When Evolution Creates Information

Summary

Mutation and selection have been watched producing things that did not exist before.  The claim that mutations never add information is false as it is usually stated, and this site says so.  This article gathers every case in the Observed Evolution series where something new was produced, gives each its numbers, its score and its grade, and links the article that argues it.  It re-argues none of them.

1. The strong claim is on record. Lee Spetner wrote in Not by Chance! that "Not even one mutation has been observed that adds a little information to the genome."¹ Werner Gitt wrote that "There can be no increase in information."² The Institute for Creation Research's first museum told visitors that mutations "have never been observed to add new information,"³ and Answers in Genesis has said that mutation and selection "only lead to a loss of functional systems."⁴
2. The careful claim is also on record, from creationists. Creation Ministries International lists "There are no beneficial mutations" among arguments creationists should not use.⁵ Robert Carter calls "Mutations cannot create new information" "almost a mantra" and says he does not agree.⁶ John Sanford does "not doubt there are beneficial mutations that create information" and argues they are far too rare.⁷ The defensible claim is that gains happen rarely, slowly and against a background of loss. The strong claim, "never," is wrong.
3. The design writers make a different claim, a size limit. Michael Behe's "first rule" predicts that adaptation will "Break or blunt" working parts first and most often, and he grants the sickle-cell gain.⁸ Stephen Meyer grants that "small amounts of specified information can occasionally arise by random processes" and predicts that more than 500 bits will not.⁹ One counter-example cannot refute either claim.
4. Whether a case counts depends on the definition. The series scores each change as a gain, loss, modification or transfer of function by the rules in Measuring Biological Information, and counts the nucleotides that were originated rather than copied as a separate ledger. Measured as functional information, Jack Szostak's measure in bits, every unguided gain that can be bounded carries at most about 43 bits, a tenth of William Dembski's 400-bit bound. The record is too small to test the design threshold in either direction.
5. Four gains were watched from start to finish. E. coli in Richard Lenski's experiment began to live on citrate after a duplication put a silent transporter gene under a working promoter. Phage lambda learned to enter cells through a second receptor by four mutations, in 25 of 102 flasks. Human influenza added nine glycosylation sites to its haemagglutinin in a century. A laboratory search found a working ligase in a protein scaffold that had no trace of the activity, though an experimenter chose what to reward.
6. More gains were dated or reconstructed in nature. The malaria parasite's chloroquine transporter, sickle haemoglobin, haemoglobin E and the glycophorin DUP4 rearrangement are scored as gains by their sibling articles. One lactase variant is a candidate gain. HIV-1's Vpu protein acquired an activity against the human protein tetherin.
7. Many famous cases are real improvements that are not gains. The extended-spectrum beta-lactamases, nylonase, the other lactase variants, the dog's FGF4 retrogene and the black peppered moth are modifications. Tibetan EPAS1 is a transfer. New plant species have been watched forming by polyploidy, which copies whole genomes and originates no sequence. SARS-CoV-2 and tuberculosis produced no scored gain at all.
8. The largest novelties were not watched. The antifreeze gene of northern cod arose from DNA that was not a gene, and the Antarctic antifreeze gene arose from a digestive-enzyme gene. Both were reconstructed by comparing living fish. So were the de novo genes. Both antifreeze proteins are repeats of a three-amino-acid unit grown from a nine-nucleotide piece of DNA, not folded enzymes.
9. The tally is small. The series' benchmark table holds seven scored gains across roughly 10²³ births. Every watched gain changed between zero and four new nucleotides. No watched case produced a new gene with a shown benefit. How far that record reaches toward what common descent needs is the subject of Benchmarking Evolution, and is not re-argued here.

Verdict.  To a critic of this site:  the cases are real and conceded.  To a reader who has used "mutations never add information":  stop using it.  To an evolutionist:  the watched gains are few and small, and most fitness gains in the series are losses or modifications.  To a design advocate:  no case comes near the 500-bit line, and none tests it.  That line stays untested until someone computes the probability of a real pathway with selection included.

Part 1: The claim, in its makers' words

The strong form: never

Lee Spetner, a physicist, made the argument in its classic form.  Neo-Darwinian theory, he wrote, needs each step to "add a little information to the genome,"¹ and no observed mutation does:

"Not even one mutation has been observed that adds a little information to the genome. That surely shows that there are not the millions upon millions of potential mutations the theory demands. There may well not be any."¹

Nineteen pages later he restated it with a hedge:  "not a single one has been found that adds a little information.  It is not impossible, in principle, for a mutation to add a little information, but it is improbable."¹  So Spetner's own claim is "never observed," not "impossible."  Answers in Genesis quotes the first passage as "This surely shows";  Spetner's page reads "That surely shows."

Werner Gitt, an information engineer, stated it as a principle.  To the question "Can new information originate through mutations?" his book answers:

"This idea is central in representations of evolution, but mutations can only cause changes in existing information. There can be no increase in information, and in general the results are injurious. New blueprints for new functions or new organs cannot arise; mutations cannot be the source of new (creative) information."²

The Institute for Creation Research put the same claim on a museum wall.  Its cavefish exhibit at the original museum in Santee, California read:  "Mutations have been observed to destroy, damage, or corrupt genetic information or to be neutral, but have never been observed to add new information."³  In 2023 ICR's Randy Guliuzza quoted that exhibit and wrote that "essentially every assertion in our old explanation has been shown to be scientifically untrue."³  His reason is not the one this article gives.  Guliuzza argues that organisms adapt by designed internal mechanisms rather than by random mutation and selection, so his retraction is not a concession that random mutation adds information.

