R E S EARCH ART I C L E

MOLECULAR EVOLUT ION

Department of Chemical and Biomolecular Engineering, Johns Hopkins University,3400 North Charles Street, Baltimore, MD 21218, USA.*Present address: Editas Medicine, 300 Third Street, First Floor, Cambridge, MA 02142,USA.†Corresponding author. E-mail: oster@jhu.edu

Steinberg and Ostermeier Sci. Adv. 2016; 2 : e1500921 22 January 2016

2016 © The Authors, some rights reserved;

exclusive licensee American Association for

the Advancement of Science. Distributed

under a Creative Commons Attribution

NonCommercial License 4.0 (CC BY-NC).

10.1126/sciadv.1500921

Environmental changes bridgeevolutionary valleys

Barrett Steinberg* and Marc Ostermeier†

Dow

nloade

In the basic fitness landscape metaphor for molecular evolution, evolutionary pathways are presumed to followuphill steps of increasing fitness. How evolution can cross fitness valleys is an open question. One possibility isthat environmental changes alter the fitness landscape such that low-fitness sequences reside on a hill in alter-nate environments. We experimentally test this hypothesis on the antibiotic resistance gene TEM-15 b-lactamaseby comparing four evolutionary strategies shaped by environmental changes. The strategy that included initialsteps of selecting for low antibiotic resistance (negative selection) produced superior alleles compared with theother three strategies. We comprehensively examined possible evolutionary pathways leading to one such high-fitness allele and found that an initially deleterious mutation is key to the allele’s evolutionary history. This mu-tation is an initial gateway to an otherwise relatively inaccessible area of sequence space and participates inhigher-order, positive epistasis with a number of neutral to slightly beneficial mutations. The ability of negativeselection and environmental changes to provide access to novel fitness peaks has important implications fornatural evolutionary mechanisms and applied directed evolution.

d

on D

ecember 28, 2020

http://advances.sciencemag.org/

from

INTRODUCTION

The fitness landscapemetaphor is often depicted as a three-dimensionalhilly surface with distinct fitness peaks and valleys (Fig. 1). This repre-sentation flattens the high dimensionality of actual landscapes to aid ourunderstanding ofwhat a fitness landscape represents. Fitness landscapesdepict the change in fitness along possible evolutionary pathways (1),and thus, the shape of the fitness landscape affects the dynamics andoutcomes of evolution. Accordingly, much effort has gone into under-standing the shape of the landscape and the degree to which ruggednessmanifests in actual evolutionary scenarios (2–5). Ruggedness resultsfrom complex interactions betweenmutations, especially those exhibit-ing reciprocal sign epistasis (6). In rugged landscapes, local fitness max-ima exist as “traps,” and therefore, the evolutionary endpoint ofadaptation from a given starting point is not predetermined as it wouldbe in a smooth landscape. The endpoint of each evolutionary pathwaydepends on selective history, landscape topology, and chance. However,one way in which evolution might escape traps and cross fitness valleysis through changes in the environment that remodel the landscape (1).Natural environments are intrinsically variable, and evolutionary path-ways are contingent on past environmental interactions and selectivehistory (7). Thus, the dynamics and complexity of selective pressuresare likely to contribute to the routes of natural evolutionary pathways.There is a growing appreciation of the complex and important rela-tionship between the environment and the fitness and epistatic effectsof mutations (8, 9). Understanding the roles that environmentalchanges and landscape ruggedness play in adaptive pathways wouldallow insights into the predictability and potentiality of evolutionarypathways.

How changes in environment or selection affect evolutionary out-comes is underexplored experimentally (8). Oneway to address this topicis to ask whether increasingly stringent positive selection is the beststrategy for optimization. This strategy is ubiquitously used in directed

evolution experiments; however, on a rugged landscape, positive selec-tion alone may quickly reach a local maximum, limiting evolutionarycapacity. Studies of evolution under strong selection have indicated thatevolution can follow few mutational pathways using positive selection,supporting the predictability of evolutionary endpoints (10). How evo-lution might cross fitness valleys remains an unresolved problem (1, 2).Models and simulations have suggested that random strategies, such asstochastic tunneling or recombination, may occasionally result in thecrossing of fitness valleys (11, 12). Other studies have examined theeffects of changing the starting point of evolution to arrive at differentfitness maxima (13), but experiments rarely fundamentally change thenature of selection over the course of experimental evolution. Excep-tions such as selection on alternating antibiotics (14) and a regimenof neutral drift followed by positive selection (15–21) have met withmixed success in evolving improved function.

RESULTS AND DISCUSSION

Experimental designWe wondered whether complex selection dynamics could facilitate thecrossing of fitness valleys. Specifically, we examined whether environ-mental changes causing dynamic shifts in a protein fitness landscapewould result in evolutionary pathways ending in distinct, potentially su-perior evolutionary endpoints. We designed an experiment in whichenvironmental changes enabled selection toward sequences previouslyresiding in a low-fitness valley and tested whether such an evolutionarypath would arrive at distinct and superior outcomes (Figs. 1 and 2).We used Escherichia coli having a bandpass genetic circuit designedfor tunable selection of b-lactam antibiotic resistance as amodel systemfor systematically examining such a scenario (22). For cellswith this genecircuit, b-lactam antibiotics such as ampicillin are required to induce theexpression of tetracycline resistance and to do so even at sublethal levelsof the b-lactam.When cells are plated in the presence of tetracycline andampicillin, growth requires enough b-lactamase activity to degrade theampicillin below lethal levels, but not too much so as to remove the sig-nal that induces tetracycline resistance. This genetic circuit allowed us to

1 of 9

R E S EARCH ART I C L E

on Decem

ber 28, 2020http://advances.sciencem

ag.org/D

ownloaded from

perform selections for b-lactam resistance within a narrow window ofresistance (that is, one can select for resistance above a first threshold butbelow a second, higher threshold) (Fig. 2A). Thus, selections for low butnonzero resistance can be performed. We term such selections “nega-tive” selections in the sense that high fitness (that is, the ability of the cellsto grow) in such an environment is achieved only by alleles conferringlow levels of b-lactam resistance. Similarly, neutral selection for mainte-nance of about the original level of resistance can be performed. Thesebandpass selections are tuned by the level of the b-lactam antibioticin the environment and require the addition of tetracycline. Thus, theseenvironments reshape the fitness landscape analogously to the de-piction in Fig. 1, and our engineered bacteria are a suitable model sys-tem for examining evolution under landscape-altering environmentalchanges.

