Design of synthetic bacterial communities for predictable plant phenotypes

Abstract

Specific members of complex microbiota can influence host phenotypes, depending on both the abiotic environment and the presence of other microorganisms. Therefore, it is challenging to define bacterial combinations that have predictable host phenotypic outputs. We demonstrate that plant-bacterium binary-association assays inform the design of small synthetic communities with predictable phenotypes in the host. Specifically, we constructed synthetic communities that modified phosphate accumulation in the shoot and induced phosphate starvation-responsive genes in a predictable fashion. We found that bacterial colonization of the plant is not a predictor of the plant phenotypes we analyzed. Finally, we demonstrated that characterizing a subset of all possible bacterial synthetic communities is sufficient to predict the outcome of untested bacterial consortia. Our results demonstrate that it is possible to infer causal relationships between microbiota membership and host phenotypes and to use these inferences to rationally design novel communities.

Fig 1. Experimental strategy to design and test small consortia of bacteria with predictable host phenotypes.

I1–I3, indifferent phenotypes 1–3; leakyReLU, leaky Rectified Linear Unit; N1–N3, negative phenotypes 1–3; Pi, phosphate; P1–P3, positive phenotypes 1–3; 16S, ribosomal gene.

Fig 2. Bacteria modify the shoot Pi content in the plant.

(A) Schematic representation of the pipeline used for the binary-association analysis. For binary-association experiments, plants were germinated in axenic condition on Johnson medium, 0.5% sucrose with either 1 mM Pi (+Pi), approximately 5 μM Pi [traces of Pi from the agar] (−Pi), or 1 mM phosphite (Phi; not shown) in a vertical position for 6 days. Seedlings were then transferred to 30 μM Pi and 100 μM Pi media (without sucrose), alone or with each bacterial strain, for another 7 days. Arabidopsis thaliana plants were grown in a growth chamber in a 16-hour light/8-hour dark regime (24°C/21°C). (B) Top: Distribution of shoot Pi content in plants cocultured with individual bacterial strains (+ Bacteria) or in axenic conditions (No Bacteria) across a 2 × 2 matrix of Pi levels used for plant germination [+Pi (1 mM Pi) and −Pi (about 5 μM Pi)] and 2 Pi concentrations (30 μM Pi and 100 μM Pi), to which seedlings were transferred concomitant with application of each bacterial strain (Materials and methods 2a). Bottom: Number of strains that significantly increase or reduce the shoot Pi accumulation compared with no bacteria, after correction for multiple testing (Materials and methods 2e). Asterisks indicate an enrichment of strains with an effect greater than expected (hypergeometric test). Bacteria (n = 183) and 3 replicas (10 plants each) were analyzed per strain in 2 independent experiments. See also S3 Table. (C) Heat map of log(fold-change) in shoot Pi concentration between plants inoculated with individual bacterial strains, compared with axenically grown seedlings. Treatments are as in (B) and bacteria are sorted according to their phylogeny, as indicated by the tree on the left. Bottom bar plot shows the p-value from Pagel’s λ test for phylogenetic signal. Only 177/183 strains that were both tested in the plant–bacterium interaction assays and had a high-quality full-length 16S sequence are included. (D) Colonization capacity of 6 bacterial strains selected according to their performance in binary-association assays: 3 strains increased (green arrows) and another 3 decreased (red arrows) the shoot Pi content (see S3E Fig). CL and MF refer to the natural soils from which the strains were isolated. For this experiment, we used plants germinated with (black block) or without (no block) Pi or in the presence of phosphite (black block). Plant tissue was crushed; serially diluted, plated, and c.f.u’s per gram of original material were determined. Data points are colored by bacterial strain. Letters at the top of each panel denote statistical significance of Tukey’s post hoc analysis of a linear model. Numerical values that underlie the data displayed in the panel are in https://github.com/surh/wheelP. c.f.u, colony-forming unit; FW, fresh weight; Phi, phosphite; Pi, phosphate; 16S, ribosomal gene.

Fig 3. Synthetic communities alter plant phenotypes according to the strain makeup of the blocks from which they were composed.

