A central goal in molecular evolution is to understand the ways in which genes and proteins evolve in response to changing environments. In the absence of intact DNA from fossils, ancestral sequence reconstruction (ASR) can be used to infer the evolutionary precursors of extant proteins. To date, ancestral proteins belonging to eubacteria, archaea, yeast and vertebrates have been inferred that have been hypothesized to date from between several million to over 3 billion years ago. ASR has yielded insights into the early history of life on Earth and the evolution of proteins and macromolecular complexes. Recently, however, ASR has developed from a tool for testing hypotheses about protein evolution to a useful means for designing novel proteins. The strength of this approach lies in the ability to infer ancestral sequences encoding proteins that have desirable properties compared with contemporary forms, particularly thermostability and broad substrate range, making them good starting points for laboratory evolution. Developments in technologies for DNA sequencing and synthesis and computational phylogenetic analysis have led to an escalation in the number of ancient proteins resurrected in the last decade and greatly facilitated the use of ASR in the burgeoning field of synthetic biology. However, the primary challenge of ASR remains in accurately inferring ancestral states, despite the uncertainty arising from evolutionary models, incomplete sequences and limited phylogenetic trees. This review will focus, firstly, on the use of ASR to uncover links between sequence and phenotype and, secondly, on the practical application of ASR in protein engineering.
Proteins, along with the genes that encode them, are the product of a continuous process of mutation and natural selection. Understanding how and why proteins have evolved to work as they do necessitates an investigation of their structure and evolutionary history. While the seminal work on protein crystallography  and the idea of resurrecting extinct ancient proteins  emerged around the same time in the mid-20th century, decades passed before the first synthesis of an ancestral protein was undertaken .
The power of this approach, termed ancestral sequence reconstruction (ASR), lies in the ability to test evolutionary hypotheses by inferring the sequence of ancient proteins based on their extant descendants, then experimentally resurrecting the predicted ancestors and analysing them in the laboratory. To date, more than 50 biomolecules have been resurrected in the laboratory and tested either biochemically or biophysically. This approach has been most heavily applied to proteins, in particular to study the functional divergence of a protein family [4–9]. However, ASR has also been implemented on other ancient biomolecules such as transposable elements, transcription factors, regulatory sequences, RNA, viruses and entire genomes [10–14]. In this review, we will focus on ancestral protein reconstruction and provide a practical perspective on the use of this approach for protein engineering.
Approaches used for ASR
The resurrection of ancient proteins requires a set of related extant sequences and a phylogenetic tree that shows their evolutionary relationships. ASR has become ever more feasible in recent years due to several factors: the explosion in genome sequence data; the proliferation of bioinformatics tools for the alignment of homologous extant sequences  and the determination of the phylogeny that describes their evolutionary relationships; and the wide commercial availability of custom gene synthesis services . Once an ancestral protein is inferred, a DNA molecule encoding it can now be readily designed and synthesized, allowing it to be expressed and characterized experimentally.
Most of the ancestral sequences studied to date have been inferred using computational methods originally developed for phylogenetic analysis. The maximum parsimony (MP) method  was used in initial reconstruction studies, due to its ease of implementation. The MP method assumes that a phylogenetic tree with minimum number of substitutions should be the most likely. However, MP fails to take into account either biased amino acid substitution patterns or long branch attraction (i.e., the erroneous grouping of two or more long branches as sister groups). In this situation, rapidly evolving genes and unequal rates of molecular evolution between different lineages can potentially lead to the confusion of homoplasy (i.e. the sharing of characteristics in the descendants that were not present in the common ancestor) for homology. This problem can be minimized using methods that incorporate differential rates of substitution among lineages.
For these reasons, the probabilistic-based approaches, maximum likelihood (ML) [18,19] and Bayesian reconstruction , have been used in more recent ASR studies. ML and Bayesian methods give more reliable results than the MP method  and estimate the confidence in each inferred ancestral state, commonly indicated by the posterior probability of an amino acid at each node of a phylogenetic tree.
There are two types of ML reconstruction methods, namely marginal and joint . Marginal reconstruction compares the probabilities of different character states at an internal node at a given site and finds the amino acid that yields the ML for the tree at that site. Marginal reconstruction assigns a character state to a single node, whereas joint reconstruction assigns the most likely set of character states to all ancestral nodes. Marginal reconstruction is more suitable when one wants the sequence at a particular node (e.g. to synthesize the hypothetical nuclear receptor of the ancestor of Metazoa), whereas joint reconstruction is more suitable for counting changes at each site (e.g. the number of threonine to methionine replacements over the entire tree) . The result of joint and marginal reconstruction may differ; the marginal reconstruction, while it can be highly efficient, does not necessarily give the globally optimal ancestral state and can only be considered as an approximation to the joint reconstruction . In practice, the marginal approach is most often employed in ancestral protein reconstruction experiments. Both ML and empirical Bayesian methods assume that the phylogenetic tree and evolutionary model are known without error, but these assumptions are often not true, especially for highly divergent proteins. To account for this uncertainty, the hierarchical Bayesian approach calculates the probabilities of various ancestral states and averages the probabilities over all possible trees and models of evolution, in proportion to how likely these trees and models are in the observed data . However, Bayesian integration, has not substantially improved the accuracy of the inferred ancestral sequences  compared with ML in studies on simulated data sets, and is more challenging to implement. Very recently, an experimental phylogeny based on fluorescent proteins was used to show that Bayesian methods outperformed MP in reflecting more accurately the phenotype of ancestral nodes; however, ML methods were not examined .
