The origin of modern metabolic networks inferred from phylogenomic analysis of protein architecture

A phylogenomic tree (15) describing the evolution of 776 folds defined by the Structural Classification of Proteins (SCOP) (12) shows that folds appearing early in evolution were widely shared by proteomes in all organisms that have been fully sequenced (Fig. 1). Details on the evolutionary model used in phylogenetic analysis and the validity of rooting of our phylogenomic trees (summarized in Materials and Methods) have been described, together with limitations and biases of the reconstruction method (13–15). There were only 16 omnipresent folds, nine of which appeared at the base of the tree. Twelve of the omnipresent folds, including the nine most ancient and basal folds, contained omnipresent superfamilies that also appeared at the base of trees of superfamilies (15). These nine ancient folds represent architectures of fundamental importance (SI Table 1) undisputedly encoded in a genetic core that can be traced back to the universal ancestor of the three superkingdoms of life (20). These architectures are widespread in metabolism and are present even in parasitic organisms with highly reduced genomes and proteome complements. Phylogenomic reconstruction of evolutionary relationships between these ancestral folds showed that the P-loop-containing nucleoside triphosphate hydrolase fold (c.37) was the most ancient architecture, followed by the DNA/RNA-binding three-helical bundle fold (a.4), and then by the two most multifunctional and widely shared folds in metabolism, the TIM βα-barrel (c.1) and the NAD(P)-binding Rossmann (c.2) folds (Fig. 1). The P-loop hydrolase fold represents a single superfamily that was also basal in trees of superfamilies (15). Phylogenetic relationships in the tree of nine ancient folds were congruent with those in the global tree of architectures (Fig. 1). All of these omnipresent architectures were also widely distributed throughout metabolism. Using MANET, we identified metabolic enzymes with one or more domains having structures that match the nine ancient folds in 105 of 133 subnetworks, present in 11 mesonetworks defining core metabolism in KEGG (see SI Table 2 for data and nomenclature). The structural associations were also functional when the main enzymatic activities were linked directly to the ancient folds. These enzymes had highly diverse functions (Fig. 1), with 3–6, 8–33, 10–67, and 18–205 enzymatic activities defined at the first (class), second (subclass), third (subsubclass), and fourth (enzyme specificity) levels of Enzyme Commission (EC) classification, respectively (SI Table 3).

The accumulation of newly discovered enzymatic activities along the entire phylogenomic tree of protein architecture (Fig. 2) showed that most activities defined at different levels of EC classification were clearly associated with the first nine, and to a lesser degree, with the first 24, folds (SI Fig. 6). These trends suggest that, during evolution of ancient architectures, there was a burst of enzymatic innovation starting in primordial metabolic networks and extending throughout modern metabolism. In fact, we found noticeable patterns of innovation, such as the existence of a burst of enzymes transferring phosphorus-containing groups with an alcohol group as acceptor (EC 2.7.1) associated with the ancient c.37 fold, a subsequent burst of enzymatic diversification associated with the c.1 fold involving discovery and diversification of isomerases (EC 5), discovery of glycosidases (EC 3.2.1), and diversification of lyases (EC 4), and episodes of diversification of dehydrogenases (EC 1.1.1) and of lyases associated with the c.2 fold. Functions associated with the nine ancestral folds are described in SI Text. Remarkably, the EC 2.7.1 transferase burst of enzymes harboring the c.37 fold appeared ancient, involved 11 subnetworks, and originated in the purine metabolism subnetwork (see below). Evidently, enzymatic diversification occurred very early, ≈300 folds away from folds delimiting episodes of prokaryotic and eukaryotic-specific protein diversification and defining upper bounds for organismal diversification (Fig. 1). Indeed, at the time of appearance of superkingdom-specific folds, most enzymatic activities had been already discovered at all levels of EC classification (Fig. 2). Consequently, the common ancestor of diversified life probably had a complete metabolic toolkit.

We then focused on the presence of the nine ancestral folds in metabolic subnetworks and devised a phylogenetic method to make inferences about the history of subnetworks. For this purpose, we introduced a previously undescribed phylogenetic feature (character), the abundance or occurrence of an ancient fold in a subnetwork (see assumptions in SI Text). The phylogenetic criterion of primary homology underlying the use of these characters was the sharing of ancient protein architectures by the subnetworks resulting from enzyme recruitment processes. Analysis of occurrence and abundance of folds in enzymes of the 133 subnetworks (SI Table 2) shows that 28 subnetworks did not contain any of the nine most ancient folds and should be considered evolutionarily derived (SI Table 4). They were removed from further analysis. Nine of these lacked structural assignments and were uninformative. These 28 subnetworks belonged to seven mesonetworks, one to metabolism of other amino acids (AA2), one to metabolism of cofactors and vitamins (COF), two to energy metabolism (NRG), four to glycan biosynthesis and metabolism (GLY), six to biosynthesis of polyketides and nonribosomal peptides (POL), nine to biosynthesis of secondary metabolites (SEC), and five to biodegradation of xenobiotics (XEN). Two derived energy-linked subnetworks stand out in the list, oxygenic mitochondrial ATP synthesis (NRG 00193) and oxygenic photosynthesis (NRG 00195), suggesting these important functions appeared late in evolution, well after discovery of most enzymatic activities. This is consistent with molecular and geological records that suggest life achieved considerable complexity before the appearance of oxygen in the atmosphere, and with enzyme distribution in aerobic pathways that suggests adaptation to oxygen occurred after major prokaryotic divergences in the tree of life (21). Subnetworks with many ancient folds belonging to the remaining mesonetworks, amino acid metabolism (AAC), carbohydrate metabolism (CAR), lipid metabolism (LIP) and nucleotide metabolism (NUC), were clearly ancestral and part of the early enzymatic burst.

