How Enzyme Promiscuity and Horizontal Gene Transfer Contribute to Metabolic Innovation

Abstract

Promiscuity is the coincidental ability of an enzyme to catalyze its native reaction and additional reactions that are not biological functions in the same active site. Promiscuity plays a central role in enzyme evolution and is thus a useful property for protein and metabolic engineering. This review examines enzyme evolution holistically, beginning with evaluating biochemical support for four enzyme evolution models. As expected, there is strong biochemical support for the subfunctionalization and innovation-amplification-divergence models, in which promiscuity is a central feature. In many cases, however, enzyme evolution is more complex than the models indicate, suggesting much is yet to be learned about selective pressures on enzyme function. A complete understanding of enzyme evolution must also explain the ability of metabolic networks to integrate new enzyme activities. Hidden within metabolic networks are underground metabolic pathways constructed from promiscuous activities. We discuss efforts to determine the diversity and pervasiveness of underground metabolism. Remarkably, several studies have discovered that some metabolic defects can be repaired via multiple underground routes. In prokaryotes, metabolic innovation is driven by connecting enzymes acquired by horizontal gene transfer into the metabolic network. Thus, we end the review by discussing how the combination of promiscuity and horizontal gene transfer contribute to evolution of metabolism in prokaryotes. Future studies investigating the contribution of promiscuity to enzyme and metabolic evolution will need to integrate deeper probes into the influence of evolution on protein biophysics, enzymology, and metabolism with more complex and realistic evolutionary models.

Introduction

Twenty years ago, few people thought about enzymes when hearing the word “promiscuity.” Today’s blossoming interest in enzyme promiscuity stems from discoveries that promiscuity plays a central role in enzyme evolution and is thus a useful property for protein and metabolic engineering. In enzymes, promiscuity is the coincidental ability to catalyze reactions that are not biological functions, using a single active site. Throughout the manuscript, the term biological function refers to enzyme activities that are under positive or purifying selection because they contribute to organismal fitness. In contrast, promiscuous activities do not contribute to fitness unless environmental or genomic changes, such as mutations or gene amplification, create a new selective advantage for the activity and an opportunity to recruit the activity into a new metabolic pathway as a biological function. Promiscuous activities encompass the use of different substrates in the same type of chemical reaction (substrate promiscuity) and the ability to carry out different types of chemical reactions (catalytic promiscuity).

Building on Ohno’s thesis that gene duplication and divergence drives evolution, Ycas (1974) and Jensen (1976) proposed that modern metabolism evolved by duplication of genes encoding primordial, broad-specificity enzymes, enabling natural selection to independently optimize the specificity, efficiency and regulation of each enzyme [1–3]. Since then, dozens of studies have refined this idea by characterizing promiscuous activities, determining the structural and chemical basis of promiscuity, and using promiscuous enzymes in protein engineering as design tools and to examine mechanisms of enzyme evolution (reviewed in [4–11]).

This review develops a holistic view of enzyme evolution, beginning with evaluating the biochemical support for four enzyme evolution models that were derived from the ideas of Ohno, Ycas and Jensen. We incorporate horizontal gene transfer (HGT) into these models to explain the evolution of metabolic diversity in prokaryotes, in which HGT is a major evolutionary force [12, 13]. The second part of the review moves from individual enzymes to metabolism, by describing the combined power of enzyme promiscuity and HGT to rewire and expand metabolic networks.

Promiscuity in Models of Enzyme Evolution

Four models aim to explain the evolution of new enzyme functions: neofunctionalization; subfunctionalization by duplication, degeneration and complementation (DDC); subfunctionalization via specialization or escape from adaptive conflict (EAC); and innovation-amplification-divergence (IAD). Previous descriptions of these models exhibit some differences regarding the selective pressures and timing of the emergence of new functions [14–20]. Model descriptions in this review follow ref [19] most closely, but we elaborate on them to illustrate the role of enzyme promiscuity and HGT. Fully distinguishing among these models requires data describing the direction of selection, substrate and reaction range, ancestral activities, and historic gene copy numbers of an enzyme family. Some of this information is incomplete or missing from most studies. Nevertheless, many intriguing studies find biochemical and molecular evolutionary support that favors subfunctionalization or IAD models over neofunctionalization. The first section of this review describes this evidence, demonstrating that promiscuity is a prominent feature of enzyme evolution.

Neofunctionalization

Under the neofunctionalization model, a random gene duplication or HGT event yields redundant gene copies (Figure 1A). The original model states that one gene copy remains under purifying selection to retain the original function, while the other is subject to genetic drift, permitting mutations that confer a new activity [3]. If these mutations do not abolish the original activity, a neofunctionalized enzyme would be bifunctional or promiscuous, depending on whether its ancestral activity still contributes to fitness. Positive selection for the new activity may correlate with loss of the original activity, yielding an enzyme that is specific for the new activity. Mechanisms for specializing the activity of this bifunctional or promiscuous intermediate would be similar to those described for the subfunctionalization and IAD models, making specialization difficult to use as a criterion for distinguishing among the models.

Figure 1.

Models of enzyme evolution that incorporate promiscuity and HGT. This simplified view illustrates trajectories leading to an enzyme that is specialized for a new function. Alternative fates of gene copies that retain the ancestral function (such as retention for gene dosage, pseudogenization, gene loss, or horizontal transfer into new hosts) are not shown. Likewise, the endpoints of each evolutionary pathway are shown to be functionally specialized, although examples in the text indicate that these endpoints can be multifunctional or promiscuous for the ancestral activity. Blue genes and proteins represent the original function, while hot pink represents the new function. Promiscuous enzymes are depicted as blue with light pink stripes, while bifunctional enzymes are depicted as hot pink with blue stripes. Each arrow is labeled with an evolutionary event (duplication/amplification, HGT, or mutation) and a symbol representing selection pressure on the event. Black symbols indicate selection pressure on duplication/amplification or HGT; blue symbols indicate selection pressure on enzymes retaining the original function; hot pink symbols indicate selection pressure on enzymes specializing for the new function; and purple symbols indicate selective pressure on enzymes carrying out either the ancestral or new activities. Positive selection is indicated by (+), purifying selection is indicated by (−), and neutral selection, or genetic drift, is indicated by (Ø). A) Neofunctionalization. The stage in parentheses is not in the original description of the model [3] but illustrates how promiscuous activity could also be incorporated into a neofunctionalization model. Likewise, the upper symbols under the first two arrows indicate the selection pressure proposed by Ohno’s original model, and the lower symbols in parentheses indicate experimentally determined selective pressure on duplicated genes. B) Subfunctionalization. The expected selective pressures on function predicted by the DDC and EAC models are shown for the last stage of functional divergence. The specialization model (not shown) is similar to the EAC model, but it does not require equal degrees of positive selection to specialize each duplicate for different functions. C) Innovation-Amplification-Divergence. The diagram shows the IAD model as it would occur in a single strain. HGT (not shown) could occur at any point, as described in the text.

