This study explores the order in which individual metabolic genes are lost in an in silico evolutionary process leading from the metabolic network of Eschericia coli to that of the genome-reduced endosymbiont Buchnera aphidicola.
Simulating the reductive evolutionary process under several growth conditions, a remarkable correlation between in silico and phylogenetically reconstructed gene loss time is obtained.A gene's k-robustness (its depth of backups) is prime determinant of its loss time.In silico gene loss time is a better predictor of their actual loss times than genomic features and network properties.Simulating the reductive evolutionary process by the loss of large blocks followed by single-gene deletions, as known to occur in evolution, yields a remarkable correspondence with the phylogenetic reconstruction and the block loss reported in the literature.
An open fundamental challenge in Systems Biology is whether a genome-scale model can predict patterns of genome evolution by realistically accounting for the associated biochemical constraints. In this study, we explore the order in which individual genes are lost in an in silico evolutionary process, leading from the metabolic network of Eschericia coli to that of the endosymbiont Buchnera aphidicola.
To evaluate the in silico gene loss time, we repeated the reductive evolutionary process introduced by Pál et al (2006), denoting the in silico deletion time of a gene in a single run of the reductive evolutionary process as the number of genes deleted before its own deletion occurred. By comparing the in silico evaluations of the gene loss time to that obtained by a phylogenetic reconstruction (Figure 1), we could evaluate the ability of an in silico process to predict temporal patterns of genome reduction. Applying this procedure on a literature-based viable media, we obtained a mean Spearman's correlation of 0.46 (53% of the maximal correlation, empirical P-value <9.9e−4) between in silico and phylogenetically reconstructed loss times. In order to provide an upper bound on evolutionary necessity stemming from metabolic constraints, we searched the space of potential growth media and biomass functions via a simulated annealing search algorithm aimed at identifying an environment/biomass function that maximizes the target correlation between in silico and reconstructed loss times. Simulating the reductive evolutionary process under the growth conditions and biomass function obtained in this process, we managed to improve the correlation between in silico and reconstructed loss times to a mean Spearman's correlation of 0.54 (63% of the maximal correlation, empirical P-value <9.9e−4, Figure 3).
Examining the dependency of the predicted loss time of each gene on its intrinsic network-level properties we find a very strong inverse Spearman's correlation of −0.84 (empirical P-value <9.9e−4) between the order of gene loss predicted in silico and the k-robustness levels of the genes, the latter denoting the depth of their functional backups in the network (Deutscher et al, 2006). Moreover, in order to examine whether the relative loss time of a gene is influenced by its functional dependencies with other genes, we performed a flux-coupling analysis and identified pairs of reactions whose activities asymmetrically depend on each other, i.e., are directionally coupled (Burgard et al, 2004). We find that genes encoding reactions whose activity is needed for activating the other reaction (and not vice versa) have a tendency to be lost later, as one would expect (binomial P-value <1e−14).
To assess the scale of these results, we examined as a control how well genomic features and network properties predict the phylogenetically reconstructed gene loss times. We examined the dependency of the latter on several factors that are known be inversely correlated with the propensity of a gene to be lost (Brinza et al, 2009; Delmotte et al, 2006; Tamames et al, 2007), including the genes' mRNA levels, tAI values (Covert et al, 2004; Reis et al, 2004; Sharp and Li, 1987; Tuller et al, 2010a) and the number of partners the gene products have in a protein–protein interaction network. Remarkably, these genomic features yield considerably lower Spearman's correlation than that obtained by the in silico simulations. Moreover, multiply regressing the loss times from the phylogenetic reconstruction on the in silico gene loss time predictions and the genomic and network variables, we found that the (normalized) coefficient of the in silico predictions in the regression is much higher than those of the genomic features, further testifying to the considerable independent predictive power of the metabolic model.
Finally, simulating the evolutionary process as large block deletions at first followed by single-gene deletions as is thought to occur in evolution (Moran and Mira, 2001; van Ham et al, 2003), a remarkable correspondence with the phylogenetic reconstruction was found. Namely, we find that after a certain amount of genes are deleted from the genome, no further block deletions can occur due to the increasing density of essential genes. Notably, the maximum amount of genes that can be deleted in blocks (i.e., until no more blocks can be deleted) corresponds to the number of genes appearing in our phylogenetic reconstruction from the LCA (last common ancestor of Buchnera and E. coli) to the LCSA (last common symbiotic ancestor, nodes 1–3 in Figure 1A), as described in the literature.
A fundamental challenge in Systems Biology is whether a cell-scale metabolic model can predict patterns of genome evolution by realistically accounting for associated biochemical constraints. Here, we study the order in which genes are lost in an in silico evolutionary process, leading from the metabolic network of Eschericia coli to that of the endosymbiont Buchnera aphidicola. We examine how this order correlates with the order by which the genes were actually lost, as estimated from a phylogenetic reconstruction. By optimizing this correlation across the space of potential growth and biomass conditions, we compute an upper bound estimate on the model's prediction accuracy (R=0.54). The model's network-based predictive ability outperforms predictions obtained using genomic features of individual genes, reflecting the effect of selection imposed by metabolic stoichiometric constraints. Thus, while the timing of gene loss might be expected to be a completely stochastic evolutionary process, remarkably, we find that metabolic considerations, on their own, make a marked 40% contribution to determining when such losses occur.