Answers in Genesis has made the strong claim in two forms.  Georgia Purdom, a molecular geneticist, wrote in 2007 that "Mutation and natural selection, thought to be the driving forces of evolution, only lead to a loss of functional systems."⁴  A 2012 AiG news item on the peppered moth called macroevolution "a process that would require the never-observed acquisition of new genetic information."¹⁰

The careful form: rarely, and slowly

Other creationists have retreated from "never," and some did so long ago.  Creation Ministries International's list of arguments not to use includes "There are no beneficial mutations":

"This is not true, since some changes do confer an advantage in some situations. Rather, we should say, ‘We have yet to find a mutation that increases genetic information in ways needed for microbes-to-man evolution to be feasible. Even in those rare instances where the mutation confers an advantage they almost always cause loss of information.’"⁵

That replacement sentence still makes two claims.  "In ways needed for microbes-to-man evolution" is a claim about scale, which the series tests in Benchmarking Evolution.  "Almost always cause loss" is a claim about proportion, which the tally below tests.

Robert Carter, a CMI scientist who works on human genetics, goes further.  "The phrase, “Mutations cannot create new information” is almost a mantra among some creationists, yet I do not agree."⁶  His conclusion is "Yes, depending on what you mean by ‘information’," and then:  "Can a mutation create new information? Perhaps, but only in the most limited sense. Can it create the kind of information needed to produce a genome? Absolutely not!"⁶  On gene duplication he writes that "New functions arising through duplication are not impossible, but they are vanishingly unlikely."⁶  Carter does score two of the series' cases differently.  He calls aerobic citrate use "the loss of control of the normal anaerobic citrate digestion" and nylon digestion "a loss of substrate specificity."⁶  The series scores the first a gain and the second a modification, for reasons given in Parts 3 and 5.

Kevin Anderson and Georgia Purdom, writing for Answers in Genesis in 2008, used "many" and "often" rather than "never":  "Many of the phenotypes result from loss of pre-existing systems. Other phenotypes involve activation of inactive systems, but this activation often involves simultaneous inactivation of another system."¹¹

John Sanford, a plant geneticist, holds both forms at once.  He writes that he is "still not convinced there is a single, crystal-clear example of a known mutation which unambiguously created information."  In the next breath he writes:  "I do not doubt there are beneficial mutations that create information, but it is clear they are exceedingly rare – much too rare for genome-building."⁷  His working claim is that mutation, "even when coupled with selection, cannot generally create new information."⁷

Michael Behe, a biochemist at Lehigh University and an intelligent-design writer, states the careful form as a rule about rates.  His 2010 review of four decades of laboratory evolution found that "only two gain-of-FCT events have been reported" outside experiments that deleted a gene and let it be restored.⁸  From that he drew "The First Rule of Adaptive Evolution":

"Break or blunt any functional coded element whose loss would yield a net fitness gain."⁸

He calls it "a rule of thumb," "a heuristic, useful generalization, rather than a strict law," and on the same page grants a gain in nature:  "Nonetheless, the sickle hemoglobin mutation did arise and spread in a regional population."⁸  Behe's rule predicts that losses and modifications will arrive first and most often.  It does not say gains never come.

The threshold form: not above about 500 bits

The intelligent-design writers make a third claim, and it is the strongest one on the design side.  It does not deny that mutation adds information.  It sets a size limit.  William Dembski, a mathematician and philosopher, counts a pattern as a mark of design only when its probability under every relevant chance process falls below a fixed bound, and he names "Darwinian and other material mechanisms" among those processes.¹²  In his 2005 statement:  "since χ must be greater than 1 for T to be a specification, this means that P(T|H) must be less than 10⁻¹²⁰."¹²  That is about 400 bits.  His earlier bound, one in 10¹⁵⁰, is about 500 bits.¹³  Stephen Meyer, a philosopher of science, uses the 500-bit line in Signature in the Cell and says plainly what it allows:

"My theory acknowledges that small amounts of specified information can occasionally arise by random processes, but that the amount of information that can be generated is limited by the probabilistic resources of the universe."⁹

His prediction, on the same page, is that "large amounts of new functionally specified information (over 500 bits) will not accumulate as a result of random or undirected natural processes."⁹  His conclusion about cells is narrower still:  intelligence is "the only known cause of specified information (at least starting from a nonbiological source)."⁹  That is a claim about the origin of the first life, not about mutation in living things.

None of the three design claims is refuted by one counter-example, because none of them says "never."  The threshold form can be tested only by measuring how much information a gain carries, in bits.  Part 2 does that, as a ceiling.

Who holds which

The strong form is testable by one counter-example.  The careful form is a claim about rates and proportions, and it needs a count.  The threshold form is a claim about size, and it needs a measurement in bits.  The rest of this article supplies all three, the last as an upper limit.

Part 2: What counts as new

Measuring Biological Information sets the rules, and every case below was scored by them in its own article.  Each change is asked four questions in order.  Did the sequence arrive from another lineage?  Then it is a transfer.  Did a working element stop working with nothing new in its place?  Then it is a loss.  Does a coded element, binding site, activity or control relationship now work that did not exist in this genome?  Then it is a gain.  Otherwise it is a modification, marked up, down, shifted or unmeasured.  The test follows Michael Behe's 2010 definitions, with transfer given its own class.  A gain is about what a molecule can now do, not about how many letters changed, so a second ledger counts the nucleotides that were originated rather than copied.  A duplication that joins a promoter to a silent gene can be a gain of function and still add zero new nucleotides.

