Introduction

Proteins, when properly configured, have a remarkable capacity for carrying out complex molecular processes with extreme precision and efficiency. If humans could master protein engineering at this level, the implications would be staggering. Where do we stand in this learning process? The answer depends on our point of reference. Relative to the starting point, it is important to ask how far we have come and to keep track of significant accomplishments. To that end, there has been a steady succession of thorough reviews covering all aspects of protein design, selection, and modification (Bolon et al. 2002; Cherry and Fidantsef 2003; Khersonsky et al. 2006; Hibbert and Dalby 2005; Johannes and Zhao 2006; Rubin-Pitel and Zhao 2006; Wong et al. 2006; Butterfoss and Kuhlman 2006; Pleiss 2006; Bommarius et al. 2006). Equally important, though, is an assessment of progress from the other point of reference—the end point. From this perspective, the question becomes: How far have we come? This review will aim to answer this and to assess the most significant obstacles to progress.

Structural or mechanical functions of proteins depend both on the overall shape and on the material properties (e.g., rigidity, elasticity, and adhesion) in ways that enzymatic functions typically do not. Conversely, enzymatic functions are uniquely dependent on local geometry and chemical environment within their active sites. A number of other categorical distinctions present themselves, such as between the mechanical roles of extended fibrous structures like that of collagen and those of globular structures like flagellin or between such disparate nonmechanical roles as photon capture, electron conduction, ion transport, chemical catalysis, signal transduction, and chaperone-assisted protein folding. The physics of these processes is so varied that we should expect considerably different design rules to apply. Enzymes pose a particular challenge, in that all the action appears to be happening in a small part of the structure. It is much easier to understand the roles of a handful of active-site residues than it is to understand the functional requirements of the scaffold that gives shape to the active site. In this review, we shall concentrate mainly on enzymes.

Competing engineering principles: Diversity and specificity

Ultimately, the objective is to make proteins perform for us as well as they perform in life. The variety of methodological approaches aiming to achieve this may be categorized, roughly, along a spectrum defined by two extremes (Fig. 1). At one end is an approach commonly referred to as a rational design, which aims to understand the principles of protein structure and function well enough to apply them in designing new properties or even novel proteins using de novo design. The value of this approach in purely scientific terms is indisputable. However, because the difficulty is likewise indisputable, any approach that might succeed sooner is worth exploring. That realization has motivated work at the other end of the spectrum, where the emphasis is on finding what works rather than predicting what works. Darwinian evolution is the inspiration behind this. In the extreme form, this means avoiding protein design principles altogether and relying instead on huge sequence libraries and carefully designed selection methods.

Fig. 1
figure 1

Rational design goes from function to structure to sequence, while random approaches try to create a sequence that forms a structure producing the desired function

Full size image

Random search

An average-sized protein has about 300 amino acids and can be put together in an enormous number of different ways. The number of different possibilities is beyond comprehension (20300). This enormous sea of possibilities forms the sequence space. The idea that life-like properties might be reasonably common among random polypeptides dates back to the 1950s, with the work of Sidney Fox on proteinoid microspheres (Fox 1980). Where does this idea stand now, 50 years later?

The most prominent recent study arguing this position is that of Keefe and Szostak (2001). Using a powerful in vitro selection method that links protein chains to their encoding mRNA (Roberts and Szostak 1997), an initial library of nearly 1013 protein sequences was cycled through eight rounds of selection on an ATP-affinity column. Then, after several rounds of mutagenesis and 17 more cycles of affinity selection, the authors ended up with a protein that shows significant ATP-binding activity despite having only marginal structural stability. A shortened version of that protein, called artificial nucleotide-binding protein (ANBP), was found to have sufficient structural integrity for crystallization, resulting in the first reported structure of a protein with a random-library pedigree (Lo Surdo et al. 2004). However, when compared to ATP-binding enzymes, the ANBP fold has very limited potential as an enzyme, even considering the possibility of amino-acid substitutions near the binding site.

A good example of the great distances between functional islands in sequence space comes from the studies with β-lactamases. A natural variation of β-lactamase has occurred in response to different penicillin derivatives, cephamycins, and four generations of cephalosporins. Resistance to all the different types of antibiotics is the result of only 13 point mutations of 290 residues (Orencia et al. 2001). When structural and genetic variation among β-lactamases from different sources was analyzed, it was concluded that out of 10199 possible structures, 10122 are active (Axe 2004). Thus, only 1 out of 1077 structures is functional.