The process of ancestral sequence inference
The first step of any reconstruction study is the generation of a multiple sequence alignment (MSA) of extant sequences (Figure 1, ). Homologous sequences are retrieved from databases such as NCBI or UniProt, using the basic local alignment search tool (BLAST) and aligned using MSA algorithms (reviewed in ) such as ClustalW , MAFFT , MUSCLE , T-Coffee  or PRANK (used for sequences with insertions ). Extant DNA or protein sequences can be used in ancestral reconstruction, depending on which models are applied for the inferences: nucleotide-, amino acid- or codon-based models. Most ASR studies have used protein sequences, however, using nucleotide sequences is less reliable when aligning regions with many insertions and deletions (indels), and does not take into account the redundancy of the genetic code and its implications for conservation of function versus nucleotide sequence. For both DNA and protein sequences, the collected sequences should represent a diverse set of homologues, generally from different evolutionary lineages or domains of life. A phylogenetic relationship is then built by means of a distance-based ML or Bayesian approach. The quality of the MSA and phylogenetic tree is crucial for the accuracy of reconstruction, thus often only the full-length sequences or sequences that contain critical residues such as active sites are included in the alignment . Any implausible sequences (e.g. sequences that do not appear to represent a true homologue) or sequences that contain many insertions or deletions (indicative of pseudogenes) may need to be culled. Often the alignment must be fine-tuned manually to improve its reliability at gap positions . The conservation of the absence or presence of gaps within a particular taxonomic group can be used as guidance in manual gap handling. Using the pruned alignment, a phylogenetic tree (gene tree) is built to determine the evolutionary relationships between sequences and compared with an accepted evolutionary tree (species tree) for the organisms represented in the sequence collection. If discrepancies are found, the MSA may require adjustment or the initial collection of sequences may need to be reassessed (e.g. by sampling from a greater number of species , investigating sequences from limited representative taxa or more stringent exclusion of implausible sequences). Masking of highly divergent regions of the alignment during generation of the tree may also improve the accuracy of the tree. However, there are many factors that lead to discordance between gene trees and species trees that cannot always be controlled [34–36].
Recently, an elegant method to overcome uncertainty in ancestral sequence prediction arising from ambiguity in tree topology (e.g. for cases in which the gene tree is not in accordance with the tree of life) has been developed. Reconciling the tree obtained from the gene of interest with a species tree, the phylogenetic tree that best describes the evolutionary relationship among various biological species derived from phylogenetic analysis of multiple alternative genes, resulted in a biochemically more realistic and kinetically more stable ancestral protein .
The other source of ambiguity comes from the evolutionary models used in the inferences. The most widely used models are the Dayhoff , JTT , WAG  and LG  substitution matrices. Allowing rate variation among sites, in the form of a discrete γ-distribution, provides a substantial improvement in fits of models of protein evolution . The discrete-γ model has since been used to improve the accuracy of phylogenetic analysis.
Using the MSA and phylogenetic tree as inputs, software packages such as PAUP (Phylogenetic Analysis Using Parsimony; [43,44]), PAML (phylogenetic analysis by maximum likelihood) , MrBayes , ANCESCON  or FastML  can then be used to find the most probable sequence at interior nodes in the phylogenetic tree. Table 1 provides a comparison of the available software packages, their respective features and examples of their usage. While many options are now available for inference of ancestral states, a general limitation of most is in the interpretation of gaps and insertions, often necessitating potentially subjective interpretations to be made by the investigator as to the relative likelihood of alternative explanations for the evolution of specific sequence features . Moreover, each software package has an effective upper limit to the number of sequences that can be used for the reconstruction, constrained by the available computational power. For the more user-friendly algorithms such as FastML, this may be as few as ∼250–300 average-length sequences, which restricts the information that can be exploited for the inference.
Resurrection in the laboratory
Once an ancestral protein sequence is inferred, it must be back-translated to derive a DNA sequence that can be used to express the recombinant ancestral protein. Most initial ASR studies used site-directed mutagenesis of a cloned extant gene to obtain an open reading frame (ORF) encoding an ancestral protein [3,49]. However, this approach is limited to more recent ancestors, where the evolutionary distance between the extant and ancestral forms is relatively short. However, as older proteins have been resurrected in more recent studies, gene synthesis or assembly of gene fragments has become the method of choice [50–52].
