Orphan Genes

Summary

An orphan gene is a gene that looks like it has no relatives. Search the databases for something similar in another species, find nothing, and the gene is called an orphan.

That is a statement about the search, not about the gene's history. Three very different things produce the same negative result:

1. The gene really is new — it arose from non-coding DNA in this lineage.
2. The gene is old, but has changed so much that the search software can no longer recognise its relatives.
3. It is not a gene at all — just a stretch of DNA that happens to look like one.

Most published orphan counts never test between these three. That is the core problem, and it is the reason the numbers keep falling every time somebody checks them carefully.

The main conclusions of this article:

1. Several published counts shrink under reanalysis. Claudio Casola re-audited 874 published mouse "de novo genes" in 2018. Of 381 retained candidates, 152 passed his checks: about 40% of that subset and 17% of the original 874. Of the 874, 491 had already been dropped from the standard mouse gene databases entirely.¹ In humans, Clamp and colleagues found experimental evidence of a protein for 12 out of 1,177 orphans.²
2. How carefully a study is done predicts how much of it survives. In Casola's re-audit, the one study that used synteny and enabler substitutions kept 70% of its genes. The two that used sequence-similarity searching alone kept 42% and 25%.¹
3. New reading frames are not scarce, so "too fast" arguments cannot start there. A calculation done for this article finds roughly 677,000 chance open reading frames of 100 codons or more in a human-sized genome. Finding an open reading frame is not the hard part.
4. The measured rate of de novo gene birth is slower than ordinary gene duplication. Wu and colleagues report 9.83–11.8 new genes per million years in humans, then convert it themselves to 0.00033–0.00039 per gene per million years and note this is slower than duplication.³ Casola puts mouse de novo birth at 5.4–9.1 times slower than duplication.¹ In yeast, Moyers and Zhang count 2 verified de novo births against about 144 functional duplications.⁴
5. The strongest argument that orphan counts are too high is made by mainstream genomicists, not by critics of evolution. Clamp's team rejected 1,177 human orphans partly because the implied birth rate was "wholly implausible."² Casola calls the same problem the "gene turnover paradox."¹
6. "Orphan numbers keep rising" is true only of the quantity that carries no information. The pooled total across all sequenced genomes rises because it is a running sum. The count in a single well-studied genome falls: of Clamp's 1,177 human orphans, only 308 were still annotated in GENCODE 12.⁵ And the paper most often cited for the rising count says in its own words that "improved taxonomic sampling is reducing the overall percentage of orphans."⁶
7. The best-documented case of a genuinely new gene is real, and also the least impressive one available. The antifreeze protein of northern cod has a life-or-death function, a purified protein, and a full ladder of outgroup species showing the ancestral DNA was non-coding.⁷ But the protein is a tandem repeat of three amino acids, and its coding stretch needed only one single-nucleotide change plus copying,⁷ though the whole gene also needed a promoter and a deletion, about four to seven mutation events in all (see Antifreeze Proteins).
8. The design-side arguments most often repeated are wrong in a specific, checkable way. They apply Douglas Axe's figure of 1 in 10⁷⁷ — an estimate for a 153-residue folded enzyme — to de novo genes whose median length is 76 amino acids and whose function is mostly unproven.⁸ That is not the same object.
9. One design-side argument is right, and mainstream sources gloss over it. Gene annotation is partly circular. Hahn and colleagues discarded 23,070 single-species gene families as "artifacts of the annotation process" — by definition, before testing.⁹
10. Correcting for missed relatives shrinks the counts but does not remove orphans, and a proven job is still rare. After the best synteny correction, divergence explains at most 45% of yeast, 20% of fruit-fly and 24% of primate lineage-restricted genes, and the authors write that the persistence of orphans "demands an explanation."¹⁰ Thousands of young human reading frames are now known to be translated, but a 2022 review that demanded firm evidence of both new origin and function found eight de novo genes that qualify.¹¹¹¹

The verdict. The claim that orphan genes appear too quickly to be explained by evolution does not survive. It fails at the supply step, because reading frames are abundant. It fails again at the rate step, because the measured rates are slower than a route nobody disputes. And it fails under the filter this article set out to apply: once orphans explained by loss or by divergence are excluded, what remains is tens to a few hundred candidates per lineage, arising more slowly than duplicates, and only a handful of those have a demonstrated function.

What survives is smaller and more interesting. The large orphan counts in popular circulation are inflated, the field has pruned its own gene catalogs partly by assuming the answer, and nobody has demonstrated function for the overwhelming majority of the genes at issue.  A real residue of lineage-restricted genes remains in every lineage studied after correction.  None of the de novo genes that pass strict tests is an enzyme or a molecular motor;  their known jobs are binding, blocking or regulating machinery that already exists.  Where enzymes and motors came from is a question the de novo cases do not reach.

What "orphan" actually means

The term dates to the 1990s, when the first genomes were sequenced and a large fraction of the predicted genes matched nothing known. They were called orphans, later taxonomically restricted genes (TRGs).

The method is almost always BLAST — software that searches for similar sequences. If BLAST returns nothing from other species, the gene is scored as lineage-specific.