Functional native-like proteins seem thus to be very rare in random sequence libraries, and if there are proteins that show folding, solubility tends to be a problem (Doi et al. 2005). Even when randomly generated, sequences were fused 50/50 to an N-terminal half of the cold shock protein CspA, a very low amount (1 in 107) of folded proteins was obtained (Riechmann and Winter 2000). In conclusion, a fully random search is hardly a way to create novel proteins for biotechnological use. This does not, however, prevent us from utilizing a random search in focused areas.

Randomizing within a local sequence space

Protein engineers can point to solutions employed in life, but we have hardly begun to ask how it is that one protein fold lends itself to a certain class of applications but not to others. We have categories for fold shapes but neither like a theory of structural mechanics or dynamics for proteins nor a theory of active-site design for enzymes; although, we have information on the reaction chemistry of the active site of enzymes. Justifiably, the first priorities have been the structural and functional characterizations of biological proteins and protein complexes. At this point, though, data collection has progressed far beyond data assimilation, making the need of a theory increasingly important. This means that we do not have a theoretical framework for identifying the right protein fold to accomplish a new functional objective. This is the reason why most enzyme engineers prefer directed evolution approaches in improving enzyme activities.

Random or directed evolution methods involve a set of techniques to create random changes in a protein structure by genetic methods and selection of new protein variants. The gene of interest is diversified through mutations, and the created library of mutated genes is tested against a specific selection pressure, for example a particular property of a protein. The best variants are iteratively subjected to a new round of mutations. The selection is directed towards a desired activity, hence the name-directed evolution. The diversity of the library can be produced with random point mutations, recombination of genes (homologous or nonhomologous), or a combination of both. Arnold (2007) describes the requirements for successful directed evolution as follows:

  1. 1.

    The desired function must be physically possible.

  2. 2.

    The function must also be biologically or evolutionarily feasible. In practice, this means that there exists a mutational pathway to get from here to there through ever-improving variants.

  3. 3.

    You must be able to make libraries of mutants complex enough to contain rare beneficial mutations.

  4. 4.

    You must have a rapid screen or selection that reflects the desired function.

Requirement no. 2 (There must be a mutational pathway through ever-improving variants) limits the search to a functional/structural island. Therefore, the starting point is always a functional protein. Bearing this in mind, some impressive practical achievements (Table 1) have been done using directed evolution methodologies.

Table 1 Examples of protein engineering
Full size table

Enzymes by design

A rational design of an enzyme catalyst from scratch is a demanding task because the catalyst must have a stable soluble scaffold, and it has to be able to bind and orient both substrates and transition states to catalytic groups. Recently, some remarkable successes have been reported. Kaplan and DeGrado (2004) succeeded in creating de novo-designed diiron proteins with phenol oxidase activity from helical peptides using computational methods and screening of a large number of variants. Kuhlman et al. (2003) used computational strategy that iterated between sequence and structure prediction and designed a stable 93-residue α/β protein fold not found in nature. Both of these studies demonstrate that it is becoming possible to create de novo protein designs, although the created structures still fall far behind the properties of life-like enzymes. Walter et al. (2005) constructed an active enzyme from only nine different amino acids, and de novo proteins have been found from a designed combinatorial library with a quite high frequency (e.g., Wei et al. 2003). Hecht et al. (2004) reviewed the progress in de novo protein design. As interesting as these developments are, practical applications most likely require much more work than the use of traditional methods starting from already functional enzymes.

Another way forward is the rational redesign approach, which means putting a new function on a natural scaffold. The TIM barrel, which was first discovered in triose phosphate isomerase, is a good example of this approach (Sterner and Höcker 2005). About 10% of all known protein structures contain at least one TIM barrel domain in which the active site is formed of or surrounded by connecting loops from eight beta strands and eight helices. Using a computer-based rational redesign, Dwyer et al. (2004) turned ribose-binding protein into a triose phosphate isomerase enzyme, and Bolon and Mayo (2001) redesigned a catalytically inert 108-residue thioredoxin scaffold to function as a p-nitrophenyl acetate hydrolyzing enzyme. Despite of some successes in altering enzyme activities, it is often very difficult to attain the results predicted by the rational design. Corey and Corey (1996) listed several failed attempts to produce de novo-designed biocatalysts.