Due to cost and time constraints on the characterization of ancestral proteins in the laboratory, the sequence of the single, most probable ancestor is typically synthesized for each node to be studied. The most probable ancestor is obtained by assigning to each position in the alignment the amino acid inferred with the highest posterior probability for that position. However, it is often difficult to conclusively assign an amino acid at every position in a protein; the inference is typically less definitive at some positions in the alignment than others. Therefore, to resurrect variants at ambiguous sites (e.g. sites at which a similar probability is obtained for two or more ancestral states in a marginal reconstruction), a conservative approach can be taken; an ancestor can be synthesized first which encodes the most probable amino acids at each position and then site-directed mutagenesis can be used to introduce alternative ancestral states at ambiguous positions. Another way of introducing variability is by limited saturation mutagenesis, using oligonucleotides that are degenerate at ambiguous sites in the gene synthesis steps or by synthesizing oligonucleotides encoding different possible ancestral states and mixing them at the frequencies that reflect the posterior distribution . If there is no significant difference in the biochemical properties (e.g. ligand- or substrate-binding characteristics, enzyme specificity, turnover rate, thermal stability) between ancestral sequences sampled from posterior distribution and the most likely sequence, then it is reasonable to assume that conclusions drawn about the ancestral sequences are robust to slight changes in sequence. Indeed, it may be more reasonable to consider a population of possible ancestors than a single ancestral state as is calculated by Bayesian posterior probability of all possible states.
The DNA sequences that encode the ancestral proteins can be codon optimized in the gene synthesis step to increase their expression in the chosen host organism (e.g. Escherichia coli). However, designing a gene for optimal expression, in particular eukaryotic proteins expressed in bacteria, may require additional sequence modification, particularly in the N-terminal region of the enzyme. Moreover, the N- and C-termini of homologous proteins often show a high degree of divergence relative to the rest of the sequence. Substituting the N-terminus of the inferred ancestral proteins with sequences known to increase the expression of cognate modern proteins can be a practical solution for cases where the extant sequences have poor identity in those regions. In these instances, care must be taken in the interpretation of the ‘ancestral’ properties of such hybrid sequences.
Since 1990, resurrections and functional analyses have been carried out on putative ancestral proteins belonging to eubacteria, archaea, animals, plants and fungi. Resurrection of ancestral proteins has provided direct insights into how natural selection has shaped proteins found in nature and elucidated evolutionary processes behind their stability, specificity and structure [52,54–57]. It has also provided insights into defining the environments in which the earliest forms of bacterial life lived . Figure 2 depicts the biomolecules that have been resurrected in the past 25 years relative to their evolutionary age. Particularly notable amongst these are the studies discussed below which illustrate how ASR has advanced the understanding of protein evolution in specific gene families.
ASR as a tool to probe the evolutionary history of a protein family
The first resurrection of an ancestral protein involved full-length ancestral ruminant RNAses, revealing the history of the emergence of the digestive function of ribonucleases from a non-digestive precursor. RNases resurrected from organisms more ancient than the last common ancestor of the true ruminant did not behave like digestive enzymes (having significantly lower catalytic activity against single-stranded RNA or short RNA fragments than extant forms), yet displayed an increased ability to bind and melt double-stranded DNA . Subsequent ancestral resurrections of RNA paralogues revealed the origins of the diverse physiological functions of RNases. Zhang and Rosenberg showed that the ancestor of eosinophil-derived neurotoxin (or RNase 2) and eosinophil cationic protein (or RNase 3) had weak antiviral activity and only two substitutions were required to obtain the huge enhancement of the RNAse activity in the descendants . This suggests that the extant digestive RNAses have evolved from an ancestor that performed different biological functions and that all these activities are retained to a greater or lesser extent in the diverse extant RNAse paralogues that resulted from the expansion that occurred during early mammalian evolution.
Another of the first seminal ASR studies concerned steroid receptors , which were thought to have evolved in the vertebrate lineage (being apparently absent from the invertebrate lineage). However, after the isolation of an oestrogen receptor orthologue from the mollusc Aplysia californica, and the characterization of the ancestral protein from which all extant steroid receptors evolved, it was hypothesized that this protein family already existed in the ancestor of the protostomes and the deuterostomes and was lost in the Ecdysozoan lineage . The DNA-binding domain of the resurrected ancestral steroid receptor increased transcription from an oestrogen response element, whereas the ligand-binding domain of the ancestral receptor activated transcription in the presence of oestrogens, supporting the prediction that the ancestor would function like an oestrogen receptor . Resurrection of a more recent ancestor of this protein family, the ancestral corticosteroid receptor, revealed the evolution of hormone selectivity in the glucocorticoid and mineralocorticoid receptors [59–61].
Another good example of using ASR for understanding the factors underpinning the diversity found in a protein family is the study of fluorescent, GFP-like proteins from corals . Resurrecting ancestral genes encoding the common ancestor of all extant pigment colours and the intermediate nodes demonstrated elegantly how the red colour evolved from a green ancestor through a stepwise adaptation .
ASR as a tool for testing hypotheses about protein structure–function relationships
ASR has been used to examine various aspects of protein structure and function. Traditional site-directed approaches to testing hypotheses about residues responsible for particular properties of proteins are confounded by the fact that the change is usually introduced into a different background from that in which it evolved originally [62,63]. Therefore, the results that are obtained and the conclusions that can be drawn from them are complicated by the possibility of epistatic interactions. Epistasis is the phenomenon whereby the effect of a substitution at a given site is altered by the nature of a residue at a second position, i.e. the effect of a given mutation on protein function depends on the background protein sequence into which the mutation is introduced . Put differently, where epistasis occurs, the effect of two substitutions is quantitatively or qualitatively different from the simple addition of their effects when introduced separately. Positive epistatic interactions are thought to arise during evolution where an initial mutation allows a subsequent change to be tolerated. For example, an initial mutation may increase the stability of a protein such that a subsequent destabilizing mutation can be introduced. If epistasis exists, mutations that are responsible for desirable characteristics in one branch of an evolutionary tree may not show the same effect when introduced into members of a different branch of the tree since they lack the neutral permissive mutations (i.e. substitutions of no immediate functional consequence but which are required to buffer the protein against destabilizing changes) or since the mutations introduced disrupt the protein due to unfavourable steric or other physico-chemical interactions (restrictive mutations) .
