Gene Duplication and Divergence

"I have no objection to thinking the proteins within those separate classes arose by gene duplication." — Michael Behe, 2019¹

"Gene duplication, as the name implies, involves the production of a duplicate copy of a preexisting gene, already rich in functionally specified information." — Stephen Meyer, Darwin's Doubt, 2013²

Summary

When a cell copies its DNA it sometimes copies a gene twice.  The extra copy can be lost, kept as a spare, or changed until it does a job of its own.  Since Susumu Ohno's book of 1970 this has been the standard account of where most genes came from:  copy an old gene, then let the copy diverge.  This article asks how much of the gene content of a genome that route can account for, and how well each step has been measured.

1. New copies arise fast, and almost all of them are lost. Mutation-accumulation lines, in which mutations are allowed to pile up under almost no selection, measure the raw rate directly: 3.4 × 10⁻⁶ per gene per cell division in yeast,³ and 1.2 × 10⁻⁷ to 2.9 × 10⁻⁵ per gene per generation in a nematode worm.⁴⁵ The rate inferred from the young duplicate pairs that survive in sequenced genomes is about 0.01 per gene per million years, tens to thousands of times lower, and the average copy is silenced with a half-life of about 4 million years.⁶⁶ In fruit flies 98.8% of new duplications are removed by selection.⁷ Supply is not the limit. Keeping a copy is.
2. Most genes have relatives in their own genome. A count made for this article from the Pfam database finds that 81% to 85% of human proteins, 60% to 71% of yeast proteins and 71% to 72% of Arabidopsis proteins share a domain family with another protein of the same species.⁸ Each family still needs one member that was not copied from another, and some proteins share only one domain among several. Setting aside one founder per family leaves 64% to 76% of human genes, 40% to 60% of yeast genes and 61% to 68% of Arabidopsis genes whose template copying within the lineage can supply.⁸ For those genes copying is the obvious source of at least the shared part of the sequence, and Michael Behe agrees for families whose members do similar jobs: "I have no objection to thinking the proteins within those separate classes arose by gene duplication."¹
3. Whole genomes have been doubled, and in yeast it is shown, not just inferred. Manolis Kellis, Bruce Birren and Eric Lander sequenced a yeast that branched off before the doubling and found each of its regions matched by two regions of baker's yeast, with genes lost alternately from one copy or the other.⁹ Baker's yeast keeps 551 such pairs, 19.6% of its genes.¹⁰ The two rounds of doubling proposed at the root of the vertebrates, and the many in flowering plants, rest on the same kind of evidence with older and noisier signal.
4. Families need founders, and copying cannot supply them. The same count finds about 4,100 distinct domain units in the human proteome, about 7,300 if related families are not merged, and 1,800 to 3,500 in yeast.⁸ Each needs at least one origin that was not a copy of anything Pfam can relate it to. Copying explains the size of families. It does not explain why there are thousands of them.
5. Where the divergence has been measured, the new job took a handful of changes. Three amino-acid changes, tested by reversing them in a rebuilt ancestral protein, were needed for the shift that fitted a leaf-eating monkey's digestive enzyme to its gut.¹¹ Three account for the colour difference between the red and green pigments of the human eye.¹² Two let the ancestor of haemoglobin assemble into a four-chain molecule, and fourteen give its modern α2β2 arrangement.¹³ Every one of these is a reconstruction from living species. The only divergence watched from start to finish was a laboratory experiment with a starting gene the experimenters built.¹⁴
6. The design side's quantitative argument is real but narrower than it is often used. Behe and David Snoke modelled a copy that needs several changes, none useful alone, and found that populations of at least a billion would be needed.¹⁵ Michael Lynch replied that with neutral intermediates and more than two usable sites the same change takes "a million years (or much less)".¹⁶ Behe and Snoke's own reply draws the line: their result counts against paths whose steps are harmful, and requires each step to be "either selected or at least neutral."¹⁷ The measured cases need two or three key changes. For the opsins each single change was made in the rebuilt ancestral pigment, and each gives a working pigment shifted part of the way, so that case is gradual.¹² For the monkey enzyme and haemoglobin nobody has measured whether the single steps helped, hurt or did nothing. Behe wrote in 2019 that the origin of haemoglobin was "now very much an open question";¹⁸ a reconstruction published a year later answered the structural half of it.
7. Several creationist treatments misread their own sources. Yingguang Liu and Dan Moran say Lynch and John Conery realised "the impossibility of neofunctionalization";¹⁹ the paper says duplicates "may only rarely evolve new functions" and that "the few survivors" are kept by strong selection.⁶ Rob Carter says the people who cite duplication "never tell us the rate";²⁰ Behe, whom he cites, gives it.²¹ Jean Lightner accepts duplication as part of the creation model.²²
8. On a young-earth timeline almost no duplicates accumulate, so families must be created as families. At the mainstream rate of retained copies, a human-sized genome gains between a fifth of one and two and a half kept duplicates in 4,500 to 6,000 years, against 13,000 to 15,600 copies to account for.²³ The young-earth model therefore needs gene families created as families, with copy-number changes like those in the human starch-digesting gene as later variation.²⁴ That is consistent with common design. No young-earth source found for this article explains the yeast genome doubling, or gives a duplication count for its own timeline.

The verdict.  Duplication followed by divergence can account for the template of most genes in human and Arabidopsis and of 40% to 60% in yeast, once each family's founder is set aside.  Behe grants it within families of similar genes, and Jean Lightner gives duplication a place in the creation model.  Liu and Moran, Peer Terborg, and Reeves, Gauger and Axe deny that families arose this way, without offering a test that separates their reading from common descent.  In terms of new functional sequence it accounts for little:  the measured new jobs took two to fourteen amino-acid changes laid on a copy of something that already worked.  It does not account for the founding member of each family, of which there are thousands per genome, or for the protein folds those families share.  That is the boundary of the route.  The fold question belongs to How Rare Is a Working Protein? and the no-template question to Orphan Genes.

Part 1: The route, and the models of what a copy does

Ohno's model: the spare copy is free to change

Susumu Ohno, a geneticist at the City of Hope in California, put the idea in its best-known form in 1970.  Selection polices every working gene.  A copy is redundant, so selection leaves it alone, and it can then collect changes that would have been fatal in the original:  "By duplication, a redundant copy of a locus is created.  Natural selection often ignores such a redundant copy, and, while being ignored, it accumulates formerly forbidden mutations and is reborn as a new gene locus with a hitherto non-existent function.  Thus, gene duplication emerges as the major force of evolution."²⁵  His preface puts it in one line:  "natural selection merely modified, while redundancy created".²⁵  Copying a working gene and then changing one copy is now called neofunctionalization.

