Stat.haifa.ac.il

The ISME Journal (2008) 2, 171–179& 2008 International Society for Microbial Ecology All rights reserved 1751-7362/08 $30.00 Enzyme improvement in the absenceof structural knowledge: a novelstatistical approach Yoram Barak1,4, Yuval Nov2,4, David F Ackerley1,3 and A Matin11Department of Microbiology and Immunology, Stanford University School of Medicine, Stanford, CA, USA;2Department of Statistics, University of Haifa, Haifa, Israel and 3School of Biological Sciences, VictoriaUniversity of Wellington, Wellington, New Zealand Most existing methods for improving protein activity are laborious and costly, as they either requireknowledge of protein structure or involve expression and screening of a vast number of proteinmutants. We describe here a successful first application of a novel approach, which requires nostructural knowledge and is shown to significantly reduce the number of mutants that need to bescreened. In the first phase of this study, around 7000 mutants were screened through standarddirected evolution, yielding a 230-fold improvement in activity relative to the wild type. Usingsequence analysis and site-directed mutagenesis, an additional single mutant was then produced,with 500-fold improved activity. In the second phase, a novel statistical method for proteinimprovement was used; building on data from the first phase, only 11 targeted additional mutantswere produced through site-directed mutagenesis, and the best among them achieved a 41500-foldimprovement in activity over the wild type. Thus, the statistical model underlying the experiment wasvalidated, and its predictions were shown to reduce laboratory labor and resources.
The ISME Journal (2008) 2, 171–179; published online 22 November 2007Subject Category: microbial engineeringKeywords: protein design; Nov–Wein model; directed evolution; rational design targeted changes in amino acids around its activesite; several other such structure-based improve- Improving the activity of a protein by manipulating its sequence—a process termed protein design—is of great interest in medicine and biotechnology, and protein is expensive, laborious and time consuming, has been widely practiced. However, the sequence and activity predictions based on structure are space is ‘more than astronomically’ vast limited in their success. Thus, design methods that do not rely on structural knowledge are needed, not experimentally feasible to test all possible mutants only for proteins whose structure is not known but of a protein nor is it necessary, since many of the also where structural information is available, since resulting sequences do not fold into functioning activity may be influenced by amino acids not One mutagenesis approach, termed rational de- sign, uses information about the three-dimensional An alternative to structure-based rational design structure of the protein and its target molecule to is directed evolution—a selective process that identify promising sequence changes. Thus, mimics nature, whereby a protein is ‘bred’ through successive generation of gene libraries; the members coli nitroreductase, NfsB, for prodrug reduction by of these libraries are randomly mutated andshuffled, and their resulting proteins are thenscreened for improved activity. Common methods Correspondence: A Matin, Department of Microbiology andImmunology, Sherman Fairchild Science Building, Stanford for generating such libraries include error-prone University School of Medicine, 299 Campus Drive W, Stanford, PCR and recombination between homologous re- These authors contributed equally to this work.
Received 4 June 2007; revised 8 October 2007; accepted 9 October Directed evolution is widely practiced and has produced important results, yet it typically A statistical approach for enzyme improvement necessitates expression, purification and screening tase of unknown structure, which has a wide of thousands of protein mutants. In addition, directed evolution is a ‘blind’ process, and it is several beneficial activities such as chromate and virtually impossible to mathematically predict its uranyl (U(VI)) reduction (useful in the bioremedia- A third approach to protein design models the relation between the sequence of a protein mutant reduction (useful in cancer chemotherapy and its activity (fitness) as a statistical relationship.
)). Improvement in all three activities is That is, one assigns a distribution of activity levels for each protein mutant, rather than a singlepredicted activity, and can thereby specify probabil-ities for the various activity levels. Among the models that belong to this class are the NK model Strains, plasmids, genes, primers and growth Supplementary Table 1 lists the strains, plasmids and primers used in this study. The various strains were grown at 37 1C to mid-exponential phase, the NK models are Gaussian, and most of the induced by 0.5 mM isopropyl-b-D-thiogalactoside regression-based methods are implicitly Gaussian, and incubated overnight for protein production.
as they assume Gaussian distribution of the errorswhen computing confidence intervals, P-values, etc.