The weakness is straightforward and everyone in the field acknowledges it. BLAST detects similarity, and similarity decays with time. A gene that evolves quickly, or is short, loses recognisable similarity faster. So "no relatives found" and "no relatives exist" are different claims, and the first does not establish the second.

Tomislav Domazet-Lošo and colleagues — who defend the method — concede the mechanism in their own introduction: "it was observed that BLAST has problems to detect remote homologues of short and fast-evolving sequences."¹²

The counts fall every time somebody checks

This is the most consistent finding in the whole literature, and it is not controversial.

Mouse: 874 candidates, 152 survivors

In 2018 Claudio Casola re-examined every mouse de novo gene reported by three earlier studies — 874 in total. He checked each against four independent criteria: does it have relatives elsewhere in the mouse genome, does it match genes in two or more non-rodent vertebrates, does the surrounding chromosome region line up with other mammals, and does it carry a known protein domain.

Two results stand out. First, 491 of the 874 were no longer annotated as protein-coding genes at all in GENCODE, RefSeq or the UCSC browser.¹ They had quietly dropped out of the databases in the intervening years. Second, after the annotation and coding-sequence filters, only 152 of 381 retained candidates passed his tests: about 40% of that subset and 17% of the original 874.¹

Casola's summary: 60% "share homology with genes from other vertebrates, originated through gene duplication, and/or share no synteny information with nonrodent mammals."¹

He also states plainly that his own 40% is an upper bound, because his homology criteria were stringent and he could not search for enabler substitutions in two of the three data sets.¹

Method quality predicts survival

The most useful number in Casola's paper is the breakdown by study. Murphy and McLysaght's 2012 set, which combined similarity searching with synteny and enabler substitutions, kept 7 of 10. Neme and Tautz 2013 kept 139 of 331 (42%). Wilson and colleagues 2017 kept 18 of 72 (25%). The latter two used similarity searching alone.¹

Careful studies survive re-checking. Quick ones do not. That pattern is the single best guide to which orphan numbers to trust.

A third of the recoverable relatives were invisible to BLAST

Casola found homologs for 98 of the candidates in two or more non-rodent vertebrates. 36 of those 98 were found only by phmmer, a profile method, and not by BLAST at all.¹

This matters in both directions. It is direct evidence that BLAST-based orphan counts contain false positives. It is also evidence that the fix is available and cheap — better search tools, not a new theory.

Mixing annotation methods multiplies the count

A second artifact sits upstream of the search.  Caroline Weisman, Andrew Murray and Sean Eddy (2022) found four groups of five species, cichlid fish, primates, bats and rodents, whose genomes had each been annotated by two different methods.  They counted lineage-specific genes once with one method for every species, and again with methods mixed as real studies mix them.  When the lineage was annotated one way and its outgroups another, the mixed count was typically about four times higher, ranging from 1.4 to 15 times, and in seven of eight cases more than double, "suggesting that the majority of lineage-specific genes inferred in heterogeneous annotations are artifacts of the heterogeneity."  Milder mixtures raised counts by up to 6 times.¹³  They also report that 18 of 25 lineage-specific studies in a prominent 2019 review, and 21 of 33 studies published from 2019 to 2022, relied on mixed annotations.¹³  This inflates counts made that way;  it does not touch counts made with one annotation method across all species, or by searching the raw genome in all six reading frames.

Human: 1,177 orphans, 12 with a protein

In 2007 Michele Clamp and colleagues at the Broad Institute went through the human gene catalog. After removing clear artifacts and correctable annotation errors, 1,177 candidates remained that "lack orthology, paralogy, or homology to known genes and are not obvious artifacts."²

They then searched the published literature for experimental evidence that any of these actually made a protein:

"we found articles reporting experimental evidence of an encoded protein in vivo for only 12 of 1,177 orphans, and some of these reports are equivocal"²

Twelve. About 0.06% of the gene catalog. Their conclusion cut the estimated human gene count from about 24,500 to about 20,500.²

Even the strict studies are funnels

Wu, Irwin and Zhang's 2011 study is the standard citation for human de novo genes. Their pipeline started with 584 human proteins that had no primate match, cut to 352 with proper start and stop codons, and ended with 60.³

They also excluded 69 genes because the ancestral state was not preserved in both chimpanzee and orangutan, and a further 56 for lack of peptide evidence.³ Those exclusions are methodologically correct. They also show how much the final number depends on choices made along the way.

A caution that applies to almost every popular treatment of this topic: Clamp's 1,177 and Wu's 60 are not the same quantity and must never be compared. Clamp counted non-conserved reading frames in the gene catalog. Wu counted reading frames complete in humans and demonstrably broken in two outgroups, with peptide evidence. One is a list of candidates; the other is a list of survivors.

Translated is common, a proven job is rare

Clamp's 12 of 1,177 dates from 2007, before ribosome profiling could show which short reading frames are actually read into protein.  The newer data move the picture in both directions.

