EPAS1 and the Tibetan Highlands

"Re-sequencing the region around EPAS1 in 40 Tibetan and 40 Han individuals, we find that this gene has a highly unusual haplotype structure that can only be convincingly explained by introgression of DNA from Denisovans or Denisovan-related individuals into humans." — Emilia Huerta-Sánchez and colleagues, Nature, 2014¹

Summary

Tibetans living above 4,000 metres do not thicken their blood the way visitors do.  A lowlander who moves to that altitude makes more red cells, raises haemoglobin concentration, and pays for it in blood viscosity, heart strain and lost pregnancies.  Tibetans mostly do not.  The gene most strongly tied to that difference is EPAS1, and the story of how Tibetans got their version of it is the reason this article exists.

1. The measurement in living people is real and replicated. Cynthia Beall and twenty-eight colleagues found 31 of 49 testable non-coding EPAS1 variants associated with haemoglobin concentration in 70 Tibetans at 4,200 metres, with major-allele homozygotes 0.8 ± 0.15 g/dL below heterozygotes, then replicated it in 91 Tibetans at 4,300 metres at 1.0 ± 0.14 g/dL, 26 variants significant both times in the same direction.^{2 2} That is half a standard deviation of the trait from one locus.²
2. The Tibetan version changes no protein. Its most differentiated variant is intronic, and no amino-acid-changing variant in the gene differs by more than 6 per cent between Tibetans and Han.³
3. What it does is turn a response down. Yi Peng and colleagues measured 31.3 per cent less EPAS1 expression in adaptive-haplotype endothelial cells and 27.1 per cent less in Tibetan placentas, and half-knockout mice reproduce the Tibetan physiology.^{4 4} Olivia Gray and colleagues found the high-altitude alleles weaken four enhancers, and reverting one derived allele raises activity again.^{5 5} Both papers use the words "blunting" and "disrupted".
4. The haplotype did not arise in Tibetans. A five-variant motif on the common Tibetan haplotype matches the Denisovan genome exactly, 15 of the 20 most differentiated variants match, and the alternative, old variation retained in two lineages, needs a 32.7-kilobase block to survive 200,000 years unrecombined and then vanish everywhere else on earth.^{1 1 1}
5. That makes it a transfer, and a transfer measures nothing about mutation. Under the scheme in Measuring Biological Information, an allele arriving from another lineage scores as a transfer worth zero new nucleotides in the receiving population. Whatever search produced it happened elsewhere, at a time nobody has dated directly. This article agrees with that article's table.
6. Nobody watched it spread. The introgression, the sweep and every date attached to them are reconstructions from comparing genomes, one read from a finger bone. The nearest thing to watching is a series of dated ancient genomes, which has the haplotype rising from 0.36 before 2500 years before present to 0.86 today.⁶
7. The claim of the "fastest allele frequency change observed at any human gene to date" rested on a date that has since moved twentyfold.³ Yi's 2010 paper put the Han-Tibetan split at 2,750 years and called EPAS1 possibly "the strongest instance of natural selection documented in a human population".³ Dongsheng Lu's 2016 study gives 15,000 to 9,000 years;⁷ Hao Hu's 2017 study gives 54,000, and shows by simulation that a simple two-population model of that history returns about 7,000.^{8 8}
8. The published selection coefficients are model fits, not measurements, and they disagree. Xinjun Zhang's 2021 study gives 0.018, interval 0.005 to 0.099;⁹ Haiyi Lou's 2015 study gives 0.0084 for a nearby deletion.¹⁰
9. The one Tibetan adaptation that did arise by coding mutation is disputed in direction. At EGLN1, Felipe Lorenzo's group measured a lower oxygen requirement and titled a figure "gain of function under hypoxia";¹¹ Frank Lee's group measured a broken protein interaction, built a mouse, and calls the same allele a partial loss.^{12 13}
10. Andeans and Ethiopians solved the same problem differently. Andean men average 18.1 g/dL of haemoglobin against Tibetan men's 16.9 at the same altitude;¹⁴ Ethiopians at 3,530 metres keep haemoglobin and oxygen saturation inside sea-level ranges.¹⁵ The Tibetan variants predict nothing in 260 Ethiopians tested with near-complete power.¹⁶

The short version.  This is the most celebrated case of recent human adaptation and a poor benchmark for how fast mutation finds new function, because no new function was found here.  A variant that already existed moved from one human population to another by marriage, and selection sorted it.  For the young-earth side that is comfortable on mechanism and uncomfortable on chronology, and no creationist author has published an account of the dates.

Part 1: What was measured, and in whom

Air at 4,000 metres holds about 40 per cent less oxygen than air at sea level.¹  The standard human response is to make more red blood cells, which carries more oxygen per unit of blood.  Past a point the response becomes the disease.  Chronic mountain sickness, first described among Andean highlanders by Carlos Monge, has excessive red-cell production as its defining feature, and the thickened blood brings heart failure, clots and poor pregnancy outcomes.²

Tibetans mostly escape it.  Beall's 2010 paper states the physiology plainly:

"Tibetan highlanders are particularly resistant to developing chronic mountain sickness (4, 11), and exhibit little or no increase in hemoglobin concentration with increasing altitude, even at 4,000 m (13,200 ft) and only moderate increases beyond (12, 13). … This suggests that Tibetans have evolved a blunted erythropoietic response to high-altitude hypoxia."²

Chronic mountain sickness runs at 1.2 per cent among Tibetans in the Tibet Autonomous Region against 5.6 per cent among Han lowlanders there.²

Three papers published in 2010 tied that difference to genes, and they did not agree on which gene.

