The systemic imprint of growth and its uses in ecological (meta)genomics

Abstract

Microbial minimal generation times range from a few minutes to several weeks. They are evolutionarily determined by variables such as environment stability, nutrient availability, and community diversity. Selection for fast growth adaptively imprints genomes, resulting in gene amplification, adapted chromosomal organization, and biased codon usage. We found that these growth-related traits in 214 species of bacteria and archaea are highly correlated, suggesting they all result from growth optimization. While modeling their association with maximal growth rates in view of synthetic biology applications, we observed that codon usage biases are better correlates of growth rates than any other trait, including rRNA copy number. Systematic deviations to our model reveal two distinct evolutionary processes. First, genome organization shows more evolutionary inertia than growth rates. This results in over-representation of growth-related traits in fast degrading genomes. Second, selection for these traits depends on optimal growth temperature: for similar generation times purifying selection is stronger in psychrophiles, intermediate in mesophiles, and lower in thermophiles. Using this information, we created a predictor of maximal growth rate adapted to small genome fragments. We applied it to three metagenomic environmental samples to show that a transiently rich environment, as the human gut, selects for fast-growers, that a toxic environment, as the acid mine biofilm, selects for low growth rates, whereas a diverse environment, like the soil, shows all ranges of growth rates. We also demonstrate that microbial colonizers of babies gut grow faster than stabilized human adults gut communities. In conclusion, we show that one can predict maximal growth rates from sequence data alone, and we propose that such information can be used to facilitate the manipulation of generation times. Our predictor allows inferring growth rates in the vast majority of uncultivable prokaryotes and paves the way to the understanding of community dynamics from metagenomic data.

Figure 1. Genomic signatures correlated to minimum generation time (d) for 214 prokaryotes.

Correlation between d and (A) the number of rRNA operons in the genome, (B) the relative distance from the origin of replication to rRNA genes (excluding species with no retrievable origin), 0.5 corresponds to half the replicon, (C,D) codon usage bias indices ΔENC′ and S . Spearman correlations are given (ρ) with all p-values<0.0001. Dashed lines represent the trend of the correlation.

Figure 2. Relative difference between the minimum generation time, codon usage bias indices, and G+C content of pairs of organisms and their phylogenetic distance for 214 prokaryotes.

Pairwise phylogenetic distances were computed from the matrix of the phylogenetic tree reconstruction (see Materials and Methods: phylogenetic analysis). Pairwise differences in doubling times (box-cox transform of d), codon usage bias indices ΔENC′, S and F and G+C content were normalized by the maximum observed difference in the 22791 pairs dataset (eq. 10). (A) The datapoints are represented in light gray. The red line represents a flexible spline fit (λ = 0.01). The black horizontal line represents the average relative pairwise difference. (B) The lines represent a flexible spline fit (λ = 0.01). For short distances (light gray area), the spearman correlations between phylogenetic distance and the relative difference in minimal generation times, ΔENC′, S and F and G+C content are respectively: 0.21, 0.28, 0.28, 0.29, 0.26 (all p-values<0.0001).

Figure 3. Minimum generation time (d) versus codon usage bias for 214 prokaryotes.

F (first principal component of ΔENC′ and S) and Φ_λ(d) (Box-Cox transform of d) are negatively correlated (ρ = −0.66). Line fitted by least squares regression: Φ_λ(d) = 0.8741−0.6496 F (R² = 0.47, p-value<0.0001).

Figure 4. Correlation between the residuals of the model (eq. 1) and optimal growth temperature (OGT).

Sperman correlation ρ = −0.37, p-value<0.0001. Residuals are positive for psychrophiles (OGT<15°C) and negative for thermophiles (OGT>60°C), indicating that for the former (latter) the observed minimal generation time is lower (higher) than expected from the genomic signatures. Relevant outliers: ¹ Sodalis glossinidius morsitans and ² Mycobacterium leprae.

Figure 5. Observed versus predicted minimum generation time.

The mesophilic predictor based on codon usage bias (eq. 3) was applied to the 187 mesophilic prokaryotic genomes. The diagonal black line corresponds to the identity.

Figure 6. Accuracy of the determination of composite codon usage bias (F_a.) with varying sample size.

F_a was calculated on randomly chosen samples (from 2 up to 450 genes) of all genes while using the full dataset of highly expressed genes. 100 iterations were effectuated for each sample size. The results for 3 organisms (one fast, slow and intermediate grower) are represented. The full black lines correspond to the whole genome value of F and the dashed lines to the standard deviations. Each data point is represented in gray.

Figure 7. Average predicted minimum generation time for 3 environmental metagenomes.

Crosses represent the average for the whole metagenome approach while dots represent the average for the pseudo-genome approach. All predictions were calculated with the predictor for mesophilic organisms (eq. 3). The average minimum generation time of the whole metagenome (crosses) and the respective standard deviation (open circles) were generated with 1,000 bootstraps on the dataset of all genes and highly expressed genes independently. The 3 whole-metagenome datasets are all significantly different (p-value<0.001). Minimum generation times were calculated using the whole genome of the sequenced genomes matching proteins of the metagenome (see Materials and Methods: classification of metagenomes into pseudo-genomes). The number of matching sequenced genomes are given above the average (dots) and standard deviation (bars) of the predictions. The 3 pseudo-genomes datasets are all significantly different (Tukey-Kramer: p-value<0.05).

Figure 8. Average predicted minimum generation time for the gut metagenomes of humans of different age groups.

Unweaned babies are 3, 4, 6, and 7 months old. Weaned children are 1.5 and 3 years old. Adults are between 24 and 45 years old. Groups not connected by the same letter (A or B) are significantly different (Tukey-Kramer: p-value<0.005). The full horizontal line represents the average of the predictions for all individuals.