In making his case for the role of natural selection inevolution, Darwin started by pointing to the enormousphenotypic variation that could be achieved in just a fewgenerations of artificial selection. In Darwin’s day, theimportance of good breeding practices in determining theproductivity and quality of a farmer’s stock or the cropsize of a pigeon fancier’s prize bird was clear to all, andDarwin’s artificial selection arguments provided a power-ful foundation for his idea that competition for limitedresources could similarly tailor phenotypes, and ultimatelycreate new species, by selecting for beneficial traits.

Today, we can use artificial selection to breed not justorganisms, but also the protein products of individualgenes. By subjecting them to repeated rounds of mutationand selection (a process usually referred to as “directedevolution”), we can enhance or alter specific traits andeven force a protein to acquire traits not apparent in theparental molecule. And, just as in Darwin’s day, these arti-ficial selection experiments have the potential to teach usa great deal about evolution, only now at the molecularlevel. The remarkable ease with which proteins adapt inthe face of defined selection pressures, from acquiring theability to function in a nonnatural environment to degrad-ing a new antibiotic, was largely unexpected when the firstlaboratory protein evolution experiments were performedtwo decades ago.

Directed evolution experiments can recapitulate differ-ent adaptive scenarios that may at least partially character-ize natural protein evolution. But perhaps even moreinteresting is the opportunity to go where nature has notnecessarily gone. Under artificial selection, a protein canevolve outside of its biological context. This allows us toexplore the acquisition of novel features, including thosethat may not be useful in nature. In this way, we can dis-tinguish properties or combinations of properties that arebiologically relevant and found in the natural world fromothers that may be physically possible but are not relevantand not easily encoded, and therefore are not encountered,in natural proteins.

Darwin’s great insight long predated any understand-ing of the molecular mechanisms of inheritance and evo-lution. How DNA-coding changes alter protein functionis new information that directed evolution experimentscontribute to the evolution story. With access to the entire“fossil record” of an evolution experiment, we can deter-mine precisely how gene sequences change during adap-tation and can connect specific mutations to specificacquired traits. During the last 20 years, directed evolu-tion experiments have revealed that useful propertiessuch as catalytic activity or stability can frequently be en -hanced by single-amino-acid substitutions and that sig-nificant functional adaptation can occur by accumulationof relatively few such beneficial mutations (changing aslittle as 1%–2% of the sequence). This contrasts with thelarge sequence distances—frequently 50% or more—thatseparate natural protein homologs, which have di vergedand adapted to different functions or en vi ron ments.Directed evolution can identify minimal sets of adaptivemutations, but the precise mechanisms by which adapta-tion occurs are still difficult to discern: The individualeffects of beneficial mutations are usually quite small,and their locations and identities are often surprising(e.g., distant from active sites).

Directed evolution experiments have also elucidated akey feature of the fitness landscape for protein evolution.A common expectation has been that mutational path-ways to new properties would be tortuous, reflecting a fit-ness landscape that is highly rugged. In fact, laboratoryevolution experiments have demonstrated over and overagain that smooth mutational pathways—simple uphillwalks consisting of single beneficial mutations—exist andlead to higher fitness. Many interesting and useful pro per- ties can be manipulated by the accumulation of beneficialmutations one at a time in iterative rounds of mutagenesisand screening or selection.

The role of neutral mutations in protein evolution hasalso been explored. Directed evolution has demonstratedan important mechanism whereby mutations that are func-

How Proteins Adapt: Lessons from Directed Evolution

F.H. ARNOLDDivision of Chemistry and Chemical Engineering, California

Institute of Technology, Pasadena, California 91125Correspondence: frances@cheme.caltech.edu

Applying artificial selection to create new proteins has allowed us to explore fundamental processes of molecular evolution.These “directed evolution” experiments have shown that proteins can readily adapt to new functions or environments via sim-ple adaptive walks involving small numbers of mutations. With the entire “fossil record” available for detailed study, theseexperiments have provided new insight into adaptive mechanisms and the effects of mutation and recombination. Directedevolution has also shown how mutations that are functionally neutral can set the stage for further adaptation. Watching adap-tation in real time helps one to appreciate the power of the evolutionary design algorithm.

