Artificial selection improves pollutant degradation by bacterial communities

Abstract

Artificial selection is a promising way to improve microbial community functions, but previous experiments have only shown moderate success. Here, we experimentally evaluate a new method that was inspired by genetic algorithms to artificially select small bacterial communities of known species composition based on their degradation of an industrial pollutant. Starting from 29 randomly generated four-species communities, we repeatedly grew communities for four days, selected the 10 best-degrading communities, and rearranged them into 29 new communities composed of four species of equal ratios whose species compositions resembled those of the most successful communities from the previous round. The best community after 18 such rounds of selection degraded the pollutant better than the best community in the first round. It featured member species that degrade well, species that degrade badly alone but improve community degradation, and free-rider species that did not contribute to community degradation. Most species in the evolved communities did not differ significantly from their ancestors in their phenotype, suggesting that genetic evolution plays a small role at this time scale. These experiments show that artificial selection on microbial communities can work in principle, and inform on how to improve future experiments.

Fig. 1. Selection method and its performance.

A Illustration of the selection method (see Methods for details). Each tube represents a community of 4 species (2 colors drawn for illustrative purposes): 1. Define 29 communities of randomly drawn species and inoculate each community in MWF+AA. 2. Following growth, measure degradation score as the difference in pollution load to an abiotic control, illustrated by the gray field at the top of each tube. 3. Select the communities with top 10 degradation percents (illustrated by tubes 2 and 5 here) and plate these on selective media to separate their members. Plating allows to combine degradation percent with extinctions to calculate community scores. 4. Collect viable cells of each species from the corresponding community with the highest score and freeze down. 5a. Generate 29 new communities in proportion to community scores. 5b. Randomly choose 21/29 of the new communities (illustrated with 4) for species exchange. Remove one resident species at random and introduce a new species in its place. Assemble the new communities in the lab using the frozen species from the previous round and repeat from step 2. B Degradation scores of the 5 best communities in each round for the selection (blue triangles) and random (orange circles) treatments, with lines through the average of the 5. C Community composition (x-axis) vs. degradation score (hue, color bar) for the best 56 communities (one third of all 167 tested communities, the full set is shown in Fig. S2) over the 18 rounds of selection (y-axis) in the selection (top panel) and random (bottom panel) treatments. Communities on the x-axis are ordered by increasing degradation scores (averaged over all instances of the same species composition). Note that these are degradation percentages, not final community scores (extinctions not considered). D Community composition corresponding to panel C, showing the presence (dark blue) or absence (white/grey) of each species, and illustrating the difference in composition by the Hamming distance (i.e. the number of substitutions needed to transform a given community to another) to the community with the highest degradation score at the bottom. Species abbreviations are as listed in Table 1.

Fig. 2. Number of extinctions, evenness and total population size over time.

A Number of extinctions per round (solid lines) and cumulative (dashed lines) in the 10 plated communities of the selection and random treatments. B-C Mean (lines)  ± SD (shaded areas) values of the 10 plated communities at each round where (B) shows evenness (the effective species number divided by its theoretical maximum value) and (C) total population size in CFU/ml. D-E Degradation percent plotted against (D) evenness and (E) total population size with the selection treatment in triangles and the random treatment in circles, and color representing selection rounds. Population size, growth or evenness could only be calculated for the 10 communities per treatment that we plated at each round (see Methods).

Fig. 3. Representation of species in evolved communities and factors that might explain it.

A Species representation and corresponding percentage of degradation in the last 5 rounds of the evolution experiment. As both measures can be quantified in percent, we display them on the same y-axis. The dashed line represents the average frequency at which we expect to see a given species in the last 5 rounds by chance, and the shaded area one standard deviation away from that average. Points that are outside the shaded area are more or less represented than expected by chance. The violin plots show the degradation scores of communities containing that species. B Degradation percent on day 3 in monocultures, pairwise co-cultures, top communities and 11 species together, using species taken from ancestral strains or strains isolated at the end of the random or selection treatment. Data-points are ordered according to the average degradation % and interesting cases are highlighted with a colored background and arrows corresponding to data shown in panels (D) and (E). C Experiment to determine whether Ac might be a “free-rider” (in 4 technical replicates). Data points in the top panel show population sizes (log₁₀CFU/ml) of different species and boxplots in the bottom panel show the distribution of degradation scores at day 3 of co-cultures as indicated on the x-axis. Each box shows the first, second and third quartiles and the whiskers the minimum and maximum values. Ac reduces the degradation score of the communities it is in, or increases their variance. D Matrix of degradation percentage in mono- (diagonal elements) and co-cultures of ancestral strains only (average of dots in panel (B)). E Matrix of population sizes (log₁₀CFU/ml) in mono- (diagonal elements highlighted with white squares) and pairwise co-cultures of ancestral strains only. In panels (B), (D) and (E) we highlight interesting cases in blue and light green that are further discussed in the text. Species abbreviations are as listed in Table 1.

Fig. 4. Effect of within-species evolution.

A Degradation percent or (B) population size (log₁₀CFU/ml) of each species on day 3 in three conditions: the ancestral strains before the experiment, strains harvested after 18 rounds of selection treatment (S) and after 18 rounds of the random control treatment (R). Data from mono- (alone) and pairwise co-cultures are shown (with partner). Significant differences are calculated using a generalized linear model with biological replicate as random variable and number of species in culture as an explanatory variable, significant p-values with a Bonferroni correction for multiple comparisons are shown. Treatment only had a significant effect on Ct (detailed statistics in main text). C–E Interactions between ancestral species (C), the species evolved in the random treatment (D) and the selection treatment (E) defined as the log₂ fold-change of the focal species in CFU/ml of day 3 in co-culture with the companion species vs mono-culture. Interactions that were not significant (no significant difference between growing alone or with companion species) are shaded, p-values were adjusted using the Benjamini-Hochberg method. Positive (facilitative) interactions are in blue, while negative interactions are shown in red. White squares are ones that we did not measure. Overall, we saw very few changes between ancestral and evolved species (black-bordered squares, quantified in Table S2). Species abbreviations are as listed in Table 1.

Fig. 5. Alternatives to our proposed method.

A–C Linear model analysis. We use a linear model (code taken from) that uses species presence/absence to predict degradation percent by using (A) all the data we generated as training and test sets, or (B) the mono- and pairwise co-culture data as training set and the rest of the data as test set, or (C) the larger communities as training set and the mono- and pairwise co-culture data as test set. Community richness is shown in color. Each dot is one degradation score measurement, such that biological replicates and technical replicates, if available, are all represented. R² shows Pearson’s correlation coefficient. D Proposing a new artificial selection method. Rather than disassembling communities, we propose to use the winning communities as templates to generate the offspring communities in the next round. These communities would then be seeded by taking the clonal ancestral species from the freezer, such that there would be no within-species evolution over rounds. Step 3 would be to select the top 10 communities, 4a to generate communities in proportion to their community scores and 4b to randomly choose 21/29 of the new communities (illustrated with 4) for either species removal or introduction (see white asterisks in step 4, 4a and 4b are shown in one step). Freezer icon created by SAM Designs from Noun Project.