The b-lactamase allele TEM-1 provides high antibiotic resistance toampicillin but very low resistance to cefotaxime. Clinically isolated al-leles of TEM-1 conferring increased cefotaxime resistance includeTEM-15 (E104K/G238S) and TEM-52 (E104K/M182T/G238S), withthe latter increasing resistance 4000-fold (23). Six independent directed

Steinberg and Ostermeier Sci. Adv. 2016; 2 : e1500921 22 January 2016

evolution experiments on TEM-1 that imposed positive selection for ce-fotaxime resistance found these same three mutations in the highest-resistance alleles (13, 24–28). The GKTS allele, which adds the A42Gmutation, has the highest resistance observed in these experiments(10, 27, 29), though other combinations of amino acids at these po-sitions can confer slightly higher resistance (30). These directed evolu-tion studies have used a typical methodology: mutation of the gene,positive selection for higher resistance, and iteration of these processesover several rounds.

Evolution experimentsWe subjected TEM-15 to eight rounds of experimental evolution usingfour different selection strategies: positive selection, neutral selection,negative selection, and oscillating selection (Fig. 2, B to D). As a controlfor evolution from low-resistance alleles using only positive selection,we also subjected TEM-1 to positive selection. The differences in theselection strategies occurred in the first three rounds of evolution andwere followed by five rounds of positive selection to climb toward localfitness maxima. The positive selection experiments represent evolutionin a fixed environment with a static landscape and are the typicaldirected evolution strategy. Because of the bandpass selection, our neu-tral selection avoids selection for alleles conferring resistance far abovethe starting allele, unlike previous directed evolution studies using neu-tral drift (15–21). During the negative selection steps in the oscillatingand negative selection schemes, we set the level of resistance needed forgrowth above the background level of resistance exhibited by E. coliwithout a b-lactamase gene.We reasoned thatmutations that complete-ly deactivate a proteinmight be exceedingly common (such as nonsensemutations), whereas maintenance of above-background level of activitywould better reflect the desired scenario that the gene is still required forcell viability and that it encodes a protein with some enzyme activity.

We constructed five plasmidswithdistinctDNAbarcodes identifyingthe five experiments tomonitor and ensure against cross-contaminationbetween the five different selection experiments. One round of evolutionconsisted of comprehensive codonmutagenesis on the plasmid-borne b-lactamase genes (isolated from the previous round), transformation ofthe newly mutated plasmid library into fresh E. coli cells, and selectionon agar plates for the desired level of cefotaxime resistance (Fig. 2A).At least 106 library members were generated and subjected to selec-tion in each round. Selections were tuned such that typically >100 andsometimes thousands of independent clones appeared at each step tolimit the effects of evolutionary bottlenecking on the experimental out-come (data S1). Populations under negative and oscillating selectionrecovered from negative selection to resistance above that ofTEM-15after one round of positive selection (Fig. 2, D and E). This result sup-ports the idea of the commonality of compensating mutations or sup-pressor mutations in restoring protein activity (31, 32) but might alsoreflect a reversal of deleteriousmutations in combinationwith beneficialmutations.

Evolution outcomesAt the end of the experiment, all strategies resulted in bacteria with a highcefotaxime resistance phenotype. The coding sequences of the genes from47 of the most resistant colonies from each selection strategy were se-quenced (data S2 to S6). All selection strategies showed sequenceclustering, indicating that each strategy, for themost part, reacheddistinctareas of sequence space (Fig. 3A). This result emphasizes how much theenvironment and the nature of selective pressure can shape evolutionary

A

C

B

Fig. 1. Simplified model representation of how evolution in alter-nating environments that modulate the fitness landscape can lead tothe crossing of fitness valleys. The xy plane represents sequence space,whereas the z axis represents fitness. The solid red arrow indicates an evo-lutionary pathway following increasing positive selection pressure. (A) Asequence (circle) sits at an optimum separated from global optima by afitness valley. (B) A change in the environment alters the landscape such thatevolution brings the gene to a new sequence that previously resided in thevalley. (C) Upon return to the original environment, the sequence resides inthe valley but can now evolve to the global optimum, which lies uphill.

2 of 9

R E S EARCH ART I C L E

on Decem

ber 28, 2020http://advances.sciencem

ag.org/D

ownloaded from

outcomes. Sequences resulting from negative selection showed both thehighest average genetic distance from TEM-15 (Fig. 3A and fig. S1) andthe highest diversity (fig. S2). We discerned no apparent pattern in thelocation of mutations on the protein structure (fig. S3). We chose sevendistinct sequences from each selection experiment (based on diversitywithin the tree) to subclone and analyze the effects of mutations on thecoding sequence as opposed to those elsewhere on the plasmid. Thecoding sequences were ligated into a fresh plasmid backbone andtransformed into fresh E. coli cells for resistance testing. All librariesproduced alleles that conferred resistance higher than that conferred byGKTS, though an approximately eightfold drop in resistance upon sub-cloning of the genes indicated that mutations elsewhere on the plasmidscontributed to the high resistance observed after round 8. We supposedthat mutations elsewhere likely consisted of those that increased plasmidcopy number or promoter strength. Strikingly, the negative selectionstrategy evolved alleles that conferred the highest resistance, with fiveof the seven alleles providing an eightfold resistance increase over thatof GKTS (Fig. 3B). This result is interesting in light of theoretical predic-tions that a low starting fitnessmay dramatically enhance the chances fora population to reach the global maximum (33).

Deep sequencing of populations during the course of thenegative selection experimentWe next sought insights into how negative selection produced this out-come. Deep sequencing of the population at each round of the negativeselection regimen revealed a high expansion of genetic diversity duringnegative selection, consistent with our expectation that deleterious muta-tions are common (fig. S4). In contrast, during the subsequent positive

Steinberg and Ostermeier Sci. Adv. 2016; 2 : e1500921 22 January 2016

selection steps, increasing purifying selection is apparent (fig. S4). In par-ticular, the A184I mutation was highly enriched in the population (figs.S5 and S6). However, a large amount of diversity was maintainedthroughout the experiment. A total of 5836 unique sequences werefound in deep sequencing data after the first positive selection, demon-strating that, despite rapidly changing modes of selection, we main-tained a high diversity of alleles following the negative selectionregimen. In addition, 85% (40 of 47) of the sequences from the finalselection were unique, with an average difference of 6.1 AA mutationsbetween unique mutants.