Cold Spring Harbor Symposia on Quantitative Biology, Volume LXXIV. ©2009 Cold Spring Harbor Laboratory Press 978-087969870-6 41

tionally neutral but stabilize the protein’s three-dimen-sional structure can set the stage for further adaptation byproviding the extra stability that allows functionallyimportant but destabilizing mutations to be accepted. Bythis mechanism, stability contributed by functionally neu-tral mutations promotes evolvability. In addition, it hasbeen demonstrated that accumulating mutations which areneutral for one function can lead to the appearance of oth-ers—a kind of functional “promiscuity”—that can serveas a handle for evolution of new functions such as the abil-ity to bind a new ligand or to catalyze a reaction on a newsubstrate.

The molecular diversity on which artificial selection actscan be created in any number of ways in order to mimic nat-ural mutagenesis mechanisms: Directed evolution experi-ments use random (point) mutagenesis of a whole gene ordomain, insertions, and deletions, as well as other, morehypothesis-driven mutagenesis schemes. Another impor-tant natural mutation mechanism is recombination. Wehave explored how recombination can contribute to mak-ing new proteins, by looking at its effects on folding andstructure as well as function. Recombination of homolo-gous proteins is highly conservative compared to randommutation—a protein can acquire dozens of mutations byrecombination and still fold and function, whereas similarlevels of random mutation lead to loss of function. Al -though the mutations made by recombination are less dis-ruptive of fold and function, they can nonetheless generatefunctional diversity. Experiments have shown that recom-bined, or “chimeric,” proteins can acquire new properties,such as increased stability or the ability to accept new sub-strates, through novel combinations of the mostly neutralmutations that accumulated during natural divergence ofthe homologous parent proteins.

In the remainder of this chapter, I describe how directedevolution experiments performed in this laboratory on amodel enzyme, a bacterial cytochrome P450, have pro-vided support for these lessons. This is a personal account,and I apologize in advance for making no attempt to coverthe large relevant literature and contributions from otherlaboratories.

DIRECTED EVOLUTION: A SIMPLEMOLECULAR OPTIMIZATION STRATEGY

Directed evolution starts with a functional protein anduses iterative rounds of mutation and selection to searchfor more “fit” proteins, where fitness is defined by theexperimenter via an assay or some other test (e.g., agenetic selection). The parent gene is subjected to muta-tion, and the mutants are expressed as a library of proteinvariants. Variants with improved fitness are identified,and the process is repeated until the desired function isachieved (or not). Directed evolution usually involves theaccumulation of beneficial mutations over multiple gen-erations of mutagenesis and/or recombination, in a simpleuphill walk on the protein fitness landscape (Romero andArnold 2009).

Directed evolution relies on proteins’ abilities toexhibit a wider array of functions and over a wider range

of environments than might be required for their biologi-cal functions. This functional promiscuity, even if only atsome minimal level, provides the jumping-off point foroptimization toward that new goal. A good starting pro-tein for directed evolution exhibits enough of the desiredfunction that small improvements (expected from a singlemutation) can be discerned reliably. If the desired behav-ior is beyond what a single mutation can confer, the prob-lem can be broken down into a series of smaller ones,each of which can be solved by the accumulation of sin-gle mutations, for example, by gradually increasing theselection pressure or evolving against a series of interme-diate challenges.

