Neanderthal DNA in your genome: how much there is, and how it is found

About 2 percent, depending on the reference genome

The first estimate came with the first Neanderthal genome. In 2010 Green and colleagues compared a draft sequence, assembled from bones of three Neanderthals from Vindija Cave in Croatia, with the genomes of five present-day people, and concluded that Neanderthals contributed 1 to 4 percent of the genomes of non-Africans^[3]. Every estimate since has fallen inside that range and narrowed it.

In 2014 a high-coverage genome from a Neanderthal found in the Altai Mountains of Siberia put the share outside Africa at 1.5 to 2.1 percent^[4]. In 2017 a second high-coverage genome, again from Vindija, raised it to 1.8 to 2.6 percent: 2.3 to 2.6 in East Asians and 1.8 to 2.4 in western Eurasians^[2]. Nothing about living people changed between those papers. The Vindija Neanderthal is more closely related to the population that mixed with modern humans, so more of the inherited sequence could be recognised, about 10 percent more per person: 40.4 million bases against 36.3 million in Europeans^[2].

The method matters as much as the reference. Chen and colleagues built one, IBDmix, that does not assume any living population is free of Neanderthal ancestry. Across the 1000 Genomes samples it found about 51 million bases of Neanderthal sequence per European, 55 million per East Asian and South Asian, and about 17 million per African, where earlier work had found well under a million^[5]. More than 94 percent of the African signal was shared with non-Africans, which the authors explain by migrations back into Africa, mostly by the ancestors of Europeans, and by earlier modern human gene flow into Neanderthals.

The rows measure different things. The genome papers use a ratio of shared variants that asks how close a population is to being entirely Neanderthal^[2]; the mapping studies add up the segments a model could assign, person by person^[6].

Short segments, a different set in each person

A child of a Neanderthal and a modern human carried one complete Neanderthal copy of every chromosome. In every generation after that, recombination swapped stretches between the two copies before one was passed on, so the Neanderthal blocks were cut shorter each time. Genomes of people who lived at Ranis, in Germany, about 45,000 years ago carry segments from the same mixing as everyone outside Africa today, which dates it to about 45,000 to 49,000 years ago^[7]. After that many generations, Sankararaman and colleagues expected a Neanderthal segment to be about 0.05 centimorgans long, roughly 50,000 bases^[6].

The cuts fall in different places in every family line, so two people carry largely different pieces, and pooled across many people they add up to far more of the Neanderthal genome than anyone has alone. Vernot and Akey recovered sequence spanning about 20 percent of the Neanderthal genome from 665 Europeans and East Asians^[8]; the archaic fragments found in 27,566 Icelanders covered 38 to 48 percent of the human genome that could be analysed^[9].

The pieces are not spread evenly. Regions dense in functionally important sequence carry less Neanderthal ancestry than the rest, genes most active in the testes are especially depleted, and the X chromosome carries about five times less than the other chromosomes. Sankararaman and colleagues read this as evidence that some Neanderthal variants lowered male fertility in a modern human background and were removed by selection^[6].

Denisovans, known mostly from their DNA

In 2010 a finger bone from Denisova Cave in Siberia yielded the genome of a group that shared a common origin with Neanderthals but was distinct from them. Reich and colleagues named the group Denisovans and estimated that it contributed 4 to 6 percent of the genomes of present-day Melanesians^[11].

Later work lowered the figure and widened the map. Vernot and colleagues measured Denisovan admixture of 1.9 to 3.4 percent in their Melanesian samples^[12]. Browning and colleagues found Denisovan segments in East and South Asian populations as well, in two groups that differ in how closely they match the sequenced Denisovan, which they read as at least two separate episodes of mixing^[13]. Even in Iceland, 3.3 percent of the archaic fragments look Denisovan rather than Neanderthal^[9].

The best-known Denisovan segment with an effect sits at EPAS1, a gene in the body's response to low oxygen and the strongest signal of selection found in Tibetan genomes. The selected version matches the Denisovan sequence, and outside Tibetans it turns up only at very low frequency among Han Chinese^[14].

How a segment is recognised as archaic

Every method asks whether a stretch of DNA differs from other modern human sequence in the way an archaic inheritance would. They differ in what they compare it with, and that decides what they can find.

Sankararaman and colleagues trained a statistical model, a conditional random field, on three signals: a variant seen in Neanderthals but absent from the Yoruba of West Africa, a stretch very different from every Yoruba sequence and close to the Neanderthal one, and a length that fits interbreeding 37,000 to 86,000 years ago^[6]. The S* statistic, introduced by Plagnol and Wall^[16], needs no archaic genome at all. It looks for the mark introgression leaves: many variants absent or rare in Africans, inherited together on one long stretch, because Neanderthals carried many variants of their own and the stretch has had little time to break up^[13]. Vernot and colleagues used S* to find candidate segments and only then compared each with the Neanderthal and Denisovan genomes to decide which group it came from^[12].