The statistical approach to protein design circum- vents the need to decipher a protein’s structure and Routine DNA manipulations were performed as promotes identification of promising mutant candi- dates, thus significantly reducing the number of was carried out by miniprep (Qiagen Inc., Valencia, in which the activity of bacterial halohydrin Corporation, CA, USA, using appropriate primers dehalogenase was significantly improved to meet design criteria in the commercial production ofatorvastatin (Lipitor), a cholesterol-lowering drug.
The enzyme was optimized through a statistical Directed evolution of the chrR gene for improving analysis method termed protein sequence activity relationship, combined with directed evolution and Error-prone PCR was used to introduce random We report here a successful first empirical using the GeneMorph II Random Mutagenesis kit application of a novel method belonging to the last (Stratagene Corporation, La Jolla, CA, USA). For- mentioned class; the method is based on a statistical ward and reverse chrR primers (Supplementary model for the sequence–activity relationship pro- Table 1) were used to amplify full-length hybrid posed by Nov and Wein (hereafter referred to as ‘the model’), whose theoretical and mathematical details The shuffled genes were ligated into the pET28a þ plasmid, and transformed into E. coli BL21 (DE3) Briefly, this model is additive, in the sense that it (Invitrogen Inc., Carlsbad, CA, USA) to allow over- assumes that after proper transformation of the data, expression. Recombinants were selected on plates the change in activity caused by a multiple-residue containing kanamycin (50 mg mlÀ1). High-throughput mutation roughly equals the sum of the activity screening of 7000 recombinants was performed by changes caused by the corresponding single-residue inoculating colonies into individual wells of 96-well mutations; the degree of non-additivity is captured microtiter plates, containing 200 ml Luria–Bertani through one of the model’s parameters. The model is medium and kanamycin. After growth to stationary sparse in parameters, and is mathematically tract- phase (overnight incubation, final A660, 1–1.5), 20 ml able, conveniently allowing one to update the aliquots from each well were used to inoculate a activity distributions of the yet-unexplored mutants second series of plates, using M9 minimal medium from the sequence–activity data of tested mutants.
(Sigma Inc., St Louis, MO, USA). Each well received In addition to their sequence–activity relationship the same initial inoculum. The first set of plates was model, Nov and Wein suggested an optimization stored at À80 1C after addition of glycerol. Cells in module for selecting promising mutant candidates; a the second inoculation series were allowed to grow variant of this module was used in this study. The to mid-exponential phase and then exposed to relevant aspects of the model used in this study are 0.5 mM isopropyl-b-D-thiogalactoside to induce the presented in the Materials and methods section.
recombinant gene expression. After overnight in- The improvement efforts targeted the E. coli cubation, cells were lysed by addition of 30 ml enzyme ChrR, an NAD(P)H-dependant oxidoreduc- BugBuster (Novagen Inc., San Diego, CA, USA), The ISME Journal
A statistical approach for enzyme improvementY Barak et al incubated for 20 min at room temperature, and Table 1 First phase: the effect of sequence changes on chromate centrifuged for 20 min at 3000 g. Supernatant reductase activity (Vmax) of the E. coli ChrR protein (100 ml) was mixed with 10 ml solution of the 2 mM NADH, 100 mM Tris-HCl (pH 7) and ddH2O and chromate reduction wasassayed as described below.
The most efficient enzymes for Cr(VI) reductase activity were purified on nickel columns, as previously obtained from the frozen plates. Protein concentrations were determined with the Bio-Rad Dc protein assay kit, using bovine serum albumin as a standard.
were used for site-directed mutagenesis. These were designed to create single-codon mutations following the desired mutations had been generated was obtained by sequencing. Proteins encoded by the modified genes were generated as described above.
Abbreviation: WT, wild type.
Mutants ChrR6–ChrR20 were obtained through directed evolution, and ChrR21 was produced through site-directed mutagenesis.