We used this phylogenetic method to generate rooted trees of subnetworks for each mesonetwork. We focused on mesonetworks because the global tree of subnetworks was poorly resolved. Trees reconstructed from fold abundance in subnetworks (SI Fig. 7) were generally congruent with those reconstructed from fold occurrence but carried more phylogenetic information (not shown). Clearcut subnetwork candidates of origin for each mesonetwork were identified at the base of individual trees, and these subnetworks were used to generate a tree of ancient subnetworks (Fig. 3). This tree was congruent with a tree describing the evolution of mesonetworks (SI Fig. 8), providing further confidence in statements of subnetwork evolution. In the tree of ancient subnetworks, the two subnetworks of the nucleotide metabolism mesonetwork, purine metabolism (NUC 00230) and pyrimidine metabolism (NUC 00240), were placed at its base. These subnetworks were followed by the porphyrin and chlorophyll metabolism subnetwork (COF 00860). This is noteworthy because nucleotides, and to a lesser extent selected cofactors in the COF mesonetwork, should be considered linked to RNA, conserved throughout life (18), and important components of an ancient RNA world (22). Two subnetworks were clearly derived, the polyketide sugar unit biosynthesis (POL 00523) and the stilbene, coumarine, and lignin (SEC 00940) subnetworks. These subnetworks belong to POL and XEN, mesonetworks that also harbor the largest number of subnetworks lacking ancestral folds. Other interesting evolutionary patterns were evident. For example, the citrate cycle (CAR 000200) subnetwork is derived in the CAR mesonetwork (SI Fig. 7), and CAR is quite derived within mesonetworks (Fig. 3 and SI Fig. 8). However, scenarios for the prebiotic evolution of metabolism suggest that the citric acid cycle was one of the first pathways to evolve (23, 24). Consequently, our results suggest prebiotic pathways evolved in a sequence unrelated to the pattern of subsequent enzymatic takeovers.

Because recruitment erases historical patterns of enzymes in networks, we used “subnetwork wheels” to reveal patterns of origin and evolution in metabolism. For each fold, these graphs represent subnetworks as vertices (nodes) and sharing of enzymatic activities (EC numbers at different levels of classification) as edges (lines connecting nodes). We assume that in network evolution, enzymes take over ancient or prebiotic reactions. In this process, a copy of a protein domain used in one metabolic context (donor site) begins functioning in a new context (host site), performing that function de novo or taking it over from the previous catalyst at the host site. This process overlaps with the invention of new architectures, beginning with the most ancient one, each new one contributing novel functions and new opportunities for recruitment. Although extant donor and host domains may differ, we assume successful recruitment results in evolutionary lockin at a structural level [structural canalization (25)] necessary to guarantee the maintenance of the fold architecture. Similarly, we consider that change is costly, and that takeovers are more plausible among sublevels within each EC classification level. Given these assumptions, four criteria were used to reveal evolutionary patterns of recruitment between subnetworks: (i) the abundance of the fold in each subnetwork, (ii) the ancestry of each subnetwork derived from trees of subnetworks, (iii) the sharing of enzymatic activities by subnetworks at different levels of EC classification, and (iv) phylogenomic superfamily relationships of the shared enzymes. These criteria provided weights to the vertices and edges of the subnetwork wheels that helped establish direction of enzyme recruitment.

Fig. 4 shows a subnetwork wheel for the most ancient architecture, the P-loop hydrolase fold. Twenty-nine subnetworks had enzymes that shared this fold, and a tree of these subnetworks again had purine metabolism, pyrimidine metabolism, and porphyrin and chlorophyll metabolism at its base. Fold abundance was also maximal in these three subnetworks. Purine metabolism appeared as the fundamental vertex of enzymatic sharing in the c.37 wheel, judged by the high degree of connectivity of this subnetwork at different levels of EC classification and the direction of enzyme recruitment. It is noteworthy that highly weighted connectivities were also established among these three most ancient subnetworks, especially at subclass level, most notably between the nucleotide metabolism subnetworks. There was also significant enzymatic sharing between purine metabolism and both sulfur (NRG 00920) and selenoamino acid metabolism (AA2 00450), but these two subnetworks had low fold abundance and were clearly derived in the set. We believe these instances of sharing represent late recruitment processes.

The ancestral enzymes in nucleotide metabolism were probably phosphotransferases transferring P-containing groups with an alcohol (EC 2.7.1) or a phosphate group (EC 2.7.4) as acceptors, hydrolases acting on P-containing acid anhydrides (EC 3.6) and perhaps ligases forming C–N bonds (EC 6.3.4) (SI Tables 5 and 6). It is likely that these enzymes were not part of ancient purine and pyrimidine biosynthetic pathways. Instead, they were involved in nucleotide interconversion, distribution (storage and recycling) of chemical energy in acid-anhydride bonds of nucleotides, and terminal production of nucleotides and cofactors. In this regard, enzymatic activities shared between the purine metabolism and the porphyrin and chlorophyll metabolism subnetworks involved phosphotransferases (e.g., that phosphorylate adenosylcobinamide; EC 2.7.1) and ligases that form C–N bonds (EC 6.3).