Evolutionary pathway reconstruction using comprehensivefitness measurementsTo determine whether the environmental change resulting in negative se-lectionwas a key factor in the evolutionary history of an allele conferring ahigh resistance,we attempted to reconstruct all possible evolutionary inter-mediates between TEM-15 and one of the fittest alleles resulting from thenegative selection regimen (BS-NEG-4) (fig. S7). This allele contained ninemissense mutations relative to TEM-15 [A11S, A42G, L49M, Q90R,A184I, I208W, A224V, F230S, and R241F in Ambler notation (34)] andconferred an eightfold higher cefotaxime resistance thanGKTS. Amongthemutations in BS-NEG-4, A11S, A42G, L49M, andA224Vhave beenreported in otherTEM alleles, though none were present together in thesame allele (35). We created a library designed to contain all 512 com-binations of these ninemutations to determine the protein fitness (w) ofall possible evolutionary intermediates in the evolution of BS-NEG-4(assuming no mutation reversals or multiple mutations at the sameamino acid position). Protein fitness quantifies the relative ability of

β-Lactam resistance

Sel

ectio

n co

effic

ient

Negative PositiveNeutralSelection

Comprehensive saturation mutagenesis

Round0 2 4 6 8

Sel

ecte

d ce

fota

xim

ere

sist

ance

(µg

/ml)

Sel

ecte

d ce

fota

xim

ere

sist

ance

(µg

/ml)

Sel

ecte

d ce

fota

xim

ere

sist

ance

(µg

/ml)

Sel

ecte

d ce

fota

xim

ere

sist

ance

(µg

/ml)

10–2

100

102

Positive

TEM-15TEM-1

Round0 2 4 6 8

10–2

100

102

Neutral

Round0 2 4 6 8

10–2

100

102

Oscillating

Round0 2 4 6 8

10–2

100

102

Negative

A CB

D E

Fig. 2. Schematic and course of experimental evolution on the b-lactamase gene. (A) General schematic. After a comprehensive saturation muta-genesis, selection proceeds in one of three manners depending on the environment. Positive selection (blue) enriches for sequences that provideresistance above a certain level. Neutral selection (yellow) enriches for sequences conferring about the same resistance as the starting gene. Negativeselection (red) enriches for sequences conferring very low but above-background resistance. The latter two selection schemes are enabled by an engi-neered bandpass gene circuit present in the cell (22) and require the addition of the antibiotic tetracycline and different levels of cefotaxime, whichconstitute the difference in environment from the positive selection. (B to E) Course of the evolution experiments. Experiments differ in the first threerounds in which the bacteria underwent (B) positive selection, (C) neutral selection, (D) negative selection, or (E) oscillating selection. Each round repre-sents bothmutagenesis (line segments) and selection at the indicated level of cefotaxime (endpoints). Red circles indicate bandpass selections, and bluecircles indicate positive selections. The dashed horizontal line indicates the resistance level conferred by TEM-15. Graphed resistance values indicate alower limit on the resistance values of the entire population during selection (that is, the concentration at which the selection was performed).

3 of 9

R E S EARCH ART I C L E

on Decem

ber 28, 2020http://advances.sciencem

ag.org/D

ownloaded from

the protein to confer a cefotaxime resistance phenotype: the higher theresistance, the higher the fitness. We determined the fitness of 98.8%(506 of 512) of all possible combinations of these nine mutations in ahigh-throughput manner involving bandpass selection and deep se-quencing (data S7) (36). This method uses the bandpass circuit. Cellswith a given level of cefotaxime resistancewill grow best within a certainrange of cefotaxime concentrations. A cell with higher resistance requireshigh concentrations, and a cell with lower resistance requires lower con-centrations. Fitness is thus determined by plating the library on platescontaining tetracycline and different cefotaxime concentrations, deep se-

Steinberg and Ostermeier Sci. Adv. 2016; 2 : e1500921 22 January 2016

quencing the b-lactamase gene on each plate, counting how many timeseach allele appears on each plate, and calculating the center of the con-centration range at which the allele allows growth. This center concentra-tion can be thought of as a proxy for theminimum inhibitory concentration(MIC) for cefotaxime. Fitness is expressed as this center concentrationnormalized to the center concentration of a reference (in this case, theresistance conferred byTEM-15). Onemutation, F230S, was particularlydeleterious (wF230S = 0.09± 0.01,with the fitness of TEM-15 set to a valueof 1.0) and remained so even in combination with one to two other mu-tations, making it a highly likely starting point on the evolutionary path-way resulting in BS-NEG-4.

Analysis of fitness landscapes and evolution pathways usingselection-weighted attraction graphingWesought amethod to visualize thehigh-dimensional space of the fitnesslandscape composed of the 512 potential evolutionary intermediates be-tween TEM-15 and BS-NEG-4. For this purpose, we introduce selection-weighted attraction graphing (SWAG), a force-directed graphingmethodinwhich nodes represent sequences and edges connect nodes differing byonly onemissensemutation (Fig. 4). Forces between connected nodes areweighted proportionally to the selection strength, as determined by thenodes’ respective fitness values. These forces are applied on the xy plane,and the graph can be depicted in three dimensions by adding fitness asthe z axis. In the graph, edges represent possible single steps in the evo-lutionary pathway, and the distance between connected nodes on the xyplane is inversely proportional to the likelihood of taking that step underpositive selection.

SWAGanalysis of the intermediates betweenTEM-15 and BS-NEG-4resulted in eight major clusters of sequences representing local fitnesspeaks (Fig. 4, A and B). Such clustering was robust to the weighting ofselection strength in the forces. No clustering was observed when fitnessvalues were randomly reassigned to the nodes. Clusters comprise groupsof sequences with many interconnections (fig. S8) that tend to lead insuccession to ever-increasing fitness values. The graph canbe seen tohavetwo “ranges”of four clusters each separatedby a large gapdevoid of nodes(Fig. 4). The two ranges differ by whether their sequences contain F230S(Fig. 4C). The range containing F230S lies far from the start point TEM-15 and contains the endpoint BS-NEG-4. The proximal range containsTEM-15andnearbynodes. Interrange edgesprimarily connect low-fitnessalleles, typically with the lower fitness node lying in the F230S range.