Ohno's account has a cost built in.  A copy that selection ignores is also a copy that selection does not protect.  Most changes to a protein damage it, so the commonest fate of an unprotected copy is to be broken.  Michael Lynch and John Conery estimated this from six genomes in 2000 and found that "the vast majority of gene duplicates are silenced within a few million years, with the few survivors subsequently experiencing strong purifying selection."⁶  The models that followed are attempts to explain why any copies survive at all.

The rivals: split the job, use the extra dose, or start with two jobs

Subfunctionalization.  In 1999 Allan Force, Lynch, John Postlethwait and colleagues proposed that a gene with several jobs, for example expression in several tissues, can lose different jobs in each copy.  Then both copies are needed, and both are kept, with nothing new made.  They called it the duplication-degeneration-complementation model and claimed "the usual mechanism of duplicate gene preservation is the partitioning of ancestral functions rather than the evolution of new functions."²⁶  Its probability depends on numbers it does not measure.  At their own example rate the chance of preservation is 0.1, and it passes 50% only "if the total rate of subfunctional mutations ... exceeds the null rate in the coding region by more than approximately fourfold".²⁶

Gene sharing, and escape from adaptive conflict.  Austin Hughes argued in 1994 that the order is usually the reverse of Ohno's.  One protein first does two jobs, and the copy then lets each daughter "specialize for one of the functions of the ancestral gene."²⁷  If the two jobs pull the single gene in different directions, the copy relieves the conflict.  The yeast galactose genes in Part 5 are the best-documented case.

Dosage.  Sometimes two copies help simply because the cell needs more of the product.  People whose ancestors ate starchy food carry more copies of the salivary amylase gene, and more copies make more enzyme:  "individuals from populations with high-starch diets have on average more AMY1 copies than those with traditionally low-starch diets."²⁴  The reverse also holds.  Genes whose products work in fixed proportions, such as the parts of a ribosome, are harmed by a single extra copy but kept after a whole genome doubles, because the proportions are preserved.  Lynch and Conery suggested this to explain why whole-genome copies survive far better than single copies.⁶  Balázs Papp, Csaba Pál and Laurence Hurst tested it in yeast in 2003:  "single-gene duplication of protein subunits is expected to be harmful, as this, too, leads to imbalance.  As then expected, we find that members of large gene families are rarely involved in complexes."²⁸  In humans, Takashi Makino and Aoife McLysaght found that copies left by the two vertebrate doublings are less often found duplicated again, or varying in copy number, than other genes:  "Over 60% of ohnologs (63.6%; 4,638/7,294) are free of SSD and CNV, compared to 32.4% (4,412/13,613) of nonohnologs".²⁹

Amplify, then diverge.  In 2012 Joakim Näsvall, John Roth, Dan Andersson and colleagues watched bacteria under selection amplify a gene with a weak second activity into many copies, after which "the amplified gene copies accumulate mutations that provide enzymatic specialization of different copies and faster growth."¹⁴  Here the extra copies are useful from the start, for dosage, so they are not exposed to decay while they wait.

Which fate is common

Nobody has counted the fates of duplicates directly.  They are inferred from patterns of sequence change and gene expression, and the answers depend on the method and the organism.  Hideki Innan and Fyodor Kondrashov, reviewing the models in 2010, classed them "into four major categories according to the mode of selection in the early phases of the evolution of the duplicated copies", and warned that the usual statistic for detecting selection "is not informative for distinguishing between different models."³⁰

• In yeast, Kellis, Birren and Lander compared each pair kept from the genome doubling with an outgroup species and found that "Strikingly, 95% of cases of accelerated evolution involve only one member of a gene pair".⁹ One copy kept the old job and the other changed, as Ohno predicted. Only 76 of the 457 pairs showed accelerated change at all.⁹
• In fruit flies, Raquel Assis and Doris Bachtrog used expression across tissues as a stand-in for function in 281 young duplicate pairs. They scored 183 as neofunctionalized, 3 as subfunctionalized, 42 as specialized and 53 as unchanged.³¹ Their method measures where a gene is switched on, not what its protein does.
• In Atlantic salmon, Sigbjørn Lien and colleagues compared the copies left by the salmon genome doubling with the single-copy northern pike and counted 3,028 cases they class as neofunctionalization against 167 of subfunctionalization.³² Like the fly study, the classes come from expression patterns.
• Xionglei He and Jianzhi Zhang found that neither model alone fits yeast protein interactions and human expression data, and proposed a mixture: "rapid SF, accompanied by prolonged and substantial NF in a large proportion of duplicate genes".³³
• The galactose genes, the leaf-monkey enzyme and the opsins in Part 5 are each one gene family studied in depth. They show that each route happens. They do not show how often.

Part 2: How often genes are copied, and how often the copies stay

The raw rate, measured directly

The direct way to measure a mutation rate is to let mutations pile up and count them.  In a mutation-accumulation experiment, lines of an organism are passed through a single individual each generation, so selection can remove only the most crippling changes, and the genomes are sequenced at the end.

The worm studies disagree with each other by a factor of 8 to 85, and the human figure lumps duplications with deletions and other rearrangements.  Even so, every direct measurement is high.  Lynch's yeast lines produced more gene duplications per genome than base substitutions:  "the expected number of gene duplications per genome is 0.019 ... which exceeds the expected number of base substitutions per genome per generation (0.004)."³  Kendra Lipinski and colleagues put the worm's rate "two orders of magnitude greater than the spontaneous rate of point mutation per nucleotide site in this species".⁴  Copies of a gene are not scarce.