The primary defining criterion of the original version of this model, genetic drift of one gene duplicate and purifying selection of the other, has little experimental support. Maintaining multiple copies of a gene without selection poses a fitness cost that reduces the probability of neofunctionalization [21]. Also, deleterious mutations, which are more frequent than beneficial mutations, accumulate in the absence of selection [14, 22, 23]. Furthermore, several studies show that duplicated genes are both typically under purifying selection, although selection may be relaxed relative to non-duplicated orthologs [14, 24]. Indeed, a library of TEM-1 β-lactamase variants rapidly lost function in the absence of selection for ampicillin resistance, but selection of the library under moderate selective pressure (i.e. ampicillin concentrations lower than that tolerated by the wild type enzyme) increased the accumulation of mutations that conferred a new function (cefotaxime resistance) [23]. Thus, a revised neofunctionalization model would propose that increasing gene dosage via duplication can compensate for mutations that are moderately deleterious for the original activity; this relaxation of selection pressure increases the number of mutations that can accumulate and thus increases the probability of neofunctionalizing mutations. This is arguably indistinguishable from the IAD model discussed below, which is sometimes classified as a neofunctionalization model [19].

As presented here, however, the major distinction between the neofunctionalization and other models is the inability of the ancestral enzyme to catalyze the new reaction prior to gene duplication. By this criterion, lactate dehydrogenase (LDH) activity evolved from malate dehydrogenase (MDH) in trichomonads by neofunctionalization (Figure 2A). In this lineage, the common ancestor of LDH and MDH was specific for malate. LDH activity emerged after gene duplication and increased over time, yielding the broad-specificity Trichomonas vaginalis LDH, which has poor discrimination among pyruvate, lactate and other 2-ketoacids [25].

Figure 2.

Reactions catalyzed by enzymes discussed in the text. Blue atoms and bonds are lost or rearranged in the reactions. Red atoms and bonds denote the shared reaction intermediate of the catalytically promiscuous NSAR/OSBS. The R represents a hydrophobic amino acid side chain. A) Lactate and malate dehydrogenase. B) o-succinylbenzoate synthase and N-succinylamino acid racemase. C) HisA (N’-[(5’-phosphoribosyl) formimino]-5-aminoimidazole-4-carboxamide ribonucleotide (ProFAR) isomerase) and TrpF (N’-(5’-phosphoribosyl)anthranilate (PRA) isomerase). D) Two reactions catalyzed by Hevea brasiliensis hydroxynitrile lyase and a reaction catalyzed by Arabidopsis thaliana hydroxynitrile lyase.

Subfunctionalization

Subfunctionalization circumvents the problem of accumulating deleterious mutations in the absence of selection by proposing that new functions arise prior to gene duplication or amplification as promiscuous activities (Figure 1B) [14]. Two classes of subfunctionalization models have been proposed, which differ in their mechanism of functional specialization [19]. Duplication-Degradation-Complementation (DDC) proposes that each copy of a multifunctional gene can lose all but one function by genetic drift because other copies retain the lost functions. The specialization and Escape from Adaptive Conflict (EAC) models propose that positive selection drives duplicated enzymes to specialize in one activity of a multifunctional ancestor. Critically, both models require a multifunctional ancestor, which is proposed to be derived from a promiscuous enzyme. In an environment in which a promiscuous activity confers an advantage, positive selection of beneficial mutations would yield a bifunctional (or multifunctional) enzyme. Thus, new functions can arise prior to and independently of gene duplication [14, 26]. Multifunctionality, or gene sharing, is a property of many extant proteins. Lens crystallins, which are derived from a variety of enzymes, are the classic example. Piatigorsky [26] and Copley [27] have described many other multifunctional enzymes, most of which exhibit non-enzymatic, “moonlighting” functions. Other examples of multifunctional enzymes that carry out multiple chemical reactions in the same active site are described below.

Multifunctional enzymes are intermediates (or starting points) for subfunctionalization. Duplication, amplification or HGT of multifunctional genes is proposed to be under neutral selection but would enable specialization of each paralog by mechanisms described below (Figure 1B) [19, 20]. Alternatively, horizontally transferred multifunctional genes could be retained in the new host by positive selection if any one of their activities contribute to the recipient’s fitness. Because this review focuses on the contribution of promiscuity to the evolution of new enzyme functions, Figure 1B illustrates that if only the new, promiscuity-derived function of a horizontally transferred bifunctional enzyme offers a selective advantage, the ancestral activity could remain as a promiscuous activity or be lost, resulting in specialization for the new function.

