Directed Evolution and Protein Engineering

"Directed evolution can innovate when the innovation is already present." — Frances Arnold, Nobel lecture, 2018¹

Summary

Directed evolution is evolution run on purpose.  A chemist makes random changes in one gene, tests the proteins, keeps the best and repeats.  Frances Arnold shared the 2018 Nobel Prize in Chemistry for making it routine.  Forty years of it are the best-measured record there is of what random change plus selection finds in protein sequence space, and at what cost.  Each result here is scored with the taxonomy of Measuring Biological Information.

1. The searches were small and the steps few.  Keqin Chen and Arnold's 1993 enzyme came from about 4,000 colonies in three rounds.²  The largest search, Keefe and Szostak's, began with six trillion random proteins.³  Finished enzymes differ from their start by 3 to 23 amino acids, at most just over 2% of the sequence.⁴

2. Nearly every new enzyme activity started from a trace of itself.  The unevolved cytochrome c that became the first carbon-silicon enzyme already made the product.¹  Arnold says her method works because an active site is often "already largely there".¹  By the series taxonomy these are modifications, up, however new the chemistry.

3. One search made a new catalytic activity from nothing detectable, with a new fold.  Burckhard Seelig and Jack Szostak selected an RNA-joining enzyme from 4 × 10¹² variants of a zinc-finger domain that does no chemistry.⁵  Its structure is not found in nature.⁶  That is a gain, and it contradicts the claim that folds are isolated from each other.  It is also small, floppy, zinc-dependent and slow.

4. Binding is common and catalysis rare.  Keefe and Szostak's random proteins bind ATP and catalyse nothing.  Donald Hilvert's group put an enzyme from fully random sequence at a library of about 10²⁴, "many orders of magnitude larger" than a binder needs.⁷  The rarity question itself belongs to How Rare Is a Working Protein?.

5. Hall's ebg experiment did not rebuild an irreducibly complex system.  Barry Hall deleted one gene.  A feeble relative took over after two substitutions, while an added chemical switched on the lactose pump the cell already had.^{8 9}  Kenneth Miller's "we watched it happen" overstates it.¹⁰  But Hall also recorded a new inducer and a new enzyme activity, both small gains, and Michael Behe's reply leaves them out.

6. The hard cases cut both ways.  Ann Gauger and Douglas Axe could not turn one enzyme into its structural cousin, even among 70% of all 6.5 million double mutants.¹¹  Critics, including the young-earth creationist Todd Wood, reply that no one claims one living enzyme evolved into another.¹²  Joakim Näsvall and colleagues watched a weak side activity become two specialised genes in under 3,000 generations.¹³

7. The laboratory's edge is not search size.  Counted in mutations per gene, as Benchmarking Evolution requires, Lenski's flasks made over 1,600 times as many mutations in a 1,000-letter gene as Chen and Arnold screened variants.¹⁴  The laboratory wins by choosing the target, rewarding every small step and starting from a trace.  So it prices only the cheap end of the series' ledger, steps that each pay.  Where chemists tried to cross a gap that did not pay step by step, they failed or projected libraries of 10²⁴ to 10⁷⁰.

Verdict.  The design reading, that selection optimises but does not invent, fits nearly every case, in the practitioners' own words, but not the Seelig and Szostak ligase.  The searches reached at most 23 substitutions in an enzyme that kept its fold, and one small new fold from 4 × 10¹² variants.

Part 1: What directed evolution is

The first experiment selected for speed and got a loss

In 1967 D. R. Mills, R. L. Peterson and S. Spiegelman put the RNA of a small virus, Qβ, into a test tube with the enzyme that copies it and nothing else.  Every so often they moved a drop to a fresh tube, giving the advantage to whatever copied fastest.  They framed the question with a verse:  "What will happen to the RNA molecules if the only demand made on them is the Biblical injunction, multiply, with the biological proviso that they do so as rapidly as possible?"¹⁵

The answer was deletion.  "By the 74th transfer, 83 per cent of the original genome had been eliminated."¹⁵  The short molecule copied 15 times faster than the full one, and the authors add that these molecules "are no longer able to direct the synthesis of virus particles."^{15 15}  By the series taxonomy this is a loss.  The test tube rewarded copying speed and nothing else, and selection delivered exactly that.