Translation of young reading frames is now well documented.  Clara-L. Sandmann and colleagues (2023) traced 7,264 translated short reading frames from a GENCODE reference catalog.  They report that 89.6% are young, that 4,101 arose from DNA that was non-coding in the ancestor, and that 162 of those arose in the human lineage.¹⁴  So "no protein" is no longer a fair description of young human reading frames as a class.

A measured effect is much rarer, and mostly measured in a dish.  Nikolaos Vakirlis and colleagues (2022) traced 715 translated human reading frames and found 155 of de novo origin.  Disrupting the reading frame with CRISPR had a significant effect on the growth of cultured cells for 44 of the 155, the same rate as for reading frames of uncertain origin.  Nineteen of the 44 arose within the last 43 million years, and two after the human-chimpanzee split.¹⁵  Those are cell-line fitness effects, not organism-level functions.  The test that separates a protein's effect from the RNA's, restoring growth with a copy that has its start codon and failing without it, covers only two young de novo cases in their set.¹⁵  The authors state the limit themselves: "Additional, more targeted experimental work is now needed to conclusively demonstrate both the functional role of these elements in human and the absence thereof in other species."¹⁵

Strict tests of both origin and function leave a handful.  Caroline Weisman (2022) reviewed only genes with positive evidence of absence in at least two outgroups, not just a failed search, and a knockout or knockdown effect traceable to the gene's product.¹¹  Eight genes qualified.  Six code for protein:  the cod antifreeze, MDF1 and BSC4 in brewer's yeast, and PBOV1, NCYM and MYEOV in humans.  Two are RNA genes:  ELFN1-AS1 in humans and Poldi in mice.¹¹  All four human genes were studied mainly in cancer cells, which she suspects partly reflects which systems are easy to study.  She calls her criteria conservative and leaves out well-studied genes that "can only be called putatively de novo,"¹¹ so eight is a count of the best-documented cases, not of all real ones.

The honest tally is thousands of translated candidates, dozens with a measurable effect in cultured cells, and eight, in Weisman's strict 2022 review, with firm evidence of both new origin and a job.  That supports the design side's claim that function is unproven for almost all orphans.  It also means the design case cannot rest on a claim that no de novo gene works:  a few clearly do.

Primates: most orphans are recycled, not new

Marta Toll-Riera and colleagues took 270 human orphan genes in 2009 and asked where each came from. Their partition is the most complete answer available for primates, and it is not what either side usually reports:¹⁶

Over half of primate orphans are built partly from transposable elements — mobile DNA already present in the genome.¹⁶ That is a third mechanism, distinct from both de novo birth and gene duplication, and it is the largest single category. Only 15 genes of 270 had positive evidence of arising from non-coding DNA.

They also report that only 6 of the 270 had direct protein evidence.¹⁶ Independent group, stricter definition, same picture as Clamp's 12 of 1,177.

A figure to stop repeating. This paper is often summarised as finding that "around 60% appeared to be new," a claim that traces to press coverage rather than the paper. No such number appears in it. The partition is 24% / 53% / 23%, and the paper's own estimate of de novo origin is 5.5%.¹⁶

And one thing this paper must not be used for. Toll-Riera's 270 were deliberately selected to have chimpanzee and macaque homologs, to raise confidence they are real genes. It is a conservative subset, not a competing estimate of Clamp's 1,177, and the two cannot be set against each other.¹⁶

Reading frames are not the bottleneck

Any argument that new genes appear "too fast" has a step before the one usually argued. Before selection can do anything with a new gene, an open reading frame has to exist. If long stretches of DNA without stop codons are rare, then de novo gene birth is limited by supply, and a rate argument can be built on that scarcity.

They are not rare. A calculation done for this article settles it.¹⁷

In a human-sized genome of 3.1 billion base pairs, counting both strands, at the genome's average composition of 41% GC:

About 677,000 chance reading frames of protein-gene length. Two independent methods — a closed-form calculation from base composition, and a simulation over generated sequence — agree to three significant figures.

Six orders of magnitude more raw material than any published birth rate requires. Whatever is difficult about making a new gene, finding a reading frame is not it. An argument that de novo genes appear too quickly has to locate its improbability somewhere else — in function, not in supply.

Clamp's team said the same thing in 2007, in one sentence: "a random GC-rich sequence (50% GC) of 2 kb has a ≈50% chance of harboring an ORF ≈400 bases long."²

A caution about the GC argument

Clamp reports that the 1,177 human orphans have a GC content of 55%, against a genome average of 39%.² That gap looks striking, and it is sometimes quoted as though it settled something.

It does not, in either direction, and the reason is worth stating carefully because this article's own calculation cuts against the obvious skeptical reading as much as it cuts against Clamp's.

All three stop codons — TAA, TAG, TGA — are AT-rich. Raising GC content lowers the chance of hitting a stop, which makes long chance reading frames more common. At 55% GC, chance reading frames of 100 codons or more are 6.6 times as abundant as at genome-average composition, and at 200 codons, 68 times as abundant. So elevated GC is exactly what one would expect if the orphans were chance reading frames sitting in GC-rich regions.