The rate of copies that last

The indirect way is to count the young duplicate pairs in a sequenced genome and divide by an estimate of their age, taken from how many silent changes they carry.  Lynch and Conery did this for six species in 2000 and arrived at "a conservative estimate of the average rate of origin of new gene duplicates on the order of 0.01 per gene per million years, with rates in different species ranging from about 0.02 down to 0.002."⁶  Li-zhi Gao and Hideki Innan argued in 2004 that gene conversion, which keeps copies looking alike, makes old pairs look young.  Correcting for it they put the yeast rate at "approximately 0.01 to 0.06 per gene per billion years", a hundred times lower again.³⁶

The two kinds of rate measure different things.  The direct rate counts copies when they form.  The indirect rate counts copies that lasted long enough to be seen as pairs in a genome.  Lynch's group noted the gap themselves:  their yeast rate "is substantially higher than previous indirect estimates for eukaryotes ... which are generally in the neighborhood of 10⁻⁸ per gene per generation."³  The yeast figure is about 340 times that, and the worm figures 34 to 2,900 times.²³  Read the other way, roughly one new copy in 34 to one in 2,900 lasts long enough to be counted as a pair, between 3% and 0.03%.  Most new copies are gone within a few generations.  In fruit flies Daniel Schrider and colleagues compared new mutations with those found in wild flies and estimated "that 98.8% of duplications and 99.2% of deletions in D. melanogaster are removed by purifying selection".⁷

Half-lives

Among copies that survive long enough to be counted, Lynch and Conery found "the average half-life of a gene duplicate is approximately 4 million years".⁶  Their own conclusion was that "50% of all of the genes in a genome are expected to duplicate and increase to high frequency at least once on time scales of 35 to 350 million years."⁶  In Arabidopsis the estimated half-life is longer, 17.3 million years.³⁷  All of these times come from silent-site clocks calibrated on the standard timescale.  The direct rates in the table do not.

What this part settles.  Copies arise constantly, in every organism measured, and almost all are lost.  The mainstream account does not lack copies to work with.  On its own rates, a genome of 20,000 genes gains about 200 lasting duplicates every million years and holds about 1,200 young ones at any moment.²³  The bottleneck is keeping a copy long enough for it to become useful, which is what the models in Part 1 try to explain.

Part 3: Whole-genome duplications

A whole genome can be copied at once, as when a cell fails to divide after copying its chromosomes, or when two species hybridise and keep both parents' chromosomes.  Plants do this often today.  The claim tested here is that it also happened in the deep past of yeasts, vertebrates and flowering plants, and left gene families behind.

Yeast: the doubling that is shown, not just inferred

Kenneth Wolfe and Denis Shields proposed in 1997 that baker's yeast descends from an ancestor that doubled its genome.  They found 55 duplicated regions in which the same genes sit in the same order, together spanning "50% of the genome".³⁸  Critics replied that the model "explains only a minority of the gene redundancy in the genome and defines putative sister regions covering less than 50% of the genome."⁹

Kellis, Birren and Lander settled it in 2004 by sequencing Kluyveromyces waltii, a yeast that branched off before the doubling.  If baker's yeast is a doubled genome, each region of K. waltii should match two regions of baker's yeast, with the genes lost after the doubling missing from one copy or the other.  That is what they found:  "We identify 145 paired regions in S. cerevisiae, tiling 88% of the genome and containing 457 duplicated gene pairs."⁹  They inferred that about 88% of the doubled genes were later lost, mostly in small deletions affecting both copies about equally.⁹  Kevin Byrne and Wolfe's curated list counts 551 pairs still kept, so "19.6% of the genes in S. cerevisiae (1102 of 5516) are members of an ohnolog pair".¹⁰  Marina Marcet-Houben and Toni Gabaldón later argued that the doubling followed a hybridisation between two species.³⁹  That changes how it happened, not whether.

This is the strongest evidence in the subject because it makes a prediction about a genome nobody had read, and the prediction came true.  Behe accepts it and calls it "yet more evidence for common descent", adding that the doubling "seems not to have done a whole lot for its recipients".²¹

Vertebrates: two rounds at the root, inferred from gene order

Ohno argued that a set of genes with linked functions, scattered over several chromosomes, can be copied together "only by becoming tetraploid", that is, by doubling the genome.²⁵  The proposal that the vertebrate genome was doubled twice near its root grew from that idea.  Counting genes did not settle it.  Paramvir Dehal and Jeffrey Boore found in 2005 "that there remains little signal of these events in numbers of duplicated genes, gene tree topology, or the number of genes per multigene family."⁴⁰  The case rests on gene order instead.  When they mapped only the human paralogs that arose before the fish and land vertebrates split, the copies fell into blocks of four.  Nicholas Putnam and colleagues, comparing the human genome with the lancelet amphioxus in 2008, found that "Over 90% of the human genome is encompassed within these linkage groups, which display a tell-tale fourfold redundancy that is consistent with whole-genome quadruplication on the vertebrate stem."⁴¹  The same paper concedes that "the number, the timing and even the genomic scale of the duplication events, and their consequences for subsequent genome evolution, are poorly understood".⁴¹  The genes still kept in pairs or quartets from these doublings, called ohnologs after Ohno, "amount to about 20 to 35% of genes in the human genome" according to Param Priya Singh, Jatin Arora and Hervé Isambert.⁴²

Robert Friedman and Austin Hughes dissented, and Dehal and Boore quote their objection:  "less than 5% of homologous gene families follow the 4:1 rule".⁴⁰  The map evidence is the answer the field accepted.  The Hox clusters, one in the lancelet and four in mammals, are the textbook example of the four-block pattern.

Later doublings are better supported because they are younger.  Salmon and trout doubled again about 80 million years ago on the standard dating, and Sigbjørn Lien and colleagues report that "55% of the Ss4R duplicates are retained as two functional copies in Atlantic salmon".³²

Plants: doublings everywhere

In flowering plants genome doubling is common enough to be counted by the hundred.  The One Thousand Plant Transcriptomes Initiative "inferred 244 ancient WGDs across Viridiplantae" from gene-family trees, and checked the method against 65 doublings already inferred from whole genomes, finding 6 missed and none falsely called.⁴³⁴³  In Arabidopsis about 60% of genes have a paralog in a block derived from one of three doublings.³⁷

What this part settles.  The yeast doubling is as well shown as anything in comparative genomics can be.  The vertebrate and plant doublings rest on the same kind of evidence, gene order preserved in blocks, read through more loss and rearrangement.  None was watched.  All of them copied whole sets of existing genes, and so, under the bookkeeping of Measuring Biological Information, added nothing new until the copies changed.

Part 4: How much of a genome sits in families

The question in this article's title has a countable half.  If a gene shares a recognisable domain with another gene in the same genome, a copying event can in principle explain that domain, and most of the gene when the two share their whole arrangement of domains.  If it shares nothing, copying inside that genome cannot.  And every family, however large, needs at least one member that was not a copy of anything else in it.