Determination of Cr(VI) reduction rates by cellextract preparation and chromate reductase assayswere conducted as described previously were collected after incubation for the specified time. A 120 ml sample was mixed with 130 ml reagent of enzyme activity were performed at pH 7 and at mixture containing 5:1:1:1:5 proportion of complex- ing solution, TAC (2-(2-thiazolyazo-p-cresol)), Triton 1C. Each assay was conducted four times unless X-100 (0.15 M), CTAB (N-cetyl-N,N,N-trimethyam-monium bromide) and triethanolamine buffer (pH6.5). The method depends on the TAC binding to U(VI), which is aided by Triton and CTAB. After Reductive prodrugs become strong killing agents of 15 min of color development, the samples were biological cells upon reduction. The capacity of the read at A588 nm using a Micro-Plate Reader (ASYS mutant enzymes to carry out this reduction was determined with minor modifications as previouslydescribed Briefly, prodrugreduction mixtures contained mitomycin C, CB 1954 (5-aziridinyl-2,4-dinitrobenzamide) or 17-AAG (17-allylamino-17-demethoxygeldanamycin) at a con- centration of 15 mM, 10 mg ml–1 pure enzyme, 50 mM NADPH and Dulbecco’s modified Eagle’s medium estimation of the model parameters, as well as other scientific programming, was carried out through Following prodrug reduction for 30 min at 37 1C, MATLAB (The MathWorks Inc., Natick, MA, USA).
0.5 ml of JC breast cancer cells (B0.5–1 Â 105) wereadded and the cells were incubated for additional24 h. After the latter incubation, 20 ml of the color reagent, CellTiter 96 AqueousOne (Promega Inc., The model has four parameters: the drift m, which is Madison, CA, USA) was added to 100 ml aliquots the expected change in fitness due to introduction of of the reaction mixture. Following 1 h of further a new, arbitrary mutation (a negative number, as mutations more often decrease than increase fit- ness); the site variance sS, which is the variance ofthe change in expected fitness contribution due to amutation across sites; the residue variance s2R, which is the variance of the fitness contribution of a For selected mutant enzymes, uranyl reductase specific single-residue mutation within a site; and activity was also determined. This was carried out the non-additivity variance s2N, which captures both the degree of non-additivity and the level of The ISME Journal
A statistical approach for enzyme improvement measurement noise (as in all additive models, these two effects cannot be distinguished from one another). For a thorough presentation of the model, pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi below, a variant of the model, with only three substrate converted per milligram protein per min- where F1 is the log-transformed r-vector (r ¼ 15) of ute) was taken to equal log10(Vmax/vwt), where vwt is the 15 Vmax values of the mutant proteins the Vmax value of the wild-type enzyme (which was excluding the wild type and ChrR17), m1 is its mean vector (computed according to Equation (1)) and S1 goodness-of-fit of the data to the model, and set is its r  r covariance matrix (computed according the fitness of the wild type to 0, as required by the to Equations (2) and (3)). The resulting estimates S ¼ 0.4861 and sN ¼ 0.1478. The estimate of The 16 mutant proteins sequenced in the first the third parameter, m, was positive, in contrast to phase involved mutations in n ¼ 11 sites (A44, the model’s assumptions. This finding was ex- D103, V120, Y128, G150, Q153, N154, T160, Q175, pected: the sequences obtained in the first phase Q184, K187; Only one of these sites, Q175, were not a random sample from the sequence space, had more than one substituent amino acid—Q175L in which a priori it is expected that most mutations and Q175H—of which the latter appeared in only are deleterious (corresponding to a negative m); one sequence, ChrR17. To improve the numerical rather, these sequences were chosen by the selective stability of the estimation computations, ChrR17 directed evolution process due to their improved was omitted from the data, so that only 15 sequences fitness, and thus carry seriously distorted informa- were used; otherwise, the parameter s2R would have tion about m. Therefore, for fitness prediction appeared in only two entries of a 16  16 covariance purposes (see below), only the two estimated matrix. It is for this reason that a three-parameter version of the model was used, employing the value of m was varied, in jumps of size 0.2, across the model is a Gaussian random field F ¼ {Fs}, where For the second application of the model predic- the index set of F consists of all 211 ¼ 2048 sequences tions, after the activity information of the first that may be generated from the genetic diversity of round’s five mutants became available, the para- the 15 mutants found in the first phase. The joint meters were