Nature has navigated efficiently through the epistatic protein space, producing functional proteins at each node on a phylogeny, despite large numbers of accumulated mutations. Thus, ancestral proteins can be used as a suitable background for studying the effects of mutations on functional diversification in the context in which they originally arose, and prior to the introduction of any subsequent, confounding epistatic interactions. While it is never possible to be certain that the inferred ancestor represents the historically accurate evolutionary intermediate, it is nevertheless likely to be closer to that antecedent than any contemporary form.
Using ancestral protein resurrection, the Thornton group [60,64] demonstrated that two permissive and five restrictive mutations played important roles in the loss of aldosterone sensitivity in the modern glucocorticoid receptors. Zhang and Rosenberg  revealed how neutral substitutions, Arg to Ser at site 64 and Thr to Arg at site 132, may play constructive roles in the subsequent evolution of ribonuclease genes of higher primates. By introducing five conserved amino acids that were different in red and green vertebrate opsins into the ancestral background, Yokoyama et al.  successfully recapitulated the shift in the opsin absorbance spectrum from red to green, whereas previous mutagenesis studies using modern proteins had resulted in contradictory results concerning the functional importance of key mutations. When deconvoluting each of these five mutations (in single, double or triple mutants), 27% of the λmax shift from red to green was found to be attributable to epistatic interactions rather than the direct effect of individual mutations.
Structural characterization of ancient proteins
To date, few ancestral protein structures have been characterized. Initial studies on lysozymes of modern game birds in 1990 resulted in seven structures of ancestral intermediate variants . However, as these ancestral variants only differed from the contemporary ones in three positions, no significant structural changes were observed among them, except for a correlation between the side chain volume of the triplet and the thermal stability of the protein . The first truly ancient domain or full-length ancestral protein structures to be resolved were the ancestors of the ligand-binding domain of a steroid receptor  and of fish galectins , respectively. Since then, 46 crystal structures of ancestral proteins have been deposited in the Protein Data Bank (PDB) (Table 2). The oldest ancestral protein for which the crystal structure has been solved was the last bacterial common ancestor of thioredoxin, hypothesized to have existed 4 billion years ago . The sequence identity of these ancient proteins ranges from around 50 to 70% in comparison with the contemporary ones, and they fold into structures that are similar to the modern descendants.
The study of ancestral protein structures has confirmed that functional divergence in a protein family can emerge either as a result of structural modification or changes in protein dynamics. The evolution of substrate specificity in the congerin and lactate/malate dehydrogenase families paralleled structural changes in dimerization and α-helix or β-sheet conformation [66,67]. Conformational epistasis, which was defined by Ortlund et al.  as allosteric mutations that reposition other residues (usually active site residues) and therefore change the effect of substitutions at that site, was shown to remodel the receptor ligand contacts in the ancestral glucocorticoid receptor and switch the oligomeric state of PyrR proteins [60,68]. Conformational dynamics without significant protein backbone rearrangements were the underlying mechanism for green-to-red photoconversion of the GFP-like proteins and the ancestral promiscuity of β-lactamase [69,70].
Limitations of ASR
As with other prediction methods, ASR has a common pitfall in that there is no guarantee that the sequences inferred are historically correct. In many cases, a single, most probable, ancestral sequence cannot be unambiguously identified and multiple near equally probable amino acids might be assigned to a site in the alignment. Also, when an incorrect model of amino acid substitution is used, inaccurate inferences can occur, yet with high probability. These ambiguities in ASR arise from external factors like limited available sequence data, especially from key evolutionary intermediates, and intrinsic factors such as the evolutionary model that was used to infer the ancestral sequences. Ambiguous gap placement in MSAs and uncertainty of the phylogenetic tree joining the protein family members will also contribute. Uncertainty in inferring ancestral residues is normally greatest at sites that have high sequence divergence among the contemporary proteins, and which generally have little or no effect on the properties of the protein, i.e. sites that reflect neutral genetic drift. Posterior probability of the marginal reconstruction at each residue can be used as a measure of prediction accuracy .
To account for the uncertainty in ancestral reconstruction, in theory, all possible combinations of the ancestral states at the ambiguous sites need to be resurrected. For example, Jermann et al.  synthesized a variety of alternative candidate ancestral RNase sequences to cover the ambiguity in the tree topology. This solution is generally not feasible, however, in particular when there is a substantial number of ambiguous sites, since the number of possible permutations of ancestral sequences rises exponentially with the number of uncertain positions. Uncertainty in the phylogenetic tree (e.g. unresolved connectivity of deep branches) can be clarified by getting more sequence data, especially from organisms representing poorly sampled branches of the evolutionary tree. For example, Zhang and Rosenberg  sequenced additional genes from various primates to better reconstruct ancestral sequences of RNAse paralogues. Improving the phylogenetic tree reconstruction leads to a more precise parsimony analysis or alters the posterior probability in an ML analysis and thereby increases the confidence in the ancestral inference.