Epistatic interactions occur when the presence of onemutation affects the contribution of another. These nonad-ditive interactions lead to curves in the fitness landscapeand constrain evolutionary searches. Mutations that arenegative in one context but become beneficial in anotherare a ubiquitous feature of protein landscapes, where theycreate local optima that could frustrate evolutionary opti-mization. Directed evolution, however, does not find allpaths to high fitness, only the most probable paths. Thesefollow one of many smooth routes and bypass the morerugged, epistatic routes. Hundreds of directed evolutionexperiments have demonstrated that such smooth paths tohigher fitness can be found for a wide array of protein fit-ness definitions, including stability, ability to function innonnatural environments, ability to bind a new ligand,changes in substrate specificity or reactivity, and more(Bloom and Arnold 2009).

CYTOCHROME P450 BM3: A MODEL ENZYME FOR DIRECTED EVOLUTION

The cytochrome P450 enzyme superfamily provides asuperb example of how nature can generate a whole spec-trum of catalysts from a single shared structure and mech-anism (Lewis and Arnold 2009). More than 10,000 P450sequences have been identified from all kingdoms of life,where they catalyze the oxidation of a stunning array oforganic compounds. These enzymes all recruit a cysteine-bound iron heme cofactor responsible for this activity. Thewidely varying substrate specificities of the P450s aredetermined by their protein sequences, which accumu-lated large numbers of amino acid substitutions as theydiverged from their common ancestor. Despite differencesin up to 90% of the amino acid sequences, the P450s allshare a common fold.

P450 BM3 from Bacillus megaterium (BM3) is partic-ularly attractive for laboratory evolution experiments. Itis one of only a handful of known P450s in which theheme domain and the diflavin reductase domains (FMNand FAD) required for generation of the active oxidantare fused in a single polypeptide chain. Furthermore, it issoluble and readily overexpressed in Escherichia coli, anexcellent host for directed evolution experiments. Thesubstrates of BM3 are largely limited to long-chain fattyacids, which it hydroxylates at subterminal positions athigh rates (thousands of turnovers per minute). Duringthe past decade, we and others investigators have used di -

42 ARNOLD

rected evolution to alter the specificity of this well-behaved bacterial P450 family member so that it canmimic the activities of widely different P450s, includingsome of the human enzymes. These experiments havedemonstrated that dramatic changes in substrate speci-ficity can be achieved with just a few mutations in thecatalytic (heme) domain (Landwehr et al. 2006; Rent -meister et al. 2008; Lewis and Arnold 2009; Lewis et al.2009).

Is a P450 Propane Monooxygenase Physically Possible?

One of our early directed evolution goals was to gener-ate a P450 that could hydroxylate small, gaseous alkanessuch as propane and ethane. In nature, these are substratesof methane monooxygenases, enzymes that are mecha-nistically and evolutionarily unrelated to the cytochromeP450 enzymes. A P450 had never been reported to acceptpropane, ethane, or methane as a substrate. We were curi-ous as to whether a P450 heme oxygenase was capable ofbinding and inserting oxygen into ethane or methane,whose C–H bond strengths are considerably higher thanthose of the usual natural P450 substrates.

P450 BM3 hydroxylates the alkyl chains of fatty acidscontaining 12–16 carbons and has no measurable activityon propane or smaller alkanes. We have never found anysingle mutation that confers this activity. To make a ver-sion of BM3 that hydroxylates propane, we therefore firsttargeted activity on a longer alkane (octane), a substratethat the wild-type enzyme does accept, albeit poorly(Glieder et al. 2002). We reasoned that variants of BM3having enhanced activity on octane might eventuallyacquire measurable activity on shorter alkanes and thusthat further mutagenesis and screening on progressivelysmaller substrates could ultimately generate enzymeswith good activity on the gaseous alkanes. This reasoningassumed that the problem was mainly one of substraterecognition and that there is no inherent mechanistic lim-itation to hydroxylation of small alkanes at the heme iron.