Beall and colleagues ran a genome-wide comparison of 35 Tibetans from Yunnan at 3,200 to 3,500 metres against 84 Han from the HapMap panel.  Eight variants near EPAS1 reached genome-wide significance, and the haplotype they form stood at 46 per cent in the Tibetan sample against 2 per cent in the Han.²  Separately, they genotyped 103 non-coding EPAS1 variants in 70 Tibetans at 4,200 metres in Mag Xiang.  Of the 49 common enough to test, 31 were significantly associated with haemoglobin concentration, and the more common allele of every significant one went with lower haemoglobin.  Sex-adjusted, homozygotes for the major allele averaged 0.8 ± 0.15 g/dL below heterozygotes.²  In 91 more Tibetans at 4,300 metres in Zhaxizong Xiang, 32 of 45 testable variants were associated and the gap was 1.0 ± 0.14 g/dL.²  Twenty-six variants were significant in both samples with the same direction of effect.  Conditional regression in each sample left a single variant carrying the signal, which fits one causal variant that the rest merely tag.²

Yi and colleagues sequenced the exomes of 50 Tibetans from two villages at 4,300 metres or higher, against 40 Han from Beijing and 200 Danes.  A three-population branch statistic put EPAS1 first in the genome, and none of a million simulations under their fitted demographic model matched its value.³  They then tested their top variant against blood traits in the same 50 people:  no relation to oxygen saturation, but haemoglobin concentration at P = 0.00131 and red-cell count at P = 0.00141, with the Tibetan-common allele on the low side of both.^{3 3}

Tatum Simonson and colleagues published the third scan.  Its own abstract ties the phenotype to different genes:  "Positively selected haplotypes of EGLN1 and PPARA were significantly associated with the decreased hemoglobin phenotype that is unique to this highland population."¹⁷  EPAS1 is in its candidate list but is not the gene it connects to haemoglobin.  Anyone who writes that three papers in 2010 found EPAS1 has overstated the agreement.

Two cautions belong here, and both are the papers' own.  Neither group was sure haemoglobin was the target of selection.

Yi's group warned that the blood measurements may not be what selection acted on:  "the hematological differences observed here may not represent the phenotypic target of selection, and could instead reflect a side effect of EPAS1-mediated adaptation to hypoxic conditions."³  Beall's group said the same in different words, noting that EPAS1 affects many things and "selection may have operated on some other aspect of the phenotype", and pointing at the gene's roles in placental and embryonic development, with pre- and postnatal mortality threefold higher among Han than Tibetans on the plateau.^{2 2}

And the samples are small.  Seventy people, then ninety-one.  Fifty exomes.  These are the numbers the whole literature rests on, and they were enough only because the effect is unusually large:  about half a standard deviation from one locus, "several fold larger than for any of the determinants of hemoglobin concentration in populations residing at low altitude".²

Part 2: What the variant does, and what remains guessed

Nothing in the Tibetan EPAS1 haplotype changes a protein.  Yi's group put it flatly:  the variant with the largest frequency difference "was intronic (with a derived allele at 9% frequency in the Han and 87% in the Tibetan sample; Table S4), whereas no amino acid-changing variant had a population frequency difference of greater than 6%."³  So whatever the haplotype does, it does by changing how much EPAS1 protein gets made, or when.

Two later papers measured that, and they agree on the direction.

Peng and colleagues, 2017.  They picked out 32 variants across the 97-kilobase gene with a frequency difference above 0.5 between Tibetans and lowlanders worldwide.  In a reporter assay, the Tibetan-enriched A allele at rs149594770 bound nuclear protein more weakly and drove a weaker promoter.⁴  In endothelial cells grown from Tibetan umbilical cords, three lines homozygous for the adaptive haplotype expressed 31.3 per cent less EPAS1 than three wild-type lines over seven days at 1 per cent oxygen (P = 0.0063).  In 63 Tibetan placentas, the 34 adaptive homozygotes expressed 27.1 per cent less than the 3 wild-type homozygotes, with heterozygotes in between.⁴  Then they took mice with one EPAS1 copy knocked out, which have about half the normal transcript, and held them at 10 per cent oxygen for four weeks.  The half-knockout mice lost less weight, carried lower haemoglobin, had lower right-ventricle pressure and less heart hypertrophy than their wild-type littermates, and did not undergo the lung vascular remodelling the wild types did.⁴