The answer depends on which definition is used, and that is why the two sides talk past each other.  Shannon's measure counts symbols and rises with any random insertion.  Andrei Kolmogorov's algorithmic measure takes "the minimal length … of the “program” p for obtaining y from x."¹⁴  It also rises with random noise, and it barely rises with a copy, because "repeat that stretch" is a short instruction.  Gitt's measure requires meaning and purpose and cannot be applied to a sequence without first deciding what it is for.  Jack Szostak's functional information, the mainstream measure, asks for the probability "that a random sequence will encode a molecule with greater than any given degree of function," and turns it into bits.¹⁵  Dembski's specified complexity adds a threshold to that idea, as Part 1 showed.  The working definition this series uses, "unique sequences of nucleotides that contribute to function," is stricter than Shannon's or Kolmogorov's:  a duplicated gene adds nothing until the copy changes, and then only the changes count.  Under a loose definition every case below is new information.  Under the strictest, several of them add zero.  The honest statement is that function has been gained by mutation, repeatedly, and that the amount of new sequence each gain required is small.  Both halves of that sentence matter.

How many bits.  Functional information has not been measured for any case in the series, because that needs every alternative sequence tested.  It can be bounded from above.  Szostak's own range runs "from 2n (the amount needed to specify a given random RNA sequence) to 0 bits,"¹⁵ so a nucleotide that had to become one particular base adds at most 2 bits, and an amino acid that had to become one particular residue adds at most about 4.3.  Where other bases or residues would also have worked, the true figure is lower.  By that ceiling the lambda receptor switch carries at most 8 bits, sickle haemoglobin and haemoglobin E at most 2 each, one new influenza glycosylation site at most about 13, the HIV Vpu change at most about 9, and chloroquine transport in the malaria parasite at most about 43.  The guided ligase is capped by its 21 randomised positions at about 91 bits.  The citrate join and DUP4 originated no new sequence, so the ceiling does not reach them;  their information is in the arrangement.¹³  For scale, Szostak's ATP-binding RNA, the measured example in his paper, carries "about 37 bits."¹⁵

So every unguided gain in the tally that can be bounded at all sits at or below about a tenth of Dembski's 400-bit bound, on a ceiling that is generous to it.  The antifreeze and de novo genes cannot be bounded this way;  Part 6 gives what is known of their sequence.  The comparison cuts both ways.  The watched record does not refute the threshold form, because nothing in it comes near the line.  It does not confirm it either, and here the design side is weak.  Bits gained in separate steps add up when selection keeps each step, so a gene of 400 bits built in fifty paying steps of 8 bits each would break no rule of Dembski's.  His bound applies to the probability of the whole outcome under the full process, selection included, and for a biological structure Dembski wrote in 2005 that "The precise calculation of P(T|H) has yet to be done."¹²  Whether real genes lie at the end of paths whose every step paid is the open question.  The Avida objection below and Gene Duplication and Divergence take it up.

Part 3: Gains watched from start to finish

"Watched" is the strictest of the four grades in Observed, or Only Inferred?:  the population was recorded before and after, and the change was seen to arise.  Four gains meet it.

Citrate in Lenski's flasks

Twelve populations of E. coli have grown in a glucose medium that also contains citrate since 1988.  E. coli cannot take up citrate when oxygen is present.  "No population evolved the capacity to exploit citrate for >30,000 generations, although each population tested billions of mutations.  A citrate-using (Cit+) variant finally evolved in one population by 31,500 generations."¹⁶  Genome sequencing found the cause:  "The Cit+ trait originated in one clade by a tandem duplication that captured an aerobically-expressed promoter for the expression of a previously silent citrate transporter."¹⁷  The duplicated stretch was 2,933 base pairs.¹⁷  Later mutations added more copies and tuned the citrate pathway, as the Long-Term Evolution Experiment article details.

• Score: gain of function, with zero new nucleotides. The transporter gene and the promoter both existed. What was new was the join, a control relationship that did not exist in this genome. Behe set the test in 2010, before the mechanism was known: if the phenotype came from "the addition of a novel genetic regulatory element, gene-duplication with sequence divergence, or the gain of a new binding site, then it will be a noteworthy gain-of-FCT mutation."⁸ In 2019 he scored it as a modification: "no new functional coded element was gained or lost, just copied."¹⁸ The series keeps it as a gain by his 2010 test and records the nucleotide count separately. Measuring Biological Information gives both readings at full strength.
• Grade: watched. Frozen samples every 500 generations let the origin be dated and replayed.
• Search: about 7.3 × 10¹³ bacterial births over 82,000 generations in all twelve flasks, for one origin in one flask. Under a different regime, with citrate as the only food, the same trait has been re-evolved in weeks. Both figures are in Benchmarking Evolution.

Phage lambda's second receptor

Justin Meyer and colleagues grew phage lambda with E. coli that were evolving resistance by making less of the phage's usual receptor, LamB.  In a first run one of six populations evolved to enter through a different outer-membrane protein, OmpF.  In the main experiment, "As before, only some phage populations (24/96) evolved the ability to use a second receptor," with a median of 12 days.¹⁹  That is 25 independent origins in 102 flasks.  Four mutations in the phage's host-recognition gene J were needed:  "phage having only three of the four cannot use OmpF."¹⁹  Each of the four had been favoured on its own, because each improved binding to the old receptor.