The method, as Arnold built it

Arnold's version aims at a useful protein.  Her first clear success, with Keqin Chen in 1993, made subtilisin E, a bacterial enzyme that cuts proteins, work in dimethylformamide, a solvent that normally cripples it.  They started from a variant called 4M, which already carried four helpful substitutions, and ran three rounds.  Each round copied the gene with deliberate errors, grew the bacteria on plates of casein in 35% solvent, and kept the colony with the biggest clear halo.  "In all, ≈4000 colonies were screened."²

The winner, PC3, carries ten amino-acid substitutions, four inherited and six new:  "Of 16 base substitutions generated during the three sequential mutagenesis cycles, 6 resulted in amino acid replacements".²  It is 256 times more efficient than wild type in 60% solvent, 130 times in 40%, and only three times better than 4M in 40% solvent.  In plain water it is no better than 4M.^{2 2}  And "Seven of the 10 amino acid substitutions in PC3 are found in other natural subtilisins."²  The reaction never changed;  the enzyme kept doing its own job in a harsher liquid.  The score is modification, up.

Why it works with so few tries

Arnold's Nobel lecture explains the constraint that shaped the method.  Her laboratory could test only what fit on plates:

"To measure what mattered, we were limited to monitoring the few thousand protein variants we could express and array in readily available 96-well plates or on a petri dish. Therefore, we could only search deeply those sequences one or two mutations away from the starting protein."¹

Yet "It often takes only a few mutations for an enzyme to acquire the targeted trait", and "Useful traits could emerge in less than ten, or even five, generations."^{1 1}  She contrasts this with nature, which "takes a circuitous route to achieving new properties, combining neutral or even negative mutations with beneficial ones.  Those paths can involve hundreds of changes.  Our approach, however, collected only adaptive mutations that yield steep changes in function."¹

Jesse Bloom and Arnold summed up what a random mutation does to a protein across many studies.  About 30 to 50% are strongly harmful, 50 to 70% are about neutral, and 0.5 to 0.01% help.⁴  At one beneficial change in 200 to one in 10,000, a few thousand clones per round is enough to find one.

Three features make the laboratory search efficient.  The experimenter chooses the target.  The screen rewards any small step toward it.  And the search starts from a protein that already folds and already does something close to the job.  Whether nature supplies equivalents is the question of Part 6.

Part 2: Chemistry new to nature

Carbon and silicon

"Enzymes that catalyze carbon–silicon bond formation are unknown in nature, despite the natural abundance of both elements."¹⁶  So begins the 2016 paper by Jennifer Kan, Russell Lewis, Kai Chen and Arnold that made the first one.  They offered a synthetic reagent, a diazo ester no cell makes, to iron-containing catalysts.  Free heme, the iron-holding group itself, did it at "very low levels, a total turnover number (TTN) of 4", with no preference between the two mirror-image forms of the product.¹⁶  P450 and myoglobin proteins did better, with almost no control over the form.  A cytochrome c from Rhodothermus marinus, a bacterium from Icelandic hot springs, "catalyzed the reaction with 97% ee," meaning nearly all of its product came out in one mirror-image form.  The authors add that bacterial cytochromes c are "electron-transfer proteins that are not known to have any catalytic function in living systems".¹⁶  Arnold's lecture gives its output as 40 turnovers, 40 product molecules per protein molecule.¹  Kan adds that "Other bacterial and eukaryotic cytochrome c proteins also catalyzed the reaction, but with lower selectivities."¹⁶

The team then changed one position at a time, each chosen from the crystal structure because it sits next to the iron.  The first round found M100D, "a 12-fold improvement over the wild-type protein" at 550 turnovers.¹⁶  Two more rounds added V75T and M103E.  The triple mutant "catalyzed carbon–silicon bond formation in >1500 turnovers and >99% ee.  This level of activity is more than 15 times the total turnovers reported for the best synthetic catalysts for this class of reaction".¹⁶  It also runs the reaction ">7-fold faster" than the wild type.¹⁶

Three substitutions in three rounds turned a trace side reaction into a catalyst better than anything chemists had made, a gain of about thirty- to fortyfold in turnovers.¹⁷  The library sizes are in a supplement that could not be obtained.

Rings from carbenes

In 2013 Coelho and colleagues in the same laboratory reported P450 variants that make cyclopropane rings from styrenes and diazo esters, a reaction that "has no biological counterpart."  Their own verb is "adapt existing enzymes".¹⁸  Arnold's lecture reports what came first:  "we discovered that iron-heme proteins do indeed promote cyclopropanation when provided with diazo carbene precursors and a suitable alkene substrate, in water."¹  Only the abstract could be read, so its library sizes and substitution counts are blank.