Whereas statisticians have developed better evolutionary models to address bias in phylogenetic inferences, protein engineers have tried to overcome this issue more effectively by targeting ambiguous sites that have low marginal probability by making libraries of possible alternative ancestors. Ancestral libraries, which contain multiple candidate ancestral sequences, can be functionally characterized to determine whether the properties subject to biological interpretation are robust with respect to the ambiguity  (Gumulya et al., unpublished). Uncertainties in the reconstruction process have been shown to result in apparently inactive, yet thermostable proteins, such as the ancestor of the glutaminase subunit (HisH) of the imidazole glycerol phosphate synthase , which, despite having a melting temperature at ∼79°C, showed no activity towards the hydrolysis of glutamine. Unlike the ancestral reconstruction of the cyclase subunit (HisF) that resulted in a thermostable yet catalytically active enzyme, the resurrection of the glutaminase subunit (HisH) was much more prone to uncertainty. The MSA of extant HisH forms contains several insertions and deletions and only 21 out of 226 residues are conserved in the alignment. Taken together, these features suggest that the inferred sequence for the HisH ancestor is largely ambiguous and may be inaccurate at residues that underwent frequent changes during evolution (due to insertions and deletions). However, the possibility cannot be excluded that the antecedent simply lacked the expected catalytic activity. This explanation is somewhat less likely, however, in proteins that have an essential and fundamental highly conserved physiological function, e.g. in a primary metabolic pathway.
As the above studies exemplify, one can be more confident that a given ancestral sequence is historically correct if a convincing evolutionary narrative can be put together when the properties of the ancestor and extant proteins are considered in the context of their respective environments. In this regard, improved character mapping within phylogenies should be helpful . However, it is important to avoid subjectivity in the evaluation of ancestors (i.e., to avoid what has been called an evolutionary ‘just so story’ ), especially since it is, at best, not trivial, and in most cases not possible, to independently verify assumptions about co-evolving proteins and environmental conditions at the time a given ancestor existed.
The emergence of ASR as a tool in protein engineering
The field of ASR initially developed due to the interest of evolutionary biologists in understanding how biomolecules adapt their function optimally in changing environment. It is not surprising, therefore, that such an approach can be adapted to the purposeful alteration of protein function to altered environments, such as required for industrial processes. Concern over the inability to validate whether resurrected proteins are the ‘true’ antecedents of modern sequences is less relevant where the ancestor is simply used as a starting material for protein engineering. ASR has already been shown to be a highly useful tool for protein engineering for drug discovery [76,77] and has great potential for use in other applications, such as the development of biocatalysts.
Whereas directed evolution is seen as moving forwards in evolution, ASR can be considered as travelling back to the origin. Importantly, traversing back and forth along an evolutionary timeline between extant and ancestral enzymes can enhance the search for proteins with novel or enhanced properties in areas of sequence space enriched in functional proteins, i.e. sequences that have survived natural selection. Importantly, this approach allows more sequence diversity to be assessed compared with, e.g. random mutagenesis or recombination of extant forms, where the introduction of deleterious mutations or adverse epistatic interactions is more likely to compromise the viability of the libraries created .
Extant specialist enzymes have been hypothesized to have evolved from ancestral generalist enzymes that catalysed many related reactions [52,78]. Moreover, based on the assumption that the Earth experienced periods of elevated ambient temperature in primordial times [79,80], it has been predicted that ancient proteins may have been generally more thermostable than extant forms, a prediction that has been supported by several studies [50,52,81–83]. Delivering dramatic changes in sequence, ASR has been shown to be more effective than any other currently available protein engineering method for increasing enzyme stability while preserving catalytic activity [51,52,54,84]. A handful of ancestral resurrection studies have revealed that at least some of the ancestral proteins from the deep past show both thermostability and high catalytic activity, making them highly suitable and evolvable departure points for engineering new traits not present in contemporary enzymes.
The ancestral mutation method (AMM)
The first ASR-related approach implemented for protein engineering was termed ancestral mutation , and involved introducing ancestral residues into a modern protein sequence. Four resurrected proteins, namely 3-isopropylmalate dehydrogenase [86,87], isocitrate dehydrogenase (IDH) , glycyl-tRNA synthetase  and β-amylase , were moderately increased in their stability, supporting the idea of a thermophilic common ancestor of the extant life on Earth. The ancestral mutation approach can be useful for combining desired properties of modern and ancestral proteins. Ancestral variants of serum paraoxonases and cytosolic sulfotransferases are highly active and functionally diverse and with just a few ancestral mutations the enzyme active site was reshaped to adopt new specificity . However, the lack of strategies for identifying which ancestral residues should be selected for substitution makes the approach less amenable to wider implementation. Importantly, epistatic interactions between the ancestral and non-ancestral residues could affect the property of interest, compromising the success of this approach.