• Score: gain of function, four new nucleotides. A binding that did not exist now works. Phage Lambda and the OmpF Switch scores it a hard case, because Behe's rule for a changed ligand gives modification on one untested condition.
• Grade: watched.
• Search: about 1.9 × 10¹⁰ to 1.2 × 10¹³ phage births across the 96 flasks. The four mutations were not simultaneous. They were climbed one paying step at a time, which is why the experiment does not test the harder case of steps that pay nothing until all are present.

Glycosylation sites in human influenza

A glycosylation site is a short motif on a protein where the cell attaches a sugar chain.  A sugar chain on the head of the influenza haemagglutinin hides it from antibodies.  Meghan Altman and colleagues measured the head glycans of 72 viruses by gel migration.  Human H1 went from one head glycan to five and H3 from two to seven.  Their summary:  "glycans are added to H1 and H3 HAs at discrete 5-to-7-year intervals, until they reach a functional glycan limit."²⁰  The newest site, added in the 2015-2016 season, needed "three mutations, namely, N162 itself, K163Q, and, crucially, I216T" for complete glycosylation.²⁰

• Score: gain of function. A new protein-processing site is on Behe's own 2010 list of gains. One to three amino acids per site, nine sites in a century.
• Grade: watched. The sites were dated in sequenced seasonal strains and one was rebuilt by reverse genetics.
• Search: 3 × 10²¹ to 3 × 10²³ virus genomes over 108 years, the largest search in the series. After the limit was reached, H1N1 "failed to evolve a widely circulating strain with 6 glycans" in at least 33 of 35 seasons.²⁰ The Influenza article treats that ceiling as the case's most useful finding.

A ligase from a scaffold with no ligase activity

Burckhard Seelig and Jack Szostak randomised 21 positions in two loops of a small zinc-finger protein domain and screened "4×10¹² unique mRNA-displayed proteins" for the ability to join two pieces of RNA.²¹  The winners sped the reaction at least 2 × 10⁶-fold.  "For the wildtype hRXRα protein domain, we could not detect any ligated product."²¹

• Score: gain of function, with a new fold. Directed Evolution and Protein Engineering finds that nearly every other laboratory case improved an activity the protein already had in trace, which it scores as modification. This one started from nothing detectable. The same article scores two small gains in Barry Hall's ebg experiments, a new inducer and the use of lactobionate.
• Grade: watched, and guided. An experimenter chose the target, built the library, and selected for the chemistry at every round. It shows that such activities exist in sequence space at a findable density. It does not show that an unguided population would find them.

Part 4: Gains dated or reconstructed in nature

The next cases happened outside a laboratory.  Their effects are measured in living populations, but their origins were bracketed between two dated samples or reconstructed from living genomes.

Chloroquine transport in the malaria parasite

Chloroquine kills the malaria parasite by building up in the compartment where it digests haemoglobin.  Resistant parasites pump it out through a transporter called PfCRT.  Robert Summers and colleagues showed that the mutant protein "mediates saturable CQ transport, whereas the wild-type form of the protein from CQ-sensitive (CQS) strains does not."²²  Every resistant form carries K76T and is "otherwise diverse, containing at least 4 and up to 10 mutations."²²

• Score: gain of function, with a growth cost. The protein carries a molecule it could not carry before. Where chloroquine was withdrawn, the resistant form declined, as Plasmodium falciparum documents.
• Grade: bracketed. Resistance appeared between 1957 and 1960 in dated clinical records. The mutational steps were not watched.
• Search: 1.5 to 7.5 × 10²¹ parasites in 75 years of drug pressure, and about 10²⁰ parasites per independent origin. This is the only gain in the series that needed more than one step before it paid, and it is the most expensive entry in the benchmark table.

Sickle haemoglobin, haemoglobin E and the Dantu blood group

Sickle haemoglobin changes one letter of the beta-globin gene.  Behe scores it as "an example of an adaptive gain-of-FCT mutation because a codon helping to specify a new, albeit weak … protein binding site has appeared."⁸  The new binding is the one that makes haemoglobin polymerise, and two copies cause sickle-cell disease.  Carter lists sickle cell among examples of "genetic degradation."⁶  Both descriptions are true of different things:  the molecule gained a contact, and the person with two copies is harmed.  The series scores the molecule.

Haemoglobin E is scored gain with loss:  one letter makes a new cryptic splice site and lowers globin output.  The glycophorin rearrangement DUP4 goes further.  Ellen Leffler and colleagues found "a complex structural rearrangement involving the loss of GYPB and gain of two GYPB-A hybrid genes, which encode a serologically distinct blood group antigen known as Dantu.  This variant reduces the risk of severe malaria by 40%."²³  It codes a protein no human had before, made by joining pieces of two existing genes.

• Scores: HbS gain, one nucleotide. HbE gain with loss, one nucleotide. DUP4 gain with loss, zero nucleotides originated and one new junction. All three are argued in Malaria Resistance Alleles, which also scores HbC, G6PD deficiency, Duffy negativity and others as modifications or losses.
• Grade: reconstructed. The protection is measured in living people. The origins predate any record.
• Search: 3 × 10⁸ to 4 × 10¹⁰ human births under malaria, on an assumed population size, so no rate is printed.