re-estimated in a similar way, using distribution of the elements of F is given by the r ¼ 15 þ 5 ¼ 20 in Equation (4) and appropriately modified F1, m1 and S1. The resulting estimates were s ¼ 0.4361 and s ¼ 0.1961, very similar to those By the additivity of the model, the conditional expected fitness of a sequence s given the data, where d(s, sˆ) is the number of sites in which a E(Fs|F1), is the sum of the conditional expected sequence s differs from the wild-type sequence sˆ fitness contributions from each of the 11 mutated and M(s, s0) is the number of sites in which both sites. The contribution from a site having the wild- sequences s and s0 differ from the wild-type type residue is 0 (and hence so is its expected sequence. In this three-parameter form (but not in contribution), and that of a site i with a non-wild- the full four-parameter form), the model is similar to type residue is a random variable fi. The conditional a regression model with random coefficients (some- expected value of the vector f ¼ (f1, ., fn) is an intercept, in which the predictors are binary variables, indicating the presence or absence of a mutation, their coefficients are N(m, s2S) randomvariables, and the variance of the error terms is where m is a constant n-vector, having all of its As no prior distribution is assumed over the element equal to m; the matrix S1 is the inverse of parameters, the model is not Bayesian.
S1 and S2 is an n  r matrix, having sS as its (i, j)thentry, if mutant j had a mutation at site i, and0 otherwise.
In the first round of the second phase, the The parameters of the model were estimated by the maximum likelihood method. Specifically, m, s2S pond to a proportion of non-additive variance of and s2N were initially estimated to be the maximizers 0.1478/(2 Â 0.4861 þ 0.1478) ¼ 0.132 among double The ISME Journal
A statistical approach for enzyme improvementY Barak et al mutants, which is low enough to allow reliable mutants isolated in this phase, so an additional predictions. As mentioned above, the value of m mutant, containing the single mutation Y128N, was not estimated from the data, and was varied was generated through site-directed mutagenesis.
from À1.5 to À0.1. For each value of m, the n-vector E(f|F1) was computed according to Equation (5) (see mutants in chromate reductase activity, exhibiting Supplementary Table 2), and the conditional ex- a 500-fold improvement over the wild type. An pected fitness values of all possible n(nÀ1)/2 ¼ 55 additional (fourth) round of directed evolution, double mutants were calculated. Among these, the using DNA from ChrR6 to ChrR21 as template and five double mutants with the highest expected screening around 1000 variants, did not yield fitness (averaged across all m, and not including sequences already in the data set) were identified, The second phase of the study consisted of applying the model to the sequence–activity data column). The sequences for the second round were chosen in the same method, with the appropriate estimated from the entire information of changes to r, F1, m1 and S1. Since only triple-residue and the sequences of the five most promising double mutants were considered in this round, the mutants mutants (that is, the five mutants that possess the chosen were the top ones, in terms of conditional highest conditional expected activity, among those expected fitness, among all n(nÀ1)(nÀ2)/6 ¼ 165 differing from the wild type in two amino acids) were mathematically identified (ChrR22 to ChrR26;These mutant proteins were generated inpure form by site-directed mutagenesis and nickel column purification as described and their Vmax for chromate reduction was A two-phase strategy for ChrR improvement was employed: a ‘blind’ directed-evolution approach in ChrR23, exhibited a Vmax of 258 000, corresponding the first phase and the model-based predictions to to an 876-fold improvement in activity over the wild obtain further improvement in the second. In the type, around fourfold improvement over ChrR13 first phase, ChrR protein mutants were obtained by (the best mutant obtained in four rounds of directed subjecting the chrR gene to three rounds of error- evolution, which necessitated screening of 7000 prone PCR. Each round was followed by screening mutants), and a 1.75-fold improvement over ChrR21 the resulting mutant proteins for chromate reductase (the best mutant isolated in the first phase). In activity, using a colorimetric method that provides addition, the average Vmax of mutants ChrR22– an approximate indication of the degree of improve- ChrR26 was significantly higher than the average ment in this activity. Around 6000 mutants were Vmax of the first-phase mutants ChrR6–ChrR21 screened. The top 15 mutant proteins were purified (104 000 vs 26 000; P ¼ 0.0084 in a one-tailed Mann–Whitney test for median comparison).