The evolution of N-succinylamino acid racemization (NSAR) activity in one branch of the o-succinylbenzoate synthase (OSBS) family illustrates the subfunctionalization model (Figure 2B, Figure 3). In the NSAR/OSBS subfamily, the earliest branching enzyme known to possess NSAR activity is from Exiguobacterium sp. AT1b. Thus, the origin of NSAR activity can be traced to the common ancestor of Exiguobacterium OSBS and the rest of the NSAR/OSBS subfamily [28]. OSBS activity is required for menaquinone biosynthesis, and the Exiguobacterium OSBS is encoded in an operon with other menaquinone biosynthesis genes. Its efficient OSBS activity and genome context strongly indicate that OSBS is its biological function, although the effect of this enzyme on organismal fitness (the true measure of biological function) has not been directly measured through gene deletion experiments. In contrast, Exiguobacterium OSBS’s NSAR activity has been inferred to be promiscuous because it is very inefficient compared to typical efficiencies of metabolic enzymes [29], and genome context does not suggest that NSAR activity is a biological function [28]. Using the same rationale, the biological functions of other members of the NSAR/OSBS subfamily (and other enzymes discussed later in this manuscript) have been inferred from genome context and enzyme activity, because fitness contributions of enzyme activities are generally not available for non-model organisms. Like Exiguobacterium OSBS, most enzymes that diverge near the base of the NSAR/OSBS phylogeny are encoded in menaquinone operons, indicating that NSAR activity first emerged as a promiscuous activity and was later recruited by natural selection to be a biological function. In Geobacillus kaustophilus, the NSAR/OSBS gene is in an operon that encodes a pathway which converts D-amino acids to L-amino acids (Figure 3C). In agreement with the gene sharing and subfunctionalization models, this NSAR/OSBS enzyme is actually bifunctional, since G. kaustophilus also requires OSBS activity for menaquinone synthesis [30]. Genome context comparisons show that other NSAR/OSBS enzymes are likely to be bifunctional for the same reason. In a third set of species, only NSAR (or an unknown activity) is inferred to be a biological function, because these species do not require OSBS to synthesize menaquinone [31]. Most of these species acquired their NSAR/OSBS via HGT and represent the final step in the subfunctionalization model (although characterized NSAR enzymes have retained OSBS as a promiscuous activity) [31–33].

Figure 3.

Evolution of the N-succinylamino acid racemase/o-succinylbenzoate synthase subfamily. In A and B, colored circles indicate gain and loss of NSAR activity or inferred biological function, while open and black circles indicate characterized enzymes that catalyze OSBS activity only, or both NSAR and OSBS activities, respectively. NSAR activity might have also been recruited into a different pathway in the purple groups below the Bacillus subtilis branch. A) Phylogeny of the NSAR/OSBS subfamily, showing evolution of NSAR activity from a promiscuous activity to a biological function, as inferred from genome context. Genome context predicts that the biological function of proteins on blue branches is OSBS activity, the biological function of proteins on red branches is NSAR activity, and the biological functions of proteins on purple branches are both NSAR and OSBS activities. Adapted from ref. [31]. B) Phylogeny of the NSAR/OSBS subfamily, showing HGT to diverse phyla. C) D- to L-amino acid conversion pathway from G. kaustophilus [30]. The R represents a hydrophobic amino acid side chain. D) Genome context of selected NSAR/OSBS enzymes. Blue arrows denote genes involved in menaquinone biosynthesis. Red arrows denote genes that are known or predicted to be involved in pathways that use NSAR activity. Purple arrows denote genes that are predicted to be bifunctional nsar/osbs genes. Abbreviations: menA-H, menaquinone biosynthesis genes; AH, amidohydrolase superfamily; M20, M20 peptidase family; GNAT, GCN5-related N-acyltransferase; M78, M78 peptidase family; α/β, α/β hydrolase superfamily; blaC, β-lactamase class C; blaA, β-lactamase class A; ABC, subunits of dipeptide/oligopeptide ABC transporter; bla/pi, β-lactamase/prolyl oligopeptidase; pspC, phage shock protein C transcriptional regulator; NRPS AD, nonribosomal peptide synthase adenylation domain; NRPS CD, nonribosomal peptide synthase condensation domain;?, hypothetical protein.

Subfunctionalization has also occurred multiple times in the HisA family, whose enzymes catalyze N′-[(5′-phosphoribosyl)formimino]-5-aminoimidazole-4-carboxamide ribonucleotide (ProFAR) isomerization. Reconstructed ancestral HisA enzymes and extant HisA enzymes promiscuously catalyze the analogous TrpF (phosphoribosylanthranilate (PRA) isomerase) reaction (Figure 2C) [34]. Although most species have separate HisA and TrpF enzymes, the bifunctional PriA enzyme from Actinobacteria evolved from these promiscuous HisA enzymes to catalyze analogous steps in histidine and tryptophan biosynthesis [35]. PriA has respecialized a number of times, as discussed below [13, 36, 37].

Subfunctionalization via Duplication-Degradation-Complementation

Duplication-Degradation-Complementation (DDC) proposes that each copy of a multifunctional gene becomes specialized by losing all but one function. Conceptually, loss of function (degradation) is neutral, but measuring selective pressure via the nonsynonymous-to synonymous rate ratio (d_n/d_s) is likely to indicate relaxed or weak purifying selection, because only the subset of sites that contribute solely to the lost activity could evolve neutrally, while positions that contribute to the retained activity would still be under purifying selection [20]. This model was developed with gene regulation in mind, proposing that mutations in regulatory regions alter the expression patterns of gene duplicates [15]. In enzymes that catalyze multiple reactions in the same active site, however, mutations are more likely to affect protein folding and stability than other properties, and these multifunctional enzymes share a subset of catalytic residues that are required for all activities [38–41]. These constraints limit the fraction of sites where neutral mutations that affect only one activity could occur, suggesting that DDC might occur at a similar frequency to the specialization or EAC models described below, which rely on positive selection for mutations at an equally limited set of sites that differentially affect only one activity. For example, the DDC model does not appear to apply to a group of Actinobacteria which acquired a tryptophan operon containing a trpF by HGT. In these species, the bifunctional PriA lost its TrpF activity and became a specific HisA enzyme, but it was under strong positive selection [13, 42]. Also, genome context indicates that several members of the NSAR/OSBS subfamily only utilize NSAR as a biological function, but their promiscuous OSBS activity is still efficient; this is inconsistent with the idea that degenerative mutations which affect only one activity are common in these enzymes [31, 33, 43].

Other instances of enzyme evolution are better explained by the DDC model. Two enzymes in which OSBS is the only inferred biological function lost their ancestral promiscuous NSAR activity [31]. In a different set of Actinobacteria that acquired a separate trpF gene, the TrpF activity of PriA became less efficient in species which had an efficient TrpF enzyme, but PriA enzymes retained higher efficiency and were bifunctional in species with inefficient TrpF enzymes. This functional tradeoff is consistent with compensation for neutral activity loss by DDC in one gene copy [36]. The last example partially fits the DDC model. In a third group of Actinobacteria, both activities of PriA seem to be declining due to genetic drift. These Actinobacteria have experienced extensive gene loss. Loss of one of PriA’s activities correlates with the loss of the corresponding amino acid synthesis pathway, but the efficiency of the remaining activity also declined, suggesting that both activities could eventually be lost [37].