Although we cannot exclude the possibility that BS-NEG-4 could re-sult from positive selection, we can say that the evolution of BS-NEG-4 isextremely unlikely under positive selection. Of the 9! = 362,880 possiblepathways between TEM-15 and BS-NEG-4, we could analyze 347,616(the missing pathways are due to six alleles not being frequent enoughin the library to analyze fitness data on those alleles). Of the 347,616 path-ways analyzed, only 16 involve increasing fitness as each mutation isadded (Fig. 5). Furthermore, evolution under positive selection is morelikely to terminate at intermediates that are local maxima than to reachBS-NEG-4 (Fig. 6). The SWAG analysis pictorially depicts this result.First, eight of the possible nine initial mutations (the exception beingF230S) are located in the proximal range (fig. S9). Second, there aremanymore steps within a cluster than between clusters (fig. S8A). Finally, stepswithin a cluster tend to be fitter sequences than steps between clusters (fig.S8B),making intercluster steps even less likely. Thus, starting fromTEM-15, the most likely mutational pathways in a constant, positive selectionenvironment progress first into one of the clusters of the proximal rangeand then preferentially climb within that cluster to local fitness maxima.

1. 0

Oscillating

NegativePositive(TEM-1)

Positive(TEM-15)

NeutralA

B

Fol

d in

crea

se in

cef

otax

ime

resi

stan

ce o

ver

TE

M-1

0

2000

4000

6000

8000

10,000

12,000

14,000

TEM-15TEM-1

Positive Positive Neutral Negative Oscillating

Fig. 3. Diversity and conferred resistance of alleles after the eighthround of evolution. (A) Unrooted tree of Hamming distances between ran-domly selected alleles. Branches are color-coded by the evolution experi-ment as indicated. Scale bar indicates 1 unit of Hamming distance (that is,onemissensemutation). (B) Increase in resistance conferred by seven allelesfrom each selection strategy as compared to TEM-1. Resistance wasmeasured using the MIC assay. The seven alleles were selected from amongthose depicted in the tree to represent diverse sequences. The dashed linerepresents resistance conferred by theGKTS allele, whereas the TEM-15 alleleconferred a 64-fold higher resistance relative to TEM-1.

4 of 9

R E S EARCH ART I C L E

on Decem

ber 28, 2020http://advances.sciencem

ag.org/D

ownloaded from

However, when the environment changes such that low resistance is ben-eficial (that is, under negative selection), mutational steps to low-fitnessvalues in the depths of the F230S range become advantageous (Fig. 4D).Then, upon a reversion back to the original environment (that is, back topositive selection), sequences tend to evolve up from the depths in theclusters of the F230S range, enabling access to BS-NEG-4. The environ-mental change can cause interrange bridges to form across gaps in theSWAG landscape that end in “gateway” sequences having F230S,allowing access to the sequence space of that range of peaks.Without thisbridge, evolution of BS-NEG-4 is extremely unlikely. We conclude thatevolution of BS-NEG-4 is extremely unlikely to occur under positive se-lection but can be accessed through a series of environmental changesthat modulate the selective pressure on the gene.

A deleterious gateway mutation allowing passage through afitness valley via epistatic effectsWenext turned our attention to the role of F230S in the evolution of BS-NEG-4. F230S is very deleterious to TEM-15 for cefotaxime resistancebut is only mildly deleterious to ampicillin resistance in TEM-1 (36).The different effects of F230S on TEM-1 and TEM-15 might originatefrom a strongnegative pairwise epistatic interactionwithG238S, which

Steinberg and Ostermeier Sci. Adv. 2016; 2 : e1500921 22 January 2016

lies at the opposite end of the same b-strand. Because we suspectedF230S to be key in the evolution of this allele, we compared the epistaticand fitness effects of mutations of evolutionary intermediates containingF230S to intermediates lacking F230S (Fig. 7). Although alleles withF230S tend to have lower fitness, especially among intermediates withfew mutations (Fig. 7A), several intermediates demonstrate higherfitness with the addition of F230S, including the global optimumamongthese sequences, which contains all but the Q90R andA184Imutations.We individually verified the MIC conferred by several of the fittest var-iants that contain F230S and observed up to a 32-fold improvement inresistance over that provided by GKTS (data S8). Seven of nine of themutations of BS-NEG-4 are slightly beneficial as the sole mutation inTEM-15. However, the higher-order epistatic effect of combinations ofthesemutations in the absence of F230S is increasingly negative (Fig. 7B),consistent with previous observations that beneficial mutations usuallyinteract with negative epistasis, creating diminishing returns (37–39). Incontrast, higher-order epistatic interactions in intermediates containingF230S are almost always positive [252 of 256 (98%)] and become in-creasingly positive as mutations accumulate up to about four to fivemutations, where the higher-order epistasis plateaus at e = +0.7. Thisepistasis value means that the average fitness of proteins having the

BA

DC

w

w w

TEM-15

–F230S

+F230S

TEM-15

BS-NEG-4

BS-NEG-4

Fig. 4. SWAG of the possible evolutionary intermediates between TEM-15 and BS-NEG-4. Nodes are sequences, and edges connect nodes that differby onemissensemutation. Forces between connectednodes areweightedproportional to the selection strength, as determinedby thenode’s fitness values.(A) Two-dimensional SWAG. Edges are curved to aid visualization. (B toD) Three-dimensional SWAG has protein fitness as the z axis and is color-coded by (B)cluster, (C) whether the sequence contains F230S, and (D) protein fitness value.

5 of 9

R E S EARCH ART I C L E

on Decem

ber 28, 2020http://advances.sciencem

ag.org/D

ownloaded from

F230Smutation is five times higher than expected based on the effectsof the individualmutations in TEM-15.We emphasize that this strongpositive epistasis with F230S only occurs in the higher-order epistasisterms. The benefits of F230S are not so apparent when one examinesthe effects of F230S on pairwise epistasis among the other eight muta-tions, though F230S tends to have a slightly positive effect on the

Steinberg and Ostermeier Sci. Adv. 2016; 2 : e1500921 22 January 2016

pairwise epistasis between certain mutations (fig. S10). We suggestthat F230S is more than just a key gateway mutation for traversingthe fitness valley. F230S also serves to alleviate the negative interactionsbetween beneficial mutations, creating the pathway to the final high-activity allele.