Rational design has been mainly used in biotechnology to improve the properties (especially thermostability) of natural enzymes. The analysis of proteins for rational design involves sequence comparison of the protein family, which gives information about the structurally important amino acids and variability at each amino acid site. In engineering, e.g., thermostability, earlier mutational studies, comparison of amino acid sequences, total amino acid content, and crystal structures between mesophilic and thermophilic enzymes may give information on the key factors behind the elevated stability (Hakulinen et al. 2003; Eijsink et al. 2004). When the protein structure is known, computer simulations can be used to study the active site properties, substrate binding, thermostability, and unfolding of the enzymes. Simulations provide information that is useful in planning mutations. For example, molecular dynamics simulations have been used to identify flexible regions in proteins (Daggett and Levitt 1993; Pikkemaat et al. 2002), and subsequently, the protein stability has been increased considerably by introducing a disulfide bridge into such a region.

One of the simplest stabilization methods is actually the construction of a disulfide bridge into a protein structure. It has to be done by rational design because random mutations are not likely to form the required simultaneous double mutation. The thermostability problems of an industrial enzyme have been solved efficiently by this method (Fenel et al. 2004), resulting in over 20°C thermostability increase when two disulfide bridges and other mutations were combined (Xiong et al. 2004). One of the most impressive thermostability increases has been obtained by a single point mutation (Glu→Gln); the thermostability of triosephosphate isomerase increased by 26°C (Williams et al. 1999).

A new semirational method to improve enzyme stability is to make a consensus sequence to the protein family. This approach is based on the assumption that conserved amino acid properties have been selected in nature because of their impact on protein stability. The consensus method efficiently explores the local sequence space. About 30°C increase in thermostability has been achieved by this method (Lehmann and Wyss 2001). In limited areas, rational design methods can be very effective in improving the properties of industrial enzymes.

Engineering the protein fold

Basic structural scaffolds (on the order of few thousands) represent only a tiny fraction of the sequence space. Although alteration of enzyme function has become a routine operation, the reshaping of the active center by modifying the basic fold is a different story. Very few experiments have been carried out trying to change one fold or basic scaffold to another.

Blanco et al. (1999) studied experimentally the sequence space between two different small proteins having different folds. One was a 62-amino acid protein that folds as an eight-stranded orthogonal β-sheet sandwich and another was a 57-amino acid protein that has a central α-helix packed against a four-stranded β-sheet. The authors designed a gradual series of mutants in trying to understand whether there would be an evolutionary path from one fold to another. The conclusion of their study was that the sequence space between the two proteins is enormous. The results suggested that only a small fraction of this space would have adequate properties for folding into a unique structure. The sequence spaces of the two small proteins did not overlap and a change from one fold to another could not be reached within a valid evolutionary trajectory.

However, there are some studies that demonstrate a fold change as a result of a single or few mutations. Cordes et al. (2000) described a single point mutation that changed a part of a small 39-amino acid protein fold from β-sheet to α-helix. However, it is not evident what the selective advantage of a mutation causing structural heterogeneity in natural environment would be. Structural intermediate is not necessarily the same thing as evolutionary intermediate. Meier et al. (2007) studied a small 27-amino acid cysteine-rich polypeptide. The secondary structure of this small peptide is limited to a single helix turn, and the structure is held together by two disulfide bonds. Alternative disulfide patterns can form different folded structures in this type of atypical polypeptide; the formation is controlled by a key amino acid site. Such abrupt folding changes are more likely in small peptides than in large proteins, in which “misfolding” typically leads to a nonfunctional protein.

Alexander et al. (2005) studied two different bacterial IgG binding domains that were nonhomologous and had different folds but were similar in size and function. Their sequence identity is 14%. By directed evolution approach, the authors changed the sequence of one of the domains so that the sequence identity was finally 59%. Both domains, however, still retained their original fold, although the binary sequence space separating these two domains had decreased almost eight orders of magnitude. The authors concluded that the folding information must be in the nonidentical parts of the domains.

A change in folded structure is sometimes possible through a short walk in the sequence space. When occurring at the posttranslational level, this kind of event is usually called miss-folding. To get something meaningful from genetically guided “miss-folding,” it must give selective advantage to the organism. At the moment, we can only expect that huge sequence libraries are needed to find novel functions through fold change events.