But Clamp reports a third number that is usually dropped: real, conserved protein-coding genes are 53% GC.² The orphans' 55% is much closer to real genes than to the genome average, and Clamp's own explanation is that orphans tend to occur in gene-rich regions.

So the GC figure is predicted by both hypotheses and discriminates between neither. Clamp's reading-frame-conservation evidence does discriminate; the GC number does not. Quoting "55% versus 39%" without the 53% overstates the case.

The rate argument, and why it fails

The design side's own statement of the argument is Ann Gauger's, from 2013.  She lists five explanations for orphans, the fifth being "Perhaps we see so many species- and clade-specific orphan genes because they are uniquely designed for species- and clade-specific functions," and closes:

"If common descent is true, the apparent rate of generation of new proteins is astonishing by anyone's expectation. What now needs to be determined is whether or not naturalistic processes known to be operating are actually capable of generating so many new proteins."¹⁸

Her claim is about the raw orphan counts, and she leaves the question open.  So the test this article applies is her question asked of the cleaned-up numbers:  after excluding orphans explained by loss or divergence in other species, do the survivors appear too quickly for evolution to account for?

The answer is no, and the numbers come from the field itself.

Humans. Wu and colleagues report "the rate of origin of de novo genes is ∼9.83–11.8 genes per million years."³ That sounds fast. The arithmetic is simply 59 fixed genes divided by 6 and by 5 million years. The same authors immediately convert it:

"when the rate is expressed in terms of per gene, ∼0.00033–0.00039 per gene per million years, it is still a lower rate than the rate of new gene origin by gene duplication"³

Mouse. Casola's revised figure is 11.6 genes per million years, down from Neme and Tautz's 65. His comparison: de novo genes "emerged at a pace that is at least about 5.4–9.1 times slower compared with gene duplicates."¹

Yeast. Moyers and Zhang count two verified de novo births against roughly 144 functional duplications since the common ancestor of the sensu stricto yeasts — "gene duplication appears to surpass de novo gene birth by 2 orders of magnitude."⁴

In all three systems, de novo gene birth is slower than gene duplication. Duplication is a mechanism no one on any side of this debate disputes. A rate that is slower than an undisputed mechanism cannot be evidence that the mechanism is impossible.

The rate argument already exists, and mainstream scientists made it

The strongest version of "these numbers are too high" is not a creationist argument. It is Clamp's, in 2007:

"If the orphans represent valid human protein-coding genes, we would have to conclude that the vast majority of the orphans were born after the divergence from chimpanzee. Such a model would require a prodigious rate of gene birth in mammalian lineages and a ferocious rate of gene death erasing the huge number of genes born before the divergence from chimpanzee. We reject such a model as wholly implausible."²

Casola makes the same argument eleven years later and gives it a name — the "gene turnover paradox." Since 75% of mouse protein-coding genes have human homologs, gene content is broadly stable, so a high birth rate demands an equally high loss rate. His conclusion: "Either most de novo genes are rapidly lost after their origin or the rate of de novo gene birth is overestimated."¹

This is worth sitting with. The rate argument that a critic of evolution would want to make has already been made, twice, by mainstream genomicists. And both times the conclusion drawn was that the orphan counts are wrong, not that evolution is in trouble.

Loss versus gain: the objection that works

There is a much better objection than the rate argument, and it is this. When a de novo gene is claimed, the evidence is usually that a matching stretch of DNA exists in related species but is broken there — a premature stop codon, a missing start. The standard reading is that the sequence was non-coding and became a gene.

But the same evidence fits the reverse: the gene existed in the common ancestor and was disabled in the others. Losing a gene is easy. Building one is not.

This objection is not a fringe position. It is published, by the leading pro-de-novo group. Nikolaos Vakirlis, Anne-Ruxandra Carvunis and Aoife McLysaght describe exactly this mechanism:

"an event akin to incomplete pseudogenization could be taking place, wherein a gene loses functionality through some disruption, thus triggering rapid divergence due to absence of constraint. After a period of evolutionary 'free fall'... this would eventually lead to an entirely novel sequence."¹⁰

They also show that shortening is a signature of divergence, not of youth: genes that diverged beyond recognition "tend to be much shorter" than the relatives they came from.¹⁰ That directly undercuts the common inference that short genes are young genes.

How strong is the loss explanation?

Two large studies have tried to measure it, and they disagree in an instructive way.

Vakirlis and colleagues (2020) used synteny — the order of genes along a chromosome. If a gene sits in the same neighbourhood in two species, the two are probably related even when their sequences no longer match. Their estimate: divergence beyond recognition accounts for 20.6% of orphans on average, ranging from 0% to 57% depending on the species pair.¹⁰

Their answer moves with the method, which is a genuine soft spot. Tightening the synteny criterion drops the estimate to 10.9%; loosening it raises it to 23%.¹⁰

Caroline Weisman, Andrew Murray and Sean Eddy (2020) built a statistical model of how similarity decays, fitted it to each gene's known relatives, and asked how likely BLAST would have been to find a relative in the outgroup if one existed. Their result: 55% of yeast lineage-specific genes and 46% of Drosophila ones have less than a 5% chance of being detected even if relatives are there.¹⁹

But their paper must not be cited as saying orphans are artifacts, and the authors are explicit about why. "Explained by homology detection failure" means "not ruled out by it" — not "shown to have relatives." And they name the problem:

"it may be difficult for de novo genes in particular, which have been hypothesized to be short and fast-evolving, to reject the null model."¹⁹

A genuine de novo gene is precisely the kind their test cannot distinguish from a diverged one. So their 55% is not an upper bound on real novelty.

