Ghost Populations: The People We Only Know From Their DNA

Start with the idea, because it is stranger than it sounds. A ghost population is a human group whose existence is inferred entirely from statistical signatures in modern or ancient genomes, not from any bone, tool, or grave. The concept has to exist because of a brutal limitation at the heart of paleoanthropology: the fossil and ancient-DNA record samples only a tiny sliver of all the populations that ever actually lived. Most people who ever existed left nothing we can dig up. But some of them left something we can read, written into the genomes of their descendants, and that is enough to prove they were here.

Here is the catch, worth stating early and plainly. A ghost population is defined by what it is NOT. It is the missing piece a demographic model needs in order to fit the real genetic data, the shape of an absence rather than a face. Nobody has ever observed the population directly. They have observed the hole it left behind and worked backward to its size and its age. Hold onto that, because it is exactly the line between the parts of this story that are rock solid and the parts that are genuinely guesswork.

And here is the framing that snaps the whole field into focus. Everyone has heard the headline that modern humans leaving Africa interbred with Neanderthals and Denisovans, the archaic cousins we cover in the sister file on the Neanderthals still inside you. Ghost populations are that same story with the lights off. The humans who stayed in Africa appear to have interbred with their own distinct archaic hominins, and the humans who spread across Asia met still older lineages, but for these partners there is no fossil at all. Same kind of event. No face on the other side of it.

In 1933 a massive human skull was recovered from the ground in Harbin, in Heilongjiang Province in northeast China. It was hidden away for decades and only brought to scientific attention in 2018, then formally described in 2021: a colossal cranium with a brain capacity of roughly 1,420 cubic centimeters, dated by uranium-series methods to at least 146,000 years old. The team led by Qiang Ji argued it was so distinct it deserved a brand new species name, Homo longi, and claimed it sat closer to modern humans on the family tree than Neanderthals do. Others, the veteran human-origins researcher Chris Stringer chief among them, suspected something bolder still: that this face, with its size and its geography and its age, was the long-missing physical form of a Denisovan. For years it stayed a standoff, because the one thing nobody could pull from the skull was DNA.

Then, on June 18, 2025, the standoff ended. Two companion papers landed the same day, one in Science and one in Cell, led by Qiaomei Fu and by Qiang Ji himself, the skull's original describer. Direct attempts on a tooth and on the skull's petrous bone failed, as they had for years. So the teams went after the mineralized plaque on the teeth, the dental calculus, and it worked: they recovered mitochondrial DNA that falls squarely inside known Denisovan variation. The proteomic analysis was richer still, identifying more than 20,000 peptides and confirming 95 endogenous proteins, three of them carrying variants unique to Denisovans, tying the skull directly to the original Denisova finger-bone specimen. One of the Cell paper's authors is Svante Paabo, whose Neanderthal genome work anchors our sister file. This was the first time a virtually complete human skull had been definitively linked to the Denisovan lineage. A people known only from a finger bone, a handful of teeth, and a jawbone suddenly had a face.

The way it got resolved is worth pausing on, because it is a small masterclass in how science is supposed to work. Back in 2021 Chris Stringer had warned that naming Harbin as an entirely new species risked inflating the apparent diversity of the human family tree, when the skull might simply belong to an already-recognized group that just happened to have no DNA of its own. That was not timidity. It was calibrated caution, and the 2025 evidence proved it well-placed: Harbin was not a new branch after all, it was the missing face of a branch we already knew from its genes. The lesson for everything that follows is blunt. A ghost with a name is not the same as a ghost with a body, and it pays to keep the two apart.

The size of it is startling. Depending on the population and the modeling assumptions, an estimated 2 to 19 percent of these West Africans' genomes appears to descend from an unknown archaic hominin. The best estimate for when that ghost lineage split off from the common ancestor of modern humans, Neanderthals, and Denisovans is roughly 625,000 years ago, with a confidence range running from about 360,000 to over a million years. In plain terms, something very deeply divergent, older than the Neanderthal split, is woven into the DNA of hundreds of millions of living people. It has no name and no fossil.

The interbreeding itself came much later than the split, and the timing is easy to garble, because the figure often quoted is really just the upper edge of a wide range. Checked against the original paper, the actual best estimate for when the mixing happened is around 43,000 years ago, with a 95 percent range of roughly 6,000 to 124,000 years. So the ghost lineage branched off more than half a million years ago, then rejoined the human story, inside Africa, around the same era our own species was spreading out of it.

And these are not inert fragments. The archaic-derived stretches of DNA in West African genomes cluster near genes involved in hormone regulation and immune function, the same functional signature seen for Neanderthal DNA in the sister file on the Neanderthals still inside you. Whoever these people were, pieces of them may still be doing a job.

So who were they? Here the honesty has to kick in, because nobody knows. No fossil has ever been definitively matched to the West African ghost. There are candidates, skulls with archaic-looking features like Iwo Eleru in Nigeria, roughly 13,000 years old, and Ishango in the Democratic Republic of Congo, but neither has yielded a scrap of recoverable ancient DNA, so neither can be confirmed or ruled out. The ghost stays a ghost.