This article counted both, using the Pfam database of protein domain families (release 37.0, May 2024) and UniProt's reference proteomes with one protein per gene (release 2026_03).⁸  A "unit" is either a Pfam clan, a group of families Pfam judges to share ancestry, or a family that belongs to no clan.  A protein is "in a family" when it shares at least one unit with another protein of the same species.

Three things follow.

Most genes have relatives in their own genome.  Between 60% and 85% of the proteins in each species share a domain family with another protein in the same species, depending on how strictly a family is drawn.  Between 45% and 72% share an identical arrangement of domains.  For these genes duplication is the obvious source of the bulk of the sequence;  for a gene that shares only one domain among several, only of that domain.  Behe grants it for families whose members do the same kind of job.  Mariclair Reeves, Ann Gauger and Douglas Axe read the same resemblance as shared design instead (Part 6).

Families are not founders.  The human proteome uses about 4,100 distinct units, or about 7,300 if Pfam's clans are not merged.  Each needs at least one origin that was not a copy of any gene in this genome that Pfam can relate it to.  Some of those origins are shared with yeast and plants and lie far back on the evolutionary account;  some are not.  Duplication explains the size of families, not the existence of them.

The rest is not counted.  Proteins with no Pfam domain, 9% in human and 26% in Arabidopsis, may have relatives that Pfam's models miss, or none.  Part of the Arabidopsis figure may reflect the gap between the 2023 protein set Pfam was built on and the 2026 set counted here.  These genes are the subject of Orphan Genes.

Published counts use stricter tests of relatedness and come out lower.  Jianzhi Zhang's 2003 review tabulated duplicate genes as 30% of yeast genes, 38% of human, 41% of fruit fly, 49% of nematode and 65% of Arabidopsis genes, and added that "It is almost certain that these proportions are underestimates, because many duplicated genes have diverged so much that virtually no sequence similarity is found."⁴⁴  Domain families reach further back than whole-gene similarity, which is why the Pfam count is higher.  Neither count says how the families arose.  Both say how many genes have relatives to arise from.

Under the bookkeeping this site uses in Measuring Biological Information, a copy adds nothing until it changes, and then only the changes count.  The 60% to 85% includes one founding member of every family, and copying did not make the founders.  Setting one founder per family aside leaves the share of genes whose template duplication within the lineage can supply:  64% to 76% in human, 40% to 60% in yeast and 61% to 68% in Arabidopsis.⁸  What copying cannot supply is the founders, about 1,800 to 7,300 per genome, and the changes that gave each copy a job of its own.  Part 5 counts those changes where they have been measured.

Part 5: The worked cases, counted

The models in Part 1 are tested in a handful of cases where the duplicate, its parent and an outgroup are all sequenced and the proteins have been made and measured.  Each case is a reconstruction:  the ancestral sequence is inferred from living species, and the old protein is rebuilt in a laboratory and tested.  None was watched in nature.  The useful number for each is how many changes the copy needed to do its new job.

Leaf-monkey ribonuclease: three changes, tested

Colobine monkeys eat leaves and ferment them with bacteria in a foregut, as cows do.  They then digest the bacteria, including the bacteria's RNA.  Jianzhi Zhang, Ya-ping Zhang and Helene Rosenberg found in 2002 that the douc langur has two copies of the pancreatic ribonuclease gene where every other primate they sampled had one.⁴⁵  Reconstructing the ancestral sequence put every one of the 12 coding differences on the new copy's branch:  "the 12 nucleotide differences between the coding regions of douc langur RNASE1 and RNASE1B all occurred in the RNASE1B lineage."⁴⁵  Ten of the twelve change an amino acid.  "There are nine amino-acid substitutions in the mature peptide of RNASE1B, and seven of them involve charge changes."⁴⁵  All seven make the protein more negative.  The made protein works best at pH 6.3, against 7.4 for the parent, which matches the more acid small intestine of a leaf-eater, and "At pH 6.3, RNASE1B is about six times as active as RNASE1 in digesting RNA."⁴⁵  The copy also lost a job:  it no longer breaks down double-stranded RNA.⁴⁵

In 2006 Zhang found that the African guereza colobus has two extra copies of its own, which work best at pH 6.7.  Three of the replacements are the same in the Asian and African copies.  He rebuilt the ancestral African protein, then reversed just those three:  "Thus, the three backward mutations restored the original optimal pH of the enzyme, indicating that the three forward mutations were necessary for the decrease of the optimal pH from 7.4 to 6.7".¹¹  That is the cleanest count in the duplication literature:  three amino-acid changes were needed for the measured shift in function.  Zhang adds that other charge-changing replacements on the two lineages "may also contribute to the decrease of the optimal pH",¹¹ so three is the number shown to be necessary, not a full count of the changes involved.

Whether the Asian and African monkeys duplicated the gene separately is still argued.  John Schienman and colleagues found the gene "has duplicated more than once in both the African and Asian colobine monkeys", with gene conversion homogenising the copies' noncoding DNA.⁴⁶  Li Yu and colleagues, with ten more colobine species, concluded the opposite, "independent duplication hypothesis", and found "No obvious evidence of gene conversion".⁴⁷  A 2019 study of howler monkeys, which duplicated the same gene on their own and show replacements parallel to the colobines', still calls the colobine history "not yet been completely resolved".⁴⁸  The dispute moves the count of duplication events.  It does not touch the three tested replacements.