(A) Heat map showing strains (n = 78) tested in binary association and that were selected because they cause positive (P), negative (N), or indifferent (I) effects on shoot Pi content in the growth conditions defined in Fig 2A. Strains are sorted within each group according to their mean effect on shoot Pi accumulation. Color scale shows log(fold-change) of shoot Pi content with respect to axenically grown plants. Bars and labels at the bottom show the 9 bacterial blocks used for the design of synthetic communities. Log(fold-change) is calculated from 6 pools of 10 plants in 2 independent experiments. See also S3 and S4 Tables. (B) Schematic representation of the synthetic communities designed using pairs of blocks. Sections in the circle are the 9 bacterial blocks from (A); black curved segments represent synthetic communities. Outer curved segments and curves inside the circle represent synthetic communities made of adjacent and nonadjacent bacterial blocks, respectively. (C) Heat map shows the scaled effect of each synthetic community on 4 plant phenotypes: Pi content (Pi), primary root elongation (Main), shoot area (Area) and total root network (Net) across the 4 growth conditions defined in Fig 2A. (D) Similar to (C) for individual bacterial functional blocks. In both (C) and (D), the values correspond to the scaled coefficients from a linear model. The values have been scaled through dividing by the standard deviation of all coefficients for the same phenotype and condition (each column in the plots). In all cases, 0 (white) represents no change with respect to axenically grown plants. The method to estimate the block and synthetic community effects are described in Materials and methods section 3j, and statistical significance (p-value < 0.05) is indicated with an “X” inside each tile, while the results of testing for significance for changes in Pi content are presented in S5 Table. Area, shoot area; Pi, phosphate; SynCom; synthetic community.

Fig 4. Synthetic communities additively modulate plant phenotypes.

Additive contributions of bacterial blocks explain synthetic community effects on all plant phenotypes. Comparisons between measured changes (x-axis) in plant phenotypes caused by synthetic communities, with respect to axenically grown plants, and expected changes (y-axis) from purely additive effects of each block, while ignoring bacterial relative abundances. In each plot, the 4 panels represent the 4 media conditions tested, with germination conditions as rows and Pi treatment as columns. Each point represents a synthetic community (n = 14); the x-axis corresponds to the color scale in Fig 3C, and the y-axis shows the result from adding the individual main effects estimated for each block (Fig 3D). The standard error from both the measured and estimated change is shown for each point. The blue line represents the least squares regression on the points from each panel, and the grey shade indicates the 95% confidence interval on the regression line. R² is shown on each panel. For all axes, 0 represents no change with respect to axenically grown plants. The values for Pi content and shoot area are indicated as log(fold-change) with respect to axenically grown plants. The values for primary root elongation and total root network represent the difference with respect to axenically grown plants. Additivity is evidenced by agreement between predicted and measured phenotypic changes. Numerical values that underlie the data displayed in the panels are in https://github.com/surh/wheelP. Pi, phosphate; R², coefficient of determination.

Fig 5. Synthetic communities additively induce the expression of phosphate starvation response marker genes in the plant.

(A) Average expression of a core of 193 phosphate starvation response marker genes in plants cocultured with synthetic communities in 4 growth conditions defined in Fig 2A. (B) Additive contributions of bacterial blocks explain synthetic community molecular phenotypes. Comparisons between the phosphate starvation response marker gene expression in (A) caused by synthetic communities, with respect to axenically grown plants and expected changes (y-axis) from purely additive effects of each block. In the plot, the 4 panels represent the 4 media conditions tested, with germination conditions as rows and Pi treatment as columns. Each point represents a synthetic community (n = 14), and the standard error for the measured and predicted change is shown. The blue line represents the least squares regression on the points from each panel, and the grey shade indicates the 95% confidence interval on the regression line. R² is shown for each condition. For both axes, 0 represents no change with respect to axenically grown plants. Numerical values that underlie the data displayed in the panel are in https://github.com/surh/wheelP. Pi, phosphate; PSR, phosphate starvation response; R², coefficient of determination.

Fig 6. Synthetic communities modify plant transcriptional profiles.

