to ginsenoside-diversifying CYP716 [64]), requiring validation. For antimicrobial compound production: Plec’s chalcone synthase and Clad’s stilbene synthase gene clusters may synthesize antibacterial flavonoids/stilbenes [65], facilitating colonization and keystone status. For host growth modulation: Plec’s tryptophan dimethylallyl transferase (DMAT synthetase, involved in indole alkaloid biosynthesis) suggests potential synthesis/modification of growth-regulating indole compounds (e.g., IAA) [66], a direction for future study. Notably, discussions on mevalonate pathway balance (HMGR/HMGL), CYP-mediated saponin diversification, and IAA-related synthesis are based on genomic predictions and require experimental validation.