Red and green opsins: three changes, and in one lineage the variety came first

Humans and other Old World primates see red and green with two X-linked opsin genes that sit side by side.  Kanwaljit Dulai and colleagues traced the join between them in 1999 and placed "the duplication event at the base of the catarrhine lineage ∼40 mya".⁴⁹  Three amino-acid changes separate the two pigments' colour:  "Three amino acid changes (S180A/Y277F/T285A) in the red pigment together shift the λmax from 563 to 530 nm, explaining fully the λmax of the green pigment (531 nm)."¹²  Shozo Yokoyama's group adds a caution of its own:  the mutagenesis results "do not necessarily reflect the true effect of S180A/Y277F/T285A on the λmax shift that actually occurred in nature."¹²  Their answer was to make the changes in the rebuilt ancestral pigment instead.  Put in one at a time, each of the three moves the colour part of the way, and every intermediate pigment works (Part 6).¹²

The opsins also show that "copy, then change" is not the only order.  Most New World monkeys have one X-linked opsin gene with several versions in the population, so only females carrying two versions see in three colours.  Howler monkeys made a second copy on their own, and Gerald Jacobs reads their gene arrangement as showing that "at least in this one case M/L pigment polymorphism was preliminary to routine trichromacy".⁵⁰  There the differing versions existed first, as variants of one gene, and the duplication put two of them on the same chromosome.  For the Old World line Dulai's group says the reverse:  the spectral shift "occurred subsequent to duplication and did not draw on different allelic forms".⁴⁹

Haemoglobin: two changes for a new structure, fourteen for the modern one

Haemoglobin is two α and two β chains, coded by genes that came from one ancestor.  Michael Behe wrote in 2019 that whether the α and β chains could have arisen "by a Darwinian process ... is now very much an open question."¹⁸  A year later Arvind Pillai, Joseph Thornton and colleagues rebuilt the ancestral globins.  The protein before the duplication was a pair, not a foursome, and held oxygen tightly.  After the duplication "On the Ancα branch, there were only 3 changes, of which none were at IF2.  On the Ancβ branch, there were 42 changes, including 5 at IF2 and 4 others at IF1".¹³  Putting just two of the β-branch changes into the ancestor made it assemble into foursomes:  "Reintroducing just two post-duplication historical substitutions into the ancestral protein is sufficient to cause strong tetramerization".¹³  Those two also lowered oxygen affinity and gave "weak cooperativity".  They were not enough for the modern α2β2 arrangement:  "The IF2 substitutions are not sufficient to yield specific occupancy of the α 2 β 2 architecture".¹³  That took all nine interface changes plus five more for solubility, fourteen in all, out of 45 on the two branches.

Behe's question was answered for the structure and only in part for the function.  The modern molecule's full cooperativity and its regulation were not rebuilt from the two changes, and the globin fold itself is older than the duplication.

Yeast GAL1 and GAL3: a split job, not a new one

Baker's yeast has two related genes in the galactose switch:  GAL1 makes the enzyme, and GAL3 senses galactose and turns the pathway on.  Its relative Kluyveromyces lactis, which did not go through the yeast whole-genome duplication, has one gene that does both.  Chris Todd Hittinger and Sean Carroll tested the two with precise gene swaps and a competition assay.⁵¹  GAL3 cannot act as an enzyme because it lacks two amino acids:  "Addition of a Ser-Ala (SA) dipeptide to the degenerated active site of Gal3 has been shown to be necessary and sufficient to confer galactokinase activity".⁵¹  Otherwise, "aside from the complete degeneration of the galactokinase activity of Gal3, most of the functional divergence between these paralogues has been regulatory."⁵¹  The promoter of GAL1 evolved a binding-site arrangement that would have been bad for the single ancestral gene.  This is the best case for escape from adaptive conflict.  Nothing new was made:  both jobs already existed in one protein, one copy lost the enzyme activity, and the other copy's switch was retuned.

Bacteria in the laboratory: the one case watched

In 2012 Joakim Näsvall, Lei Sun, John Roth and Dan Andersson watched a weak side activity become two specialised genes in Salmonella.  The starting gene carried two mutations that gave it a weak second activity, and the experimenters placed it where "Duplications and amplifications of this region are frequent and have low fitness cost".¹⁴  Copies then multiplied and diverged:  "This rapid process, completed in fewer than 3000 generations, shows the efficacy of the IAD model".¹⁴  Each copy that specialised in the new job differed from the starting gene by one to three amino-acid changes.¹⁴  The design side does not count it as a new function.  Reeves, Gauger and Axe call it "a genuine change of substrate" but "not a genuine conversion of function, according to our definition, because the TrpF and HisA functions are represented by the same generalized chemical reaction."⁵²  That definition is stricter than the one Behe used in 2010, under which the divergence of a duplicated gene's activity counts as a gain.⁵³  The experiment, its set-up and its limits are covered in Directed Evolution.

Antifreeze, Hox clusters and the rest

The Antarctic fish antifreeze glycoprotein came from part of a duplicated trypsinogen gene, and the eelpout type III antifreeze from a duplicated sialic acid synthase gene whose ice-binding was already "rudimentary".⁵⁴  Both are counted in Antifreeze Proteins.  The Antarctic glycoprotein gene took about four or five mutation events, then copying of a nine-letter unit;  the six or seven events counted there for the northern cod gene belong to a gene made from non-coding DNA, not from a copy.

The Hox genes, which lay out the body plan, sit in one cluster in the lancelet and in four in mammals, and the four are taken as products of the two vertebrate genome duplications in Part 3.  No step in their divergence has been rebuilt and tested.  They show that the copies exist and where they sit.  They do not show how many changes each copy needed.

The cases side by side

In every case but the antifreeze, the copy started from a protein that already folded and already did something close to the new job;  the antifreeze is a short repeat that needs no fold.  The changes that made the difference are two to fourteen amino acids.  Under the bookkeeping of Measuring Biological Information those changes are the whole of the new sequence:  a few dozen nucleotides at most, laid on hundreds that were copied.

Part 6: The design and creationist side, tested

The writers below are stated first in their own words, then checked against the sources they cite and against the measurements in Parts 2 to 5.

Behe and Snoke: a copy that needs several changes at once

Behe and the physicist David Snoke built a model in 2004 of a duplicated gene that must acquire a feature needing two or more amino acids, such as a disulfide bond, where the intermediate steps do not help and most other changes would break the copy.  They concluded "that, in general, to be fixed in 10⁸ generations, the production of novel protein features that require the participation of two or more amino acid residues simply by multiple point mutations in duplicated genes would entail population sizes of no less than 10⁹."¹⁵  They stated the model's limits themselves:  "results bearing on the efficiency of this one pathway as a conduit for Darwinian evolution say little or nothing about the efficiency of other possible pathways."¹⁵

Michael Lynch answered in 2005 that the model "is non-Darwinian", because its intermediate steps are useless, and that with neutral intermediates, more than two usable sites and the protection a linked pair of copies enjoys, "Numerous simple pathways exist by which adaptive multi-residue functions can evolve on time scales of a million years (or much less) in populations of only moderate size."¹⁶  Behe and Snoke replied that every one of Lynch's objections came from his assuming the intermediates neutral where they had assumed them harmful.¹⁷  They also said plainly what their result does and does not show.  Their model "assumes intermediate states to be deleterious" and so "isn't a plausible evolutionary pathway";  what follows is "a strong requirement—that not only the end products, but steps along the way to a multi-residue function, must be either selected or at least neutral."¹⁷  The waiting times on both sides, and the later exchange with Rick Durrett and Deena Schmidt, are scored in Haldane's Dilemma.