Five generations of random mutagenesis of the hemedomain, recombination of beneficial mutations, andscreening for activity on an octane surrogate led to BM3variant 139-3, which contains 11 amino acid substitutionsand is much more active on octane (Glieder et al. 2002).The improved octane activity was in fact accompanied bymeasurable activity on smaller alkanes, including pro -pane. Further rounds of mutation and recombination ofbeneficial mutations further enhanced activity on pro -pane. Variant 35E11, with 17 mutations relative to BM3,was highly active on propane and even provided modestconversion of ethane to ethanol (Meinhold et al. 2005).Breaking down the more difficult problem of obtainingactivity on very small substrates by first targeting octaneand then propane lowered the bar for each generation andallowed the new activities to be acquired one mutation ata time.

This enzyme, however, was still not as efficient athydroxylating the alkanes as is the wild-type enzyme withits preferred fatty acid substrates. Finely tuned conforma-

tional rearrangements within and among the heme andreductase domains mediate electron transfer and effi-ciently couple BM3-catalyzed hydroxylation to consump-tion of the NADPH (nicotinomide adenine dinucleotidephosphate) cofactor. When these processes are disrupted,either by mutations or by introduction of novel substrates,catalysis is no longer coupled to cofactor consumption:NADPH consumption instead produces reactive oxygenspecies that eventually cause the enzyme to self-destruct.To retune the whole system for oxidation of propane, wetherefore also targeted the FMN and FAD domains ofvariant 35E11 for mutagenesis (individually, but in thecontext of the holoenzyme) and continued to screen forincreased ability to convert propane to propanol. We thencombined the optimized heme, FAD, and FMN domainsto generate P450

PMO(Fasan et al. 2007, 2008). This

enzyme displayed activity on propane comparable to thatof BM3 on fatty acids and 98% coupling of NADPH con-sumption to product hydroxylation. P450

PMO thus became

as good an enzyme on propane as the wild-type enzymeis on laurate with a total of 23 amino acid substitutions,amounting to changes in less than 2.3% of its (>1000amino acid holoenzyme) sequence.

Creation of P450PMO

, a complex, multidomain enzymefinely tuned for activity on a substrate not accepted by thewild-type enzyme, demonstrates the remarkable ability ofthe cytochrome P450 to adapt to new challenges by accu-mulating single beneficial mutations over multiple gener-ations.

The P450PMO Evolutionary Trajectory

Studying the evolutionary intermediates along the lin-eage of P450

PMO revealed interesting features of adapta-

tion to propane. Activity on propane first emerged in139-3, a variant that is active on a wide range of sub-strates. But by the time the enzyme became highly activeand fully coupled on propane, it had lost its activity onlaurate—a more than 1010-fold change in specificity, justfrom 139-3 to P450

PMO. Thus, becoming a good propane

monooxygenase in P450PMO

came at the cost of the nativeenzyme’s activity on fatty acids, even though this prop-erty was not included in the artificial selection pressure.This is apparently the easiest route to high activity onpropane.

Substrate specificity changes for selected variantsalong the lineage to P450

PMOwere also investigated on

alkanes having one to 10 carbons. These activity profilesrevealed that intermediate variants (e.g., 139-3, 35E11)acquired activity on a range of alkanes before ultimatelyrespecifying for propane (Fig. 1) (Fasan et al. 2008).P450

PMOis highly specific compared to its precursors: Its

activity drops precipitously on alkanes having just onemore or one less methylene group. Only positive selection(for high activity on propane) had been used to obtainP450

PMO; there was no selection against activity on any

other substrate. One can conclude that it is easier to obtainvery high activity on propane than it is to have high activ-ity on a range of substrates; thus, a highly active “special-ist” is easier to find than a highly active “generalist.”

HOW PROTEINS ADAPT 43

Sequencing reveals the mutations acquired in each gen-eration of directed evolution. The 21 amino acid substitu-tions in the heme domain of P450

PMO(two of the 23 are in

the reductase domain) are distributed over the entire protein(Fig. 2). Many are distant from the active site and influencespecificity and catalytic activity through unknown mecha-

nisms. The crystal structure of 139-3 (Fasan et al. 2008)revealed only small changes in the active site volume, con-sistent with its activity toward a wide range of substrates.Modeling studies, however, indicate much more dramaticreduction in the volume accessible to substrate in P450

PMO

(C Snow, unpubl.).