Gray and colleagues, 2022.  They built 24 reporter constructs covering 37 candidate variants in 11 regions and found that "high-altitude alleles disrupt the activity of four EPAS1 enhancers in one or more cell types."⁵  One of the four, which they call ENH5, is itself switched up by low oxygen.  Reverting a single derived high-altitude allele, rs375554942, back to the ancestral base raised the enhancer's activity significantly in both normal and low oxygen.⁵  Deleting the whole one-kilobase enhancer with CRISPR in endothelial cells lowered EPAS1 transcripts at 24, 48 and 72 hours of low oxygen but not in normal oxygen, and lowered a downstream target as well.⁵

What is shown, then, is that the Tibetan alleles weaken existing regulatory elements, that less EPAS1 is made under low oxygen as a result, and that halving EPAS1 in a mouse produces the Tibetan pattern of physiology.  What is inferred is that this is why Tibetans keep their haemoglobin down.  The inference is reasonable and it is not a demonstration.  Three gaps are worth naming, all of them acknowledged by the authors:

• Peng's mice are knockouts, not carriers of the Tibetan haplotype. They show what halving expression does, not what the Tibetan alleles do.
• Peng's tenfold effect in the reporter assay came from a kidney-derived cell line, and the paper says the size "may not reflect the in vivo situation".⁴
• Gray's mouse deletion removes the whole enhancer, a bigger change than any Tibetan allele, and even then "EPAS1 itself was not down-regulated at lfsr < 0.05 in any single tissue."⁵ They read the downstream targets instead. They also state that they could not test the long-term effects in a living animal.⁵

The 3.4-kilobase deletion, which is not Denisovan

There is one more piece of the locus, and it complicates the tidy picture.  In 2015 Haiyi Lou and colleagues found a 3.4-kilobase deletion 80 kilobases downstream of EPAS1, at chr2:46,694,276–46,697,683.  About 90 per cent of Tibetans carry it and about half carry two copies.  In 2,792 samples from the rest of the world, 3 per cent carry it and none are homozygous, with none at all in Africans or Europeans.¹⁰  They confirmed the breakpoints by long-range PCR and Sanger sequencing and found them identical in every carrier.  The deleted stretch carries an enhancer-associated histone mark and two DNase-hypersensitive clusters in ENCODE data, so it may well be regulatory.¹⁰

The interesting part is its pedigree.  The deletion sits in complete linkage with the five-variant Denisovan motif, yet it is not in the Denisovan genome:

"It was also in complete LD with the 5-SNP motif, which was suspected to be introgressed from Denisovans, but the deletion itself was absent from the Denisovan sequence."¹⁰

If the deletion is the functional variant, then the part of this story that happened in modern humans is the removal of a regulatory stretch.  That is a loss under the taxonomy, not a gain, and Lou's own dating treats it as standing variation rather than a new mutation:  their selection estimate starts from the 3.4 per cent frequency the deletion already had in Han Chinese.¹⁰  Whether the deletion matters or merely rides along with an EPAS1 variant was open when the paper was published, and the authors proposed the cell experiment that would settle it.  As of Peng's 2017 association work, which credits the deletion with 1.91 per cent of haemoglobin variance in 508 Tibetans, it is still association and not mechanism.⁴

Part 3: Where the haplotype came from

In 2014 Huerta-Sánchez and twenty-six colleagues re-sequenced the whole EPAS1 gene plus 30 kilobases on each side, 147 kilobases in all, in 40 Tibetans from villages at 4,300 and 4,600 metres and 40 Han from Beijing, at about 200-fold coverage.  477 variants survived filtering.  Differentiation between the two populations peaks in a 32.7-kilobase stretch holding the 32 most differentiated variants, which is where the advantageous variant must be.¹

The first surprise was inside Tibetans.  The common Tibetan haplotype carries a five-variant motif, AGGAA, inside a 2.5-kilobase window, and none of those five variants appears on any of the minority Tibetan haplotypes either.  A sweep does not normally produce that.  Simulating both a sweep from a new mutation and a sweep from standing variation, at three strengths of selection, the authors could not reproduce the number of fixed differences between the two haplotype classes:  "it is unlikely (p<0.02 under either a SSV scenario or under a SDN scenario) that the high degree of haplotype differentiation could be caused by a single beneficial mutation landing by chance on a background of rare SNPs, which are then brought to high frequency by selection."¹  That left two candidates:  the haplotype came in from a divergent population, or it is very old variation retained in more than one lineage.

The search for a donor is the part worth reading closely, because the match is not a statistical impression.  It is a list.

• The AGGAA motif is absent from all 40 Han, from every population in the 1000 Genomes Project except one Southern Han and one Beijing Han individual, and from every modern genome in the 2012 Denisovan paper's panel. The Denisovan genome carries AGGAA at those five positions: "Strikingly, the Denisovan haplotype at these 5 sites (AGGAA) exactly matches the 5-SNP Tibetan motif".¹
• Twenty variants in the 32.7-kilobase stretch differ by at least 0.65 in frequency between Tibetans and every 1000 Genomes population. Fifteen of the twenty match the Denisovan. Five are shared only by Tibetans and the Denisovan.¹
• Over the whole stretch, sequence divergence between the common Tibetan haplotype and the Denisovan is 0.0038, significantly lower than typical human-Denisovan comparisons at P = 0.0028.¹
• Two statistics designed to detect archaic ancestry, D and S*, are significant against several null models with no gene flow (D at P < 0.001, S* at P ≤ 0.035).¹
• The Tibetan and Denisovan copies differ at 12 sites out of 32,700. Against the published Denisovan divergence times, the expected number is 11.25, and 17 or more would have rejected the introgression model at the 5 per cent level.¹