One lactase variant

At least five single-letter changes near the lactase gene let adults keep digesting milk:  "−13910T (rs4988235), −13907G (rs41525747), −13915G (rs41380347), −14009G (rs869051967) and −14010*C (rs145946881)."²⁴  Most strengthen an existing binding site, which is a modification.  One may be more.  The −14009G probe "showed a pattern that differed more dramatically from the three other probes tested at that locus, in that a prominent additional DNA–protein complex was formed."²⁵

• Score: −14009G is a candidate gain, one nucleotide, if the new complex is confirmed as a new binding site. The other variants are modifications up. See Lactase Persistence.
• Grade: bracketed by ancient DNA for the European variant. The East African variants are reconstructed.

HIV's Vpu protein and tetherin

Human cells make tetherin, a protein that holds new virus particles to the cell surface.  The chimpanzee virus that became HIV-1 could not block human tetherin.  Paul Sharp and Beatrice Hahn wrote that this "has resulted in the reacquisition of an anti-tetherin activity by Vpu," and that "Only in the case of HIV-1 group M has adaptation been fully successful."²⁶  In group N it came at a cost:  "HIV-1 group N Vpu has evolved to become active against tetherin, but this appears to have been at a cost," the loss of Vpu's older activity against CD4.²⁶  The site's HIV article traces the group M change to two amino-acid positions in Vpu's membrane-spanning region, each one or two nucleotides away from the chimpanzee virus.

• Score: not scored by class on this site. By the series' test the new binding would be a gain in HIV-1 group M, and a gain with loss in group N. Sharp and Hahn call it a reacquisition because other monkey viruses' Vpu proteins have the activity.
• Grade: reconstructed, by comparing human and chimpanzee viruses. HIV's drug resistance, by contrast, has been watched in patients.

Part 5: Real change that is not a gain

Most of the series' famous cases are not losses, and critics are right to object when they are called losses.  But they are not gains either.  They improved, redirected or copied something that already worked.

Modifications

• Extended-spectrum beta-lactamases. Four amino-acid changes and one change upstream of the TEM-1 gene raise resistance to the antibiotic cefotaxime from 0.088 to 4.1 × 10³ micrograms per millilitre, about 46,000-fold.^{27 27} Nothing breaks, and the same substitutions arise in hospitals. Antibiotic Resistance scores it a modification, shifted, five nucleotides, and calls it "a real gain" in the ordinary sense: the enzyme now destroys a drug it barely touched. It is the site's fourth standing concession, that some antibiotic resistance is genuine gain. Grade: watched.
• Nylonase. Bacteria that digest the waste from nylon-6 manufacture are real. The enzyme NylB differs from its nearly inactive twin by 47 amino acids, and two of them do the work: "only two amino acid replacements in the EII' protein are enough to increase the activity by 1000-fold."²⁸ NylB is an esterase of the beta-lactamase fold. The famous claim that it arose by a frameshift was a proposal that was never tested. Nylonase scores it a modification, two nucleotides. Grade: reconstructed, from the two enzymes.
• The other lactase variants, modification up, one nucleotide each.
• The dog's FGF4 retrogene. A copy of the FGF4 gene, inserted at a new site, causes the short legs of dachshunds and corgis. It "has a coding sequence that is identical to that of the source gene."²⁹ Dog Breeds scores it a modification up, by copy, zero unique nucleotides, because the copy carries its parent's own start region and raises the output of a gene already active in the same tissues. Grade: reconstructed for the first copy.
• The black peppered moth. The mutation behind industrial melanism "was the insertion of a large, tandemly repeated, transposable element (TE) into the first intron of the gene cortex."³⁰ Its most likely date is 1819, from a skewed estimate with a median of 1763.³⁰ It is a new mutation, not an old allele. Peppered Moths scores it a modification up, by copy, zero unique nucleotides. Grade: the spread was watched; the origin is reconstructed.
• Laboratory yeast. Gregory Lang and colleagues sequenced forty evolving yeast populations. At the mating-pathway and Ras-regulator genes the adaptive mutations looked like loss of function. At the cell-wall genes all thirteen were missense, "suggesting that at these loci selection is for alteration or attenuation, not loss, of function."³¹ That is modification, watched. The Genetic Entropy article also cites Jana Helsen's 2020 finding that deleting some yeast genes opened new routes to adaptation, and Yong-Chao Xu's 2019 survey of Arabidopsis, in which about 1% of genes carrying loss-of-function mutations in one population showed signs of positive selection. Both show adaptive loss, not gain, and the second shows that the adaptive share of loss is small.

Transfers

Tibetan high-altitude adaptation at EPAS1 came from interbreeding:  "the selected haplotype is only found in Denisovans and in Tibetans, and at very low frequency among Han Chinese."³²  EPAS1 and the Tibetan Highlands scores it a transfer worth zero originated nucleotides, and finds that the variant lowers the gene's output.  Most clinical antibiotic resistance is also transfer, of genes such as mecA, vanA and CTX-M that existed in soil bacteria.  Transfer moves information.  It does not show where the information came from.