ConclusionsOur results are important for applied directed evolution experiments,which almost exclusively use positive selection as a strategy. In ourstudy, a selection regimen including negative selection generated thehighest activity variants analyzed and outperformed strict positive selec-tion. This result may appear counterintuitive to protein engineers butcould prove useful in escaping plateaus in protein functionality. A fewdirected evolution studies have demonstrated the advantage of neutraldrift in the evolution of activity toward alternate substrates (15–21), butour results are the first to show that a regimen involving negative selec-tion can rival or even, perhaps, outperform other strategies, includingstrict positive selection and neutral drift. Although the neutral selectionregimen performedhere also accesses alternate fitnessmaxima, negativeselection reached more distant sequences with superior properties onaverage. Thus, negative selectionmay be amore effective means of gen-erating highly diverse and effective alternative starting points for di-rected evolution experiments.

Our study lends further credence to models of rough fitness land-scapes shaped by epistasis. SWAG enabled us to visualize and gain in-sights into potential evolutionary pathways using network analysis. Ourtechnique creates statistical inference fromevolutionary data that can be

BA

Fitness

evitisoPevitageN Neutral

Fig. 5. Radial graph of all analyzed pathways from TEM-15 to BS-NEG-4. Each unit branch extending from the center indicates the gain of one of thenine amino acids fromTEM-15 until the endpoint containing the full set ofmutations is reached. (A) The line color indicates the log of the fitness change from

the addition of the mutation. This graph pictorially shows that most pathways have steps with decreased fitness. (B) Same data as in (A), but with the colorindicating the fitness relative to TEM-15 after the addition of the mutation. This graph pictorially shows that early steps in most pathways pass throughmutants with fitness values less than that of TEM-15. Data is the same in A and B; only color coding is different.

BS-NEG-4

TEM-15

w

Fig. 6. Three-dimensional SWAGof all evolutionary pathways inwhichfitness increasesmonotonicallywith eachadditionalmutation. Fitness isrepresented by height on the z axis. Color indicates clusters. Each node issized by the likelihood of arrival, as measured by PageRank algorithmweighted by selection strength between nodes.

6 of 9

R E S EARCH ART I C L E

on Decem

ber 28, 2020http://advances.sciencem

ag.org/D

ownloaded from

understood intuitively. We demonstrate SWAG visualization of a land-scape with an evolutionary valley and multiple fitness peaks. SWAGlandscape visualization should be broadly useful for depicting fitnesslandscapes and evolutionary pathways. Our experiments address thepredictions of Wright (1) regarding the importance of environmentalchange in evolutionary processes. We present evidence that environ-mental changes that alter the fitness landscape can extend access toareas of sequence space that are difficult to explore in a constantenvironment. In our particular example, this access resulted in superiorevolutionary outcomes. Whether our result is the exception or the rulewill require testing of other genes and systems.

MATERIALS AND METHODS

PlasmidsAll experiments used the expression plasmid pSkunk3 with a 9–base pair(bp) barcode indicating selection type downstream of the b-lactamasegene. The barcodes didnot alter the level of antibiotic resistance conferredby theplasmid.TheplasmidpSkunk3 contains the geneunder the control

Steinberg and Ostermeier Sci. Adv. 2016; 2 : e1500921 22 January 2016

of the isopropyl-b-D-thiogalactopyranoside (IPTG)–inducible tacpromoter andconfers spectinomycin resistance. Inductionofb-lactamaseexpression was accomplished using 300 mM IPTG and the maintenanceantibiotic spectinomycin (50 mg/ml).

MutagenesisPFunkel mutagenesis was performed as described (36, 40). Briefly,plasmids were transformed into E. coli CJ236 to prepare single-strandedtemplate DNA containing uracil, using filamentous phage R408.Mutagenic oligonucleotides with degenerate codon NNN were tar-geted to every codon in the b-lactamase gene. Following mutagen-esis, all reaction libraries were transformed into electrocompetent NEB5-alpha cells (New England Biolabs) to prepare a naïve library foreach round. All naïve libraries exceeded 106 transformants followingmutagenesis.

SelectionFor bandpass experiments, plasmid libraries were isolated from the NEB5-alpha cells and transformed intoE. coli SNO301 in the presence of plas-mid pTS40 (22). Bandpass selections were performed, as described pre-viously (22, 36), on solid LB agar plates with spectinomycin (25 mg/ml),chloramphenicol (25 mg/ml), and 300 mM IPTG at 37°C for 24 hours[except for tetracycline (10 mg/ml) and various concentrations of cefotax-ime]. Selections under non-bandpass conditions were performed in NEB5-alpha cells on LB broth agar containing 300 mM IPTG, spectinomycin(25mg/m), and various concentrations of cefotaxime at 37°C for 24hours.Roughly 106 colony-forming units were plated for each selection experi-ment. Cell stocks were not previously exposed to b-lactam antibioticsbefore selection. At the end of the experiment, the 47 alleles were pickedfor sequencing based on colony size on plates, with antibiotics at the finalselection level (that is, the largest 47 colonies were picked, first from thehighest-concentration plate and then from the next highest-concentrationplate, as necessary to reach 47).

Screening and prevention of cross-contaminationBetween rounds, the 9-bp barcodes on each plasmid were used tomonitor for contamination via a polymerase chain reaction (PCR)–based screen. Primers were designed to specifically amplify DNA con-taining each barcode to diagnose the presence of cross-contaminationbetween the selection experiments. In the event of contamination,the desired fraction of the contaminated library was amplified withthe appropriate barcodes and ligated into fresh vector to eliminate thecontaminants.

Deep sequencing of populations duringevolution experimentsBacteria were recovered from plates with LB broth containing 10%glycerol and pelleted by centrifugation. Plasmid DNA was isolatedfrom the pelleted cells. Amplicons were prepared for deep sequenc-ing via PCR from each plasmid prep with the addition of a 3-bp bar-code on either side of the amplicons indicating the round from whicheach amplicon was isolated. PacBio deep sequencing was performedon these amplicons using a three-pass circular consensus. CustomMATLAB scripts described previously (36) were used to align, analyze,and quantify reads and mutation composition. Reads were filtered forquality score, length, and quantity of insertions and deletions. Eachread was then aligned to the reference gene to identify barcodes andmutations.