The alternative, an ancient variant retained in both lineages, is ruled out by length and by geography rather than by similarity.  A 32.7-kilobase block would have to pass through at least 200,000 years of recombination intact in two separate lineages, which the authors put at p = 0.0012 at a constant recombination rate, and then it would have had to be lost independently in every African and non-African population except Tibetans and a trace of Han.^{1 1}

Two details of the pattern are easy to miss and both matter.  The haplotype is not found in Melanesians, who carry far more Denisovan ancestry than East Asians do, and it is not in the high-coverage Altai Neanderthal.¹  And because it turns up in one Southern Han and one Beijing Han individual who show no sign of being recent migrants from Tibet, the authors read it as having entered before Han and Tibetans separated, then having been selected only in Tibetans.¹

Zhang's 2021 follow-up narrowed the donor and excluded the other archaic candidate.  Applying a reference-free detection method to 38 Tibetan genomes, it found archaic segments flanking EPAS1 matching the Altai Denisovan at 72 and 46 per cent, and tied the EPAS1 haplotype to the East Asian-specific one of the two known Denisovan admixture pulses.^{9 9}  A Neanderthal origin was tested directly, by asking whether the Tibetan haplotype shares derived alleles with Neanderthals the way one Neanderthal does with another, and it does not.⁹

Hu's 2017 study added the number that any short chronology has to face.  The gene tree for this haplotype puts the last common ancestor of the Tibetan and Denisovan copies at 997,000 years, with a 95 per cent interval of 868,000 to 1,139,000, and the corresponding population split at 868,000 years.⁸  In other words, the mainstream reading is not that Tibetans got this from the Denisovan girl's own population last week;  it is that the archaic population which handed it over had been separate from the sequenced Denisovans for a long time already.

Part 4: Was any of it watched?

The Observed Evolution section admits an article only when something was watched happening.  This one passes on one count and fails on others.

Observed.  The association between EPAS1 genotype and haemoglobin in 161 Tibetans in Beall's two cohorts and 50 in Yi's, which a sceptic and a creationist should both accept.  Also observed:  the cell and mouse measurements in Part 2.

Inferred from comparing genomes.  The Denisovan origin.  The Denisovan genome itself, read from a finger bone in a Siberian cave.  The rejection of ancient shared variation.  Every date:  when the interbreeding happened, when selection started, how fast the allele rose, how long ago the Tibetan and Denisovan copies parted.  Every selection coefficient.  All of these are reconstructions, and each depends on a demographic model, a mutation rate, a recombination rate and a generation length that were assumed rather than measured in the population concerned.

Partly observed, and this is the interesting middle.  Ancient DNA gives a dated frequency series, which is as close as this case gets to watching.  In 2023 Hongru Wang and colleagues published genome-wide data from 89 individuals from 29 plateau sites, dated 5,100 to 100 years before present, and called the EPAS1 haplotype from 20 variants in anyone with at least four reads at each site.  Across the central, southeastern, southern and southwestern plateau the estimated haplotype proportion runs 0.36 before 2500 years before present (95 per cent interval 0.21 to 0.52, n = 19), 0.47 at 2400 to 1900 years (0.33 to 0.61, n = 24), 0.59 at 1600 to 700 years (0.47 to 0.71, n = 37), and 0.86 in 33 present-day Tibetans from the same region.  The rise is faster than for randomly chosen variants across the genome, at P = 0.0055, which is the argument that selection rather than migration did it.^{6 6}

The oldest known modern human on the plateau, from Zongri at about 5,100 years before present, already carries two copies.⁶  Eleven of seventeen Zongri individuals from 4,800 to 3,900 years before present carry one.

An earlier and smaller ancient-DNA study, Choongwon Jeong's 2016 work on eight individuals from the Annapurna region of Nepal, found the derived EGLN1 D4E allele in every individual it could read, including the oldest at 3,150 to 2,400 years before present, but derived EPAS1 alleles only in two of three individuals from the latest period, 1,750 to 1,250 years before present.^{18 18}  Four or five individuals do not make a frequency, and Jeong's group said so.¹⁸

What would have had to be watched.  A population sampled at intervals across the rise, with genotypes and haemoglobin measured in the same individuals, plus the interbreeding recorded as it happened.  The first half now exists in part, for the last 5,000 years, and shows a real rise.  The second half is unavailable in principle, on any chronology, because it predates writing.

Part 5: The dates, and why they disagree

Yi's 2010 paper is where the famous number comes from, and it is worth seeing exactly what produced it.  The authors fitted a two-population model to the two-dimensional frequency spectrum of synonymous sites in their 50 Tibetan exomes and 40 Han genomes:

"The best-fitting model suggested that the Tibetan and Han populations diverged 2,750 years ago, with the Han population growing from a small initial size, and the Tibetan population contracting from a large initial size (Fig. S2).  Migration was inferred from the Tibetan to the Han sample, with recent admixture in the opposite direction."³

They then used that date to reach the claim the paper is remembered for:

"Given our estimate that Han and Tibetans diverged 2,750 years ago and experienced subsequent migration, it appears that our focal SNP at EPAS1 may have experienced a faster rate of frequency change than even the lactase persistence allele in northern Europe, which rose in frequency over the course of about 7,500 years (26).  EPAS1 may therefore represent the strongest instance of natural selection documented in a human population"³

Every estimate published since is older, and by a lot.