(A) Comparison of differentially expressed genes between all the positive (green) and all the negative blocks (magenta), in all conditions (top) or at 30 μM Pi (bottom). The first column shows all differentially expressed genes sorted by their log(fold-change), while the following columns indicate different functional annotations. Numbers at the top of each column show how many genes are marked and colored asterisks indicate a significant enrichment of the function among genes more expressed by positive (green asterisk) or negative (magenta asterisk) blocks. Panels (B) and (C) compare the expression of IPS1, a gene activated by low Pi, with (B) the JA response marker VSP2 and (C) the glucosinolate biosynthesis marker SUR1. Expression values are RPKM on a log₁₀ scale. (D) Comparison of 103 differentially expressed genes between 2 negative bacterial blocks (N1, N3) under low Pi (30 μM) condition. Rows represent genes and columns specific synthetic communities under different conditions. Color in the heat map shows the average expression of the corresponding genes across 2 independent experiments (3 replicates per experiment). Hierarchical clustering dendrograms are shown for both genes and conditions. Color in the dendrogram indicates the block that is included in the corresponding condition (column) or that up-regulates the corresponding gene (columns). Darker magenta color corresponds to block N3, and lighter magenta color corresponds to block N1, as in Fig 3B. Genes involved in stress response (Stress) are indicated on the right, and the logFC in expression between blocks N1 and N3 is also indicated, with positive values indicating a higher expression in the presence of block N1. Panels (E) and (F) compare the expression of IPS1 with (E), a phosphate starvation response–induced ubiquitin-conjugating E2 enzyme, PHO2 (F), and an auxin-regulated gene, ARGOS. Expression values are RPKM in a log₁₀ scale. Ellipses highlight samples from plants inoculated with synthetic communities P3N3 (pink asterisk) and N2N3 (blue asterisk). For panels (B), (C), (E), and (F), points on the axes represent samples in which the expression of the corresponding gene was not detected. ABA, abscisic acid; ARGOS, AUXIN-REGULATED GENE INVOLVED IN ORGAN SIZE; IPS1, INDUCED BY PHOSPHATE STARVATION1; JA, jasmonic acid; logFC, log(fold-change); PHO2, PHOSPHATE2; Pi, phosphate; PSR, Pi starvation response; RPKM, reads per kilobase per million; SA, salicylic acid; Stress, stress response; SUR1, SUPERROOT1; VSP2, VEGETATIVE STORAGE PROTEIN2.

Fig 7. The effect of novel synthetic communities on plant shoot Pi content can be predicted by an NN.

(A) Schematic representation of the NN defined and applied for predictions. Nodes are neurons, and arrows are weights that are estimated from the data. (B) Cross-validation error from the 3 types of models tested for their ability to predict shoot Pi content. Each model is trained on all but one synthetic community and evaluated on that held-out synthetic community. Each dot in the plot represents the mean Pi content prediction error on the held-out synthetic community. (C) Sensitivity of Pi accumulation with respect to each biological variable for each type of model. Each dot represents the change of shoot Pi content under a specific combination of input conditions (Materials and methods 4f). (D) The 25 most significant block replacements with a positive effect on the shoot Pi concentration predicted by the NN. These block replacements involved 20 different synthetic communities. Each box represents selected replacements in a particular constant background noted at the top. Each arrow represents a replacement of the bacterial block on the left with the block on the right. Asterisks indicate the blocks that lead to maximal plant Pi accumulation in the validation experiment. (E) The shoot Pi accumulation change predicted by the NN (x-axis) and the change observed experimentally are significantly correlated (Spearman’s correlation coefficient 0.42, p-value = 0.0375). For (D–E), color represents the validation experimental result: significant increase (dark blue), nonsignificant increase (light blue), nonsignificant decrease (light green), and significant decrease (dark green). (F) Prediction error on all tested block replacements for the LM, INT, and NN. The mean prediction error values ± standard errors are indicated above each box. The validation prediction error on NN is significantly smaller than LM (p-value = 5.25 × 10⁻¹⁰) and INT (p-value = 4.65 × 10⁻⁷). Numerical values that underlie the data displayed in the panels are in https://github.com/surh/wheelP. −NS, nonsignificant decrease; −S, significant decrease; +NS, nonsignificant increase; +S, significant increase; INT, linear model with interaction; leakyReLU, leaky Rectified Linear Unit; LM, linear model; NN, neural network; Pi, phosphate.