The measured cases in Part 5 bear on this directly, and they do not all bear the same way.  The leaf-monkey enzyme needed three changes, the opsins three, the haemoglobin tetramer two.  The question that decides between Behe and Snoke's model and Lynch's is whether each single step helped, did nothing, or hurt.

• The opsins meet Behe and Snoke's condition. Yokoyama's group put each of the three changes alone, and in pairs, into the rebuilt ancestral pigment and measured every one. Each single change made a working pigment shifted part of the way toward green: 5, 10 and 17 nanometres, and the pairs and the triple come within 3 nanometres of the sum of the singles.¹² So the colour shift is not a feature that needs several residues at once. Every intermediate works, which is exactly the "selected or at least neutral" path that Behe and Snoke's reply allows. This fits Behe's own 2019 view that new opsins "might fit in by chance plus selection".¹ What was measured is colour, not fitness.
• The monkey enzyme and haemoglobin are open. Zhang reversed the three monkey changes together, not one at a time.¹¹ No single-change result for the two haemoglobin interface changes was found in Pillai's paper. For these two cases the path is not known.
• The laboratory case was selected at every step. In Näsvall's experiment each copy that specialised in the new job differed from the starting gene by one to three amino-acid changes, and every evolved gene tested alone made the bacteria grow faster.¹⁴

So the reconstructions show that small multi-residue changes happened on copies.  In the one case where every intermediate was measured, the opsins, the path was gradual.  In the others the path is not shown.

Behe: duplication grants a spare part, not a new machine

In The Edge of Evolution (2007) Behe accepted the rate:  "Gene duplications also seem to occur at about the same frequency" as point mutations, so in a large population "each gene is also duplicated in someone, somewhere."²¹  He accepted the yeast genome doubling as evidence of common descent.²¹  His objection was to what a copy can become:  "although gene duplication can help in limited circumstances, like Darwinian processes in general it doesn’t take us very far."²¹

In Darwin Devolves (2019) he generalised from Joseph Thornton's work on steroid receptors.  That work, he wrote, "calls into grave question whether the others, even the most familiar, could have developed that way either.  As just one example, the alpha and beta chains of hemoglobin are universally thought to have arisen in the distant past from a myoglobin-like precursor by gene duplication and divergence."¹⁸  Whether that "could have happened by a Darwinian process ... is now very much an open question."¹⁸  He also kept a door open:  "Some may have been, but many others not."¹⁸

Three checks:

1. Behe's own definition counts these cases as gains. In 2010 he defined a gain of a functional coded element to include "the divergence by mutation of the activity of a previously duplicated coded element."⁵³ The leaf-monkey enzyme, the green opsin and the haemoglobin chains are each the divergence of a duplicated element's activity. By his test they are gains. What he disputes is that unguided processes made them.
2. The opsins are the kind of case he concedes. Answering reviewers in 2019, he wrote that for opsins and odour receptors "My guess is that new odorant receptors and opsins might fit in by chance plus selection."¹ Three changes for a 30-nanometre colour shift fits that concession.
3. The haemoglobin question was half answered within a year. Pillai and colleagues found that two changes in a rebuilt ancestral globin give a four-chain molecule with weak cooperativity, and fourteen give the modern arrangement.¹³ No reply from Behe to that paper was found for this article. It does not show that every step was useful, or rebuild the full regulation of modern haemoglobin.

One quotation in Darwin Devolves is shortened in a way that matters.  Behe writes that Thornton's team "predict that further work 'will support a molecular version of Dollo's Law.'  That is, 'as evolution proceeds, shifts in protein structure-function relations become increasingly difficult to reverse.'"¹⁸  The sentence in the paper continues:  "whenever those shifts have complex architectures, such as requiring conformational changes and multiple epistatically interacting substitutions."⁵⁵  The condition limits the prediction to some proteins.  Behe calls the authors "confident that the results will be quite general".

Gauger and Axe: converting one enzyme into its neighbour

Ann Gauger and Douglas Axe of the Biologic Institute tried in 2011 to turn the E. coli enzyme Kbl2 into its structural relative BioF2.  They chose the closest pair in the family on purpose, "a case where the necessary structural change is known to be small relative to the changes commonly attributed to paralogous divergence."⁵⁶  Even this pair is only "34% identical over 381 aligned positions".⁵⁶  They concluded that "successful functional conversion would in this case require seven or more nucleotide substitutions."⁵⁶  In 2014 Mariclair Reeves, Gauger and Axe screened single mutants of nine genes in the family, and 70% of all possible double mutants of two of them, and found no BioF2 activity in living cells.⁵²  One of the two, BIKB, had been reported to do both jobs in the test tube, so it is the kind of two-job starting enzyme the mainstream route proposes.  Their best case, two changes for a weak activity and a third to make more of the enzyme, "would require about 10¹⁵ years in a natural population".⁵²  They read the family's resemblance as design:  "structural similarities among enzymes with distinct functions are better interpreted as supporting shared design principles than shared evolutionary histories."⁵²

The first objection came from the young-earth side.  Todd Wood, a creationist biologist, reviewed the 2011 paper on his blog:  "Instead of ancestral reconstruction, Gauger and Axe focused directly on converting an existing enzyme into another existing enzyme.  That left me scratching my head, since no evolutionary biologist would propose that an extant enzyme evolved directly into another extant enzyme."⁵⁷  He would have rebuilt the likely ancestor first, because that "would provide me with the 'best case'".⁵⁷  Reeves, Gauger and Axe answered in 2014 that the modern enzymes "are both the things to be explained and the things we can observe", and that they had asked the question in the present tense on purpose:  "We are asking whether the enzymes now present in life are as amenable to functional conversion as ancient enzymes must have been in order for evolution to work."⁵²

Each side has a real point.  The design side is right that modern enzymes of one family resist conversion into each other, and the critics agree:  the cell biologist Martin Poenie, answering Axe, writes that "Approaches similar to what Axe have done have failed repeatedly."⁵⁸  It is also right that the successful cases in Part 5 run through ancestors inferred from living species, not observed ones.  The design answer is weak in one place.  The evolutionary account does not predict that two modern enzymes 66% apart should convert easily.  The steroid-receptor work quoted above finds that changes in a protein's job "become increasingly difficult to reverse",⁵⁵ and turning Kbl2 into BioF2 means undoing one lineage's history before redoing the other's.  So the failed swap is what both views expect, and it cannot decide between them.  The experiment that could, rebuilding the Kbl-BioF ancestor and testing it, is the one Wood asked for, and nobody has published it.  The experiments, and their place among the other laboratory searches, are covered in Directed Evolution.  For this article the point is narrower.  The successful cases in Part 5 all went through a rebuilt ancestor or a copy of one gene, not across from one modern enzyme to another, so the Kbl2 result and the monkey result do not contradict each other.