They also report a hard residual: in all six lineages studied, 10–20% of lineage-specific genes cannot be explained by detection failure.¹⁹ And here is the part that cuts against both sides — the residual mostly is not de novo either. Of 25 yeast genes rejecting their model, 12 turned out to have relatives outside the lineage found by synteny, and 8 showed evidence of duplication. Their preferred explanation for the rest is meiotic drive, not de novo birth.¹⁹

What is left after the correction. Read the other way round, the same two studies say how many lineage-restricted genes survive the best available check for missed relatives:

Vakirlis, Carvunis and McLysaght put the conclusion in their title, that divergence "is not the main source of orphan genes," and in their discussion: "The persistent presence of orphans and TRGs in almost every genome studied to date despite the growing number of available sequence databases demands an explanation."¹⁰ This is the strongest secular support for the design side's starting point.  Orphans are not just a search artifact that better software will make disappear.  A residue survives the best correction in every lineage studied.

What the residue does not show is how it arose.  Vakirlis's group reads it as de novo birth and gene transfer, by subtraction:  whatever divergence cannot explain is assigned to the other mechanisms, without tracing any single gene.¹⁰  Weisman's firm residue is mostly not de novo either, as shown above.  A designer, de novo birth, gene transfer, and divergence too fast for synteny to catch all remain on the table for the leftover genes, and only gene-by-gene evidence of the kind described below can sort them.

Where this leaves the loss objection. It is real, it is published, and it explains a substantial minority of orphans — somewhere between about 11% and 55%, depending on method. It does not explain all of them, and no serious participant claims it does. It is a strong objection to inflated counts and a weak objection to de novo gene birth as such.  It is also a weak objection to the existence of orphans:  between about a tenth and four-fifths of them survive it, depending on lineage, method, and how firm a residue is demanded.

What the best-documented case looks like

The strongest single case in the literature is the antifreeze glycoprotein of northern cod, worked out by Xuan Zhuang and colleagues in 2019. It deserves to be taken seriously, and then looked at closely.

What is strong about it. Seven gadid species were sequenced, chosen for their positions on the tree: three with the antifreeze gene, four without, at three different depths.⁷ The ancestral non-coding state was confirmed by experiment, not just by computer — Northern blots on pancreatic RNA found no transcripts in the species lacking the gene.⁷ The enabling mutation is identified: a single-nucleotide deletion causing a frameshift that joined a signal peptide to the repeat region. The undeleted ancestral state is present in two independent taxa.⁷ And the function is not inferred — the mature glycoprotein was purified from fish blood decades earlier, and it prevents the fish freezing to death.

That is a complete case. Almost nothing else in the literature has all of those parts.

What is weak about it, and this is not a small point. The protein is a tandem repeat of the tripeptide Thr-Ala-Ala, repeated anywhere from 22 to more than 550 times. The smallest working version has four repeats.⁷ The authors' own account of its coding sequence is that a single nucleotide change in an existing alanine-rich repeat created the founding codons.⁷

This is the lowest-information protein sequence one could design. Its coding stretch needed one substitution plus the copying of a short repeat — a process that happens constantly — and the whole gene also needed a promoter and a deletion, about four to seven events in all. So the best-documented case of a new gene is also, by some distance, the least demanding one.

Anyone arguing that this case shows the power of evolution to build complex new proteins is overreading it. Anyone arguing that it is improbable is describing a different molecule.

The two strongest kinds of evidence never appear together

There are two ways to show that a stretch of DNA was non-coding in the ancestor. One is conserved synteny — the neighbourhood matches across species, so the same location can be compared directly. The other is a shared disabling state — the same break, at the same position, in several independent outgroups, which makes repeated parallel loss implausible.

Across the cases examined for this article, no single case has both.

The cod antifreeze gene has the shared disabling state, confirmed in two independent taxa and backed by experiment. But it has no conserved synteny at all — the authors report "a complete lack of microsynteny" and have to invoke an unobserved translocation to explain how the gene reached its promoter.⁷ Conversely, the synteny-based studies that do compare matched locations, such as Vakirlis and colleagues, do not identify a single de novo gene; their worked examples are cases of divergence.¹⁰

This is not fatal to de novo gene birth. It does mean the evidence is thinner than the volume of literature suggests, and it explains why careful studies keep producing smaller numbers than quick ones.

The methods dispute is real, and it is not settled

Behind all of this is an unresolved argument in the mainstream literature, and any honest treatment has to report it as unresolved.