44 ARNOLD

Figure 1. (A) Total turnovers catalyzed by selected variants along the P450PMO

lineage on propane and ethane. (B) Relative activitieson C

n(n = 1–10) alkanes. (Reprinted, with permission, from Lewis and Arnold 2009 [© Swiss Chemical Society].)

Figure 2. P450 BM3 heme domain backbone, showing locations of 21 of 23 mutations that convert P450 BM3 to a highly active, fullycoupled propane monooxygenase (P450

PMO).

STABILITY PROMOTES EVOLVABILITY: A ROLE FOR NEUTRAL MUTATIONS IN

ADAPTIVE EVOLUTION

It is useful to consider when this simple adaptive walkmight fail. Of course, it will fail if the functional bar is settoo high—this happens when the fitness improvementsrequired to pass the screen or selection are not reached bysingle mutations. It also fails when the protein is not robustto mutation (Bloom et al. 2005, 2006). At one point duringthe evolution of P450

PMO, in fact at mutant 35E11, we could

find no additional mutations that further enhanced theenzyme’s activity on propane. Upon characterizing 35E11and its precursors, the reason for this became clear: Theenzyme had become so unstable that it simply could nottolerate any further destabilization and still function underthe expression and assay conditions. Most mutations aredestabilizing, and most activating mutations are also desta-bilizing, possibly more so than the average mutation. Theprocess of enhancing P450’s activity on propane had desta-bilized it so much that 35E11 simply could not accept anyfurther destabilizing mutations. Once we incorporatedmutations that stabilized the structure (but were neutral ornearly neutral with respect to activity), directed evolutionof activity could continue as before, and significant addi-tional improvements were achieved (Fasan et al. 2007).Sta bi liz ing the structure made it robust to further mutationand opened up the ability to explore a whole spectrum ofmutational paths that were previously inaccessible.

We demonstrated this key role of functionally neutralbut stabilizing mutations in adaptive evolution withanother experiment that directly compared the frequencywith which a marginally stable and a highly stable cyto -chrome P450 enzyme could acquire activities on a set ofnew substrates upon random mutation (Bloom et al.2006). A markedly higher fraction of mutants of the sta-ble protein were found to exhibit the new activities. Thisincreased evolvability could be traced directly to theenzyme’s ability to tolerate catalytically beneficial butdestabilizing mutations.

Directed evolution has thus shown the crucial role thatstability-based epistasis can have in adaptive evolution. Aprotein that has been pushed to the margins of tolerablestability may lose access to functionally beneficial butdestabilizing mutations. But this protein is still not stuckon a fitness peak, because it can regain its mutationalrobustness and evolvability via a neutral path, by accu-mulating stabilizing mutations that do not directly affectfunction. In natural evolution, such a process mightrequire stabilizing mutations to spread by genetic drift(Bloom and Arnold 2009).

ADAPTIVE EVOLUTION RELIES ONFUNCTIONAL PROMISCUITY, WHICHCHANGES WITH NEUTRAL MUTATIONS

A well-recognized feature of proteins is their functionalpromiscuity. Enzymes, for example, often catalyze amuch wider range of reactions, or reactions on a widerrange of substrates, than are biologically relevant.

Directed evolution experiments have shown that proteinactivities or functions present at a low level can often beimproved via an adaptive pathway of sequential benefi-cial mutations. Protein functional promiscuity thus pro-vides a stepping stone for generation and optimization ofnew functional molecules by adaptive evolution.