The trace that was already there

Both cases have the same shape.  The reaction was new to biology.  It was not new to the protein.  Arnold's lecture says this is the rule, not a coincidence:

"For a long time I could not see how my conservative directed evolution strategy of accumulating one or two beneficial mutations per generation would create entirely new enzyme active sites. Unless, of course, an active site is already largely there…"¹

She adds that "When innovating, Nature does not invent new active sites de novo", and that "If a new catalytic activity is already present, even at a low level, our conservative process of accumulating beneficial mutations can mold it into a new enzyme."¹  The idea behind this is enzyme promiscuity.  Khersonsky and Tawfik's 2010 review, which Arnold cites, opens:  "Many, if not most, enzymes can promiscuously catalyze reactions, or act on substrates, other than those for which they evolved."  It then addresses "the hypothesis that promiscuous enzymatic activities serve as evolutionary starting points".¹⁹

The practitioners also say where the method stops.  Philip Romero and Arnold wrote in 2009 that stepwise search "has proven to be very effective at optimizing a function or property that already exists or can be reached through a series of intermediate steps", but that some functions "simply can not be reached through a series of small uphill steps".  Their examples include "the appearance of a new catalytic activity or activity on substrate for which the parent and its single mutants show no measurable activity."  Their remedy is to "create a new starting point, a parent protein with at least some minimal function".^{20 20}  Bloom and Arnold say the same:  failures "come when the bar is set too high, such as searching for activity on a new substrate on which the parent protein is completely inert."⁴

The standard way over a high bar is to lower it in stages.  Arnold's group turned a P450 into an enzyme that works on propane by first selecting for activity on octane.  The result carries 23 substitutions, "only a little >2% of the amino acid sequence."  On the way, the changes that raised activity lowered stability, and further progress needed a round of deliberately selected stabilising mutations.^{4 4}  A staged route works because each stage creates a trace of the next target.

How these score

By the series procedure, a change is a gain only if an activity now works that did not exist before.  Both activities existed before any mutation, so both score modification, up.  Both are hard cases, because no cell ever used the starting activity.  Behe's 2010 definition agrees.  His modification class covers mutations "shifting its activity somewhat (such as allowing a protein to bind a structurally-related ligand at the same site as its normal substrate)".²¹  What the chemistry shows is that a protein can carry a useful activity nobody knew about, and a handful of changes can raise it many times over.

Part 3: Function from random sequence

A second line of work, led by Jack Szostak, started from sequences that had never been proteins.

A binder from six trillion random proteins

In 2001 Anthony Keefe and Szostak made "a library of 6 × 10¹² proteins each containing 80 contiguous random amino acids" and kept those that stuck to ATP.  Eight rounds of selection "yielded four new ATP-binding proteins that appear to be unrelated to each other or to anything found in the current databases of biological proteins."³  They then mutated one family at 3.7% per amino acid per round for three rounds and selected for six more.  By round 18, "four amino-acid substitutions had become predominant in the selected population", and 16 others had been enriched.^{3 3}  The best protein binds ATP tightly, with a dissociation constant of 100 nanomolar.³  The authors suggest it folds around a zinc ion held by four invariant cysteines.³

The structures came later.  Lo Surdo and colleagues crystallised a version of it in 2004 and found "a novel alpha/beta fold bound to its ligand," the first protein structure from laboratory evolution and the first nucleotide-binding protein held together by zinc.²²  Mansy and colleagues solved a folding-optimised version by NMR in 2007.  It "differs from its ancestral sequence through the accumulation of 12 amino acid mutations".²³

This protein binds.  It does nothing to what it binds.  What its frequency says about the rarity of working proteins is argued in How Rare Is a Working Protein?.  Here it is a search size and a class:  6 × 10¹² random sequences, 18 rounds, 12 substitutions from its random ancestor, and a gain, a binding site where there was none, in a test tube outside any genome.

How much rarer an enzyme is

In 2001 Sean Taylor and colleagues in Donald Hilvert's group at ETH Zürich rebuilt a natural enzyme, a chorismate mutase, from simple modules of only eight amino acids.  Active enzymes came out, with "biophysical and kinetic properties typical of the natural enzyme even though ≈80% of the protein originates from the simplified modules".⁷  But they were rare, even with the fold fixed and the key active-site residues placed by hand:  "There is, nevertheless, a low probability of finding catalysts, even when both position and identity of all critical active site residues are determined in advance."⁷

From their hit rates of one in 4,500 to one in 17,500, they estimated that "if every position in the protein had been randomized, a library of ≈10²⁴ members would have been needed".  They add:  "The size of such a library is many orders of magnitude larger than that needed to identify noncatalytic ATP-binding proteins from random sequences", citing Keefe and Szostak.^{7 7}  Their conclusion is that enzymes cannot be pulled "from unbiased random libraries in a single step", and that design will need "incremental strategies" in which a scaffold comes first and catalytic groups are added and optimised.⁷  The 10²⁴ is an extrapolation for one fold and one reaction, not a measurement, and should not be set against Keefe and Szostak's binding frequency as if the two measured the same thing.