So reference-free means two different things. S* and Sprime are free of an archaic reference, which lets them find segments from groups nobody has sequenced. IBDmix is free of a modern one, which is why it found Neanderthal DNA in Africans that methods using Africans as the baseline could not see^[5]. None of them finds everything: in simulations, Browning and colleagues could detect about half of the introgressed material, the rest being in segments too short to call with confidence^[13].

A percentage from a chip is a model's output

The estimates above come from whole genomes. A chip reads a fixed list of positions: the ten exports we measured listed between 563,320 and 955,958 each^[17]. Spread evenly over the 3.1 billion bases of the reference genome^[18], that is one position every 3,200 to 5,500 bases, so a segment of 50,000 bases, the size of the chromosome 3 COVID-19 haplotype, would hold about 9 to 15 of them, by our own arithmetic.

To turn that into a percentage, a model has to separate the two chromosome copies, fill in the positions the chip never read from a reference panel of other people, and choose what to compare each stretch with. The published figures moved with those choices even on whole genomes. A chip-based figure makes them with far less data, which is why we do not compute one.

Named segments with measured effects

Large cohorts made it possible to test Neanderthal variants against traits. Dannemann and Kelso, using 112,000 UK Biobank participants, found Neanderthal alleles affecting skin tone, hair colour, height, sleeping patterns, mood and smoking status, with some alleles pushing skin and hair lighter and others darker^[19]. Skov and colleagues then tested 271 traits in 27,566 Icelanders, found five associations driven by archaic variants, and concluded that most associations reported before were better explained by variants that are not archaic^[9].

The chromosome 3 segment is the clearest case. In 2020 the COVID-19 Host Genetics Initiative's analysis of 3,199 hospitalised patients found a single region with genome-wide significance, on chromosome 3, at an odds ratio of 1.6 for hospitalisation^[15]. Zeberg and Pääbo showed that the risk variants sit on a core haplotype of 49.4 thousand bases that closely matches the Vindija Neanderthal^[15]. In gnomAD rs10490770, a variant on the same core haplotype that consumer chips read, sits on 8.72 percent of Finnish chromosomes against 7.50 percent in other Europeans^[23], which works out to about 17 percent of Finns carrying at least one copy, by our arithmetic and assuming the two copies are inherited independently.

A second Neanderthal segment, on chromosome 12 around the OAS1, OAS2 and OAS3 genes, points the other way. Zeberg and Pääbo found it associated with a relative risk of needing intensive care about 22 percent lower per copy^[20]. The likely working part is a splice variant in OAS1 whose protective form is the ancestral one, kept on the Neanderthal haplotype^[20], and that form also occurs in people of African ancestry who did not get it from Neanderthals^[24].

Mendez and colleagues described the STAT2 haplotype, around an immune signalling gene, in 2012 as a candidate for positive selection in Papua New Guinea^[21]. The CYP2C8*3 and CYP2C9*2 alleles, known in pharmacogenomics for slowing the breakdown of warfarin and phenytoin, travel together on a haplotype of about 300,000 bases that Haeggström and colleagues traced to Neanderthals^[22]. The CYP2C9 gene page and the warfarin page cover what those alleles mean for prescribing; their origin changes none of it.

What Aimosti's report shows

The report has an archaic ancestry section, and it gives no percentage. It reads a fixed panel of 9 named segments, all of them Neanderthal; no Denisovan segment is on the panel yet. Each is marked by one or more positions whose archaic letter was checked against the sequenced Neanderthal genomes, and each card cites the study behind it^[25]. The chromosome 3 COVID-19 segment, STAT2 and the CYP2C pair are on the panel. The chromosome 12 OAS segment is not.

The section says how many of the segments you carry out of how many your file could be read for, and gives for comparison the average across all ancestries, computed from population frequencies. That average treats the segments as independent, and the report says they are not quite.

• Chip export: at most 4 of the 9. The other 5 are marked at positions no consumer array we measured carries, or that cannot be read reliably from one. A segment your chip lacks is named as unread and left out of the count.
• Whole-genome gVCF: all 9, because its reference blocks show which positions were read and matched the reference.
• Plain VCF: only the segments you carry. It lists only where you differ from the reference, and every archaic marker on the panel is the non-reference letter, so a missing record could mean read and not carried, or never read. The report counts it as unread.
• BAM or CRAM through Deep Read: nothing extra here. The section comes from the VCF or gVCF uploaded with the reads; Deep Read's own panel is pharmacogenes.

The Neanderthal DNA page lists the 9 segments one by one, with what each is linked to and what it does not mean. With a list that short, a percentage would claim a precision the panel does not have.