Directed evolution experiments with P450 BM3 havealso demonstrated that promiscuous activities can emergeon mutations that are neutral with respect to a main (bio-logical) function (Bloom et al. 2007). We performed akind of neutral evolution by random mutagenesis andselection for retention of catalytic activity on a fatty-acid-like substrate. The variants containing these “neutral”mutations were then examined for activity on several othernontarget substrates. In many cases, the neutral mutationshad led to changes in these promiscuous activities. Neutralmutations can also set the stage for adaptation by explor-ing a varied set of evolutionary starting points, at little orno cost to the current biological function.

I already discussed how neutral mutations can enhancea protein’s stability, thereby increasing its tolerance forsubsequent functionally beneficial but destabilizingmutations. Neutral mutations can also lead to changes infunctions that are not currently under selective pressurebut can subsequently become the starting points for theadaptive evolution of new functional proteins. A processthat generates large numbers of mostly neutral mutationsis recombination (of homologous proteins), whichexploits the genetic drift that underlies the divergence oftheir sequences. As we discuss below, swapping thesemutations in the laboratory can generate proteins differ-ent from the parent proteins, including those that are morestable or exhibit activities not present in the parents.

NOVEL PROTEINS BY RECOMBINATION

Recombination is an important mutation mechanism innatural protein evolution. We have studied the effects ofmutations made by recombination of homologous pro-teins (that share a three-dimensional structure but maydiffer at hundreds of amino acid residues) by making andcharacterizing large sets of “chimeric” proteins. Theprobability that a protein retains its fold and functiondeclines exponentially with the number of random muta-tions it acquires—random mutations are quite deleteriouson average. By quantifying the retention of function withmutation level in chimeric β-lactamases made by swappingsequence elements between two homologous en zymes, weshowed that the mutations made by re combination aremuch more conservative, presumably because they hadalready been selected for compatibility with the lacta-mase folded structure (Drummond et al. 2005). Re com -bination can generate proteins that have a high probabilityof folding and functioning despite having dozens of muta-tions compared to their parent sequences. Thus, recombi-nation is conservative. But does it lead to new functionsor traits?

We generated a large set of recombined P450 hemedomains by swapping sequence elements among three nat-ural P450 BM3 homologs sharing ~65% sequence identity

HOW PROTEINS ADAPT 45

(Otey et al. 2006). A sampling of the functional P450sshowed that they exhibited a range of activities, includingactivity on substrates not accepted by the parent enzymes(Landwehr et al. 2007). The chimeric P450s also exhibited arange of stabilities, with a significant fraction of them morestable than any of the parent enzymes from which they wereconstructed (Li et al. 2007). Like many proteins, P450s areonly marginally stable, never having been selected for ther-mostability or long-term stability. Depending on the degreeto which stability has already been maximized in the parentsequences, recombination can generate proteins that are lessstable or more stable than the parent proteins.

Recombination shuffles large numbers of mutationsthat individually have little or no effect on function. Ourexperiments have shown that these mutations can gener-ate proteins of widely varying stabilities and with a widerange of promiscuous activities, both of which can opennew pathways for further functional evolution.

CONCLUSIONS

Directed evolution does not necessarily mimic naturalevolution: Laboratory proteins evolve under artificialpressures and via mutation mechanisms that usually dif-fer significantly from those encountered during naturalevolution. These experiments nevertheless allow us toexplore protein fitness landscapes, the nature of the evo-lutionary trajectories, as well as the functional features ofindividual protein sequences. Anything created in the lab-oratory by directed evolution is also probably easily dis-covered by natural evolution. Thus, knowing whatfunctional features are accessible to evolution helps us tounderstand what biology cares about, i.e., what featuresare retained and encouraged by natural selection, andwhat biology tends to throw away. Laboratory evolutionexperiments beautifully demonstrate that biological sys-tems, themselves the products of millions of years of evo-lution, readily evolve to meet new challenges.