Number of mutations

0 2 4 6 8 10

Epi

stas

is o

f eac

h gr

oup

of m

utan

ts

–0.8

–0.6

–0.4

–0.2

0

0.2

0.4

0.6

0.8

1

1.2

Number of mutations

0 2 4 6 8 10

Fitn

ess

of e

ach

mut

ant

0

1

2

3

4

5

6

7

8

9F230S absentF230S present

F230S absentF230S present

A

B

Fig. 7. Protein fitness and epistasis along possible evolutionary path-ways between TEM-15 and BS-NEG-4. (A) Protein fitness values of all possi-

ble evolutionary intermediates. (B) N-order epistasis between all mutations asa function of the number ofmutations. Proteins containing F230S are noted inred, whereas all other proteins are shown in blue. Points are jittered for clarity.Largerpoints indicate themean foreachdata set,withbars extending to theSD.

7 of 9

R E S EARCH ART I C L E

on Decem

ber 28, 2020http://advances.sciencem

ag.org/D

ownloaded from

Deep sequencing to determine protein fitnessFitness values were determined essentially as previously described (36)and summarized here in a way that includes the minor differences.Bandpass selection selects for cells with activity at or near a particularlevel, as determined by the cefotaxime concentration. Thus, plating at avariety of different cefotaxime concentrations partitions the library intobins of relative fitness levels. Thus, by deep sequencing the plasmidDNA to determine the frequency of each allele on each plate, the fitnessfor each mutant can be determined.

To determine fitness values, we performed selections over 11 platescontaining doubling levels of cefotaxime (from 0.01 to 10.24 mg/ml) un-der bandpass conditions [300 mM IPTG, spectinomycin and chloram-phenicol (25 mg/ml), and tetracycline (10 mg/ml) on LB broth agar at 37°Cfor 24 hours]. Selected libraries were swept with LB brothwith 10% glyc-erol, cells were pelleted, and plasmidDNAwas isolated. Amplicons weregenerated directly from each plasmidDNAprep. During PCR, 3-bp bar-codes indicating the cefotaxime concentration on which each sequencewas found were added at both ends. After amplifying the library fromeach plate, PacBio deep sequencing data were generated for each selec-tion condition. We used custom Matlab scripts to align, analyze, andquantify reads and amino acid mutation composition. Reads werefiltered for quality score (reported average quality score > 30, or averagereported probability of error = 10−3), length (length of read less than1100 bp butmore than 930 bp), and quality of alignment to the referencegene (entirety of reference gene aligned to read) and the barcode (perfectmatch accepted only). We tabulated allele counts for each plate. In ourprevious work (36), the counts were used directly in the fitness calcula-tion. Here, we adjusted the counts to reflect the desired proportion ofsequencing reads from each plate as determined by the frequency ofcolonies appearing on that plate. Next, for each protein sequence, theplate with the highest counts was found, and the counts in that plateand in the surrounding two plates (five in total) were used to calculatethe final fitness measurement as before (36). Only fitnesses for proteinsequences with five or more counts were considered. Fitness values (w)were defined relative to the TEM-15 fitness.

Phylogenetic analysisWe used the Hamming distance between any two given sequences at theamino acid level to construct a distance matrix using custom scripts inMATLAB (fig. S1). Next, we applied a neighbor-joining algorithm inMATLAB to generate the tree from all combined sequences and clusterbranches. The distance tree was visualized using FigTree (41).

Selection-weighted attraction graphingSWAG network analysis was accomplished using Gephi (42) andcalculated fitness values. Forces between nodes were weighted by thecalculated relative selection strength between nodes (sr, the ratio of fit-nesses wN and wN+1 between node N and node N + 1, which differ byone amino acid mutation)

sr ¼ wNþ1

wNð1Þ

Nodes that differed by only one amino acid mutation were connected.The ForceAtlas2 algorithm (43) was used to generate graphs. Clusteringwas accomplished using the modularity algorithm weighted by selectionstrength. PageRank (44) was implemented and also weighted by selectionstrength.

Steinberg and Ostermeier Sci. Adv. 2016; 2 : e1500921 22 January 2016

Epistasis calculationsPairwise epistasis between mutation A with fitness wA and mutation Bwith fitness wB was calculated as

eA•B ¼ log10wABwo

wAwB

� �ð2Þ

in which wo is the wild-type fitness (45). Higher-order epistasis oforder N with mutations i,j,k,…,n was calculated using Eq. 3

ei;j;k;…;n ¼ log10wi;j;k;…;nwN−1

o

Pni wi

!ð3Þ

Among three mutations A, B, and C, the amount by which a third mu-tation changes the epistasis is the same regardless of which mutation isconsidered the third mutation

De3 ¼ eA•BjC−eA•B ¼ eA•CjB−eA•C ¼ eB•CjA−eB•C ð4Þ

in which ei•j|k refers to the pairwise epistasis between i and j in the con-text of an allele containing mutation k. This calculation was used infig. S10.

MIC testingMIC testingwas performed inNEB 5-alpha cells by growing 1ml of eachculture to be tested overnight in LB broth in the presence of spectinomy-cin (50 mg/ml). Saturated growths were diluted 1:100 and grown untilabove OD600 (optical density at 600 nm) = 0.3. All samples were then di-lutedback toOD600=0.003, and1ml was spottedontoLBbroth agar plateswith 300 mMIPTGand various concentrations of cefotaxime in twofold in-crements.Growthwasevaluatedafter24hoursof incubationat37°C.TheMICwas the lowest concentration of cefotaxime at which no growth was visible.