reduced per milligram protein per minute) was To further improve the ChrR enzyme, we con- ducted a second screening round according to the model predictions. The parameters of the model reduction (425-fold improvement) compared to the were re-estimated, using the sequence–activity data wild-type enzyme, the best, ChrR13, showing a V of 67 500, corresponding to about 230-fold improved seven most promising triple mutants were identi- fied. One of these mutants could not be generated, Sequence analysis revealed that the Y128N sub- but the remaining six were produced as described stitution was common to almost all of the improved above, and their chromate reductase Vmax valueswere measured Strikingly, one of these,ChrR30, exhibited 1554-, 6.6- and 3.1-fold improve-ments over the wild type, ChrR13 and ChrR21, Table 2A First round of second phase: sequence and Vmaxactivity of five mutants predicted by the Nov–Wein model to respectively. Thus, by screening just a few mutants, have improved chromate reductase activity a multifold enhancement was obtained in anenzyme already improved to a large degree. The aggregate average Vmax of the 11 mutants ChrR22– ChrR32 (117 000) was also significantly higher than that of the first-phase mutants ChrR6–ChrR21 Previous results had shown a positive correlation between chromate reductase activity and other We therefore examined the activity ofthree of the most active mutants in chromate Abbreviation: WT, wild type.
Predictions were based on the sequence–activity data of reductase—ChrR21, ChrR23 and ChrR30—in two The ISME Journal
A statistical approach for enzyme improvement additional respects, namely, prodrug and U(VI) reductase activity for each mutant; ChrR30 being reduction. The capacity of the mutants to reduce prodrugs was determined by the efficiency with The three mutants also exhibited improved uranyl which they killed cells of the JC breast cancer cell reductase activity compared to the wild-type en- line. Three prodrugs, namely, mitomycin C, CB 1954 and 17-AAG, were used. All three mutants were activity, no further improvement in this activity was more potent than the wild-type enzyme in activating shown by the other mutants over ChrR21.
each of the drugs, and in causing the drug-mediatedkilling of the cells (This activity corre-lated, by and large, with improved chromate Directed evolution has resulted in successful gene-ration of many improved proteins, but this approach Table 2B Second round of second phase: sequence and Vmax is blind, laborious and time consuming. Typically, activity of six additional mutants chosen according to the the improvements achieved in the early rounds of directed evolution are significant, but in laterrounds, even when a large number of further mutants is screened, improvements become smaller and less frequent. For example, and were able to improve the thermostability of lactate oxidase by 18-fold after screening around 3000 mutants, but had to screen more than 20 000 additional mutants for a twofold further improvement. Since mutations are more often deleterious to protein activity, it hasbeen thought that increased mutational rate was likely to correlate with loss of function, and 1–3 mutation rate per gene was considered desirable. Recently,however this notion has been questioned. and have shown that higher mutation rate libraries (15–30 per gene) have a better probability of generating im-proved employed in our experiment was low (1–5 per gene)and therefore the fourth round of directed evolution resulting with no improvement might be explained Viability (%)
by ‘masking’ of deleterious mutations over bene-ficial ones.