An enzyme from a scaffold with no catalytic activity

In 2007 Burckhard Seelig and Szostak did what Taylor's group said had to be done incrementally, in one campaign.  They took a small zinc-finger domain from a human hormone receptor, a protein that binds DNA and catalyses nothing, and randomised two of its loops.  They screened "4×10¹² unique mRNA-displayed proteins" for joining two RNA pieces held side by side on a template strand.⁵  After 8 rounds they recombined and mutated the survivors and continued to round 17, cutting the reaction time from overnight to five minutes.  By then "just 16% of the clones (8 of 49) retained the cysteine pattern as originally designed and were free of major deletions."⁵

The best ligase is "at least 2 × 10⁶-fold faster than the uncatalyzed reaction."  For the starting domain, "we could not detect any ligated product".⁵  The reaction needs zinc.⁵  In 2013 Fa-An Chao and colleagues in Seelig's laboratory solved its structure by NMR.  "This artificial enzyme lost its original fold and adopted an entirely novel structure with dramatically enhanced conformational dynamics".⁶  Compared with natural enzymes it has "substantially fewer secondary structure elements such as α-helices and β-strands, and instead exhibited increased flexibility."⁶  The authors call it "the first new protein structure emerging simultaneously with a novel enzymatic function."⁶

This is the mainstream side's strongest result.  The start had no detectable trace of the activity, so the class is gain, with a fold not found in nature.  Its limits are as plain.  The template strand holds the two RNA ends together, doing much of an enzyme's positioning work.  The enzyme is slow, loosely packed and zinc-dependent, and it came from a designed selection of 4 × 10¹² variants of a domain that already folded.

A random segment inside a living virus

Yuuki Hayashi and colleagues ran the closest thing to natural selection in this set.  They replaced an essential domain of a coat protein of the virus fd, the part that attaches to the host's pilus, with a random 139-amino-acid segment.  "The initial defective phage fd-RP showed little infectivity, indicating that the random polypeptide RP3-42 contributes little to infectivity."²⁴  Twenty rounds of random substitution and selection for infection, with about ten clones per round at first and up to a million later, raised infectivity "1.7×10⁴-fold".^{24 24}  The evolved segments look nothing like the natural domain.²⁴

The authors' model priced the rest of the climb.  The evolved virus still infected about 2,000 times less well than the natural one.^{24 25}  Reaching wild-type fitness "requires a library size of 10⁷⁰ with 35 substitutions", which "is impractical", while "the maximum library size that can be prepared is about 10¹³".^{24 24}  They conclude that the natural domain must have evolved by other means as well, "such as homologous recombination", which they did not test.  Their summary is balanced:  random substitution climbs "with relative ease up to the middle region of the fitness landscape from any primordial or random sequence, whereas an enormous range of sequence diversity is required to climb further up".²⁴  Nobody studied what the evolved segment does.  It may bind the pilus or may only have stopped getting in the way, so the class is unmeasured.

Part 4: Hall's ebg system and the acid test

The two claims being tested

Michael Behe's Darwin's Black Box defined the property at issue:  "By irreducibly complex I mean a single system composed of several well-matched, interacting parts that contribute to the basic function, wherein the removal of any one of the parts causes the system to effectively cease functioning."  Such a system, he argued, "would have to arise as an integrated unit, in one fell swoop, for natural selection to have anything to act on."²⁶

Kenneth Miller, a cell biologist at Brown University, proposed a test in Finding Darwin's God (1999):  "The most direct way to do this would be with a true acid test — by using the tools of molecular genetics to wipe out an existing multipart system and then see if evolution can come to the rescue with a system to replace it."¹⁰  He said Barry Hall of the University of Rochester had passed it.  Hall deleted the gene for the lactose-splitting enzyme in E. coli, and "Before long, however, mutant strains appeared that could handle lactose nearly as well as the originals."¹⁰  Miller's conclusion:

"Irreducible complexity. What good would the permease be without the galactosidase? What use would either of them be without regulatory genes to switch them on? And what good would lactose-responding regulatory genes be without lactose-specific enzymes? No good, of course. By the very same logic applied by Michael Behe to other systems, therefore, we could conclude that the system had been designed. Except we know that it was not designed. We know it evolved because we watched it happen right in the laboratory!"¹⁰

What Hall did, in Hall's words

Hall's 66-page review of 1982 settles most of the dispute.  He built the system to test one thing:  "The EBG system focuses evolutionary pressure more precisely by directing the evolution of a single, specific new function:  the ability to hydrolyze lactose."  The host lacked the enzyme gene lacZ but kept the lactose pump, lacY, and everything downstream, so "Such a strain is unable to hydrolyze lactose, but is able to carry out all other functions involved in the lactose degradative pathway."⁹  The pump was switched on by a chemical, IPTG, in every lactose medium:  "The sole function of the IPTG is to induce synthesis of the lactose permease, and thus to deliver lactose to the inside of the cell.  Neither the constitutive nor the inducible evolved strains grew on lactose in the absence of IPTG."⁹

The replacement enzyme came from a second gene, ebgA, whose product is "a catalytically feeble enzyme" in the wild type.²⁷  Growth on lactose needed "two mutations:  one in ebgA to improve the efficiency of the enzyme, and the other in ebgR to increase the amount of enzyme that is synthesized."⁹  Hall later showed the enzyme step has only one road.  Ebg "differs from other members of its clade at only 2 of the 15 active-site residues, and the two mutations required for full Ebg beta-galactosidase activity bring Ebg into conformity with the other members of its clade."  "The evolutionary potential of Ebg is thus limited to those two replacements."⁸  When both lacZ and ebg were deleted, repeated selection and chemical mutagenesis "failed to yield any lactose-utilizing mutants".⁹

Hall classed his own results.  The enzyme changes "do not result in truly new activities; they simply improve old functions", because the unevolved enzyme's weak activity on those sugars "can be detected and measured in vitro".⁹  The repressor already answered to lactose:  "In a wild-type strain, induction by lactose leads to a 100-fold increase in the expression of ebgA", though only to 5% of the maximum.⁹  A repressor mutation that simply stops repressing is, in Hall's words, "equivalent to a loss of regulatory function rather than evolution of a new function."⁹

Two results go further.  First, a later class of repressor mutation, ebgR^+L, made lactulose an inducer as well as lactose, so that "lactose ceases to be the sole powerful inducer of ebg enzyme synthesis."⁹  Second, lactobionate, a sugar acid, is split neither by the ancestral enzyme nor by the evolved classes I and II, not even in a test tube.  Growth on it "requires three mutations in the ebgA gene", and a trace of the activity appears in the two-substitution class IV enzyme before the third makes it useful.^{9 9}  Hall also reported, from unpublished work, that one evolved enzyme converts lactose into an inducer of the lac genes.⁹

Behe's reply and Miller's answer

Behe answered in "A True Acid Test" (2000).  His central point was that, measured by Miller's own test, "only one component of a multipart system was deleted."  He added that the IPTG meant "the system was being artificially supported by intelligent intervention", and that Hall had shown only "a series of micromutations stitched together by intelligent intervention."^{28 28 28}  His factual claims check against Hall 1982:  one gene was deleted, the pump was never built by the experiment, IPTG was always present, and the unmutated enzyme and repressor both already acted on lactose.

One sentence of his does not check.  Behe wrote that "When two or more components were deleted, or when the bacterium was cultured in the absence of an artificial chemical (called IPTG), no viable bacteria could be recovered", with no citation.²⁸  The nearest experiment in Hall's review deleted lacZ and ebg, two alternative sources of the same enzyme, not two parts of one system.

Miller replied that three things had to evolve after the deletion:  "a new galactosidase enzyme, second, a new lactose-sensitive control region, and third, a new way to switch on the lac permease gene."  He pointed out that ebg is "only 34% homologous" to lacZ.  On IPTG:  "In short, he screened for mutants, he didn't produce them as Behe implies."^{29 29 29}  The last point is right:  IPTG decided which mutants could be recovered and did not make them.  The 34% answers a question Behe did not ask.  His point was the active site, and there Hall's figures cut both ways:  Ebg matches its own clade at 13 of 15 active-site positions, while the Ebg and LacZ subclasses differ at 12 of 15.²⁷  So Behe's phrase "a spare near-copy of the original component" is wrong about LacZ, and Miller's 34% is beside the point.²⁸  Miller also cites Hall's 1999 paper as "FEMS Microbiology Letters 178: 1-6";  it is volume 174, pages 1 to 8.^{29 27}

How it scores

Miller overstated the case.  No multipart system was rebuilt.  One enzyme was replaced by improving a feeble relative in two steps, with its repressor loosened, while the experimenter switched on the pump.  But Behe's reply stops at lactose.  Hall also recorded a repressor that learned a new inducer and an enzyme activity the ancestor lacked, reached in three steps.  Neither is irreducible complexity.  Both are small gains.