What keeps this from being one lab's fluke is that the result does not stand alone. Independent teams keep hitting the same wall from different directions. Michael Hammer's group found introgressed archaic haplotypes in Central African hunter-gatherers back in 2011. Joseph Lachance's team found archaic-introgression signals in African whole genomes in 2012 that had nothing to do with the Neanderthal or Denisovan signal. In 2019 Aaron Ragsdale and Simon Gravel showed that formal demographic modeling of African population history simply refuses to fit unless you add at least one ghost archaic population. And Hsieh's team found deeply divergent haplotypes clustered at the MUC7 salivary-mucin gene, a gene under balancing selection. Different populations, different genes, different methods, one conclusion: there was somebody else in Africa.

Now the story folds in on itself. The Denisovans, the archaic humans who gave Dragon Man his face, turn out to carry a ghost of their own. When Kay Prufer's team sequenced the high-coverage Denisovan genome from the Altai Mountains in 2014 and published it in Nature, they found that roughly 1 percent of it came from a population that had split from the common ancestor of modern humans, Neanderthals, and Denisovans more than a million years ago. A superarchaic lineage. A ghost inside a ghost.

How much and how old is where a single study pushes further than the rest. In 2020 Alan Rogers and colleagues used Approximate Bayesian Computation to argue the superarchaic contribution might run as high as roughly 5 percent, and that the source population itself split off around 2 million years ago, deep enough to be Homo erectus or a close relative. That is a bolder claim on wider modeling assumptions than the original 1 percent, and it is worth flagging as one team's extension rather than settled fact.

The consequence is almost dizzying. In 2019 Guy Jacobs and colleagues documented in Cell what they called introgression of introgression. The Denisovan ancestry carried by living people in New Guinea includes variants that had first entered the Denisovan lineage from that superarchaic source. Which means people alive today carry DNA that passed through a lineage more than a million years divergent from us, not by meeting it, but by meeting someone who met it. A genetic relay across a million years, ending in a person breathing right now.

And the superarchaic source itself? Unknown, of course. The candidates are Homo erectus, which was present in Asia for something like 1.8 million years, or Homo heidelbergensis or its eastern equivalent, or a hominin nobody has named yet. We will probably never confirm it genetically, because no ancient DNA has ever been recovered from any Homo erectus fossil, and DNA is thought not to survive much beyond a million years except under permafrost. This ghost may be permanently faceless.

Ghosts are not only deep-time archaic hominins. Some are far more recent, and one of them helped shape modern Europe. In 2014 Iosif Lazaridis and colleagues found that you cannot explain the genomes of Europe's first farmers without invoking a population they named the Basal Eurasians, a group that contributed somewhere between 25 and 45 percent of the ancestry of the early Neolithic farmers who spread across the Near East and into Europe. It is a foundational, heavily cited result.

What makes the Basal Eurasians distinctive is what they lack. They carry less Neanderthal ancestry than other populations outside Africa, which suggests they branched off the main out-of-Africa line before the big Neanderthal admixture event, or caught much less of it. When their ancestry flowed into early farmers, it diluted the Neanderthal signal, and that is the currently favored explanation for a real and slightly odd fact: modern Europeans carry somewhat less Neanderthal DNA than modern East Asians do. A ghost is the reason your Neanderthal fraction depends on which continent your ancestors farmed.

And yet, for all their importance, nobody has ever found them. As of the careful current statement in our source file, no ancient genome has been identified representing a pure, unmixed Basal Eurasian population. Every ancient Near Eastern genome tested, the Natufians, the Pre-Pottery Neolithic farmers, the Epipaleolithic populations of the Maghreb, carries Basal Eurasian ancestry only blended with other components. The source itself has never been sampled. It remains, in the most literal sense, a true ghost.

Once you know the signature, you start seeing ghosts everywhere. In 2015 Pontus Skoglund's team found that two Amazonian populations, the Surui and the Karitiana, share a genetic affinity with Australo-Melanesian peoples that no other Native American group shows. They nicknamed the source Population Y. It hints at either an early migration wave into the Americas that genuinely carried Australasian ancestry, or a distinct southern coastal route that the other first Americans did not travel. The statistical signal is firmer than any one story told to explain it.

Not every ghost stays faceless, and the Ancient North Eurasians are the encouraging case. This population contributed an estimated 14 to 38 percent of Native American ancestry and about 20 percent of European ancestry, the latter carried in much later by the Yamnaya steppe migration. For a while they too were purely statistical. Then real ancient DNA turned up to match them: the Mal'ta boy and the Afontova Gora specimens from Siberia. The Ancient North Eurasians are a genuine, if partial, de-ghosting, proof that a ghost can be dragged into the light when the right frozen bones are found.

Others are barely sketched. Joseph Pickrell's team, in 2012, found a signature of admixture from an uncharacterized East African source inside Bantu-speaking populations of southern Africa, sometimes called the Xhosa ghost, read as an unknown group absorbed during the great Bantu expansion. It is a single study, and the source population has no identity yet.