While we could have obtained further improve- ment in ChrR activity using the directed evolution process by screening a large number of additionalmutants (perhaps 20 000 or more) in later rounds, The effect of mitomycin C ( ), CB 1954 (5-aziridinyl- 2,4-dinitrobenzamide) ( ) and 17-AAG (17-allylamino-17-de- our use of the Nov and Wein model clearly afforded methoxygeldanamycin) ( ) on the killing of JC breast cancer a significant saving in screening effort in these later cells in the presence of the wild-type or the evolved enzymes stages. To provide a perspective: it was necessary to (10 mg mlÀ1). The concentration of the drugs was 15 mM. The screen around 7000 mutants in four rounds of enzymes were incubated with the drug for 30 min (37 1C), directed evolution (the last of which yielded no followed by the addition of the cells. After 24-h incubation(37 1C), cell viability was determined as described in the Materials additional improvement) to obtain a 230-fold in- crease in ChrR activity; in contrast, the model made Table 3 Uranyl reduction kinetics of selected evolved mutants The ISME Journal
A statistical approach for enzyme improvementY Barak et al it possible to improve the enzyme significantly mutant activity (such as and thus to isolate further (46-fold improvement over the best mutant mutants that otherwise may have been difficult to obtained by directed evolution and 41500-fold improvement over the wild type) by screening only All mutants designed in the second phase based 11 targeted new mutants. This saving in screening is on the model predictions are built from mutations especially attractive when the screening cost is high generated through directed evolution in the first compared to the cost of producing site-directed phase. Thus, in principle, it was possible to obtain the new mutants through additional rounds of directed evolution, without using the statistical how a statistical model can augment directed model. However, as directed evolution is a blind evolution to significantly improve the cyanation process governed by chance, it is not clear screening of how many additional mutants would have been Although both studies employed additive statistical required to achieve an improvement comparable models coupled with traditional techniques, their to that which the application of the model made results do not permit easy comparison, since (a) possible; it should be kept in mind that the enzyme activity was measured differently in the two last round of directed evolution yielded no studies and (b) it is not known which of the two enzymes is more amenable to optimization. How- The model postulates that mutations are approxi- ever, Fox et al. improved activity by B4000-fold mately additive. Is this assumption supported by the after screening more than 500 000 mutants in 18 data? Based on the second, more complete estimate rounds, while in the present work, we achieved of the parameters, the fraction of the total variance of B1500-fold activity improvement after screening the fitness of a double mutant that is due to non- B7000 variants in six rounds. Furthermore, as the additivity (and measurement noise) is 0.1961/ structure of the halohydrin dehalogenase enzyme is known, some of the diversity in Fox et al. experi- the fraction is 0.1961/(3 Â 0.4361 þ 0.1961) ¼ 0.13.
ments was generated through rational design. In this These relatively low numbers indicate that the data work, no structural knowledge was used, as the are not particularly noisy, there are no strong structure of the ChrR enzyme is unknown. Both epistatic effects and that the mutational effects are studies demonstrate the power of statistical model- mostly additive. Two additional points regarding ing in protein design, and both permit beneficial use additivity are noteworthy. First, additivity is as- of information gained from mutants with reduced sumed to apply to the transformed activity measure- ments, rather than to the raw data. For example, The genetic diversity spanned by the directed ChrR31 is the combination of ChrR11 and ChrR25, and the deviation from perfect additivity in 3000 possible combinations, among which (after the raw data (202 778 vs 1000 þ 147 222) is much omitting ChrR17) 55 are double mutants and 165 are greater than that in the transformed data (2.83 vs triple mutants. As diversity increases, the numbers 0.53 þ 2.70). Second, as often happens in statistical grow exponentially: when one considers 15 mutated analysis, even when approximate additivity holds, positions with two possible mutations in each, there some considerable exceptions occur; this can be are 4107 possible combinations (420 double mu- seen in our data when comparing ChrR10, whose tants, 3640 triple mutants); and with 20 mutated transformed activity is 1.81, against ChrR11 com- positions and three possible mutations in each, bined with ChrR21, whose sum of transformed there are 41012 possible combinations (1710 double mutants, 430 000 triple mutants). Thus, exhaustive Producing designed mutants through site-directed search in a laboratory, even only for double and mutagenesis, as our approach required, is not triple mutants, does not scale well, and system- always simple, as certain designed mutants are atically producing and screening all of them would difficult to generate in a laboratory. A potential be an extensive and highly laborious feat. The remedy for this problem is to create in the second predictions of the model allow one to screen instead phase, after statistically analyzing sequence–activity only a few targeted mutants, and still improve data from the first phase, combinatorial libraries containing only putatively beneficial mutations.