SUPPLEMENTARY MATERIALSSupplementary material for this article is available at http://advances.sciencemag.org/cgi/content/full/2/1/e1500921/DC1Fig. S1. Heat map of the matrix of Hamming similarity between unique proteins found from 47of the most resistant alleles from each selection regimen following round 8.Fig. S2. Comparison of the differences in sequences obtained using the different selectionregimen.Fig. S3. Structural mapping of mutations found in each selection scheme following round 8.Fig. S4. Heat map of the frequency of mutations observed during the negative evolutionregimen, as analyzed by deep sequencing.Fig. S5. Stacked frequency distribution of mutations found by deep sequencing at each step inthe negative evolution regimen as a function of codon position over eight rounds.Fig. S6. Stacked distribution of mutations found in the top 47 alleles from all selectionstrategies following round 8 as a function of codon position in b-lactamase.Fig. S7. Structural mapping of mutations found in the BS-NEG-4 allele, which confers a highresistance to cefotaxime.Fig. S8. Pathways between clusters found in SWAG analysis.Fig. S9. Two-dimensional SWAG landscape of pathways between TEM-15 and BS-NEG-4 withthe nine single mutants indicated.Fig. S10. Change in pairwise epistasis (De3) between two mutations on the pathway from TEM-15 to BS-NEG-4 in response to the addition of F230S.Data S1. Evolutionary history of selection schemes.Data S2. Sequences resulting from the positive selection regimen (TEM-1).Data S3. Sequences resulting from the positive selection regimen (TEM-15).Data S4. Sequences resulting from the neutral selection regimen.Data S5. Sequences resulting from the negative selection regimen.Data S6. Sequences resulting from the oscillating selection regimen.Data S7. Fitness and epistasis values for BS-NEG-4 pathways.Data S8. Additional MIC data.

8 of 9

R E S EARCH ART I C L E

on Decem

ber 28, 2020http://advances.sciencem

ag.org/D

ownloaded from

REFERENCES AND NOTES1. S. Wright, The roles of mutation, inbreeding, crossbreeding and selection in evolution.

Proc. 6th Int. Congr. Genet. 1, 356–366 (1932).2. S. Gavrilets, Fitness Landscapes and the Origin of Species (Princeton Univ. Press, Princeton,

NJ, 2004).3. J. A. G. M. de Visser, J. Krug, Empirical fitness landscapes and the predictability of evolu-

tion. Nat. Rev. Genet. 15, 480–490 (2014).4. M. S. Breen, C. Kemena, P. K. Vlasov, C. Notredame, F. A. Kondrashov, Epistasis as the

primary factor in molecular evolution. Nature 490, 535–538 (2012).5. A. I. Podgornaia, M. T. Laub, Pervasive degeneracy and epistasis in a protein-protein

interface. Science 347, 673–677 (2015).6. F. J. Poelwijk, S. Tănase-Nicola, D. J. Kiviet, S. J. Tans, Reciprocal sign epistasis is a necessary

condition for multi-peaked fitness landscapes. J. Theor. Biol. 272, 141–144 (2011).7. S. J. Gould, Wonderful Life: The Burgess Shale and the Nature of History (W. W. Norton &

Company, New York, 1990).8. K. M. Taute, S. Gude, P. Nghe, S. J. Tans, Evolutionary constraints in variable environments,

from proteins to networks. Trends Genet. 30, 192–198 (2014).9. V. Mustonen, M. Lässig, From fitness landscapes to seascapes: Non-equilibrium dynamics

of selection and adaptation. Trends Genet. 25, 111–119 (2009).10. D. M. Weinreich, N. F. Delaney, M. A. DePristo, D. L. Hartl, Darwinian evolution can follow

only very few mutational paths to fitter proteins. Science 312, 111–114 (2006).11. D. M. Weinreich, L. Chao, Rapid evolutionary escape by large populations from local fitness

peaks is likely in nature. Evolution 59, 1175–1182 (2005).12. C. S. Gokhale, Y. Iwasa, M. A. Nowak, A. Traulsen, The pace of evolution across fitness

valleys. J. Theor. Biol. 259, 613–620 (2009).13. M. L. M. Salverda E. Dellus, F. A. Gorter, A. J. M. Debets, J. van der Oost, R. F. Hoekstra, D. S. Tawfik,

J. A. G. M. de Visser, Initial mutations direct alternative pathways of protein evolution. PLOSGenet. 7, e1001321 (2011).

14. M. F. Schenk, S. Witte, M. L. M. Salverda, B. Koopmanschap, J. Krug, J. A. G. M. de Visser,Role of pleiotropy during adaptation of TEM‐1 b‐lactamase to two novel antibiotics. Evol.Appl. 8, 248–260 (2014).

15. J. D. Bloom, P. A. Romero, Z. Lu, F. H. Arnold, Neutral genetic drift can alter promiscuousprotein functions, potentially aiding functional evolution. Biol. Direct 2, 17 (2007).

16. J. D. Bloom, F. H. Arnold, In the light of directed evolution: Pathways of adaptive proteinevolution. Proc. Natl. Acad. Sci. U.S.A. 106, 9995–10000 (2009).

17. W. S. Smith, J. R. Hale, C. Neylon, Applying neutral drift to the directed molecular evolutionof a b-glucuronidase into a b-galactosidase: Two different evolutionary pathways lead tothe same variant. BMC Res. Notes 4, 138 (2011).

18. K. L. Petrie, G. F. Joyce, Limits of neutral drift: Lessons from the in vitro evolution of tworibozymes. J. Mol. Evol. 79, 75–90 (2014).

19. R. D. Gupta, D. S. Tawfik, Directed enzyme evolution via small and effective neutral driftlibraries. Nat. Methods 5, 939–942 (2008).

20. S. Bershtein, K. Goldin, D. S. Tawfik, Intense neutral drifts yield robust and evolvable con-sensus proteins. J. Mol. Biol. 379, 1029–1044 (2008).

21. G. Amitai, R. D. Gupta, D. S. Tawfik, Latent evolutionary potentials under the neutral mu-tational drift of an enzyme. HFSP J. 1, 67–78 (2007).

22. T. Sohka, R. A. Heins, R. M. Phelan, J. M. Greisler, C. A. Townsend, M. Ostermeier, Anexternally tunable bacterial band-pass filter. Proc. Natl. Acad. Sci. U.S.A. 106, 10135–10140(2009).

23. C. Poyart, P. Mugnier, G. Quesne, P. Berche, P. Trieu-Cuot, A novel extended-spectrumTEM-type b-lactamase (TEM-52) associated with decreased susceptibility to moxalactamin Klebsiella pneumoniae. Antimicrob. Agents Chemother. 42, 108–113 (1998).

24. M. Barlow, B. G. Hall, Predicting evolutionary potential: In vitro evolution accurately repro-duces natural evolution of the TEM b-lactamase. Genetics 160, 823–832 (2002).

25. G. Kopsidas, R. K. Carman, E. L. Stutt, A. Raicevic, A. S. Roberts, M.-A. V. Siomos, N. Dobric,L. Pontes-Braz, G. Coia, RNA mutagenesis yields highly diverse mRNA libraries for in vitroprotein evolution. BMC Biotechnol. 7, 18 (2007).