Hu's group did not merely report a different number.  They explained the old one.  A method that reads pairwise coalescence rates cannot distinguish a genuinely recent split from an old split with heavy migration, and when they simulated a 54,000-year split with the gene-flow rates their model inferred, the coalescence method failed to see it.  Running the frequency-spectrum method on the simplified history instead reproduced a separation of about 7,000 years.⁸  Yi's estimate came from a model without that much migration built in, and Yi's own sentence notes that migration was inferred in both directions.  A young split date is what a simple model returns when two populations have been exchanging people for a long time.

The dates for when selection started are separate, and they disagree too.

Jeong's group, writing in 2016 without a stake in any of these, summed up the state of play:  "dating the onset of selection for the derived EPAS1 haplotype has proved more controversial…and dates ranging from 2,750 ya to 18,250 ya have been proposed."¹⁸

And the selection coefficients:

The two real estimates differ by a factor of about two, and Zhang's own interval spans a factor of twenty.  They are not measurements of the same thing by the same method, and averaging them would produce a number nobody holds.

What the arithmetic says

Two calculations follow from the published inputs, and neither side has printed them.

How long does the sweep need?  Take the deletion's 3.4 per cent frequency in Han as the starting point, as Lou did, and today's 0.86 as the endpoint, as Wang measured.  Under the standard deterministic model with additive fitness, the time is (2/s)·ln[p₁(1−p₀)/(p₀(1−p₁))].  At Zhang's s = 0.018 that is 574 generations, about 14,300 years at 25 years per generation.  At Lou's s = 0.0084 it is 1,229 generations, about 30,700 years.  Only at the top of Zhang's credible interval, s = 0.099, does it fall to 104 generations, about 2,600 years.¹⁹

What does the ancient DNA imply?  Wang's series goes from 0.36 to 0.86 in about 2,500 years.  Running the same equation backwards gives s = 0.048 at 25-year generations or 0.055 at 29-year generations, roughly three times Zhang's point estimate, though inside its interval.¹⁹  Two readings are open.  Either the sweep was stronger in the last 2,500 years than the genome-wide fit suggests, or part of Wang's rise is population movement rather than selection, which the P = 0.0055 comparison argues against but does not eliminate.

One cross-check is reassuring about the mainstream account's internal consistency.  Zhang's own figures, 357 generations of selection at s = 0.018, imply the haplotype was at about 20 per cent when selection began.¹⁹  Wang's ancient DNA measured 0.36 before 2500 years before present.  Two entirely different kinds of evidence land in the same range.

Part 6: Scoring it

The taxonomy in Measuring Biological Information asks four questions in order and stops at the first yes.  The first question is whether the sequence arrived from another lineage.  Here it did, so the class is transfer and the nucleotide count in the receiving population is zero.  That article's table already scores EPAS1 that way, and this article agrees with it.

The series measures how much working sequence mutation and selection originate.  A transferred allele was originated somewhere else, in a population whose size nobody knows, so EPAS1 in Tibetans supplies no data point for that question.  There is no search in it.

What was the Denisovan allele itself?

The next question is the honest one, and the answer is that nobody knows.  If the variant arose by mutation in the Denisovan lineage, then the search happened there, in a population whose size, generation time and mutation rate are all unmeasured, over a period nobody has bounded.  If it was old variation that Denisovans inherited, the debt moves further back still.

What is on record is only this.  Hu's gene tree puts the last common ancestor of the Tibetan and Denisovan copies at about 997,000 years, and the population split that produced them at about 868,000, which means the haplotype was already distinct long before the sequenced Denisovan lived.⁸  Zhang concluded that selection in Tibetans "acted on standing archaic variation", and that the conclusion holds under every demographic scenario they tried.⁹  Nobody has published an estimate of when the variant first appeared, or of how many individuals it took to find it.  Anyone who cites EPAS1 as a measured example of adaptive mutation is citing a case whose mutational history has never been dated.

EGLN1, which really did arise by mutation, and is disputed

The Tibetan adaptation that fits the "new mutation" description is at a different gene.  EGLN1 encodes prolyl hydroxylase 2, the oxygen sensor that destroys the EPAS1 protein when oxygen is plentiful.  Tibetans carry two amino-acid changes in it, D4E and C127S, in near-complete linkage, and Song's 2020 paper notes that unlike EPAS1 this allele "is not a result of introgression from Denisovans".¹³  Hu found D4E absent from every other 1000 Genomes population.⁸

Two groups have studied it for a decade and they read it opposite ways.