Part 5: The searches that failed, and the one that answered them

Gauger and Axe: converting a cousin

Ann Gauger and Douglas Axe of the Biologic Institute in Redmond, Washington, published the sharpest negative results in the field.  Their 2011 paper, edited by Behe, took two E. coli enzymes of the same structural family, Kbl2 and BioF2, and tried to make the first do the second's job.  "After identifying and testing 29 amino-acid changes," they made the active-site cavity of Kbl2 identical to BioF2's, which "nonetheless fails to produce detectable BioF2-like function in vivo."  They inferred that conversion "would in this case require seven or more nucleotide substitutions."³⁰  With a population model they put the wait for such a change at "some 10³⁰ or more generations".³⁰

Their introduction is also a fair summary of the mainstream record.  Aspartate aminotransferase "has been converted by seven base changes into an aspartate decarboxylase that is some 100,000-fold less efficient" than the natural one, and "eleven base changes were used to convert a dehalogenase into a crotonase, with only 0.005% of wild-type crotonase activity achieved".³⁰  Conversions between cousins in the laboratory are rare, weak and need several changes at once.

In 2014 Mariclair Reeves, Gauger and Axe tried random search instead.  They made single-mutant libraries of nine enzymes of the family, "None of the singly mutated genes had this function."  Then "we prepared and tested 70% of the 6.5 million possible mutation pairs for Kbl2 and for BIKB," an enzyme with both activities in a test tube.  "Again, no BioF2 activity was detected in vivo."  Their best case needs three changes after a gene duplication, which "would require about 10¹⁵ years in a natural population".¹¹  This was a random search, near-exhaustive for pairs.

The critics

The critics did not dispute the measurements.  They disputed the question.  P. Z. Myers, a biologist and atheist blogger, wrote that "they picked two extant proteins and then asked how they could have evolved from each other", and that the sensible approach is to "examine parental and grandparental states."^{31 31}  His post is abusive in tone, but the objection is sound.  Martin Poenie, a cell biologist at the University of Texas writing as a Christian critic of Axe, measured the gap:  the two enzymes differ at 250 of 381 aligned positions, "a long distance", so "It is not surprising that his efforts to swap functionality by swapping a few residues failed."¹²  He also noted that the paper never showed what the substitutions did, so "these mutations could have simply led to an unfolded protein."¹²

The most telling criticism came from a young-earth creationist.  Poenie quotes Todd Wood, whom he introduces as "The young earth creationist Todd Wood":  "no evolutionary biologist would propose that an extant enzyme evolved directly into another extant enzyme.  So they're testing a model that no one would take seriously?"¹²  The results show that Kbl2 is not two or three steps from BioF2.  They do not show that the common ancestor of the two was not.

Axe and Gauger: does a trace have to be a real active site?

In 2015 Axe and Gauger ran their own directed evolution to test the promiscuity idea.  They took two weak ampicillin-resistance proteins derived from TEM-1 β-lactamase, the enzyme covered in Antibiotic Resistance.  TEMΔ had lost much of its active site.  Basal-β kept it.  Three rounds of mutation and selection did nothing for TEMΔ.  Six rounds took basal-β past the natural enzyme:  "The strongest mutant following this final selection … has more than 500 times the activity of the basal-β starting point," with growth "above the wild-type MIC of 6,600 μg/ml at 37°C".  It differs from basal-β at 13 positions.^{32 32}  Their conclusion:  "while mutation and selection can improve the proficiency of good designs through small structural adjustments, they seem unable to convert fortuitous selectable activities into good designs."³²

The contrast is real, with two limits.  Three rounds against six is not a matched comparison.  And TEMΔ is a damaged modern enzyme, not the kind of trace the promiscuity idea is about;  Arnold's cytochrome c was an intact protein with a working iron site that happened to do a second reaction.  The basal-β result is itself a modification, up, with 13 substitutions.