New species without new sequence

Speciation has been watched, in plants, by polyploidy.  Marion Ownbey found "Four small amphiploid populations" of goatsbeard in 1949, "four recent and independent instances of the doubling of the chromosome sets," and named them as new species, Tragopogon mirus and T. miscellus.^{33 33}  Observed Speciation makes this concession without walking it back.  Polyploidy adds two existing genomes together and originates no sequence, and the young genomes lose genes.  A new species is not, by itself, new information.

Organisms with no scored gain

SARS-CoV-2 produced 1.5 × 10²⁰ to 4.4 × 10²¹ new mutations between 2019 and late 2021.  Its article scores fourteen changes:  twelve modifications and two losses.  Receptor binding rose steeply and antibody escape spread, but no change passes the gain test.  Mycobacterium tuberculosis produced about 10¹⁶ mutations in the drug era.  Its article scores ten changes:  eight modifications and two losses.  Neither organism added a scored gain.  That absence is a finding, and it belongs in the tally.

Part 6: Reconstructed gains, the largest and the least watched

The biggest novelties in the series were not watched.  They were reconstructed by comparing the genomes of living species, which is inference from common descent.  They are listed separately so that no reader mistakes them for observations, and so that no critic can say the site left them out.

Northern cod antifreeze from non-coding DNA

Cod and their relatives in Arctic waters make an antifreeze glycoprotein that stops ice crystals growing in their blood.  Xuan Zhuang, Christina Cheng and colleagues compared the antifreeze locus with the matching DNA in relatives that lack the gene.  Their conclusion:  "We experimentally proved that all genic components of the extant gadid AFGP originated from entirely nongenic DNA."³⁴  Their method was comparison:  they "constructed genomic DNA libraries for AFGP-bearing and AFGP-lacking species across the gadid phylogeny" in order to "reconstruct the gadid AFGP evolutionary path."³⁴

• Score: gain, a new gene from DNA that was not a gene. It is the site's third standing concession. The nucleotide count is left empty, because the series has no agreed way to count a gene built from repeats and a frameshift.
• How much new sequence. Less than the phrase "a new gene" suggests. By the authors' own account the coding region is many copies of one unit: "a nine-nucleotide (9-nt) element therein that supplied the codons for one Thr-Ala-Ala unit from which the extant repetitive AFGP-coding sequence (cds) arose through tandem duplications." The signal sequence was existing non-coding DNA, and "Through a 1-nt frameshift mutation, these two parts formed a single read-through open reading frame (ORF)."³⁴ The unit itself came from a non-coding stretch by "A chance 1-nt substitution."³⁴ So the protein is a short three-residue unit, repeated, made by a few point changes, copying and joins, like the citrate case on a larger scale. It is not an enzyme, and it does not test how a folded enzyme could arise. The other side of this is equally real. A gene with a promoter, a secretory signal and a working product did arise from DNA that coded for nothing, and the steps the authors list are all ordinary mutations.
• Grade: reconstructed. "Experimentally proved" refers to experiments on living fish DNA, not to watching the origin. Observed, or Only Inferred? found that the Genetic Entropy article's phrase "documented base by base" reads as if the origin had been watched, and it was not.

Antarctic antifreeze from a digestive enzyme

Antarctic notothenioid fish make an almost identical antifreeze by a different route.  Liangbiao Chen, Arthur DeVries and Christina Cheng reported that these proteins "evolved from a pancreatic trypsinogen."³⁵  The gene kept pieces of the enzyme gene at each end and grew a new repeated core in between.  The protein is again a repeat:  the antifreeze molecules are "all composed of a simple glycotripeptide repeat, (Thr-Ala/Pro-Ala)n," with a sugar on each threonine.³⁵

• Score: gain, by duplication and divergence. Carter allows that "New functions arising through duplication are not impossible."⁶ Here the new function is not a close relative of the old one. The gene kept the enzyme gene's secretory signal and its tail, deleted or silenced the rest of the enzyme's coding sequence, and grew its new core from a nine-nucleotide piece. The harder test of duplication, turning one folded enzyme into another, is argued in Gene Duplication and Divergence. There Ann Gauger and Douglas Axe report that converting the E. coli enzyme Kbl2 into its relative BioF2 "would in this case require seven or more nucleotide substitutions," and that such innovations "would be extraordinarily rare, becoming probable only on timescales much longer than the age of life on earth."³⁶ The young-earth biologist Todd Wood answered that "no evolutionary biologist would propose that an extant enzyme evolved directly into another extant enzyme."³⁷ Both points stand: the experiment measured a real distance between two living enzymes, and that distance is not the path evolution proposes.
• Grade: reconstructed.

De novo genes

Some genes in yeast, flies and primates have no relatives outside a narrow group, and their best cases sit beside a matching non-coding stretch in related species.  Orphan Genes weighs three explanations for each such gene:  real and new, old but diverged past detection, or not a gene at all.  The strongest cases rest on the presence of the non-coding version in several outgroups, not on the absence of a match.  Reported de novo genes are short and weakly expressed, and few have a demonstrated function.  The protein-rarity figures used against the origin of folded enzymes do not apply to them, and this site does not aim those figures at them.

• Score: gain where a function is shown; unscored where it is not.
• Grade: reconstructed.

The honest reading is that the reconstructed cases are the strongest evidence for new genes arising, and the weakest evidence for anyone having watched it.  Both halves hold.