So is the eastern one. Melinda Yang's 2020 modeling of ancient Chinese DNA found that the data only fits if you add an early-diverging eastern lineage, informally the Eastern Non-African or ENA ghost, feeding into both East Asian and Australasian populations, perhaps tied to an early coastal dispersal out of Africa separate from the main inland route. Model-dependent, single study, genuinely speculative, and named here so the picture is complete rather than tidy.

But the sharpest ghost of all leads straight back to Dragon Man's people. In that same 2019 Cell study, Jacobs and colleagues found that living New Guineans carry Denisovan DNA from at least three genetically distinct Denisovan-related populations, informally D0, D1, and D2. Only one of the three is represented by the actual Denisova Cave genome. The other two are, themselves, literal ghost populations with no fossil at all. So even now, with a face finally on the Denisovan lineage, at least two of its close cousins remain exactly what this whole file is about: people we know only from the DNA they left in the living.

So how do you find a people who left no bones? Entirely in the math, and the methods are worth understanding, because they are also the source of every honest doubt about the whole enterprise. The workhorse is the D-statistic, the so-called ABBA-BABA test, which asks whether two populations share more genetic variants with a third than a simple family tree would predict. It is simple, robust, and everywhere. Its weakness is fundamental: it can tell you that reality deviates from the simple tree, but it cannot tell you what the missing source actually was.

The heavier tools each buy power at a price. Approximate Bayesian Computation simulates enormous numbers of possible population histories and asks which ones reproduce the real data, which is how you get an estimated divergence time or admixture percentage, but the answer depends heavily on which models the researcher chose to test in the first place. Conditional Random Fields, a machine-learning method, can actually map introgressed segments along a person's chromosomes, but it needs reference genomes to train on. And the f-statistic frameworks, qpAdm and qpGraph, allow rigorous testing of specific admixture models, at the cost of model choices that can be somewhat arbitrary and calculations that are computationally demanding. Every one of these is real science. Every one of them is only as good as the assumptions fed into it.

And the field keeps moving. As new ancient DNA comes out of the ground, some ghosts get de-ghosted, the way the Ancient North Eurasians did when the Mal'ta boy turned up. Others go the other way and grow more complex as better math is applied to the same old data, the way a single Denisovan lineage became at least three. The whole discipline is drifting away from tidy branching family trees toward admixture graphs, tangled diagrams that can hold many overlapping mixing events at once, because that messier picture is simply closer to what happened.

The most serious challenge is that a ghost might not be a person at all. In 2018 Daniel Lawson and colleagues raised a live methodological worry: some ghost populations may be statistical artifacts thrown off by a misspecified model rather than real extinct groups. The standard answer is convergence. When several independent methods, D-statistics, Approximate Bayesian Computation, segment mapping, all point at the same ghost in the same case, that agreement is what makes it credible, rather than any single method's assumptions. It is a genuine, unresolved debate, not a settled question, and the good findings survive it precisely because more than one tool sees them.

There is a subtler trap underneath it. As far back as 2011, Pontus Skoglund and Mattias Jakobsson warned that these inferences are sensitive to which reference populations you happen to include. Leave a real but unsampled close relative of the true source out of your comparison, and the math can conjure a spurious ghost to fill the gap. The ghost you detect can sometimes be an artifact of who you forgot to invite to the analysis.

Even a clean detection tells you less than you would like. Knowing that a population split from our lineage around 625,000 years ago does not tell you what it looked like or how to classify it. It could turn out to be a species we already have a name for, like Homo heidelbergensis, that simply left no DNA of its own, rather than a genuinely new and unnamed branch. A divergence date is not an identity, and the Dragon Man story is exactly why: a ghost that looked like a brand new species turned out to be a face we already half knew.

And one hard limitation hangs over almost the entire African side of this research, stated plainly rather than hedged. Ancient DNA barely survives in tropical and equatorial Africa, where heat and humidity destroy it far faster than the cold does, a problem confirmed by our own file on the missing fossil record. So unlike the Eurasian Neanderthal and Denisovan story, which rests on actual recovered ancient genomes, nearly every African ghost is inferred from the DNA of living people alone, with no ancient specimen to check it against. That is a categorically weaker kind of evidence, and the honest move is to say so out loud.

First, ghost DNA is not alien DNA. There is a persistent fringe claim that these unidentified archaic sequences are traces of contact with non-human or non-terrestrial intelligence. The genetics says otherwise, flatly. Every ghost lineage detected so far is phylogenetically consistent with ordinary hominin evolution. They are deeply diverged branches of our own family, extinct cousins, not unknown creatures and not alien genomes. The mystery is real. Its answer is human.

Second, and more important to get right, none of this supports any biological notion of human 'races.' Ghost-population findings have been misappropriated to argue that different human groups are fundamentally different kinds of people. They are not. Every living human shares more than 99.9 percent of their genome with every other. Archaic admixture proportions are small, a few percent, and they vary gradually and continuously across geography rather than sorting humanity into types. The ghosts are shared inheritance, not dividing lines, and any reading that turns them into the latter is both wrong and harmful.