Subfunctionalization by specialization or Escape form Adaptive Conflict

Another set of subfunctionalization models, specialization and Escape from Adaptive Conflict (EAC), propose that positive selection optimizes each activity after duplication of a multifunctional enzyme (Figure 1B). Escape from Adaptive Conflict is a stringent model, which proposes that activities in multifunctional proteins cannot be optimized by natural selection, because mutations that enhance one activity are deleterious to others. After gene duplication or amplification, however, these mutations offer a selective advantage, and each duplicate is subject to positive selection that adapts them to different activities [17]. Similarly, residues that improve activities of a horizontally transferred enzyme will be subject to positive selection if they improve the host’s fitness, while residues that only affect the unused activities will be under neutral selection. From a biochemical perspective, this model is intuitive, but studies that link changes in biochemical activity to evolutionary pressure are limited. For example, data in the initial publication that described the EAC model was insufficient to support the EAC model, because of incomplete characterization of ancestral and descendant substrate specificity [17, 44]. In a more convincing study that used ancestral reconstruction to investigate evolution of substrate specificity of α-glucosidases, structural analysis, mutagenesis and activity assays suggested that steric clashes would prevent optimization of the enzymes to hydrolyze both maltose-like and isomaltose-like substrates. Furthermore, several residues near the active site that could influence substrate specificity may have experienced positive selection, although statistical challenges associated with methods for detecting positive selection prevent definitive conclusions [45]. Likewise, two of three substrate-specificity switches in a family of plant methyltransferases were attributed to EAC because 1) a promiscuous activity of the ancestral enzyme became the primary function of the extant enzymes after gene duplication; 2) positive selection occurred at the same time as a switch in substrate preference, as determined by ancestral reconstruction; and 3) mutagenesis of the identified adaptive amino acids revealed that the mutations released an adaptive conflict by improving the new activity at the expense of the ancestral activity [46].

EAC can also explain the evolution of a specialized HisA from a bifunctional PriA. PriA’s dual specificity is mediated by three active site loops that adopt different conformations for each reaction [47, 48]. Subfunctionalization occurred because these loops adopted different conformations and were less dynamic than in bifunctional PriA [42]. Other investigations of substrate promiscuity by engineered variants of PriA or HisA discovered that identifying support for a particular enzyme evolution model can be confounded by non-native reaction mechanisms [49, 50]. In these variants, absence of a suitably positioned catalytic acid was compensated by the carboxylate of the TrpF substrate phophoribosylanthranilate, which was able to serve as the acid when mutations rendered the active site more hydrophobic [50]. Although a role for substrate-assisted catalysis in the natural evolution of TrpF activity has yet to be demonstrated, it raises an interesting question regarding categorizing instances of enzyme evolution according to the models described here. Would the acquisition of a mutation that introduces a new catalytic amino acid that renders substrate-assisted catalysis unnecessary be considered neofunctionalization (since catalysis of the reaction is not an intrinsic property of the enzyme) or subfunctionalization (since the ancestral enzyme was capable of catalyzing a reaction with this substrate, as a promiscuous or bifunctional activity)?

Rauwerdink, et al. (2016) also expected EAC to explain the evolution of hydroxynitrile lyases within a family of plant esterases, because the reactive carbonyls bind in opposite orientations [51]. But because characterization of enzymes in the family was sparse, they ended up examining the ancestor of two hydroxynitrile lyase subfamilies instead (Figure 2D) [52]. Fortunately, these subfamilies have different substrate specificities and enantioselectivities, enabling them to determine that the ancestor was considerably less substrate-specific, less enantioselective, and exhibited lower reaction rates than its descendants. This data could be consistent with the EAC model, but the ancestral enzyme’s low activity with one substrate prompted the authors to reject the EAC model, which proposes a bifunctional intermediate, in favor the IAD model (described below) [51, 53]. Interpretation of these results is complicated, because biological substrates of some of these enzymes is uncertain and alternative reconstructions of the ancestral enzyme lacked hydroxynitrile lyase activity [53]. This study illustrates a pitfall in enzyme evolution research: ample sequence data facilitates evolutionary analysis, but limited experimental characterization and misannotation are barriers to biochemical insights.

Support for the EAC model could also be limited because adaptive conflicts might not be imposed equally on both activities. For example, the NSAR/OSBS subfamily enzymes from Deinococcus radiodurans and Thermus thermophilus retain promiscuous OSBS activity, even though these species do not require OSBS for menaquinone synthesis [32, 33, 54]. In fact, their efficiency for the OSBS reaction is the same as several enzymes that only have OSBS activity [33]. Furthermore, an amino acid substitution that enabled the non-promiscuous OSBS enzyme from Alicyclobacillus acidocaldarius to catalyze the NSAR reaction had no effect on its OSBS activity [31]. These results suggest that evolution of NSAR activity did not create an adaptive conflict that favored the loss of OSBS activity. Likewise, in some directed evolution experiments, increasing the rate of a promiscuous activity does not trigger an equivalent loss of the native activity [55, 56]. These results support a less stringent specialization model, which implicates positive selection as a means to adapt duplicated enzymes to distinct functions, but allows for variability in the tradeoffs between the new and old functions during the optimization of each activity [14]. This specialization model is likely to be more broadly applicable than the EAC model.