It is serendipitously possible to identify promis- These focused libraries will then be subject to ing mutants by simply ‘gazing’ at activity data and directed evolution, and are more likely to achieve detecting beneficial mutations, as was done in the improvement than straightforward directed evolu- discovery of the single-residue mutant ChrR21 in tion libraries that do not incorporate statistical this work. However, a systematic mathematical analysis in their design. This approach is pursued approach is needed to identify more complex mutants, such as ChrR30. The model is shown here One might suggest that our statistical analysis to be a valuable tool for such situations, as it allows could benefit from adopting a Bayesian approach, one to rigorously separate the expected contribution where prior distributions are set over the para- to activity from each of the mutations in a data set of meters. However, as this work is the first to study The ISME Journal
A statistical approach for enzyme improvement enzyme activity data in light of the model, we could Chen K, Arnold FH. (1993). Tuning the activity of an not use informative priors for Bayesian estimation.
enzyme for unusual environments: sequential random The proper choice of non-informative priors is mutagenesis of Subtilisin E for catalysis in dimethyl- under debate among statisticians, especially for formamide. Proc Natl Acad Sci USA 90: 5618–5622.
parameters of the type appearing in our model, Chica RA, Doucet N, Pelletier JN. (2005). Semi-rational approaches to engineering enzyme activity: combining which are not constrained to lie in a known interval.
the benefits of directed evolution and rational design.
We note, though, that when varying the value of m in our analysis, we effectively used a Bayesian-like Daugherty PS, Chen G, Iverson BL, Georgiou G. (2000).
approach with a non-informative prior for estima- Quantitative analysis of the effect of the mutation frequency on the affinity maturation of single chain Fv antibodies. Proc Natl Acad Sci USA 97: 2029–2034.
improving bacterial bioremediation and prodrug Dennett DC. (1995). Darwin’s Dangerous Idea: Evolution and the Meanings of Life. Simon & Schuster Inc.:New York, NY.
Drummond DA, Iverson BL, Georgiou G, Arnold FH.
(2005). Why high-error-rate random mutagenesis li- braries are enriched in functional and improvedproteins. J Mol Biol 350: 806–816.
We are grateful to Drs Bruno Salles, Mike Benoit and Fox RJ, Davis SC, Mundorff EC, Newman LM, Gavrilovic Ms Mimi Keyhan for their useful advice and stimulating V, Ma SK et al. (2007). Improving catalytic function by discussion. We thank Dr Stephen H Thorne for kindly ProSAR-driven enzyme evolution. Nat Biotech 25: supplying us with freshly made JC breast cancer cells. We also thank three anonymous referees whose insightful Grove JI, Lovering AL, Guise C, Race PR, Wrighton CJ, comments and suggestions greatly improved this article.
White SA et al. (2003). Generation of Escherichia coli This work was supported by Grants DE-FG03-97ER- nitroreductase mutants conferring improved cell 624940 and DE-FG02-96ER20228 from the Natural and sensitization to the prodrug CB1954. Cancer Res 63: Accelerated Bioremediation Program of US Department of Energy, and Stanford Office of Technology Licensing Kauffman SA, Levin S. (1987). Towards a general theory (1105626-100-WOAAA). YB and DFA were supported, of adaptive walks on rugged landscapes. J Theor Biol in part, by a Postdoctoral Fellowship from Lady Davis Postdoctoral Fellowship and FRST New Zealand Kuipers OP, Boot HJ, de-Vos WM. (1991). Improved site- (STAX0101) Fellowship, respectively.
directed mutagenesis method using PCR. NucleicAcids Res 19: 4558.