Lorenzo and colleagues, 2014, say gain.  The double mutant's requirement for oxygen is lower:  a Michaelis constant of 110 µM against 150 µM for the normal enzyme, about 30 per cent lower, with the same maximum rate.  Neither single mutation does this on its own, so the two act together.¹¹  In cells, the double mutant lowers EPAS1 protein under 1 per cent oxygen.  Red-cell precursors from six Tibetans homozygous for the haplotype grew poorly at 5 per cent oxygen and not at all at 1 per cent, where control cells grew better than normal.¹¹  Their Figure 2 is titled "The p.[Asp4Glu; Cys127Ser] PHD2 mutant shows gain of function under hypoxia."¹¹

Lee and colleagues, 2014 and 2020, say loss.  The two changed residues flank a zinc finger at the protein's front end that binds a partner protein called p23, and the Tibetan version binds it far more weakly:  the haplotype "strikingly diminishes the interaction of PHD2 with p23, resulting in impaired PHD2 down-regulation of the HIF pathway."¹²  They proposed in 2014 that the allele is hypomorphic, a partial loss.¹²  In 2020 they built a mouse carrying the Tibetan substitutions.  It shows a stronger breathing response to low oxygen, which is what losing this function predicts, and the same phenotype appears in mice whose p23 was altered to break the interaction from the other side.¹³  Their mouse carrying the Tibetan allele shows no haemoglobin effect at all.¹³

Song's 2020 abstract states the disagreement without softening it:  "Key issues regarding Tibetan PHD2 are whether it is a gain-of-function or loss-of-function allele, and how it might contribute to high-altitude adaptation."¹³  Their proposed resolution is that Tibetans carry both kinds of change in one pathway:  "Tibetans possess genetic alterations that both activate and inhibit selective outputs of the HIF pathway".¹³

The two results are not actually contradictory as measurements.  Lorenzo measured enzyme kinetics on a peptide substrate;  Lee measured a protein-protein interaction and a whole-animal breathing response.  An enzyme can be faster at low oxygen in a tube and less effective in a cell if it has lost the partner that puts it where it needs to be.  But the classification depends on which is decisive, and nothing published settles that.  Neither group claims a new gene, a new binding site or a new activity;  the dispute is over the direction of an existing enzyme's performance.

Part 7: Who the Denisovans were

The literature calls this "admixture with other hominin species".¹  This site does not accept that wording, for reasons of evidence:  Neanderthals and Denisovans interbred with modern humans and each other and left fertile descendants in living populations, as Archaic Admixture sets out.  On this reading the EPAS1 event is one human population marrying another.

The mainstream reading.  Denisovans are an archaic human population, known from a high-coverage genome from a finger bone in the Altai, separated from the modern human line for hundreds of thousands of years, present on the Tibetan Plateau itself by at least 160,000 years ago on the Xiahe mandible.⁶  Modern humans reaching the plateau met their descendants, or descendants of a related population, and acquired a version of EPAS1 already suited to thin air.  Huerta-Sánchez's closing sentence is the mainstream summary:  "adaptation to local environments may have been facilitated by gene-flow from other hominins that may already have been adapted to those environments."¹

The young-earth readings, from their own primaries.  Peter Line's 2019 survey of the Denisovan evidence for Creation Ministries International reaches the population question directly:  "if the evolutionary assumptions and long ages are dropped, then a simple explanation is that the Denisovan, Neanderthal and SH Homo heidelbergensis remains all belonged to populations that were fully human, and to various degrees interbred with each other."²⁰  He notes that the Xiahe mandible was found by a monk in 1980 at 3,280 metres in Baishiya Karst Cave on the plateau itself.²⁰  Robert Carter, a geneticist at the same organisation, had already read the 2010 Neanderthal genome the same way and criticised those who accept interbreeding while rejecting the species test that goes with it.²¹

The only young-earth treatment of this locus that could be found is Elizabeth Mitchell's 2014 news article for Answers in Genesis, and it is more careful than a critic might expect.  It accepts the physiology.  It accepts the mechanism, describing the variant as one "in which a normal physiologic function is limited".²²  It accepts that interbreeding, not shared ancestry, is the better explanation of the pattern:  "The pattern of the matching genetic material in this study is best explained by past hybridization between Denisovans and the ancestors of the Tibetan people."²²  It treats Denisovans as people:  "Denisovan DNA is distinguishable from Neanderthal and modern human DNA, though all are fully human of course", and "Genetic mixing of people groups, whether classified 'archaic' or modern, is no surprise as all descended from Adam."^{22 22}  It rejects the chronology:  "The dates assigned to the Denisovans come from the usual worldview-based assumptions and circular reasoning inherent in molecular clock dating."²²  And it places the Denisovans after Babel:  "As people groups became isolated, limitations in genetic variability combined with other genetic phenomena such as mutations and founder effects produced the people groups anthropologists call Neanderthals, Denisovans, and many others."²²

Two things should be said about that last sentence, and the second is a criticism of the young-earth side.