Part 7: Cases that turned out not to be gains

Four famous cases dropped out when their articles scored them.  The cecal valves of the Italian wall lizards on Pod Mrčaru have not been shown to be inherited:  they formed in 12 of 16 wild lizards and none of 20 kept on insects, so the case is not yet scoreable.  Every change watched in Darwin's finches was a shift in the frequency of old haplotypes or a transfer between species.  Sticklebacks adapted to fresh water within a decade from alleles already present in the sea, and the changes behind those alleles are a loss, a modification down and two copy-number or expression changes.  The cichlid radiations sorted many species from stored variation and were not watched at all.  All four are strong evidence for fast adaptation, and none is a watched gain.

Part 8: The tally

The table lists every case above.  "Nucleotides" means new nucleotides by the series' conversion rule:  substitutions that produced the function count, copied or joined sequence counts zero, and a new gene from non-gene DNA is left empty.  "Search" is the population the sibling article gives, where it gives one.

The sums.  Benchmarking Evolution counts seven scored gains in its table:  citrate, lambda, the influenza glycans, PfCRT, HbS, HbE and DUP4, two of them gain with loss.  Its rows hold roughly 7 × 10²² to 5 × 10²³ births between them.³⁸  Outside that table the series adds the laboratory ligase and two small ebg gains, all guided, one candidate gain in lactase, the unscored Vpu change, and three reconstructed gains that were not watched.  By grade, the gains and candidate gains listed here are five watched, counting the two ebg gains as one entry, of which four were in a laboratory and one in nature;  one bracketed;  seven reconstructed from living genomes, plus the de novo genes as a class;  and none asserted.  Every watched gain changed between zero and four new nucleotides, or one to three amino acids per glycan site.  None of them is a new gene with a shown benefit.  In bits, by the ceiling in Part 2, none of the unguided gains exceeds about 43.

Part 9: What it adds up to

For the critic of this site.  The strong claim is false.  Mutation and selection have been watched producing a new control relationship, a new receptor binding, new protein-processing sites and a transporter that carries a new molecule.  A laboratory search has found a new enzyme activity in a scaffold that had none.  Comparative genomics has reconstructed two antifreeze genes, one from DNA that was not a gene.  This site concedes all of it, and the concessions stand whatever the rest of the series finds.

For the reader who holds the strong claim.  Stop using it.  "Mutations never add information" can be refuted by one case, and the series holds several.  Spetner's "Not even one mutation has been observed" does not survive the cases above, most of which were published after he wrote it.  The argument that remains is the one CMI, Carter and Sanford already make:  that gains are rare, small and slow, set against a much larger number of losses and modifications.  That argument has to be made with counts, and the tally above is the count.

For the design advocate.  Behe's first rule fits the tally:  losses and modifications outnumber gains.  Meyer's threshold is not touched by it:  the largest unguided gain is a tenth of the line.  Neither result is a positive case for design.  The threshold argument still needs what Dembski said in 2005 had not been done, a probability for a real biological pathway with selection included.  Until then, the gains in this article show only that small steps happen, and the question is whether large things are made of small steps that each paid.

For the evolutionist.  The watched gains are small.  Each changed between zero and four new nucleotides, or one to three amino acids at a glycan site.  The largest search in the series, influenza's 10²¹ to 10²³ genomes, bought one kind of gain and then stopped at a measured ceiling.  SARS-CoV-2 and tuberculosis bought none.  Most of the fitness gains in the series are losses or modifications, as Behe's rule of thumb predicts for what arrives first.  The one gain that needed more than one unpaid step, chloroquine transport, cost about 10²⁰ parasites per origin.  None of the watched gains is a new gene with a shown benefit.

What this article does not settle.  Whether the watched rate, carried along one mammal lineage for the time common descent allows, could build what mammals have is a question about extrapolation.  Benchmarking Evolution finds that the comparison by head count fails, because a newborn mammal carries thousands of times more new mutations than a newborn parasite, and that only the per-lineage form of the argument survives.  That form depends on how much of a mammal genome needed steps that paid nothing on their own, which nobody has measured.  This article records the gains.  It does not use them to argue either way about that question, and nothing in it takes back the concession.

Objections and responses

"Citrate use is just a broken switch, a loss of regulation."  Carter describes it as "the loss of control of the normal anaerobic citrate digestion."⁶  The genome data do not show that.  The native control region of the transporter "showed no expression (above background)" in oxygen, and nothing in it was broken.¹⁷  The trait appeared when a copy of the gene was joined to a different promoter.  That is a new control relationship, which Behe's 2010 test counts as a gain.  What the objection gets right is that no new sequence was built.

"Nothing new was made, only copied, so it is not information."  Under the strictest definition, the citrate join, DUP4, the FGF4 retrogene and the moth's insertion add zero unique nucleotides, and the tally says so.  Lambda, the glycans, PfCRT, HbS and the ligase are not copies:  they are substitutions that made a function that did not exist.  The objection applies to some cases and not to others.

"These are all within kinds, so they do not matter."  The claim this article answers is not about kinds.  It is the claim that mutation never adds information.  That claim is false within kinds as well as across them.  Whether the gains can add up to the differences between kinds is the extrapolation question, handled elsewhere.

"Beneficial mutations are too rare to matter."  That is Sanford's form of the claim, and it is a claim about rates, not about existence.  The tally supports part of it:  seven scored gains across roughly 10²³ births in the benchmark table, against many more losses and modifications.  It does not settle whether the rate is too low for common descent, because that depends on the per-lineage question.