The reconstructing evolutionary adaptive paths approach (REAP)
To identify which subset of ancestral residues should be targeted for substitutions, the REAP approach  uses signatures of functional divergence found in the phylogenetic analysis of a protein family. The hypothesis behind REAP is that when a phenotypic difference (e.g. qualitative change in activity) occurs between two subfamilies in a tree, the residues that change along the branches that diverge leading to the two subfamilies may be responsible for the observed functional switch. This approach has been used to broaden the substrate scope of a DNA polymerase to include a new class of triphosphate substrates. By screening a library of only 93 variants, each having replacements at three or four of 35 sites identified by REAP, eight variants that showed improved ability to accept the new class of triphosphate substrate were obtained .
The consensus mutation method
An alternative to ancestral mutation, but one that is superficially similar and sometimes inappropriately assumed to be effectively the same, is the consensus approach, in which the residues most commonly found at a given position or positions in a protein family are introduced into an extant protein . The consensus approach is commonly used in enzyme engineering to generate protein variants with increased stability , based on the hypothesis that the residue most commonly found at a given position is likely to be the one providing greatest fitness, i.e. commonly the best stability . However, it must be stressed that no information on phylogenetic relationships is incorporated into the consensus approach, and this may compromise its usefulness [33,96]. Four independent studies [50,54,97,98] comparing the phenotypes of consensus variants versus ancient proteins resurrected using evolutionary model-based methods show the superiority of true ASR in terms of producing proteins that can be expressed at high levels and show greater thermal stability. Despite having high sequence similarity to the ancestral sequence, the consensus forms of elongation factor (EF) Tu, the β-subunit of DNA gyrase, nucleoside diphosphate kinases and β-lactamase showed a substantially lower stability than those obtained from laboratory-resurrected enzymes [50,54,97,98].
Application of ASR to specific protein engineering objectives
Common goals of protein engineering strategies are: to make enzymes more robust to industrial conditions, e.g. with respect to temperature and solvent concentration; to alter substrate or reaction specificity; to enhance enantioselectivity; to improve catalytic rates and to adapt enzymes to a different cofactor. ASR has the potential to address several of these objectives in enzyme engineering as described below. However, recent studies have also revealed improvements in properties of more particular benefit to specific classes of proteins, such as the immunogenicity of protein drugs .
Thermal stability is desirable in proteins used for industrial applications to allow reactions to be run at higher temperatures but also because thermostable proteins have longer half-lives at mild temperatures. In addition, a robust template is often required for directed evolution studies, as mutations that promote new functions are frequently destabilizing . The seminal study that indicated that ASR could be used to engineer thermostable proteins was the reconstruction of EF proteins, very loosely dated to between 3.5 and 0.5 billion years old. The thermostability of more than 25 resurrected ancestral EFs suggested that the ancestors of current organisms were thermophilic and they subsequently adapted to lower temperatures [50,81]. Several later ASR studies on thioredoxin , β-lactamase  and nucleoside diphosphate kinase  have documented similar increases in the thermal stability profile as one goes back in time, and this result has been interpreted as evidence of ancient proteins needing to survive a hot ancient global environment. However, older ancestors do not necessarily show the highest thermal stability, as was demonstrated in the resurrected ancestors of 3-isopropylmalate dehydrogenase  and ribonuclease H1 . Thus, thermostability is not exclusively a primitive trait, meaning that it can be readily gained and lost throughout evolutionary history, as proteins traverse their separate evolutionary paths to adapt to the local conditions under which they were selected and to the functions they evolved to perform. Controversially, based on a theoretical study , it has been reported that the thermal stability of resurrected proteins might be overestimated by the ML method used for inferring ancestral sequences, as it tends to disregard variants at a position that are detrimental to stability and less frequent, including those that were never fixed. The ancestor predicted by ML can thus be considered as an average of the ancestral population genotypes rather than a sequence that would be expected to be sampled from any given individual. However, the alternative Bayesian method, which sometimes chooses less probable residues from the posterior probability distribution, can eliminate this bias and thus provide more robust conclusions. However, the authors acknowledge that it is unclear whether this interpretation will prove to be correct for reconstructions of real proteins, especially where epistatic effects can affect the stability of the evolving proteins. Interestingly, in the same theoretical study, ML produced the most accurate reconstructions .
The increases in denaturation temperature of ∼30–40°C that have been obtained by resurrecting ancestors from the deep past are much larger than those typically obtained in engineering studies aimed at protein stabilization (as reviewed recently in ). Moreover, the ancestral resurrection approach allows for the stabilization of protein without requiring detailed structural information. To date, most studies have used the AMM to engineer stability into proteins of interest. However, Jackson and colleagues used the ancestral form of periplasmic-binding proteins to develop thermostable variants for the construction of robust biosensors .