Innovation-Amplification-Divergence

Promiscuous activities are often too inefficient to offer a selective advantage [4, 56–58]. The Innovation-Amplification-Divergence (IAD) model proposes that positive selection for gene amplification overcomes this barrier by increasing protein expression to compensate for low efficiency of a pre-adaptive promiscuous activity (Figure 1C) [16, 19, 59]. This model could be grouped with the subfunctionalization model because the ancestral enzyme exhibits the promiscuous activity prior to gene amplification; however, some sources consider it a neofunctionalization model because the promiscuous activity does not confer a fitness advantage until gene amplification occurs [19]. In addition to dosage, gene amplification provides more potential sites for beneficial mutations and limits adaptive conflicts, because most gene copies will retain the original activity under purifying selection. Positive selection for mutations that increase the activity of the new function in one or more gene copies reduces the need for increased gene dosage. Conversion of excess gene copies to pseudogenes and gene loss could occur neutrally or by positive selection if maintaining extra genes imposes a fitness cost. The end result is a paralog specialized to carry out a formerly promiscuous activity, which may or may not retain the ancestral activity as a promiscuous activity [21, 59–61]. Horizontal gene transfer could occur at any point in this process. HGT at an early stage would transfer a gene for a promiscuous enzyme and contribute to amplification of the gene in the recipient. In support of this idea, several studies found that horizontally transferred genes are overrepresented among duplicated or amplified genes, although the studies were unable to link this observation to functional innovation [62–64].

In support of the IAD model, the frequency of gene amplification is 4–8 orders of magnitude higher than the rate of beneficial point mutations [59]. In contrast to the neofunctionalization model, selective pressure to maintain multiple gene copies helps offset the cost of tandem gene amplification [21]. Gene amplification is often observed during natural and experimental adaptive evolution [65–68]. Finally, several studies mimicked the IAD model by overexpressing promiscuous enzymes from multicopy plasmids, increasing their expression sufficiently to complement deficient strains that required the promiscuous activity [69–73].

In a study that followed the phenotypic reversion of a hemC⁻ Salmonella typhimurium strain, fitness was restored in the three steps proposed by IAD: gene amplification increased the defective hemC’s dosage; a mutation restored HemC activity in one allele; and restoration of HemC activity eliminated selective pressure to maintain the gene amplification, which was subsequently lost [60]. Similarly, Nasvall et al. (2012) tracked the evolution of a hisA⁻trpF⁻ Salmonella enterica strain that initially relied on a promiscuous HisA for HisA and TrpF activities [61]. By 500 generations, a 4-fold increase in TrpF activity was accompanied by a 3-fold loss in HisA activity, suggesting an adaptive conflict [74]. Gene amplification also occurred in most lineages within 1000 generations. By 3000 generations, some gene copies had become specialized for TrpF or HisA activity. Specialization was accompanied by loss of gene copies in some lineages [61]. The increased efficiency of the evolved HisA and TrpF specialists were traced to changes in the dynamics of two active site loops. In the initial variant, the loops were very dynamic, adopting different conformations for each substrate. Mutations that accumulated in the TrpF specialist appeared to stabilize the loops in the correct conformation for TrpF activity, while the final HisA specialist had lost a small duplication necessary for the TrpF conformation [74]. This result parallels the mechanism for evolving a specialized HisA from a bifunctional PriA, although different mutations occurred [42].

Enzyme evolution models meet reality: blended models and challenges for identifying key distinguishing features

At face value, the models described above make distinct predictions that should clearly distinguish between them. However, we have already described how aspects of the neofunctionalization and IAD models overlap when experimental data concerning the selective pressure on gene duplicates is taken into account. Other models could overlap in a different way. Preservation of gene duplicates during subfunctionalization or IAD could increase the probability that a neofunctionalizing mutation could occur [19]. Although the DDC and specialization/EAC models make opposite predictions about the selection pressure on enzyme function, they could still occur at the same time, because the subset of sites under positive selection to enhance one enzyme activity might differ from the neutrally-evolving subset of sites that contribute to other activities. Similarly, many positions in enzymes will be under purifying selection to maintain the structure and stability of the enzyme, even if there is positive selection of a subset of sites to improve one activity. As a result, the nonsynonymous-to-synonymous codon rate ratio (d_n/d_s), which is used to measure selective pressure, often remains <1, indicating purifying or relaxed purifying selection instead of positive selection.

Consequently, additional criteria need to be used to determine which model best fits the evolution of a given enzyme. Many studies use ancestral reconstruction to evaluate changes in activity between ancestral enzymes and their descendants [25, 45, 46, 51]. For example, the neofunctionalization model predicts that the ancestor cannot catalyze the activity of the newly evolved enzyme. The specialization model predicts that one or more activities of a multifunctional ancestor will be enhanced in the descendant enzymes. And the DDC model predicts that the activity level of the subfunctionalized descendants will be equivalent to that of their multifunctional ancestor. Distinguishing the IAD model from subfunctionalization models poses another challenge, because additional gene copies that accumulated during the amplification phase are expected to be lost. Thus, experimental evolution, which occurs on a short time scale, has provided considerable support for gene amplification in the IAD model, but long-term molecular evolution analysis of enzyme families would likely be blind to gene loss.

Instead, identification of bifunctional enzymes helps to distinguish between subfunctionalization and IAD. Bifunctional enzymes like PriA and NSAR/OSBS are a central feature of the subfunctionalization models, but are not required by the IAD model (although they are not ruled out). Whether new enzyme activities evolve via subfunctionalization or IAD will depend on tradeoffs between the original and new activities. Early gene amplification (as proposed by the IAD model) could be necessary if improving a promiscuous activity is too costly, but bifunctionality could be maintained for longer if improving the promiscuous activity has less effect on the original activity [75]. As a result, investigating the prevalence of bifunctional enzymes can provide insight into the predominance of these evolutionary mechanisms.

For example, Sulfolobus solfataricus encodes three enzymes from the Entner-Doudoroff pathway, a dehydrogenase, a dehydratase, and an aldolase that have dual specificity for glucose and galactose. This is similar to the case of dual specificity of PriA, in that separate sugar-specific enzymes carry out these reactions in other microorganisms [76, 77]. As another example, E. coli D-Malate dehydrogenase can complement a defect in a leucine auxotroph that lacks isopropylmalate dehydrogenase (LeuB), even when expressed from its native locus. Because expression of this enzyme must be induced with D-malate to complement the leuB⁻ strain, it is unclear whether this activity should be considered promiscuous or bifunctional [78]. Multifunctional enzymes are particularly common in plant specialized metabolism, employing substrate permissiveness to produce wide arrays of metabolites that could be advantageous in the arms race against herbivores and pathogens [79]. Several bifunctional enzymes occur in cholesterol and phytosterol biosynthesis in tomatoes [80]. Notably, many plants have several terpene synthase enzymes, some of which are capable of converting one substrate to many different products [81].