First, it is the weaker of the two young-earth options available.  Mitchell reaches for mutation and founder effects, but the haplotype is not a mutation in Tibetans on anybody's account, including hers two paragraphs earlier.  The stronger creationist move is the one the site's creation model of rapid speciation already uses elsewhere:  a founding pair plus a flood bottleneck leaves standing variation that environments then sort, and an introgressed allele sorted by altitude is exactly that.  Zhang's own conclusion, that selection "acted on standing archaic variation", is the same shape of claim.⁹

Second, the numbers are not addressed anywhere in the creationist literature on this locus, and they are the hard part.  The Tibetan and the sequenced Denisovan copies of the 32.7-kilobase block differ at 12 sites.  Huerta-Sánchez's arithmetic, which reproduces exactly, uses a divergence time of 394,000 years, a sample age of 100,000 years and a rate of 0.5 × 10⁻⁹ per base per year to predict 11.25 differences, and gets 12.¹  Run the same equation with a short chronology and the rate is what has to give.  If the two lineages parted 4,500 years ago and the Denisovan individual died 4,000 years ago, the rate needed is 7.3 × 10⁻⁸ per base per year, about 147 times the rate used, and about 147 times the standard pedigree rate per generation.  If they parted 2,000 years ago and the sample is 1,000 years old, the factor is 245.²³  That is a specific, checkable demand, and no young-earth author has stated it, let alone met it.  Mutation rates do vary, but a hundredfold or greater acceleration confined to a nuclear intron needs its own evidence, not an assertion that clock dating is circular.

Part 8: Three populations, three answers

If the plateau posed a problem with one solution, populations facing it separately should have found the same solution.  They did not, and this is the best-documented case of the point in human biology.

The physiology differs first.  Cynthia Beall's 2007 comparison of paired Tibetan and Andean samples measured at about 4,000 metres with the same protocols found Tibetan men averaging 16.9 g/dL of haemoglobin against Andean men's 18.1 across 53 samples, and resting ventilation of 15.0 litres per minute against 10.5 across 28 samples.^{14 14}  Tibetans also carry lower oxygen saturation, more exhaled nitric oxide and lower pulmonary artery pressure;  Andean highlanders are consistently reported with pulmonary hypertension.  Her conclusion:  "the two followed different routes to the same functional outcome of successful oxygen delivery, long-term persistence and high function."¹⁴  Ethiopians are a third pattern again.  A survey of 236 residents at 3,530 metres, screened for iron deficiency and inflammation, found haemoglobin at 15.9 g/dL in men and 15.0 in women and oxygen saturation at 95.3 per cent, all inside sea-level ranges.¹⁵

The genetics differ too, and the negative test is the strongest part.  Abigail Bigham's 2010 scan of Andeans and Tibetans found that the two populations' "patterns of genetic adaptation are largely distinct from one another", with EGLN1 the one gene selected in both and carrying a different pattern of variation in each.²⁴  Jacob Crawford's 2017 whole-genome scan of Andeans found its strongest signals at BRINP3, NOS2 and TBX5, which "are associated with cardiovascular development and function but are not in the response-to-hypoxia pathway", and proposed that Andean selection mitigated the harm of thick blood rather than preventing it.^{25 25}

The Ethiopian test is the cleanest.  Gorka Alkorta-Aranburu's 2012 study took the EPAS1 and EGLN1 variants that predict haemoglobin in Tibetans and tested them in 260 unrelated Ethiopians:

"None of these SNPs were significantly associated with Hb levels in Ethiopians (Table 1).  Because we have complete or nearly complete power to detect a genotype-phenotype association in our Ethiopian samples (see Text S4 and Table S21), we infer that the SNPs associated with variation in Hb levels in the Tibetans do not make a contribution in Ethiopians."¹⁶

A different variant, rs10803083, was genome-wide significant in Amhara with an effect of about 0.83 g/dL, sitting in a gene desert with no obvious candidate within 600 kilobases, and awaiting replication.¹⁶  Their summary:  "Ethiopian and Tibetan highlanders adapted to the same environmental stress through different variants and genetic loci."¹⁶

Both sides should take something from this.  A design argument that treats an adaptation as requiring one particular sequence is refuted here:  three populations met the need three ways, and the Ethiopian way has no replicated variant in the sources read here.  An argument that the Tibetan case shows how readily mutation finds high-altitude adaptation is also weakened, because the Tibetan solution was not found by mutation and the Andean one appears to sidestep the hypoxia pathway.

Objections and responses

"Calling it a transfer is a dodge.  The allele is beneficial, selection fixed it, that is evolution."  Agreed, and this article says so.  Selection really did raise the haplotype from about a third of the plateau's chromosomes to 86 per cent, and the ancient DNA shows it happening.⁶  The transfer class is not about whether the change is real.  It is about which population's ledger the origination goes on, because the series is counting how many organism-generations it takes mutation to find new function.  This case supplies no such count, in either direction.

"The Denisovan origin is an inference from a statistical model, so it can be dismissed."  It should not be.  The evidence is a list of concrete matches, not a model output:  five variants matching exactly at named positions, fifteen of twenty matching, twelve differences over 32.7 kilobases where 11.25 were expected, and absence of the motif from every other sampled population on earth except two individuals.^{1 1 1}  Answers in Genesis accepted the hybridisation reading in 2014 and was right to.²²

"If the sequences are that similar, doesn't that fit a recent common origin rather than deep time?"  Similarity alone would, which is why the mainstream argument does not rest on it.  It rests on the pattern:  a long unbroken block, exclusive to two lineages, absent from a third archaic genome and from every modern population sampled, with a number of differences that matches a long separation.  A short chronology can accommodate the similarity, but it must then account for the 12 differences, and the rate that requires is about 90 to 250 times the measured pedigree rate.²³  That is the open problem, and it belongs to the young-earth side.