Näsvall: a side activity becomes a gene

The mainstream model of how a new enzyme begins is innovation, amplification and divergence.  A weak side activity becomes useful, extra copies of the gene raise its output, and the copies then specialise.  In 2012 Joakim Näsvall, Dan I. Andersson and colleagues watched it happen in Salmonella.  Their starting gene was a hisA that had gained a weak TrpF activity through a small internal duplication and a substitution, and so did two jobs badly.¹³  In a strain lacking both hisA and trpF, growth sped up from about five hours per division to 1.9 to 2.5 hours "Within a few hundred generations," as the gene was amplified.  By 3,000 generations copies in the amplified array had diverged into HisA specialists, TrpF specialists and improved generalists, sometimes both specialists in one cell.^{13 13}

It is the strongest answer to Gauger and Axe, and it was watched.  Its limits are in the paper.  The experimenters built the starting gene and "placed this bifunctional parental gene (dup13-15, D10G) under the control of a constitutive promoter", on a plasmid region where "Duplications and amplifications of this region are frequent and have low fitness cost".¹³  By the series procedure each specialised copy improves an activity the ancestor already had, so the class is modification, up, by copy.  Behe's 2010 definition of gain includes "the divergence by mutation of the activity of a previously duplicated coded element", so by his wording it is a gain.²¹  The full scoring belongs to Gene Duplication and Divergence.

Part 6: What the record shows

The design reading

Every enzyme success in Parts 1 and 2 was an intelligent search.  A chemist chose the target, built a screen that rewards any step toward it, and started from a protein that already folded and usually already had a trace of the activity.  Even then the steps were few and small.  Axe and Gauger put the conclusion in their title:  selection "is not inventive".³²  The practitioners agree, as Part 2 showed.  Hayashi's team priced the climb from a random sequence to a natural one at 10⁷⁰ variants.  Taylor's team priced an enzyme from fully random sequence at 10²⁴.  Stephen Meyer's Darwin's Doubt draws the general claim:  "The extreme rarity of protein folds also entails their isolation from each other in sequence space."³³

The mainstream reading

Tiny searches find new activities within a handful of substitutions, again and again.  That means function is dense near existing proteins.  Khersonsky and Tawfik's "Many, if not most, enzymes" carry side activities, and Kan found the carbon-silicon trace in several cytochromes c, not one.^{19 16}  A starting point is not a rare gift.  Bloom and Arnold argue that a protein is rarely stuck, because "a 300-residue protein can undergo 5,700 unique single amino acid mutations, each of which represents a different direction on the fitness landscape."⁴  Where a trace is missing, the Seelig and Szostak ligase shows one can be found, from a scaffold with no catalytic activity, in a single campaign.  Näsvall's Salmonella shows the natural version of the staged route, watched.

What the evidence supports

Three things can be stated as numbers.

How far the searches reached.  The finished enzymes sit 3 (Kan), 10 (Chen and Arnold, six of them new), 13 (Axe and Gauger) and 23 (the propane P450) substitutions from their starting points.  The random-sequence binder sits 12 from its random ancestor.  The largest, 23, is "a little >2%" of the sequence.⁴  No search read here moved a folded enzyme further than that from its starting point.

Whether any made a new fold.  Yes, twice.  The ATP binder's fold is new, and the ligase "lost its original fold and adopted an entirely novel structure".^{22 6}  Both are small, held together by zinc, and loosely packed.  Meyer's isolation claim is wrong for folds of that kind.  No laboratory has yet reached a large, well-packed enzyme domain from random sequence, which is the kind the design argument is about, so the claim is untested there.

Where the searches stopped.  Every attempt to cross a gap with no rewarded path failed or was projected out of reach:  Kbl2 to BioF2 by pairs, a random segment to the natural phage domain, a fully random chorismate mutase.  The ligase is the exception, and it was found by sampling 4 × 10¹² sequences at once, not by climbing.

So the design reading is right about the method and nearly every result.  The mainstream reading is right that traces are common and that one real invention has been made.  What neither side has shown is the thing that matters most:  how often nature supplies a trace that its own selection can see.  A trace that makes a cell grow faster is a starting point.  A trace against a reagent no cell meets is not.

Part 7: The ledger

Each row is one search.  "Trace" means the starting protein already showed the target activity before any mutation.  Empty cells are numbers the sources read for this article do not give.  They are left empty rather than guessed.

The Reeves total and the Hayashi ceiling are computed in a note.³⁴

Against the natural populations

Benchmarking Evolution found that comparing populations by head count misleads, because what a search explores is mutations, not bodies.  The same rule applies here, with one more step.  A laboratory library puts all its mutations into one gene.  A natural population spreads them over the whole genome.  So the fair comparison is mutations per gene.