Bifunctional enzymes also occur in a number of microorganisms with reduced genomes. Several bacteria that lack an alanine racemase homolog appear to use a horizontally transferred homolog of cystathionine β-lyase (metC) for alanine and glutamate racemization, which are required for peptidoglycan synthesis. These low-efficiency enzymes are from species with extremely small genomes, suggesting a tradeoff between maintaining extra genetic material and utilizing less efficient, broad-specificity enzymes. Such bifunctional enzymes could be more common than previously thought, particularly in species with reduced genomes [82]. For example, the obligate intracellular parasites Chlamydia trachomatis and Chlamydia pneumoniae appear to have a bifunctional glutamate racemase/diaminopimelate epimerase or alanine racemase/serine hydroxymethyltransferase, respectively [83, 84]. Other Chlamydia species uses a TrpF homolog to catalyze an isomerization that bypasses a missing folate synthesis enzyme. These species have lost other tryptophan synthesis genes, indicating that their TrpF activity is retained as a promiscuous, rather than biological function [85]. Likewise, IlvC, a bifunctional ketol-acid reductoisomerase that catalyzes analogous steps in isoleucine and valine synthesis in most species, adopts a third keto-acid substrate to replace the activity of the missing PanE (2-dehydropantoate 2-reductase) in pantothenate synthesis of the intracellular endosymbiont, Buchnera aphidicola [86]. Whether the multifunctionality of these enzymes derives from subfunctionalization or a process not adequately described by the main enzyme evolution models is worth investigating.

The evolutionary potential of underground metabolism

The previous section presented models that describe the role of promiscuity in evolving new enzyme activities. A complete understanding of enzyme evolution must also explain the ability of metabolic networks to integrate these new activities. Metabolic networks are inherently messy. Some metabolites are shared by multiple pathways, other compounds are by-products of promiscuous reactions that go unused, some reactions are uncatalyzed, and broad-specificity and promiscuous enzymes can cross-connect separate pathways. D’Ari and Casadesús (1998) coined the term “underground metabolism” to refer to the use of endogenous metabolites by promiscuous enzymes [87]. Underground reactions could contribute to metabolic pathway evolution by replacing old metabolic pathways with alternatives or integrating unused by-products of promiscuous enzymes into the metabolic network. Recruitment of broad-specificity or promiscuous enzymes from different metabolic pathways to rewire or expand the metabolic network has been described as patchwork assembly [88–90].

The evolvability of metabolism is thus tied to the pervasiveness and diversity of underground metabolism. Several experiments investigated this diversity by overexpressing genomic libraries to identify promiscuous activities that complement the deletion of an essential gene, mimicking the IAD model [69–72]. One study that complemented 104 auxotrophic E. coli strains found that 20% of the auxotrophs could be rescued. Rescue occurred not only by promiscuous activities which directly replaced the missing activity, but also by other mechanisms that rewired metabolism or altered flux, demonstrating the robustness of underground metabolism [71].

Similarly, seven overexpressed proteins rescued a pdxB auxotroph which could not synthesize pyridoxal phosphate (PLP), but none were promiscuous enzymes that simply replaced pdxB activity. Instead, they comprised three different underground pathways that rewired metabolism to synthesize metabolites downstream of PdxB. One pathway rerouted 3-phosphohydroxypyruvate from serine biosynthesis through 4-hydroxy-L-threonine to a precursor of pyridoxal phosphate in four steps, using two promiscuous or broad-specificity enzymes, an uncharacterized enzyme, and an uncatalyzed reaction (Figure 4A) [72]. This underground route to PLP is not unique to E. coli. In many bacteria, a paralog of PdxA and a kinase catalyze two steps in D-threonate catabolism, but they also convert 4-hydroxy-L-threonine to a precursor of PLP. These activities appear to be promiscuous because most species that have this pair of enzymes also have a known PLP synthesis pathway [91, 92]. Underground metabolism in Bacillus subtilis is also capable of rewiring PLP synthesis, using a plasmid encoding the last two enzymes of E. coli’s PLP synthesis pathway and two uncharacterized genomic suppressor mutations. This experiment demonstrated the contributions of underground metabolism, horizontal transfer, and gene amplification of a suppressor in metabolic pathway evolution [93].

Figure 4:

Underground metabolism circumvents metabolic defects. A) An underground pathway for PLP synthesis in a ΔpdxB E. coli strain [72]. 3-Phosphohydroxypyruvate (3PHP) from serine biosynthesis is converted to 4-hydroxythreonine (4HT) and 4-phosphohydroxythreonine (4PHT). Purple arrows indicate steps catalyzed by promiscuous enzymes. B) Two underground pathways for the synthesis of β-alanine in a ΔpanD strain [95]. One pathway alters flux through the uracil synthesis and degradation pathways (blue box). The other pathway re-routes aspartate semialdehyde, an intermediate in amino acid synthesis (orange box). Thick arrows indicate increased flux due to a mutation in a repressor. Red asterisks and dashed arrows indicate mutations that decrease enzyme activity. Purple arrows indicate steps catalyzed by promiscuous enzymes, and the purple asterisk indicates a mutation that increased a catalytically promiscuous activity.

Another study demonstrated an underground link between branched chain amino acid metabolism and proline metabolism in Streptomyces coelicolor. The IlvC enzymes from several Streptomyces species catalyze the reaction of ProC much less efficiently than their native reactions, and they complement a proline auxotroph of S. coelicolor. Notably, deleting ilvC1 and ilvC2 in addition to proC was necessary for constructing the S. coelicolor proline auxotroph, suggesting that the underground activity of IlvC could contribute to metabolic flux of proline in this species [94].

In a different approach, Pontrelli et al. (2018) used serial passage of an E. coli mutator strain to identify underground routes that bypass a defect in the synthesis of β-alanine, a precursor of Coenzyme A. Three different routes were accessible by only a few mutations (Figure 4B). One strain upregulated uracil synthesis and degradation to produce sufficient quantities of a substrate for a broad-specificity aminotransferase that can make β-alanine. Another route utilized the substrate promiscuity of two enzymes and acquired a mutation to enhance the catalytic promiscuity of a third enzyme [95].