Gauger and colleagues made a second point with a measured result in 2010.  A copy that must change before it helps carries a cost first, and mutations that cut the cost can win and block the path:  "if a duplicate gene requires modification in order to provide the new function, the cost will be present before the functional benefit emerges.  Since early reductive mutations also confer a benefit in such cases, they may sweep the population".⁵⁹  That is the same objection as Behe and Snoke's, measured in one bacterial construct.  It does not apply to copies that are useful from the start, by dose, as in Näsvall's experiment and the amylase genes.

Meyer: the copy already carries the information

Stephen Meyer's objection is the cleanest, and this article agrees with it as far as it goes.  "Gene duplication, as the name implies, involves the production of a duplicate copy of a preexisting gene, already rich in functionally specified information."²  Duplication and the rest are "'causes now in operation.'  No one disputes that.  But have these processes demonstrated the capacity to produce the effect in question".²  Nicholas Matzke replied that duplications are observed and ancestral genes are reasonably inferred.⁶⁰  That answers whether copying happens.  It does not answer where the first copy's information came from, which is Meyer's question.  The family count in Part 4 puts a number on the gap:  thousands of founding families per genome that copying does not explain.

Young-earth writers

• Lightner accepts duplication within the creation model. "It is important to understand the role of gene duplication within the creation model."²² Of the subfunctionalization model she writes, "This certainly fits within the creation model".²² Of the leaf-monkey enzyme she writes that "it appears highly unlikely that natural selection can account for this non-random pattern in monkeys", and proposes "impressive genomic programming".²² Zhang agrees the pattern is not random; that was his point, and he attributes it to selection in a more acid gut.¹¹ Neither side offers a measurement that would tell selection from programming here.
• Liu and Moran argue that duplicates "are usually silenced", that the regulation of gene families "is irreducibly complex", and that Lynch and Conery, "Realizing the impossibility of neofunctionalization", fell back on a passive role for duplicates.¹⁹ Lynch and Conery's abstract says that duplicates "may only rarely evolve new functions" and that "the few survivors" are then held by strong purifying selection.⁶ "Only rarely" is not impossible.
• Carter writes that those who cite duplication "never tell us the rate of duplication necessary" and that new functions by duplication "are not impossible, but they are vanishingly unlikely".²⁰ Lynch and Conery give the rate, and so does Behe, whose book Carter cites for the same argument.²¹ Carter gives no number for "vanishingly".
• Purdom says of the proposal that a duplicated photosystem became a second one, "This will only lead to the loss of information."⁶¹ The measured cases in Part 5 are counterexamples by Behe's own definition of a gain.
• Jerry Bergman accepts that "polyploidy is a mechanism that produces variety within created kinds".⁶² Nathaniel Jeanson's Replacing Darwin does not treat gene duplication as a source of genes; the one passage found on DNA duplication asks whether the large genomes of amphibians and lungfish are "the result of accidental DNA duplication".⁶³

None of the young-earth sources read for this article gives a duplication rate, a count of retained copies expected on its own timeline, or an account of the yeast genome doubling.

Part 7: The evolutionary side, tested

"Duplication explains gene families" is several claims of different strength.  Sorted by how they are known:

The field states several of its conclusions more strongly than its evidence.  Ohno called duplication "the major force of evolution" in 1970;²⁵ that was a judgement before any genome was read.  Force and colleagues wrote that partitioning old jobs is "the usual mechanism" of keeping copies;²⁶ Assis and Bachtrog, and Lien's salmon team, later found the reverse by their methods, and none of the three counted fates directly.  He and Zhang wrote that "Our results demonstrate that enormous numbers of new functions have originated via gene duplication";³³ what they measured was divergence in protein interactions and expression, which shows that copies came to differ, not that each difference is a new function.  On the other side of the ledger, Matzke wrote that even Behe accepts new genes "which cause death if deleted",⁶⁰ and he was right.  Writing in 2013 about Umbrea, a fruit-fly gene copied from another and now essential, Behe said "I have no reason to doubt this scenario" and called it "a gain-of-function mutation in my view", adding that "such gains of function do happen sporadically by Darwinian means".⁶⁴

Where the evidence is strongest, it supports something narrower than the textbook line.  The reconstructed cases are re-tunings and splits of old jobs, made by two to fourteen changes on a copy that already folded and already worked.  That fits Behe's 2019 concession for opsins and odour receptors as well as it fits the textbook.  Nothing in this literature shows a copy reaching a new fold, or a family arising from no family.

Part 8: What each history needs

The mainstream history has more than enough copies.  On Lynch and Conery's rate a 20,000-gene genome gains about 200 lasting duplicates per million years;  over 500 million years that is about 100,000 duplications kept long enough to count, most later lost.²³  The constraint is founders and the divergence of copies into new jobs, and the evidence on divergence is the handful of tested cases in Part 5.

A young-earth history has almost no time for copies to accumulate.  At the same rate the human genome would gain between 0.2 and 2.5 kept duplicates in 4,500 to 6,000 years, and with Gao and Innan's lower yeast rate far fewer.²³  Against that, 13,000 to 15,600 human proteins belong to families with another member.  So a young-earth model must hold that gene families were created as families, which is the common-design reading, and that duplication since then has added only copy-number variation like the amylase genes.  Lightner's position, duplication as a within-kind mechanism, fits this.  Nobody on that side has published the count.

