Rhizosphere metabolome detection

To identify rhizosphere metabolites that could drive microbiome assembly in the rhizosphere of diseased plants, rhizosphere metabolites were extracted with four pairs of samples (BD/BH, CD/CH, LD/LH, WD/WH) and analyzed according to our previous method with some modifications [61]. Briefly, rhizosphere soils were extracted twice with methanol solution (V_methanol: V_H2O = 3:1) and ethyl acetate. The extractions were combined for drying by adding 20 μL methoxyamination hydrochloride, followed by incubation for 30 min at 80 °C before being treated with 30 μL of BSTFA (bis (trimethylsilyl) trifluoroacetamide) reagent (1% trimethylchlorosilane, v/v). The mixture was then incubated for 1.5 h at 70 °C and finally analyzed with a gas chromatograph (Agilent 7890) coupled with a GC-TOF-MS (Shanghai Biotree Biotech Co. Ltd.). Raw peak analyses were performed as reported by Wen et al. [36].

For the differences among groups, relative abundances were used to standardize the metabolite profiles, and Bray-Curtis similarity matrices were prepared using the R package “vegan.” Permutational multivariate analysis of variance (PERMANOVA; Adonis, transformed data by Bray-Curtis, permutations = 999) was used to determine significant differences in beta diversity, and principal coordinate analysis (PCA) plots were generated from Bray-Curtis similarity matrices using “ggplot2” in R. Network analyses were performed using the R package “ggClusterNet” [62].

In order to determine metabolites that may drive the process of microbial community assembly in the diseased rhizosphere, machine learning was used to distinguish the rhizosphere metabolites associated with diseased and healthy rhizosphere soils. Because we found lower model accuracies when the models were built with all of the detected rhizosphere metabolites, we then trained a series of random forest models based on cutoff values for enriched metabolites characterized by relative abundances ranging from 1 to 90% and found the greatest accuracy in those trained with metabolites at > 3% (Supplementary Table 4). To avoid omitting important metabolites, we also trained a series of models from low-abundance metabolites (< 3%) and found that the greatest accuracy occurred with metabolite abundances at < 1% (Supplementary Table 4). The “important” metabolites were selected by cross-tabulations in R with “randomForest.” Wilcoxon tests (“stat” package) were conducted in order to detect the differences in rhizosphere metabolites between the diseased and healthy samples. Metabolites deemed as “important” from the classifiers and those that were significantly different between the two groups were selected for correlation analysis (the maximum-entropy approach) with their relative abundance and βNTI value in the diseased samples. Those metabolites significantly associated with the process of microbial community assembly were selected for further confirmation.