Enzymes are astoundingly good catalysts: they allow reactions to occur billions of times faster than would be possible without them, at temperatures much lower than those required by typical synthetic catalysts. But enzymes have evolved to accelerate only biological reactions, under the narrow set of conditions that are compatible with life. Two papers from the same group, one in this issue (Röthlisberger et al.1, page 190) and another in Science (Jiang et al.2), show how these limitations can be overcome. They describe a method for designing enzymes that catalyse unnatural reactions, and demonstrate its use for two different chemical transformations.

Enzymes work by lowering the activation energy of reactions, specifically by confining substrates in binding sites that stabilize the highest-energy arrangement of atoms in the reaction pathway (known as the transition state). They also shield the reactants, thus preventing possible side reactions. The idea behind the latest work1,2 is simple — model the transition state for a reaction, stabilize it by surrounding it with carefully placed chemical groups, graft the resulting active site into an existing protein and then alter the amino-acid sequence of the protein to accommodate the changes. In practice, this is a complicated procedure. For starters, building an accurate model of a transition state requires a detailed understanding of the reaction's mechanism, which isn't always available. Furthermore, transition states are modelled using quantum-mechanical calculations, but currently available methods can handle only a limited number of atoms, and are often inadequate for modelling enzyme reactions.

Designing a protein that folds into a given structure is equally challenging. For a protein made of 100 amino acids, there are about 10130 possible sequences, each of which can adopt many different conformations. The thermodynamic stability of every sequence and conformation must therefore be calculated to find the lowest-energy structure (that is, the one most likely to form). Some simplifications can be made using advanced computational methods to quickly eliminate unfavourable combinations. This has resulted in several notable accomplishments, such as the complete redesign of a protein consisting of 28 amino acids3, the design of an amino-acid sequence that forms a structure not found in nature4, and the engineering of naturally occurring proteins into biosensors for trinitrotoluene (TNT) and other small molecules5.

With these precedents, you might think that designing catalytic proteins should be straightforward, but success has been limited. Catalytically inactive proteins have been converted into modestly catalytic ones for two different reactions, but the observed enhancements of rate6,7were only about a millionth of those produced by naturally occurring enzymes. It is also sometimes difficult to prove that designer enzymes are truly catalytic on the basis of biochemical observations, and some exciting claims have been found to be flawed.

But some reports of catalysis by designed enzymes have fared rather better — especially those that are based on sound crystallographic evidence6,7,8. An essential step in demonstrating the success of a designer enzyme, therefore, is the determination of a high-resolution crystal structure for the protein, to verify that the designed catalytic features are present. The results of Röthlisberger et al.1 and Jiang et al.2 are remarkable in the spectacular agreement between their computationally predicted enzyme models and the experimentally determined structures (Fig. 1).

Figure 1: Enzymes by design.
figure 1

Röthlisberger et al.1 have computationally designed and prepared the first enzyme capable of catalysing a non-biological reaction. Here, the computational model (grey) is overlaid with the crystal structure of the actual protein (green); the two overlap almost perfectly. The substrate is shown at the centre of the structure. The design process involved modifying the amino-acid sequence of a naturally occurring protein. Residues selected computationally to form the active site are shown as purple spheres. Additional mutations that were introduced in vitro to optimize the enzyme's performance are shown as green spheres.

Röthlisberger et al.1 made an enzyme that catalyses the Kemp elimination reaction (see Fig. 1a on page 190 for a reaction scheme). The Kemp elimination is initiated by the removal of a hydrogen ion from a carbon–hydrogen bond in the substrate; the minimum requirement for catalysis of the reaction is the presence of a base to perform this step. The authors therefore identified two amino acids — aspartic acid and histidine — that have side chains that can act as bases under physiological conditions, and used these as the starting points of their putative active sites. They decorated models of the proposed active sites with other chemical groups found in proteins, choosing those that could interact favourably with groups in the substrate. They then used state-of-the-art quantum-mechanical methods to precisely place all the groups in the models to maximize stabilization of the transition state of the substrate. The authors thus obtained a large ensemble of designs for catalytic sites in enzymes.

Next, Röthlisberger et al. selected about 100 proteins that could be used as scaffolds for their proposed active sites. The criteria for selection were the availability of high-resolution crystal structures and the presence of pre-organized cavities, with a preference for proteins that behave well in experiments (that is, those that have good solubility, are expressed easily in cells, and so on). The authors then used computational methods to search each of the proteins for specific regions that could accommodate the sites, narrowing down the vast number of possibilities to about 100,000 promising leads. These were whittled down further using an automated modelling technique to find the optimal amino-acid sequence in defined shells around the active site, selecting sequences that maintained protein stability and integrity.

This computational screening method picked out 59 candidate enzymes, which the authors expressed in cells and evaluated for their ability to catalyse the target reaction. Only eight of the proteins had measurable catalytic activity. The team then used in vitro evolution to further optimize one of their successful leads (designated KE07), mutating the amino-acid sequence in both random and directed locations. After several rounds of mutation and screening, Röthlisberger et al. obtained improved enzymes that were up to 200 times more active than KE07. The best two of these mutants accelerate the rate of the Kemp elimination reaction to about a million times that of the uncatalysed version.

The strategy used by Röthlisberger et al.1 promises to be general, as the same group2 has successfully applied the procedure to another chemical transformation known as the retro-aldol reaction, which is very different from the Kemp elimination. The complexity of the design procedure is underlined by the number of interdisciplinary groups involved in the work, and by the huge amount of computational power required to solve the problem — donated from hundreds of thousands of idling computers around the world as part of a project known as Rosetta@home9.

Those in the know might say that the performance of the designed enzymes is far from impressive — the reaction-rate enhancements for typical, naturally occurring enzymes are anywhere between 10,000 and 1 billion times higher than those of the artificial enzymes described in these papers1,2. Furthermore, the chosen reactions are relatively easy targets. The Kemp elimination is accelerated by several catalysts, including various synthetic compounds, catalytic antibodies and even serum albumin. Similarly, the retro-aldol reaction is catalysed by antibodies10 and by various peptides11,12. Indeed, the rate enhancements reported by Röthlisberger et al.1 are equivalent to those of only the most sophisticated catalytic antibodies13,14; the enhancements obtained by Jiang et al.2 for the retro-aldol reaction are even more modest.

Another limitation of the design process is that, although naturally occurring enzymes have evolved to optimize steps other than just catalysis (such as the binding of substrates and the release of products), the model used by the authors1,2 to design their enzymes doesn't attempt to address these factors. This is understandable, because many of the finer features that provide enzymes with their unique properties are not yet understood. For example, the mutations introduced by the authors into their enzymes by directed evolution did not modify the active site itself, but occurred at neighbouring positions (Fig. 1). The effect of some of these mutations can be easily understood with hindsight, but others are much less obvious. It was therefore wise of the authors to let nature lend a helping hand in their designs.

Nevertheless, these results1,2 are a milestone in biochemistry. For the first time, artificial enzymes have been designed for non-biological reactions, providing rate accelerations that are about 1,000 times faster than previous examples of computationally designed enzymes. Biochemists have long wanted to build artificial enzymes to identify and validate the minimal requirements for enzyme-like catalysis. These reports provide an accurate framework for this enterprise to which further features can be added. As Röthlisberger et al.1 note, the ability to design enzymes will truly test our understanding of enzyme catalysis.