Lejon T, Strom MB, Svendsen JS. (2001). Antibiotic activity of pentadecapeptides modeled from aminoacid descriptors. J Pept Sci 7: 74–81.
Mee RP, Burton TR, Morgan PJ. (1997). Design of active analogues of a 15-residue peptide using D-optimal Ackerley DF, Barak Y, Lynch SV, Curtin J, Matin A. (2006).
design, QSAR and a combinatorial search algorithm.
Effect of chromate stress on Escherichia coli K12.
Minagawa H, Hiroki K. (2000). Effect of double mutation Ackerley DF, Gonzalez CF, Park CH, Blake R, Keyhan M, on thermostability of lactate oxidase. Biotechnol Lett Matin A. (2004). Chromate reducing properties of soluble flavoproteins from Pseudomonas putida and Minagawa H, Yoshida Y, Kenmochi N, Furuichi M, Escherichia coli. Appl Environ Microbiol 70: 873–882.
Shimada J, Kaneko H. (2007). Improving the thermal Aharoni A, Gaidukov L, Khersonsky O, Gould McQS, stability of lactate oxidase by directed evolution. Cell Roodveldt C, Tawfik DS. (2005). The ‘evolvability’ of promiscuous protein functions. Nat Gen 37: 73–76.
Nov Y, Wein LM. (2005). Modeling and analysis of protein Aita A, Husimi Y. (2000). Adaptive walks by the fittest design under resource constraints. J Comput Biol 12: among finite random mutants on a Mt. Fuji-type fitness landscape. J Math Biol 41: 207–231.
Park C-H, Keyhan M, Wielinga B, Fendorf S, Matin A.
Arnold FH. (1998). Enzyme engineering reaches the (2000). Purification to homogeneity and charcteri- boiling point. Proc Natl Acad Sci USA 95: 2035–2036.
zation of a novel Pseudomonas putida chromate Arnold FH. (2006). Fancy footwork in the sequence space reductase. Appl Environ Microbiol 66: 1788–1795.
Park H-S, Nam SH, Lee JK, Yoon CN, Mannervik B, Barak Y, Ackerley DF, Dodge CJ, Lal B, Cheng A, Francis Benkovic SJ et al. (2006). Design and evolution of new AJ et al. (2006b). Analysis of novel soluble Cr(VI) and catalytic activity with an existing protein scaffold.
U(VI) reductases and generation of improved enzymes using directed evolution. Appl Environ Microbiol 72: Qian Z, Lutz SJ. (2005). Improving the catalytic activity of Candida antarctica lipase B by circular permutation.
Barak Y, Thorne SH, Ackerley DF, Lynch SV, Contag CH, Matin A. (2006a). New enzyme for reductive cancer Sambrook J, Fritsch EF, Maniatis T. (1989). Molecular chemotherapy (YieF) and its improvement by directed Cloning: a Laboratory Manual, 2nd edn. Cold Spring evolution. Mol Cancer Ther 5: 97–103.
Harbour Laboratory Press: Cold Spring Harbor, NY.
Chatterjee R, Yuan L. (2006). Directed evolution of Stemmer WP. (1994). DNA shuffling by random fragmen- metabolic pathways. Trends Biotech 24: 28–38.
tation and reassembly: in vitro recombination for The ISME Journal
A statistical approach for enzyme improvementY Barak et al molecular evolution. Proc Natl Acad Sci USA 91: Teixeira LSG, Costa ACS, Ferreira SLCM, Freitas LM, Carvalho S. (1999). Spectrophotometric determination Suzuki FC, Christians B, Kim A, Skandalis MEB, Loeb LA.
of uranium using 2-(2-thiazolylazo)-p-cresol (TAC) (1996). Tolerance of different proteins for amino acid in the presence of surfactants. J Braz Chem Soc 10: Supplementary Information accompanies the paper on The ISME Journal website The ISME Journal

Source: http://stat.haifa.ac.il/~yuval/papers/2008-ISME-enzyme-improvement.pdf