"Tibetans have low haemoglobin because they are iron-deficient or malnourished, not because of genetics."  Beall's cohorts were screened, excluding anyone anaemic by the standard thresholds, hypertensive, febrile, pregnant or with poor lung function.²  The Ethiopian survey applied the same kind of screen.¹⁵  And the effect is genotype-linked within a single population living in one place, which nutrition does not explain.

"Since the adaptation is only a loss of regulation, it is evidence that evolution only breaks things."  It is evidence about one locus, and this site does not argue the general claim.  Measuring Biological Information concedes genuine gains, including the citrate result in Richard Lenski's long-term experiment and the origin of an antifreeze gene from non-coding DNA in cod.  At EPAS1 the published functional work does describe weakening of existing elements, and at EGLN1 one of the two research groups describes a gain of enzyme function.¹¹  Reporting only the half that suits the argument would be dishonest.

"Yi's 2,750-year date is still the published figure, so the fastest-sweep claim stands."  It does not.  Later studies give 15,000 to 9,000, about 7,000, and 54,000 years, and Hu's group showed why a simple fit returns a young date under heavy migration.^{7 8 8}  Yi's sentence was hedged with "may", and its model was the best available in 2010.

"A short chronology has no room for any of this, so the whole reconstruction collapses."  The reconstruction does not collapse;  it just conflicts.  A young-earth reader has to reject the 997,000-year gene tree, the 48,000-year introgression date and the radiocarbon dates on Wang's 89 burials.  The first is a rate question that can be argued.  The last is not:  the ancient-DNA frequency series is a set of dated remains, and a chronology that compresses it has to compress the archaeology of the plateau with it.  Saying so plainly is better than not mentioning it.

Issues and unanswered questions

1. Which variant actually does the work is unknown. Beall's conditional regression points to one causal variant that the rest merely tag; Peng points at rs149594770; Gray points at rs375554942 inside ENH5; Lou points at the 3.4-kilobase deletion. Nobody has tested them against each other in one system.
2. Whether the Gray and Peng variants are among the Denisovan-shared sites was not established here. Huerta-Sánchez's Table S3 lists the 32 differentiated positions; the supplementary tables were not obtained, so the overlap was not checked. It matters, because if the functional variants are not the introgressed ones, the transfer scoring changes.
3. Simonson et al. 2010 was not obtained in full. It is behind the Science paywall, absent from PubMed Central, and not in any open repository Unpaywall knows. Only its abstract is cited.
4. Aldenderfer's 2011 archaeological answer to the genetic dating was not obtained (High Altitude Medicine & Biology, closed at the publisher, not in PubMed Central, blocked on the Wayback Machine). The chronology dispute above is therefore told from the genetics, plus the plateau occupation dates reported in Wang 2023.
5. No creationist technical treatment of this locus exists, as far as searching found. Mitchell 2014 is a news article. Line 2019 and Carter 2010 do not mention EPAS1. The sweep arithmetic in Part 5 and the mutation-rate demand in Part 7 are the calculations that side needs and has not published.
6. The selection coefficient implied by the ancient DNA, about 0.05, has not been published by anyone, including here beyond the note. Somebody with Wang's per-individual genotypes and a proper likelihood should do it properly; the calculation above is a deterministic approximation ignoring drift, and the plateau's effective population size was small.

Verdict

EPAS1 in Tibetans is the most celebrated case of recent human adaptation and one of the least useful benchmarks for how fast evolution builds anything.  The trait is real and replicated in 161 Tibetans.  The mechanism, as far as cells and mice show, is a weakened enhancer and less of an existing transcription factor:  a response turned down, not a capability added.  And the variant arrived from a Denisovan or Denisovan-like population by interbreeding, which is a transfer worth zero originated nucleotides in Tibetans.

For the young-earth side the mechanism is comfortable and the chronology is not.  Nothing here requires a new gene, binding site or activity, and the one Tibetan adaptation that did arise by coding mutation, at EGLN1, is disputed between a modest gain of enzyme efficiency and a partial loss, with the loss reading supported by a knock-in mouse.  Against that, every date is deep:  a gene tree of about a million years, an introgression around 48,000 years, a sweep beginning 9,000 to 13,000 years ago.  The one creationist treatment rejects those dates without pricing them.  The price is a mutation rate at this locus about 90 to 250 times the pedigree rate, and until somebody defends that number, this locus is a problem for a short chronology, not a support for it.

The most durable result in the whole file is the least argued over.  Tibetans, Andeans and Ethiopians met the same shortage of oxygen and answered it three different ways, using largely different genes, and the Tibetan variants do nothing measurable in Ethiopians who had the statistical power to detect them.  Whatever the plateau demanded, it did not demand one particular sequence.

Sources