ACKNOWLEDGMENTS

The author thanks all of her coworkers that have con-tributed to the work described here, and especially thanksJesse D. Bloom, Phil Romero, Jared C. Lewis, and RudiFasan. Support is from the Jacobs Institute for MolecularMedicine, the Department of Energy, the U.S. Army,DARPA, and the National Institutes of Health.

REFERENCES

Bloom JD, Arnold FH. 2009. In the light of directed evolution:Pathways of adaptive protein evolution. Proc Natl Acad Sci106: 9995–10000.

Bloom JD, Silberg JJ, Wilke CO, Drummond DA, Adami C,Arnold FH. 2005. Thermodynamic prediction of protein neu-trality. Proc Natl Acad Sci 102: 606–611.

Bloom JD, Labthavikul ST, Otey CR, Arnold FH. 2006. Proteinstability promotes evolvability. Proc Natl Acad Sci 103: 5869–5874.

Bloom JD, Romero PA, Lu Z, Arnold FH. 2007. Neutral geneticdrift can alter promiscuous protein functions, potentially aidingfunctional evolution. Biol Direct 2: 17.

Drummond DA, Silberg JJ, Meyer MM, Wilke CO, Arnold FH.2005. On the conservative nature of intragenic recombination.Proc Natl Acad Sci 102: 5380–5385.

Fasan R, Chen MM, Crook NC, Arnold FH. 2007. Engineeredalkane-hydroxylating cytochrome P450

BM3exhibiting native-

like catalytic properties. Angew Chem Int Ed 46: 8414–8418. Fasan R, Meharenna YT, Snow CD, Poulos TL, Arnold FH. 2008.

Evolutionary history of a specialized P450 propane monooxy-genase. J Mol Biol 383: 1069–1080.

Glieder A, Farinas ET, Arnold FH. 2002. Laboratory evolution ofa soluble, self-sufficient, highly active alkane hydroxylase. NatBiotechnol 20: 1135–1139.

Landwehr M, Hochrein L, Otey CR, Kasrayan A, Bäckvall J-E,Arnold FH. 2006. Enantioselective α-hydroxylation of 2-ary-lacetic acid derivatives and buspirone catalyzed by engineeredcytochrome P450 BM-3. J Am Chem Soc 128: 6058–6059.

Landwehr M, Carbone M, Otey CR, Li Y, Arnold FH. 2007.Diversification of catalytic function in a synthetic family ofchimeric cytochrome P450s. Chem Biol 14: 269–278.

Lewis JC, Arnold FH. 2009. Catalysts on demand: Selective oxi-dations by laboratory-evolved cytochrome P450 BM-3. Chimia63: 309–312.

Lewis JC, Bastian S, Bennett CS, Fu Y, Mitsuda Y, Chen MM,Greenberg WA, Wong C-H, Arnold FH. 2009. Chemo en zy -matic elaboration of monosaccharides using engineeredcytochrome P450 B

M3demethylases. Proc Natl Acad Sci 106:

16550–16555. Li Y, Drummond DA, Sawayama AM, Snow CD, Bloom JD,

Arnold FH. 2007. A diverse family of thermostable cytochromeP450s created by recombination of stabilizing fragments. NatBiotechnol 25: 1051–1056.

Meinhold P, Peters MW, Chen MY, Takahashi K, Arnold FH.2005. Direct conversion of ethane to ethanol by engineeredcytochrome P450 BM3. Chembiochem 6: 1765–1768.

Otey CR, Landwehr M, Endelman JB, Hiraga K, Bloom JD,Arnold FH. 2006. Structure-guided recombination creates anartificial family of cytochromes P450. PLoS Biol 4: 0789–0798.

Rentmeister A, Arnold FH, Fasan R. 2008. Chemo-enzyme fluo-rination of unactivated organic compounds. Nat Chem Biol 5:26–28.

Romero PA, Arnold FH. 2009. Exploring protein fitness land-scapes by directed evolution. Nat Rev Mol Cell Biol 10: 866–876.

46 ARNOLD