Counted that way, the laboratory searches were not large.  Lenski's twelve flasks of E. coli made about 6,500 mutations at every base pair over the experiment, or about 6.5 million in a gene of 1,000 letters.^{35 14}  That is more than 1,600 times Chen and Arnold's 4,000 colonies, and comparable to Hayashi's whole search.  Only the two Szostak libraries, at 4 to 6 × 10¹² random sequences, exceed it, and those are samples of random sequence, not mutants of a working gene.  The same article finds that one line of mammal descent tries each single-letter change 3.6 × 10⁵ to 8 × 10⁸ times.  A laboratory round of a few thousand variants tries each of a 300-residue protein's 5,700 single changes about once at most.^{1 4}

So the laboratory's advantage is not size.  It is the screen.  Every round keeps the best variant for a target the chemist chose, so every step pays.  That places directed evolution at the cheap end of the series' price list:  steps that each pay are climbed fast, anywhere.  It says little about the expensive end, a change that needs several steps before any of them pays.  Benchmarking Evolution found that only the per-lineage form of the young-earth argument survives, and that it turns on how much of a genome needed such unpaid steps.  The laboratory record adds that, where chemists tried to cross such a gap on purpose, with the target known, they mostly failed.

Objections and responses

"Directed evolution is design, so it shows nothing about evolution."  The chemist chooses the target, the screen and the starting protein.  The mutations are random, and how the protein responds is not designed.  How many substitutions separate a trace from a working enzyme is a fact about proteins, and nature can use it wherever a trace helps a cell.  The design inference has to rest on what the chemist supplied, and this article names those three things.

"The ligase refutes the design argument."  It refutes two claims:  that selection never invents, and that folds are isolated from one another.  It does not reach a large, well-packed enzyme fold.  It was found by sampling 4 × 10¹² sequences at once on a scaffold that already folded, for a reaction a template strand partly positions.

"Side activities are everywhere, so new enzymes are easy."  Traces are common for the reactions chemists test.  Whether a trace helps a cell enough to be selected is a different question and has not been measured.  Reeves, Gauger and Axe also show that some neighbouring functions cannot be reached by any pair of changes.

"Gauger and Axe tested a straw man."  Mostly yes, and a young-earth creationist said so first.  But the pairs experiment is still a measurement:  no double mutant of Kbl2 or BIKB makes BioF2 in a cell.  Neither side has tested the route through the two enzymes' reconstructed ancestor, which is the experiment that would settle it.

"IPTG invalidates Hall's experiment."  No.  Miller is right that it decided which mutants could be recovered and did not make them.  It does mean the pump was never evolved, so no system was rebuilt.

"Laboratory searches are tiny next to nature's, so nature can do far more."  Per gene, the natural searches are bigger.  Nature lacks a chosen target and a screen for small steps toward it.  The laboratory shows that rewarded steps are cheap, which the series already concedes.  The open question is the price of steps that are not rewarded.

Issues and unanswered questions

• Library sizes for the carbon-silicon, cyclopropanation, propane and Axe and Gauger searches are in supplements or bodies that could not be obtained. Those cells are empty.
• Only abstracts were read for Hall 1998, Hall 1999, Khersonsky and Tawfik, Coelho, Lo Surdo and Mansy. Hall's 1982 mutation rates are printed with exponents the scan's text layer destroys, so they are not printed here.
• How often a natural trace activity is strong enough for selection to see is unmeasured, and the whole comparison with nature turns on it.
• The Kbl2 and BioF2 conversion has not been tried through a reconstructed ancestor by anyone.
• Jonathan McLatchie's reply to Poenie and Axe's book Undeniable (2016) were not obtained.
• The ebgR^+L score depends on whether Behe's ligand rule is applied, and the Näsvall score is left to Gene Duplication and Divergence.

Verdict

Laboratory search finds new activity cheaply when a protein already has a trace of it and each step can be rewarded.  The steps are few, 3 to 23 substitutions, and the searches are small next to what a natural population explores per gene.  That supports the design side's narrow claim, stated by Arnold herself, that the method refines what is already present.  It does not support the broad claim that selection never invents or that folds are isolated:  the Seelig and Szostak ligase is a real gain with a new fold, though small, zinc-bound and slow.  Hall's ebg experiment replaced one enzyme, not a system, and Miller overstated it, but it also produced two small gains that Behe's reply leaves out.  The question the laboratory has not answered is the one the series needs:  how often nature offers a trace its own selection can see.

Sources