Altered (broadened) substrate scope or novel function
It was hypothesized originally by Jensen, in 1976 , that specialized enzymes have evolved from more generalist (i.e. more promiscuous) primordial forms and that the specificity of modern enzymes has been tuned to an optimal state by natural selection. ASR studies have revealed that some ancient proteins appear to have broad substrate scope and some have either the same or a different/novel function compared with their descendants. Two to three billion-year-old β-lactamases can degrade a variety of antibiotics with catalytic efficiencies that are similar to those of an average modern enzyme, suggesting that the predecessors of β-lactamase enzymes were moderately efficient promiscuous enzymes . The common ancestor of cyclin-dependent kinases (CDKs), mitogen-activated protein kinases (MAPKs) and the related kinase, Ime2, appears to have been able to phosphorylate peptides having either a proline or an arginine residue at the +1 position, in contrast with the more stringent proline requirement of the extant enzymes . The ancestor of a fungal glucosidase showed some activity for maltose and isomaltose, yet, over time, the modern enzyme appears to have specialized to hydrolyse one substrate, with a significant loss of activity towards the other substrate . The bi-functionality of the resurrected mammalian ancestor of serum paraoxonases, the quorum-quenching and detoxifying lactonase activities, explained the overlapping specificities of some extant paraoxonases .
The broader substrate scope seen in some resurrected ancestors may originate in an intrinsic prediction bias assigned to ancestors that ‘averages’ the different specificities seen in the descendent families . Alternatively, it may indicate sub-functionalization of a gene duplication event in a protein family, where the old and new copies of the gene have each evolved to perform just one of multiple ancestral functions. Neofunctionalization, i.e. the gain of a new function, can also occur following duplication , with or without retention of the original function in either paralogue, depending on subsequent selection pressures.
‘Artificial’, prospective evolution studies have demonstrated that the ability to evolve a promiscuous function often requires an existing weak initial activity in the wild-type template . Thus, it is highly advantageous to be able to use a more promiscuous ancestral protein as a template for protein engineering, i.e., to be able to improve on a property that is already present to some extent. The results of many ancestral resurrection studies support the prevailing view that ancestral forms may be less specialized than contemporary ones, e.g. the ancestral corticoid receptor apparently had affinity for both classes of hormones (glucocorticoids and mineralocorticoids), whereas the contemporary proteins are activated selectively by only one of these types of ligands ; similarly, ancestral β-lactamases showed activity towards a broader range of antibiotics .
Ancestral proteins could, however, have different phenotypes or catalyse entirely distinct reactions to modern proteins due to gene duplication events followed by sub-functionalization or via orthologous neofunctionalization. The Filozoan ancestor of the GK protein interaction domain appears to have had different biochemical functions, such as in spindle orientation, which play important roles in forming organized tissues in multicellular organisms and which have been lost in the evolution of guanylate kinase activity in the extant forms .
While ample studies indicate that ancestral forms with broadened specificity and modest levels of activity can serve as potential templates for further optimization, to date, examples of this approach are lacking (Gumulya et al., unpublished). Only the AMM has been used to alter the specificity of an enzyme by the incorporation of ancestral mutations . Understanding the change or loss of catalytic activities in ancestral proteins can be used as a guide for designing novel enzymes. Yet characterizing amino acid residues responsible for a switch in enzyme activity in a protein family can be a challenging task. To identify such mutations in the family of seminal ribonuclease, Sassi et al.  calculated the dN/dS ratio (ratio of non-synonymous to synonymous substitutions) for individual branches in the tree and detected episodes of positive selection that led to an increase in immunosuppressive activity in modern RNases. Regardless of the evolutionary model, the choice of out-group or the tree topology, only the branch leading to the modern seminal RNase in ox underwent adaptive evolution (high dN/dS ratio), and replacement of Gly 38 by Asp accounted for the increased catalytic activity against duplex RNA. Another approach used has been to calculate the rates of amino acid substitution at the sites that differ between two protein superfamilies. By selecting residues with low rates and hypothesizing that this characteristic indicated importance in the binding of substrates, Sun et al.  were able to identify that replacement at a single position (His 47 to Asn) is sufficient to cause a nearly complete swap in specificity between two supergroups of Pax genes. Combining a non-homogeneous γ-model (which allows site-specific rate shifts) and the available structural information for a protein allows identification of sites that are likely to be involved in the change of function . The DIVERGE software has been used to detect variability in the evolutionary rate associated with Type-I evolutionary functional divergence .
Unlike approaches for improving thermal stability and broadening substrate scope, currently there is no straightforward strategy implemented in enzyme engineering for altering an enzyme's cofactor preference, for example from the more costly nicotinamide adenine dinucleotide phosphate (NADP) to nicotinamide adenine dinucleotide (NAD) in oxidoreductases. Phylogenetic analysis of the IDH family suggested that the ancestral form of IDH used NAD(H), but a switch to NADP(H) occurred ∼3.5 billion years ago . Introducing three ancestral residues (Lys344Asp, Tyr345Ile and Val351Ala) in the binding pocket switched the cofactor specificity of E. coli IDH from NADP(H) to NAD(H), from a 7000-fold preference for NADP(H) to a 200-fold preference for NAD(H) . As the sequence divergence within the IDH enzyme family is very high, it was not possible for the authors to reconstruct the ancestor without significant ambiguity. Additional data including X-ray structures of two extant proteins, NADP bound to E. coli IDH and NAD bound to Thermus thermophilus isopropylmalate dehydrogenase, were used to align conserved residues in the sequence alignment. The study highlighted that with the addition of two substitutions outside the binding pocket (Cys201Met and Cys332Tyr), the ability of the altered cofactor IDH to catalyse the decarboxylation of isocitrate was as effective as naturally occurring, NAD-dependent members of the family. The fact that there is so far only this single example of cofactor switching reported in the literature using the AMM  indicates the difficulty in identifying two enzyme families that utilize different cofactors yet are not so divergent as to compromise the feasibility of ASR.