Some studies have also investigated the potential of underground metabolism to fulfill novel metabolic functions. One study investigated the ability of overexpressed proteins to improve growth under 237 toxin-containing conditions. Growth improved in nearly 40% of the conditions, usually due to overexpression of broad specificity efflux pumps or general stress response regulators. But improved growth in some conditions was attributed to enzyme promiscuity [73]. Other studies employed high-throughput screens to evaluate enzyme promiscuity [96–98]. Screening dozens of esterases revealed that their breadth of substrate promiscuity correlated with active site volume and catalytic triad accessibility [96]. Screening a large number of haloalkane dehalogenase superfamily enzymes showed that the majority could use >5 substrates [97]. Last, screening a metagenome library identified activities that could not have been predicted based on sequence similarity to characterized enzymes [98].

Other strategies to assess the breadth of underground metabolism employed flux balance analysis, a metabolic modeling strategy to assess growth capacity under different nutrient conditions. Notebaart et al. (2014) first added known, non-native substrates of enzymes to the E. coli metabolic network. By comparing flux balance of the native E. coli metabolic network to the network expanded by promiscuous reactions, they predicted that non-native reactions would enable utilization of 19 additional nutrients. Overexpressing a library of E. coli proteins showed that in silico modeling correctly predicted enhanced growth on 4 of 9 new carbon sources. Enhanced growth on other carbon sources that were not predicted by their model suggests that additional underground activities are yet to be identified [99]. In a follow up paper, the same group used their computational model to accurately predict that an E. coli population could adapt to several non-native carbon sources via a small number of mutations. Following a short-term laboratory evolution experiment, E. coli populations gained the ability to utilize five of seven new carbon sources. This required only a few mutations in promiscuous enzymes, regulatory proteins or gene promoters [100].

Two other studies developed strategies to investigate the divergence between computational prediction and experimental results. Flux balance analysis incorrectly predicted that aspC, argD, and gltA are essential, suggesting that promiscuous enzymes or alternative pathways compensated for their loss. Deleting genes that were upregulated in these mutants identified one or more broad-specificity or promiscuous enzymes that compensated for the loss of each of these genes [101]. Conversely, flux balance analysis also predicted that some genes were non-essential, contradicting experimental results. This conflict was resolved by conducting longer-term growth experiments, in which intermediates needed for an underground reaction could accumulate once mutations that lowered flux through a native pathway occurred. In some cases, this involved gene amplification, as expected in the IAD model [102].

Some computational approaches show promise for identifying underground reactions. Two groups applied different sequence-motif search strategies trained on bifunctional PriA enzymes to identify HisA enzymes that possess promiscuous TrpF activity [34, 103]. Because these strategies required a training set of known bifunctional enzymes, generalizing these approaches could be difficult. Another strategy searched protein structures for a three-dimensional constellation of active site residues which mimic the active site geometry required to carry out the ene-reductase activity of Old Yellow Enzyme. This study identified two nonhomologous enzymes that exhibit ene-reductase activity with several substrates [104].

The preceding studies illustrate many strategies for discovering underground reactions. One challenge posed by these strategies is that methods to predict promiscuity are the same methods used to predict biological functions, making it difficult to distinguishing between the two [91, 92]. Such efforts are also likely to underestimate the extent of underground metabolism, because testing every possible activity is not feasible [105]. Despite these challenges, the discovery that metabolism can be rewired multiple ways to overcome some metabolic defects is startling.

Horizontal gene transfer works in tandem with underground metabolism

HGT is the main source of new enzyme activities in prokaryotes [12, 106]. The rate of HGT is similar to the nucleotide substitution rate, but most horizontally transferred genes are rapidly lost [107, 108]. To avoid deletion from compact microbial genomes, transferred genes must provide a useful function [106, 109]. Thus, HGT is likely to be more successful if functionally related genes are clustered in operons that can be transferred simultaneously [106, 110, 111], However, the potential for evolutionary novelty could be higher if single HGT-acquired enzymes are plugged into the metabolic network via primary or underground activities.

Several studies illustrate the combined contribution of HGT and promiscuity to metabolic pathway evolution. As mentioned above, acquisition of a horizontally transferred trpF drove specialization of PriA [42]. Another study identified a bacteriophage enzyme that altered metabolic flux and regulation. Depletion of S-adenosylmethionine (SAM) by a bacteriophage SAM hydrolase caused induction of methionine biosynthesis. The upregulated MetB (cystathionine γ-synthase) rescued an isoleucine-dependent auxotroph by promiscuously catalyzing the reaction of IlvA (threonine dehydratase) [112].

Investigating recent evolution of pathways that degrade synthetic chemicals also reveals the intersecting roles of horizontal transfer and promiscuity. Although the synthetic pesticide pentachlorophenol (PCP) was only introduced in the 1930s, a pathway for its degradation has already evolved in Sphingobium chlorophenolicum, probably by recruiting enzymes involved in degrading natural aromatic compounds (Figure 5) [89, 113]. Poor kinetic properties and broad specificity suggest that the efficiency of the PCP degradation pathway is still evolving. For example, PcpB has very low turnover, while PcpC is inhibited by its substrates [114, 115]. Furthermore, PcpA and PcpD have broad substrate specificity, while PcpC and PcpE each catalyze two sequential steps in the PCP degradation pathway [116–119].

Figure 5:

The PCP degradation pathway from S. chlorophenolicum and γ-HCH degradation pathway from S. japonicum UT26 converge at β-ketoadipate [120, 123]. PCP and γ-HCH are metabolized to β-ketoadipate (cyan box), a common intermediate in aromatic compound degradation. PCP degradation pathway enzymes are shown in blue, and γ-HCH degradation enzymes are shown in maroon. PcpA and LinE (orange) share 52% identity; PcpE and LinF (pink) share 91% identity [120].