The yeast genome doubling is the case such a model has not addressed.  Its signature, each region of a single-copy relative matched by two regions of baker's yeast with the losses scattered between them, is what a doubling followed by loss leaves, and it was predicted before the relative was sequenced.  If a young-earth reader placed the doubling within a created yeast kind after the Flood, the losses of about 88% of the extra copies would have to fit in a few thousand years.  For scale only, since no young-earth source makes this claim:  at the measured loss rate of 2.1 × 10⁻⁶ per gene per division, a lineage dividing once a day collects about three deletions per gene copy in 4,500 years, so the supply of deletions would not be the limit.⁶⁵  Whether the losses could spread through whole populations, and why related species lost different copies, has not been computed by anyone.

Objections and responses

"Duplication cannot add information.  A copy is a copy."  True of the copy, and this article counts it that way:  a duplicate adds nothing until it changes.  But copies do change, and the measured cases changed in ways that, by Behe's own 2010 definition, count as gains.⁵³  The defensible form of the objection is Meyer's:  copying does not explain the founding member of each family.  That form stands.

"Most duplicates are lost, so the route cannot work."  Most are lost, as Lynch and Conery found and the mutation-accumulation lines confirm.  The route needs only the few that stay, and on the mainstream rates there are tens to hundreds per million years per genome.  The objection is a statement about the mainstream timescale, not against it.

"The leaf-monkey and haemoglobin cases are just stories."  They are reconstructions, not observations, and this article says so.  But they are tested reconstructions:  the ancestral proteins were made and measured, and the key changes were put in or taken out to see what each did.  A test of that kind could have failed, and did not.¹¹

"Gauger and Axe showed enzymes cannot be converted."  They showed that Kbl2 could not be turned into BioF2 by the changes they tried and by any of most pairs of changes.⁵²  No one proposes that one living enzyme became another, as the young-earth biologist Todd Wood was first to point out.⁵⁷  Their answer, that modern enzymes are the things to be explained,⁵² is fair, but the mainstream account also expects two modern enzymes 66% apart to resist conversion.  The successful cases went through a rebuilt common ancestor or a fresh copy.  The result does show that some neighbouring functions are far apart in sequence, which is a real limit on how far the route reaches.

"Behe and Snoke proved multi-residue features need impossible populations."  They showed it for one pathway in which the intermediate steps are useless or harmful, and said so.¹⁵  The tested cases needed two or three changes.  In the opsins every single step has been measured and each one works, so Behe and Snoke's numbers do not apply there.¹²  In the monkey enzyme and haemoglobin the single steps are unmeasured, so neither Behe and Snoke's numbers nor Lynch's apply to them with confidence.

"Whole-genome duplication proves nothing:  Behe says it did yeast no good."  Behe's remark that the doubling "seems not to have done a whole lot for its recipients" is a judgement, not a measurement.²¹  The doubling itself he accepts.  The galactose genes, one of the retained pairs, are a case where the two copies are measurably better than one.⁵¹

"The 60% to 85% figure assumes common descent."  It does not.  It counts shared Pfam domains within one genome.  How the families arose is the interpretation, and the common-design reading is stated in Part 6.

"Young-earth creationists have no answer to gene families."  They have one:  families were created.  It costs them the ability to use duplication as an explanation of family origins, and they have not yet addressed the yeast doubling.

Issues and unanswered questions

1. The fitness of single steps. In every reconstructed case the decisive question between Behe and Snoke's model and Lynch's is whether each intermediate helped, did nothing or hurt. For the opsins the single and double changes have been made in the ancestral pigment, and each is a working pigment with a colour between the two ends,¹² though fitness itself was not measured. Rebuilding the single and double mutants of the leaf-monkey enzyme or the haemoglobin interface and measuring them would settle it for those cases. Nobody has published it.
2. How often each fate occurs. Fly and salmon studies infer mostly neofunctionalization from expression, and Kellis's yeast pairs show one-sided change in 95% of accelerated cases, but no study counts fates by testing protein function across a genome.
3. The direct and indirect rates disagree by two to three orders of magnitude, and the indirect rates themselves differ a hundredfold between Lynch and Conery and Gao and Innan. How many new copies survive to fixation in a natural population has not been measured directly.
4. The worm rates disagree with each other by a factor of 8 to 85 between Lipinski 2011 and Konrad 2018.
5. Founders. The family count in Part 4 is a floor on founders from Pfam's coverage. Pfam misses distant relatives, so the true number of founders may be lower; the 9% to 26% of proteins with no domain may add to it. A count with a profile method tuned for remote homology would tighten it.
6. The yeast doubling on a short timeline. No young-earth source places it in time or explains its pattern of reciprocal losses. A population model of loss after a post-Flood doubling, using the measured loss rate, would say whether the scenario is possible.
7. The colobine history of independent versus shared duplication is still argued.⁴⁸ It moves the count of duplication events, not the tested changes.
8. Several design-side and young-earth citations could not be traced: Liu and Moran's "only about 5% of amino acid-changing mutations" figure and Bergman's unnamed "dead issue" evolutionist. Bozorgmehr's 2011 Complexity paper on duplication and information, Lightner's 2009 paper on Old World monkeys, and Innan and Kondrashov's full review were not obtained.

Verdict

Gene duplication is real, frequent and measured:  copies form at rates far above point mutation per site, and almost all are lost.  Whole-genome doubling is shown in yeast and well inferred in vertebrates and plants.  Between 60% and 85% of the genes in each genome counted have relatives in it, and once each family's founder is set aside copying can supply the template of 40% to 76%.  The measured cases of divergence show new jobs made by two to fourteen amino-acid changes on working copies.

That is also the limit of the route.  Duplication explains how families grow, not why there are thousands of them.  Its tested successes are re-tunings and splits of old functions;  none reaches a new fold or a new family.  The design side is right that the copy's information is inherited, and wrong where it says the measured divergences are impossible, unexplained by selection, or never given a rate.  The young-earth side can treat families as created and duplication as later variation, which fits the evidence on families but leaves the yeast doubling unexplained.  The question that would move the argument most is whether the single steps of the reconstructed cases helped or hurt.  For the opsins the single steps have been measured for colour and each one works, so that case is gradual.  For the leaf-monkey enzyme and haemoglobin it is measurable and has not been measured.  The design side's enzyme-conversion experiments press a real point, that modern enzymes resist conversion into their neighbours, but they cannot decide the question until someone tests the inferred ancestor instead.

Sources