Moyers and Zhang (2015) simulated protein evolution where every gene was the same age by construction, then ran the standard method on the simulated data. It reported 13.85% of genes as younger than they were, and 2.77% as having no detectable relatives anywhere — which is what an orphan looks like.²⁰ They concluded that trends attributed to gene age might be artifacts.

Domazet-Lošo and colleagues (2017) replied by granting the error rate and arguing it is too small to matter. In real yeast data, 40% of genes lack a detectable relative in the distant species S. pombe, against the 11% attributable to error. In the three youngest age classes, 445 real cases against 14 simulated — "over 30 times more."¹² They also deleted every gene Moyers and Zhang had flagged and showed the published patterns survived.¹²

They landed a sharper blow too. Moyers and Zhang's simulation was seeded with real sequences and their real evolutionary rates, then paired with the real values of properties like expression level. So the genes that are short and fast-evolving in reality are the same ones misdated in the simulation — 76% of fly and 88% of yeast sequences called young in the simulation were also called young in the real data.¹²

Two things must be reported alongside this.

First, Moyers and Zhang published an erratum retracting one of their three headline results. A coding error invalidated their claim about germ-layer expression: "No deviation in any age group for any expression site is statistically significant."²¹ Domazet-Lošo's group caught it. Any article citing that particular result is citing a withdrawn one.

Second, the exchange does not end there. Moyers and Zhang replied again in 2017, reconfirming their error estimate using simulations that Domazet-Lošo's group had themselves suggested, and arguing that a trend claimed to be robust disappears when only error-resistant genes are analysed.²² The dispute is open.

What everyone agrees on

The disagreement is narrower than the titles suggest. All parties accept:

• BLAST misses remote relatives of short, fast-evolving proteins.
• De novo gene birth is real and happens. Moyers and Zhang: "it is clear that de novo gene birth plays some role in the formation of new genes in yeast."⁴
• Sequence-similarity searching alone cannot establish a de novo gene. Domazet-Lošo's group: the method "can only identify 'putative' de novo genes," and an "unequivocal demonstration of de novo evolution requires also invoking synteny information."¹²
• Synteny plus profile-based search is the fix.

Where the design argument goes wrong

Several arguments in the intelligent-design literature about orphan genes contain a specific, checkable error, and this article names it rather than repeating it.

The rarity error. Douglas Axe's 2004 estimate that roughly 1 in 10⁷⁷ sequences form a working fold is for a 153-residue folded enzyme domain. Ann Gauger applies it to orphan genes directly: "only 1 in 10⁷⁷ proteins form a fold with a target function... then getting an orphan gene by an evolutionary process is extremely unlikely."⁸ Jonathan Witt makes a similar move.²³

The objects do not match. The de novo genes with actual protein evidence in Ruiz-Orera's study have a median length of 76 amino acids,²⁴ and for the great majority of claimed de novo genes no function has been demonstrated at all. Axe's number is a rarity estimate for a specific, tightly folded catalytic domain. Applying it to a short peptide of unproven function is not a conservative estimate. It is a different calculation about a different thing.

The point cuts hardest when the two claims are made in the same argument. If the orphan counts are inflated because they include unproven transcripts — which is true — then those unproven transcripts are not folded enzymes, and the folded-enzyme rarity figure does not apply to them.

The premise behind the rarity argument is shared, and a mainstream answer to it is on record.  Caroline Weisman, who led the 2020 homology study above, calls it "Jacob's conundrum," after François Jacob's 1977 remark that a functional protein appearing de novo is "practically zero" in probability.  She sets it out as an argument whose first premise is the design side's own:  "Premise 1 ("Sparsity"): A very small fraction of all possible sequences would produce a biological effect beneficial to the organism."¹¹  Her answer, drawn from the eight best-documented de novo genes, is that they do "freeloader" jobs.  They bind, block or switch on machinery that already exists, and "freeloading makes function vastly more common in sequence space than we have imagined, severely violating our sparsity premise."¹¹  The one de novo protein whose structure has been studied by experiment, yeast BSC4, has a water-avoiding core but "does not have a single, stable fold."¹¹

This cuts both ways.  It backs this article's point that Axe's enzyme figure is the wrong yardstick for these genes, and Weisman says so in so many words:  "striking, self-powered examples, like molecular motors and enzymes" are what springs to mind as function, "But these are exceptions, not the norm."¹¹  It also backs the design side's narrower point.  None of her eight genes is an enzyme or a motor, and her own summary is that "most of these genes do not do much that seems novel."¹¹  So the documented de novo cases say nothing about where self-powered folds such as enzymes came from.  That question belongs to the founding protein families, which How Rare Is a Working Protein? treats.  A skeptic's best reply is that one route need not explain everything:  enzymes are old, and duplication with divergence accounts for most of their later variety.  That reply moves the question to the founders.  It does not answer it.

A correction to a claim this article previously relied on, and which is often made on both sides. De novo proteins are frequently described as intrinsically disordered, and that description was used in earlier drafts here to sharpen the mismatch. It does not hold up. In yeast, Carvunis and colleagues and Vakirlis and colleagues both found young and emerging reading frames less disordered, not more.²⁵ In mouse, Casola showed the elevated disorder of de novo proteins tracks their high GC content rather than any adaptive property — GC and disorder correlate at r = 0.92 across all mouse genes.¹ The rarity objection stands on length and unproven function alone, and does not need the disorder claim. Anyone making it should drop that part.