To identify origins of the pcp genes, Copley and co-workers compared the genomes of S. chlorophenolicum and Sphingobium japonicum, which does not degrade PCP. pcpB and pcpD are distant from pcpC on the chromosome, suggesting separate acquisition by HGT. The ancestral source of PcpB, PcpD, and PcpC is unclear, because their low sequence identity to related enzymes indicates that they did not arise from recent gene duplication and divergence [120]. In contrast, the final enzymes in the PCP degradation pathway, PcpA and PcpE, were present in the last common ancestor of S. chlorophenolicum and S. japonicum, suggesting recruitment from other pathways in these species. PcpA is nearly identical to LinEb of S. japonicum, whose biological substrate is unknown [120, 121]. PcpA and LinEb are very similar to LinE in S. japonicum, which is involved in degrading γ-hexachlorocylcohexane (γ-HCH, or lindane) another pesticide introduced in the 20^th century (Figure 5) [121, 122]. Thus, PcpA and LinE were both recently recruited to degrade synthetic pesticides, but the function of their common ancestor is unknown. Similarly, PcpE from S. chlorophenolicum and LinF from S. japonicum are 90% identical and reduce maleylacetate to β-ketoadipate, linking the origin of these enzymes to degradation of natural aromatic molecules, including amino acids [120].

Not only do degradation of PCP and γ-HCH overlap at their final step (PcpE/LinF) and use closely related homologs (PcpA/LinE) in earlier steps, but PcpC and LinD also catalyze similar reductive dehalogenation reactions. However, PcpC and LinD share <25% identity and did not diverge from a recent common ancestor [120]. Like PCP degradation genes, γ-HCH degradation genes were acquired by HGT, and most are associated with mobile genetic elements [122–124]. The distribution of lin genes among γ-HCH-degrading species is complicated. Higher conservation of lin genes relative to the rest of the genome supports the independent HGT of some genes or the whole pathway to each strain [123, 124]. Some strains do not possess all genes in the γ-HCH degradation pathway, but whether this is due to gene loss (possibly during strain propagation), catalysis of missing steps by other enzymes, or complementation of missing steps by other strains in a microbial community is unknown [124]. Together, the evolution of the PCP and γ-HCH degradation pathways demonstrates the role of HGT in patching together new pathways and integration of enzymes derived from HGT into central metabolism. However, low sequence similarity to enzymes with known functions has prevented identification of the ancestral functions of most PCP and γ-HCH-degrading enzymes, making it difficult trace the co-evolution of structure and function of enzymes in these pathways [120, 123].

Finally, one of the most striking observations about the evolution of the NSAR/OSBS subfamily is the correlation between the emergence of NSAR activity as a biological function and its horizontal gene transfer. Nsar/osbs genes were horizontally transferred from a Firmicutes ancestor to distant phyla, including Actinobacteria, Deinococcus-Thermus, and even Archaea (Figure 3B) [32]. These horizontally transferred genes rarely replace the recipient’s native OSBS. One exception is the OSBS from the Firmicutes species Alicyclobacillus acidocaldarius, which lacks NSAR activity but clusters with NSAR/OSBS enzymes from non-Firmicutes species in the phylogeny [31].

Most other HGT-derived NSAR/OSBS enzymes are predicted to use NSAR (or an unknown activity) as their biological function. Diverse genome contexts in species that do not require OSBS for menaquinone synthesis suggest that horizontally transferred nsar genes have been independently recruited into multiple metabolic pathways. In G. kaustophilus, the bifunctional NSAR/OSBS is encoded in an operon with an N-succinyltransferase from the Gcn5-related acyl-transferase (GNAT) superfamily and an L-desuccinylase from the M20 metallopeptidase superfamily, whose proteins comprise a pathway for converting D-amino acids to L-amino acids [30]. Many nsar/osbs genes are in similar (but usually not identical) operons. For example, instead of an M20 metallopeptidase gene, Roseiflexus castenholzii’s operon encodes an enzyme from the α/β hydrolase superfamily, which could also be an L-desuccinylase (Figure 3D) [31]. The NSAR/OSBS-encoding operon from Amycolatopsis mediterranei is more divergent, encoding enzymes from the GNAT, M20 metallopeptidase, amidohydrolase, and β-lactamase superfamilies. Annotation of these genes suggest a role in peptidoglycan degradation, which could use the succinylamino acid racemase activity of an NSAR [31]. An NSAR/OSBS domain is even fused to adenylation and condensation domains of a nonribosomal peptide synthase from the Cyanobacterium Planktothrix prolifica [125]. Together, genome context analysis of nsar/osbs genes illustrates the important contributions of catalytic promiscuity, horizontal transfer, and patchwork assembly to the evolution of new metabolic pathways.

Conclusions and Future Directions

In the last two decades, enzyme evolution research has moved into the mainstream of protein biochemistry and evolutionary biology. Experiments demonstrate that promiscuous activities comprise a large pool of potential ligand binding and catalytic activities, creating new questions about the accessibility of these activities to evolutionary processes. First, a small number of case studies provide experimental support for each enzyme evolution model, leaving questions about the relative frequency of each mode of evolution. Furthermore, these studies show that evolution is more complex than the models indicate, suggesting much is yet to be learned about selective pressures on enzyme activities. Second, biophysical limitations on the enhancement of weak promiscuous activities is unknown. These limitations will determine what fraction of promiscuous reactions are viable leads for enzyme evolution. Challenges associated with designing enzymes that are as specific and efficient as many natural enzymes also raises questions about biophysical mechanisms and evolutionary pressures that “optimize” enzymes. Answering these questions has practical applications in protein engineering [126]. Third, a few studies demonstrated that adjusting flux and utilizing promiscuous reactions enables metabolic networks to bypass defects, but integrating novel activities derived from HGT is largely unexplored. Optimization of regulation and flux add additional layers of complexity to understanding metabolic network evolution. Strategies to answer these questions about the contribution of promiscuity to enzyme and metabolic evolution will integrate deeper probes into the influence of evolution on protein biophysics, enzymology, and metabolism with more complex and realistic evolutionary models.

Abbreviations

HGT

Horizontal gene transfer

IAD

Innovation-Amplification-Divergence

DDC

Duplication-Degradation-Complementation

EAC

Escape from Adaptive Conflict

LDH

lactate dehydrogenase

MDH

malate dehydrogenase

NSAR

N-succinylamino acid racemase

OSBS

o-succinylbenzoate synthase

PLP

pyridoxal 5’-phosphate.

PCP

pentachloropheno

γ-HCH

γ-hexachlorocylcohexane