Obstacles in protein engineering

In view of the very substantial challenges remaining and the considerable effort expended thus far, we should pause to ask what things are most impeding our progress. Questions of this kind always reveal diversity of opinion, which is not a bad thing. In offering our opinions here, we hope merely to stimulate an important discussion that might enhance our collective progress. With that in mind, we suggest the following as the most significant obstacles to be overcome:

  1. 1.

    Lack of a theory for structure design (i.e., specifying a structure that is suited to a target function)

  2. 2.

    Lack of a general approach for sequence design (i.e., specifying a sequence that folds to a target structure)

  3. 3.

    Overreliance on the Darwinian methodology

There is information about the rules on how the catalytic amino acid residues have to be located in three dimensions for the enzymatic reaction to happen. This knowledge can be used to design active sites (Dwyer et al. 2004). Similarly, we have some ideas of protein scaffolds into which an active site can be grafted. Despite of all the progress, we are in the very beginning in designing de novo active enzymes. Thus, we are still missing general theories that would help us to design novel enzymes without a need to use methods that are based on a random search in the local sequence space.

It is often said that random genetic methods to improve enzyme properties “rely on simple but powerful Darwinian principles of mutation and selection” (Johannes and Zhao 2006). We agree. It is also said that “every protein has become adapted by step-by-step improvement and refinement of its function over millions of years” (McLachlan 1987). The present theories, however, only partly explain the protein diversity, although a recent study (Poelwijk et al. 2007) shows that even the key-lock dilemma can be resolved by the Darwinian approach when the operation field for random search is within the same protein family, and the new key-lock pair closely resembles the original (ancestral).

As discussed above, the transition from one fold to another is very problematic. Even inside a fold family, most transitions are very challenging. For example, family 10 xylanase and xylose isomerase (XI) have a TIM barrel fold, but the active site of xylanase is open and large, whereas the active site of XI is buried inside the protein so that the loops between secondary structure elements cover the active site. In xylanase, a long filament has to fit into the active site, and in XI, the isomerization of sugars has to happen in an environment in which the amount and positioning of water are controlled. These two examples demonstrate the wide diversity potential of the TIM barrel fold. It is also evident that evolutionary transitions in this fold family can be very complex. Development of loop grafting libraries might produce interesting results in TIM barrel scaffold.

Gene duplication

Gene duplication and subsequent divergence as mechanisms to create natural variety and novel structures are now decade’s old theories (Ohno 1970). Basically, directed evolution approach is an application of the gene duplication concept. Gene duplication is seen as a way to avoid random sequences in evolution, because random sequences most often are not functional. Mutations in the duplicated genes explore the local sequence space and expand the number of members in a gene family.

It appears, however, that the full potential of duplicated genes are not effectively explored in nature, as only a small number of duplicated genes do not experience a dead end (Lynch and Conery 2000). Positive selection may be needed to retain the duplicated genes in the population (Ohta 2002; Kondrashov and Kondrashov 2006; Shiu et al. 2006). Divergence after gene duplication, sometimes under lowered selection pressure, may create new activities via amino acid substitutions or small deletions and insertions. However, small amino acid changes are not likely to form a new protein fold (Bogarad and Deem 1999). This can be easily understood as even after, e.g., 70% divergence, the protein is still likely to retain its original fold (Chothia and Lesk 1986). Park et al. (2006) have demonstrated that the task is not easy even for a protein engineer; even minor changes in the scaffold need huge design efforts. Thus, gene duplication methods can be used to find novel solutions but probably mainly inside a gene family.

Exon shuffling

Exon shuffling (and shuffling of protein blocks) is often considered as a more powerful evolutionary method to create novel proteins than gene duplication. Exon shuffling basically applies a gene duplication concept but in smaller units. It seems that in practice, single units (exons or protein blocks) must be fused to a target gene sequentially and not several exons from different sources in a single event (Peisajovich et al. 2006). Too radical changes disrupt the protein function (and protection by selection) or cause damage in the genome. Because of these reasons, the exon shuffling process has to occur in reality so that the overall protein fold does not change.