The first ancestral libraries were created to address the uncertainty issues in the inferences of ancestral sequences, by experimentally generating a ‘cloud’ of probable estimates of the true ancestral sequence. Ugalde et al.  targeted ambiguous sites in the ancestral sequences of GFP-like proteins to be varied, thus creating a library to sample alternative ancestral states predicted using different evolutionary models. Several groups have resurrected alternative ancestral sequences for positions for which the predictions are ambiguous, either with site-directed or saturation mutagenesis approaches, depending on the degeneracy to be introduced. However, there is no consensus on what constitutes an appropriate ‘uncertainty’ threshold . In one study, low-confidence sites have been defined as those that have a second plausible reconstruction with a posterior probability of 20% , whereas other studies assign ambiguity to sites when the posterior probability was <90 and 80%, respectively [9,115]. A more desirable approach would be to recreate experimentally the calculated posterior the distribution of ancestral sequences and characterize the phenotypic traits of a sample of the sequences . Unfortunately, the construction of a degenerate gene in which the variants are represented at unequal, predefined proportions represents a significant technical challenge. Currently, such libraries are not routinely available via commercial gene synthesis.
In more recent studies, ancestral libraries have been used as a way to generate ‘smart libraries’ for directed evolution. Ancestral libraries, in which phylogenetic analysis and ancestral inference have been used to choose sites that are important for functional diversification, could potentially distinguish amino acid replacements that are functionally important from those that are neutral, and thus increase the chances of finding improved variants in the screening process. Tawfik's group has used such a strategy to create ancestral libraries consisting of 300 variants of serum paraoxonases and cytosolic sulfotransferase, from which mutants with up to 50-fold higher activity could be obtained . Assuming the inference is robust, the combinatorial ancestral libraries widen the protein sequence space to be explored compared with that which can be attained in conventional directed evolution libraries (Figure 3). Additionally, ancestral libraries can also be used as a way to deconvolute combinatorial mutations that have led to functional diversification in protein families. Matz and colleagues  have successfully identified epistatic interactions that resulted in the evolution of the red phenotype from a green ancestor of GFP-like proteins by generating a library of ancestral intermediates, in which each variant contains half of the 37 amino acid changes that occurred between the two ancestors.
Conclusions and future perspectives
Despite continuous debate in the field concerning the accuracy of resurrected ancestral proteins in comparison with the true antecedents, and the best methods to use for inference, ASR has been used with increasing frequency by both evolutionary biologists to understand natural evolution processes and, more recently, by protein engineers to design novel enzymes. While it may never be possible to conclusively identify a single, historically accurate ancestor, approaches involving greater sampling of predicted sequences, such as the characterization of ancestral libraries that encompass the uncertainty in predictions, can indicate whether the inferences made about ancestral characteristics of a given protein family are likely to stand the test of time. Importantly, new methods are urgently needed that better manage the treatment of insertions and deletions, ideally in an alignment-independent fashion, and that can deal with larger numbers of sequences, so as to fully exploit the copious amounts of information becoming available from whole genome sequences.
The near future should see an expansion in the use of ancestral sequences as templates for directed, prospective evolution. Combined with the ever-growing collection of sequenced genomes and the refinement in computational methods for phylogenetic analysis, ASR opens the door to areas of sequence space that were not previously accessible to protein engineers. In particular, resurrection of ancestors may provide thermostable forms of proteins that are robust to mutagenesis. Whether this is an artefact of the methods used for inference or reflects a true ancestral characteristic is somewhat irrelevant for protein engineering. While objectives such as engineering a switch in cofactor specificity may remain challenging to solve by ASR due to the evolutionary distance between homologous protein sequences, altering the substrate and reaction specificity of ancestral proteins, by directed evolution or conventional protein engineering, should be more tractable.
Ultimately ASR has the potential to allow biochemists to look, Janus-like, into both the past and future of protein evolution. The prospect of resetting the evolutionary clock using inferred ancestors and rerunning evolutionary ‘experiments’ under altered selection pressures should provide insights into how proteins have evolved in nature and also how we can better direct artificial evolution so as to produce proteins that are useful for the future.
The Authors declare that there are no competing interests associated with the manuscript.
The authors would like to thank Drs. M. Boden, V. L. Arcus and J. W. Thornton for helpful discussions during the preparation of this review.
Abbreviations: AMM, ancestral mutation method; ASR, ancestral sequence reconstruction; CDK, cyclin dependent kinase; EF, elongation factor; GASP, gapped ancestral sequence prediction; IDH, isocitrate dehydrogenase; MAPK, mitogen activated protein kinase; ML, maximum likelihood; MP, maximum parsimony; MSA, multiple sequence alignment; Mya, million years ago; NAD, nicotinamide adenine dinucleotide; NADP, nicotinamide adenine dinucleotide phosphate; NCBI, National Center for Biotechnology Information; PAML, phylogenetic analysis by maximum likelihood; PAUP, Phylogenetic Analysis Using Parsimony; PDB, Protein Data Bank.
- © 2017 The Author(s); published by Portland Press Limited on behalf of the Biochemical Society