A category slip. Gauger writes that "TRGs can be up to 10-20 percent of a taxonomic group's genome."²⁶ The published figures are percentages of genes, not of genomic DNA. Those are very different quantities.

A claim that collapses against its own source. Jeffrey Tomkins argues the human SRGAP2 genes "do not look duplicated at all."²⁷ But the paper he cites describes them as human-specific duplicates, identifies the duplicated unit as "the promoter and first nine exons," and says only SRGAP2C is likely to encode a functional protein. They are paralogs, not orphans, and the source says so.

One piece of credit that is owed. Paul Nelson's peer-reviewed chapter with Richard Buggs does not make these errors. It contains no design argument, no probability calculation, and no citation to Axe, Behe or Meyer. On the central question it says the opposite of the popular version: "We should expect that increased genomic sampling will provide matches (orthologs) for many TRGs."²⁸ Criticism of the popular arguments should not be aimed at that chapter.

Where the design argument is right

One criticism lands. It is not an ID invention, and its best source is a mainstream one.

Gene annotation is partly circular. Genes are often identified by looking for sequences conserved across species. A gene with no relatives fails that test by definition — so it can be excluded as an annotation error, and the exclusion then supports the belief that such genes are rare.

Aoife McLysaght — one of the leading researchers on de novo genes, writing in PLoS Genetics — states the problem plainly:

"A serious limitation in this approach is that it relies on existing gene lists that have been annotated using criteria that usually include the presence of a homolog in other genomes. Novel genes fail to meet this criterion by definition, thus they are usually not reliably annotated."²⁹

She then gives a concrete measure of how unstable these lists are. Of the de novo candidates in Wu's study, "over half... are not included in the more recent Ensembl gene lists they used (version 56), and by version 60 only six of these genes were still listed."²⁹

That is the same pattern Casola documented eleven years later, from inside the field, before anyone had made a controversy of it.

The concrete instance is checkable and was verified directly for this article. Hahn and colleagues, studying gene family evolution across 12 Drosophila genomes, wrote:

"we found 23,070 families that consisted of a single gene and that appeared to have evolved on a terminal lineage (i.e., they are found in only a single species). These single-gene families were regarded as artifacts of the annotation process, and were removed from further analysis."⁹

Their Table 1 carries a row headed "Annotation artifacts: 23,070." The category was defined as "families with one gene in one species" — that is, membership was decided by the very property under investigation.

In fairness to Hahn and colleagues, they went back and audited the 1,074 cases in D. melanogaster. Of those, 716 turned out to be RNA genes and 94 had no supporting evidence. But 15 did not fit the label, and they said so:

"Though we have called these genes annotation artifacts, it appears more likely that they are simply extremely rapidly evolving genes."⁹

That is honest reporting, and it should be credited. The underlying problem stands: 23,070 families were removed by definition, and only one species' worth was ever checked.

Clamp's paper carries a version of the same problem. Their reason for rejecting the orphans includes the judgement that the implied birth rate is "wholly implausible."² That is a prior belief about rates doing work in a decision about what counts as a gene. Nelson and Buggs put it accurately: "Our inability to explain the evolution of TRGs has been used as an argument to support the proposition that they are non-functional."²⁸

To their credit, Clamp's team said the pruned genes "can be restored to the catalog on a case-by-case basis," and 18 were restored in that paper.²

Do orphan numbers really keep rising?

The most repeated argument for taking orphans at face value is that their numbers have not levelled off. If they were a search artifact, the reasoning goes, better databases should have made them disappear. They have not.

Three different quantities are being run together here, and separating them settles the question.

(a) The total pooled across all sequenced genomes. This rises, always. It is a running sum of a quantity that cannot be negative — sequence another genome, add its lineage-specific genes to the pile. It would rise even if every orphan were an artifact. This is the quantity almost every citation actually refers to, and it carries no information about biology.

(b) The fraction of genes that are orphans in a newly sequenced genome. This falls. Wilson and colleagues, whose 2005 bacterial survey is the standard citation for the rising-count claim, report the series themselves: 100% at one genome, 30% at ten, 15% at 55, 12% at 122, and a projected 10% at 200.⁶ Their own summary sentence is the one that never gets quoted alongside the rest:

"although improved taxonomic sampling is reducing the overall percentage of orphans, it cannot be used to assign all orphans to known gene families"⁶

Their claim is the narrow one — denser sampling will not drive the percentage to zero. Not that it is rising.

Why the pooled curve keeps climbing is visible in their own data. Pirellula was the 100th bacterial genome sequenced and still returned 3,576 orphans, 49% of its genes — because it was the first member of its division to be done. Leptospira interrogans was the 92nd genome but only the third spirochaete, and returned 45%.⁶ Orphan fraction tracks phylogenetic isolation from what has already been sequenced, not the date. Wilson and colleagues attribute the missing plateau to undersampling in their own conclusion: it "reflects both the small proportion of the total bacterial diversity sampled to date and the widespread occurrence of orphans in almost all bacterial taxa."⁶

So the pooled curve is a graph about sampling design. Reading it as a graph about gene birth is the error.