"The directed-evolution ligase proves design, since a scientist designed the experiment."  It proves that such activities exist at a density a search of 4 × 10¹² variants can find.  The experimenter chose the target and supplied selection at every round.  It is marked guided for that reason, and it is not counted among the gains made by unguided populations.

"Digital organisms build complex functions from nothing, so information grows wherever selection acts."  Richard Lenski, Charles Ofria, Robert Pennock and Christoph Adami evolved self-copying programs in the Avida software in 2003.  "Complex functions evolved by building on simpler functions that had evolved earlier, provided that these were also selectively favoured," and "In some cases, mutations that were deleterious when they appeared served as stepping-stones."³⁹  The first program to perform the most complex function, EQU, had 60 instructions, and "eliminating any of 35 of them destroyed that function."³⁹  Their own control is the point the design side presses:  "the complex feature never evolved when simpler functions were not rewarded."³⁹  Winston Ewert, Dembski and Robert Marks read the reward table as information the programmers put in:  "Avida uses stair step active information by rewarding logic functions using a smaller number of nands to construct functions requiring more."⁴⁰  Lenski's group had already answered that a ladder of useful steps "is precisely what evolutionary theory requires."³⁹  Both are right about Avida.  It shows that mutation and selection can build a function that needs 35 specific instructions, even through a harmful step, when every rung of the ladder pays.  It cannot show that living genes have such ladders, because the programmers wrote this one.  Meyer's 2009 prediction states the design side's expectation:  "genetic algorithms simulating the power of undirected evolutionary processes will not produce more than 500 bits of new complex specified information" unless "an intelligent programmer provides necessary “active information.”"⁹  The ladder need not be complete, though.  In Ewert's own runs, with four of the eight simpler rewards removed, EQU still evolved in 222 of 353 runs.⁴⁰  Avida is digital, so it is not in the tally.  Evolutionary Computation treats it in full.

"The reconstructed antifreeze genes assume common descent."  They do.  The cod case rests on the matching non-coding DNA in relatives that lack the gene, and on the order of steps that comparison implies.  That is why they are graded reconstructed and kept out of the watched list.  The inference is still strong:  no young-earth author has published an account of why the cod gene sits where a non-coding relative predicts.

"Behe himself scores citrate as a modification, so the site is misrepresenting him."  Behe's 2010 paper set a test in advance and his 2019 book scored the result against a different reading.  Both are quoted above.  The series applies the 2010 test because it was written before the answer was known.

"The site is conceding too much and handing evolutionists a win."  The concession is limited to what was found.  The same records show that the watched gains are small and few, that the largest search in the series stopped at a ceiling, and that two organisms made no gain at all.  A reader who hears the concession from this site will trust the rest of the ledger more, not less.

Issues and unanswered questions

• The count depends on how cases are split. Nine glycan sites are one row here; two ebg gains are one entry. A different split changes the numbers without changing the pattern.
• The HIV row is not scored by class. Benchmarking Evolution gives HIV an upper bound of 5,000 nucleotides and no class counts, so HIV's gains are under-represented here. The site's HIV article also has a list of corrections pending, recorded on a separate card.
• Several scores are hard cases. Lambda and the ebg inducer are gains by the series' test and modifications by Behe's ligand rule. HbE depends on whether its protection comes from the new splice site. The lactase candidate depends on confirming the new binding factor.
• Two sibling articles disagree with the benchmark table on counts. The SARS-CoV-2 article scores twelve modifications and two losses; the benchmark table prints eleven modifications. The influenza article states seven modifications, but its own table lists six. Neither changes the gain count.
• The watched record is biased toward the cheap. Experiments and clinics record changes that arrive in days to decades. A gain that needs thousands of years in a large animal could not appear in this record even if it happens. That cuts against using the record as a ceiling.
• The de novo gene count is unsettled. Published human counts range from 60 to over 1,400, and few have a shown function. Until that settles, the reconstructed column is a class, not a number.
• The bit figures are ceilings, not measurements. They assume that each changed site had to be one particular base or residue. Where several would have worked, the true figure is lower. Nobody has measured functional information for any case in the series. Robert Hazen and Szostak's group note that the random sampling it needs "is not possible with living organisms."⁴¹
• The design threshold is untested in both directions. No watched gain approaches 400 to 500 bits, so none tests the claim that unguided processes cannot pass that line. No design writer has computed the probability of a real biological pathway with selection included, so the claim has no positive test either.
• The young-earth side has no published model for the reconstructed cases. Nobody on that side has explained the Arctic cod antifreeze locus, the Antarctic gene's trypsinogen ends, or the dated origins of the peppered moth insertion and the lactase variants in terms of a created kind and a short chronology.

Verdict

Evolution has created information, by any definition that counts a new function without a size threshold, and it has been watched doing it.  The claim that mutations never add information should be retired by everyone who uses it, and this site retires it here.  The watched gains are also small:  zero to four new nucleotides each, a minority beside the losses and modifications, and no new gene with a shown benefit in any watched population.  Measured as functional information, no unguided gain in the series carries more than about 43 bits even on a generous ceiling, about a tenth of the 400 to 500 bits that Dembski and Meyer set as the design threshold.  These cases neither refute that threshold nor confirm it.  The largest novelties, the two antifreeze genes and the de novo genes, were reconstructed from living genomes, not watched.  Whether gains of this size, at this rate, could build what common descent requires is the open question of Benchmarking Evolution, and this article leaves it there.

Sources