26. M. C. Orencia, J. S. Yoon, J. E. Ness, W. P. C. Stemmer, R. C. Stevens, Predicting the emer-gence of antibiotic resistance by directed evolution and structural analysis. Nat. Struct. Biol.8, 238–242 (2001).

27. W. P. C. Stemmer, Rapid evolution of a protein in vitro by DNA shuffling. Nature 370, 389–391(1994).

Steinberg and Ostermeier Sci. Adv. 2016; 2 : e1500921 22 January 2016

28. M. Zaccolo, E. Gherardi, The effect of high-frequency random mutagenesis on in vitroprotein evolution: A study on TEM-1 b-lactamase. J. Mol. Biol. 285, 775–783 (1999).

29. M. K. Singh, B. N. Dominy, The evolution of cefotaximase activity in the TEM b-lactamase.J. Mol. Biol. 415, 205–220 (2012).

30. E. Firnberg, M. Ostermeier, The genetic code constrains yet facilitates Darwinian evolution.Nucleic Acids Res. 41, 7420–7428 (2013).

31. K. M. Brown, M. S. Costanzo, W. Xu, S. Roy, E. R. Lozovsky, D. L. Hartl, Compensatory muta-tions restore fitness during the evolution of dihydrofolate reductase. Mol. Biol. Evol. 27,2682–2690 (2010).

32. N. Tokuriki, D. S. Tawfik, Stability effects of mutations and protein evolvability. Curr. Opin.Struct. Biol. 19, 596–604 (2009).

33. P. Hegarty, A. Martinsson, On the existence of accessible paths in various models of fitnesslandscapes. Ann. Appl. Probab. 24, 1375–1395 (2014).

34. R. P. Ambler, A. F. W. Coulson, J. M. Frère, J. M. Ghuysen, B. Joris, M. Forsman, R. C. Levesque,G. Tiraby, S. G. Waley, A standard numbering scheme for the class-A b-lactamases. Biochem.J. 276, 269–270 (1991).

35. Q. K. Thai, F. Bös, J. Pleiss, The lactamase engineering database: A critical survey of TEMsequences in public databases. BMC Genomics 10, 390 (2009).

36. E. Firnberg, J. W. Labonte, J. J. Gray, M. Ostermeier, A comprehensive, high-resolution mapof a gene's fitness landscape. Mol. Biol. Evol. 31, 1581–1592 (2014).

37. M. F. Schenk, I. G. Szendro, M. L. M. Salverda, J. Krug, J. A. G. M. de Visser, Patterns ofepistasis between beneficial mutations in an antibiotic resistance gene. Mol. Biol. Evol.30, 1779–1787 (2013).

38. H.-H. Chou, H.-C. Chiu, N. F. Delaney, D. Segrè, C. J. Marx, Diminishing returns epistasisamong beneficial mutations decelerates adaptation. Science 332, 1190–1192 (2011).

39. R. C. MacLean, G. G. Perron, A. Gardner, Diminishing returns from beneficial mutations andpervasive epistasis shape the fitness landscape for rifampicin resistance in Pseudomonasaeruginosa. Genetics 186, 1345–1354 (2010).

40. E. Firnberg, M. Ostermeier, PFunkel: Efficient, expansive, user-defined mutagenesis. PLOSOne 7, e52031 (2012).

41. A. Rambaut, A. Drummond, FigTree (tree figure drawing tool 2012); http://tree.bio.ed.ac.uk/software/figtree/

42. M. Bastian, S. Heymann, M. Jacomy, Gephi: An open source software for exploring andmanipulating networks, International AAAI Conference on Weblogs and Social Media,San Jose, CA, 2009.

43. M. Jacomy, T. Venturini, S. Heymann,M. Bastian, ForceAtlas2, a continuousgraph layout algorithmfor handy network visualization designed for the Gephi software. PLOS One 9, e98679 (2014).

44. L. Page, S. Brin, R. Motwani, T. Winograd, The PageRank Citation Ranking: Bringing Order tothe Web (Stanford InfoLab, Stanford, 1999).

45. B. Østman, A. Hintze, C. Adami, Impact of epistasis and pleiotropy on evolutionary adap-tation. Proc. Biol. Sci. 279, 247–256 (2012).

Acknowledgments: We thank the Johns Hopkins Deep Sequencing and Microarray CoreFacility for performing the PacBio deep sequencing. Funding: This work was supported bythe National Science Foundation (grants DEB-0950939 and CBET-1402101 to M.O.). Authorcontributions: B.S. and M.O. conceived the approach, analyzed the data, and wrote the paper.B.S. performed all experiments. Competing interests: M.O. is co-inventor on patent applica-tions for the mutagenesis technique. This intellectual property has been licensed to RevolveBiotechnologies, for which M.O. serves as scientific advisor and has stock options. B.S. co-founded Revolve Biotechnologies and holds equity in the company. Data and materialsavailability: All data needed to evaluate the conclusions in the paper are present in the paperand/or the Supplementary Materials. Additional data related to this paper may be requestedfrom the authors.

Submitted 10 July 2015Accepted 12 November 2015Published 22 January 201610.1126/sciadv.1500921

Citation: B. Steinberg, M. Ostermeier, Environmental changes bridge evolutionary valleys. Sci. Adv.2, e1500921 (2016).

9 of 9

Environmental changes bridge evolutionary valleysBarrett Steinberg and Marc Ostermeier

DOI: 10.1126/sciadv.1500921 (1), e1500921.2Sci Adv 

ARTICLE TOOLS http://advances.sciencemag.org/content/2/1/e1500921

MATERIALSSUPPLEMENTARY http://advances.sciencemag.org/content/suppl/2016/01/19/2.1.e1500921.DC1

REFERENCES

http://advances.sciencemag.org/content/2/1/e1500921#BIBLThis article cites 40 articles, 9 of which you can access for free

PERMISSIONS http://www.sciencemag.org/help/reprints-and-permissions

Terms of ServiceUse of this article is subject to the

is a registered trademark of AAAS.Science AdvancesYork Avenue NW, Washington, DC 20005. The title (ISSN 2375-2548) is published by the American Association for the Advancement of Science, 1200 NewScience Advances

Copyright © 2016, The Authors

on Decem

ber 28, 2020http://advances.sciencem

ag.org/D

ownloaded from