(c) The orphan count in one well-studied genome, tracked over time. This is the decisive test, and the human genome supplies it directly.

Ezkurdia and colleagues re-counted Clamp's orphans using GENCODE 12:

"In 2007, Clamp (5) identified 1177 genes as 'orphans' with features typical of non-coding RNA... The GENCODE 12 annotation still annotates 308 of these orphans, and there were 248 orphans among the potential non-coding genes. We detected peptides for just 3 of these 248 genes."⁵

869 of the 1,177, or 74%, were absent from the GENCODE 12 protein-coding annotation.  That measures a change in the catalog; it does not by itself establish why each entry was removed or prove that every removed candidate is non-coding.

For the best-annotated genome on Earth, the number of Clamp's orphans still in the catalog fell by three-quarters.  That fall is a decision by annotators, not an independent test, and part of the case for pruning was Clamp's own judgement that the implied birth rate was "wholly implausible," the circularity described under "Where the design argument is right."  The independent part of Ezkurdia's check is the peptide search:  3 of 248.

Clamp's team had already seen the same effect across species in 2007. Mouse was then credited with at least 33,000 genes and dog with about 19,000 — a 14,000-gene gap between two mammals. Their explanation was not biology but sequencing effort: the difference "is attributable to nonconserved ORFs identified in cDNA sequencing projects," and it was "currently unclear whether it reflects meaningful evolutionary differences among species or simply varying numbers of spurious ORFs in species with more cDNAs in current databases."²

A caution that cuts against over-using this. Ezkurdia's team state the limit themselves: "Not detecting peptides does not prove that the corresponding gene is non-coding because it may be a consequence of the protein being expressed in few tissues, having very low abundance, or being degraded quickly."⁵ Absence of evidence again.

And a caution against building a tidy time series. Orphan percentages from different papers are not comparable. Domazet-Lošo's 26–29% for Drosophila is measured against non-insects; Palmieri's 7% is measured against ten other Drosophila species; Weisman's 1% for human is measured against other primates. The falling numbers across the literature look like a trend and are mostly a change of yardstick.³⁰

Objections and responses

"Orphan numbers keep growing as more genomes are sequenced. Doesn't that show they are real?"

This is the most common argument in the popular literature and it deserves a full answer, which is in the next section. In short: three different quantities are being run together, and the one that rises is the one that carries no information.

"But some orphan genes are essential. Doesn't that prove they are real genes?"

The study everyone cites for this does not say what it is usually taken to say, and the point is worth getting right.

Chen, Zhang and Long knocked down 195 young Drosophila genes with RNAi and found 30% essential for viability — statistically the same rate as old genes.³¹ That is a real and striking result. But 56 of the 59 essential genes arose by duplication. Only three arose de novo.³¹

So the paper is evidence that duplicate genes quickly become essential. Its de novo sample is three. Citing "30% of new genes become essential" in an argument about de novo origination misrepresents it, and the authors themselves draw the distinction explicitly.

There is a second problem even where genes are essential. Essentiality can be acquired by a gene inserting itself into an existing network. That shows it is now required, not that it was independently useful when it arose.

"Doesn't the 12-of-1,177 figure just reflect that nobody looked?"

Partly, yes, and this is the honest weakness in leaning on that number. Absence of published protein evidence is not proof of absence. It is still telling that "the vast majority of the well studied genes have been directly shown to encode a protein" while the orphans had almost nothing.²

"If de novo genes are so hard, how did the cod antifreeze gene arise?"

The reconstructed origin is well documented.  One nucleotide substitution plus repeat copying concerns the repetitive coding stretch; the whole gene also required a frameshift deletion and acquisition of a promoter.⁷⁷  It is evidence for a route to this repetitive antifreeze protein.  It does not demonstrate the same route to a complex folded enzyme.

"Isn't this article just conceding the whole argument?"

No. It concedes that the rate argument fails, because the evidence says so. What it does not concede is the popular picture. The counts in circulation are inflated by roughly a factor of two and a half in mouse, function is demonstrated for a tiny fraction, and gene catalogs have been pruned using assumptions about what rates are plausible.

What would settle it

The field already knows what the answer looks like, and Moyers and Zhang wrote it down: establish de novo origin gene by gene, against four criteria — transcribed and translated; a homologous sequence present in the syntenic region of relatives but non-coding there; ancestrally non-coding; and under selection.⁴

Verdict

The reviewed reanalyses show that a failed similarity search is insufficient evidence of de novo origin.  The stronger cases require evidence for the ancestral non-coding state and the resulting gene's function.  Their scope cannot be inferred from the pooled count of orphan annotations.  After the best corrections for missed relatives, a residue of lineage-restricted genes remains in every lineage studied;  how it arose is decided gene by gene, and in a strict 2022 review eight de novo genes met tests of both new origin and function, none of them an enzyme or a motor.

Sources