DNA shuffling methods based on homologous recombination keep the resulting hybrid genes inside the original gene family. Nonhomologous recombination, like exon (or protein block) shuffling can, in principle, create proteins with novel folds, provided that the new proteins have such an activity that a screening/selection method can detect them. Nonhomologous random recombination has been studied in vitro, e.g., with chorismate mutase (CM; Bittker et al. 2004). Functional enzyme variants contained insertions, deletions, and rearrangements. The authors also randomly recombined CM with fumarase—another α-helical protein. The resulting active CM proteins contained the original CM core and fumarase sequences only at the termini or one loop. This experiment demonstrates that the destruction of the stable core leads to the loss of activity supporting the view that protein scaffold can change mainly by incremental modifications of the fold. DNA shuffling methods based on nonhomologous recombination may find fully novel functions, but apparently, a very large number of variants have to be screened. This creates a practical limit that may not be easily broken with current methods.

Novelty through promiscuous activities

Enzymes are not always too accurate in their substrate specificity. Pastinen et al. (1999), for instance, showed that the industrially used glucose (xylose) isomerase has a number of side activities isomerizing a large number of both natural and rare hexoses and pentoses. Such weak side activities can be improved by design (Karimäki et al. 2004) or by random methods. It has been proposed that in nature, a new function is evolved from an initially promiscuous activity (Aharoni et al. 2005). Khersonsky et al. (2006) reviewed the present views of divergence of today’s enzymes from ancestral proteins catalyzing a whole range of activities at low levels. A promiscuous activity can be a promising starting point for directed evolution methods to produce biotechnologically relevant enzymes (Bornscheuer and Kazlauskas 2004).

A way forward: hybrid approaches

Practical experience shows that directed evolution can produce remarkable changes that are, at present, not easily achieved by rational design. However, these methods have their limits as discussed above. Furthermore, selection and screening in directed evolution methods require that each step improves the enzyme. When the desired change involves a simultaneous change in several amino acids, it is not likely to be reached by the random approach (Behe and Snoke 2004). In many cases, a combination of design to create the needed structure or function and its improvement by random techniques is a better approach. As Taylor et al. (2001) put it: “The low frequency of protein catalysts in sequence space indicates that it will not be possible to isolate enzymes from unbiased random libraries in a single step.” Recently, Park et al. (2006) changed a metallohydrolase through designed deletion and insertion of several structural loops in the active site to form a new enzyme with a different catalytic function and then applied random techniques to improve the designed activity.

To focus mutations near the active site when trying to change the catalytic properties of an enzyme appears to be more effective than distant mutations (Morley and Kazlauskas 2005). Yoshikuni et al. (2006) used saturation mutagenesis and systematic recombination approach inside or near the active site to expand the synthetic capacity of γ-humulene synthase to produce different sesquiterpenes. The engineered enzyme variants used different reaction pathways to synthesize a variety of reaction products while maintaining the specific activity of the original enzyme. The remarkable results in changing the active site properties by a blind search in a limited number of amino acids indicate that there exists a number of biotechnologically interesting enzyme variants beyond the sequence space that can be easily reached by the directed evolution method in which the whole gene is a target, and tens or hundreds of thousands of variants has to be screened. Computational methods may also be increasingly used in reducing the sequence space that is screened by high-throughput screening methods (Voigt et al. 2001; Hayes et al. 2002).

Bloom et al. (2006) recently demonstrated that the stability of the protein scaffold improves its evolvability. In other words, when the stability of a protein scaffold is increased, it may be possible to make mutations that create a truly novel property, which without increased stability would be too harmful for the protein. This approach can be very promising in finding novel enzymatic solutions that otherwise cannot be detected in the functional screening.

Conclusions

Twenty years of protein engineering has resulted in an impressive array of genetic engineering and computational tools as well as several concrete results in modifying and improving the properties of enzymes. We have learned to understand, on one hand, the rarity of functional proteins in the sequence space and on the other hand, the large variation potential of biotechnologically relevant protein functions. We only need methods to dig them out. We know superficially how a large number of life-like proteins look like, but we are still far from understanding how life-like proteins are designed and even further from being able to design life-like nonnatural proteins.

In spite of the progress, we still do not have a general theory on how a sequence produces a specific structure and how a structure determines a function. Therefore, a blind Darwinian search within a known protein scaffold is often used to modify proteins. Unfortunately, blind searches have hard resource limits whereas insight has not. Therefore, in the long run, blind searches are of limited value in compensating our present ignorance. We still have a long way to go before we are able to design a suitable protein scaffold, position binding, and catalytic groups correctly into this scaffold and optimize the designed protein for life-like efficiency.