Ancient DNA | Medieval and Early Modern India | Mystery file
Roopkund, the Skeleton Lake
Hundreds of human bones lie in a glacial lake at 5,020 metres in the Himalaya. The DNA says they are not one group of victims but three, a thousand years apart, and one of them came from the Aegean.
Hero image: AI generated illustration of a frozen glacial lake with scattered bones below Himalayan peaks. It is not a photograph of Roopkund.
In 1942 a forest ranger patrolling the Nanda Devi area in what is now Uttarakhand came across a small lake high above the tree line. As the ice thawed, the shallows and the scree around it turned out to be full of human bones. British officials first suspected dead Japanese soldiers who had slipped across the mountains. They were wrong by a thousand years, and, as ancient DNA would reveal in 2019, also wrong by several thousand kilometres.
Roopkund is small, frozen most of the year, and sits on the route of the Nanda Devi Raj Jat, a Hindu pilgrimage held roughly every twelve years. Remains of more than 300 people have been found there. For decades the favourite explanation was a single catastrophe: a band of pilgrims caught by a violent hailstorm in the ninth century, which fits a local legend about King Jasdhaval of Kanauj and his court dying on the way to the goddess Nanda. Radiocarbon dates from Oxford around 850 CE and skull injuries described as blows from round objects falling from above seemed to settle it.
Then the genomes arrived, and the story fell apart.
What the 2019 study found
Harney, Nayak and colleagues sequenced genome-wide DNA from 38 skeletons and combined it with radiocarbon dating and isotopes. The individuals fell into three sharply distinct groups:
- Roopkund A, 23 people with South Asian ancestry, dated between the 7th and 10th centuries CE. They were genetically diverse, spanning much of the variation seen across India today, and their diets varied, with differing amounts of millet.
- Roopkund B, 14 people whose ancestry is typical of the eastern Mediterranean, most similar to present-day Crete and Greece. They died around 1800 CE, roughly a thousand years after group A, and ate a diet with very little millet.
- Roopkund C, a single individual with Southeast Asian ancestry, from the same late period as group B.
No close relatives were detected. Men and women were both present. The authors concluded that the lake did not receive its dead in one catastrophe, but in at least two separate episodes a millennium apart. Who the Mediterranean travellers were, and why they died on a Himalayan pilgrimage path, remains unexplained by any historical record.
A lake full of skeletons is dramatic. A lake full of Greek-looking skeletons from around 1800, in a place no European traveller is known to have reached, is a genuine mystery. We reopened the genomes to see how precisely DNA can place them.
Our data
Global25 coordinates are available for 33 of the 38 Roopkund individuals: 21 from group A, all 11 group B individuals with usable data, and the single group C individual. Coverage is good for this kind of material. Group B individuals sit between 31 and 63 percent of the Global25 SNP set, so their positions are reliable one by one. Group A ranges from about 10 to 52 percent, so the lowest coverage A samples need more caution. The Moriopoulos 2026 collection already splits group B into three profiles: Aegean Greek (8 individuals), Balkan Greek (2) and Central Anatolian Greek (1). We test whether that split holds up.
Three worlds in one lake
A principal component analysis built on modern populations from Greece, Anatolia, the Levant, South Asia, the Himalaya and Southeast Asia shows the problem at a glance.
These are not three shades of one population. The distances between the clusters are of the same order as the distance between Greece and Java. Whatever happened at Roopkund, it drew people from opposite ends of Asia.
Group B: how Cretan were the Mediterranean travellers?
We compared the average of the 11 group B individuals with every modern population in the Global25 panel. The closest match is Crete (distance 0.0093), followed immediately by Cretan Turks (0.0096), the Deep Mani in the southern Peloponnese (0.0116), the Lasithi region of eastern Crete (0.0119) and Milos in the Cyclades (0.0144). Distances below 0.01 are about as close as two populations get in Global25. This agrees with the headline of the 2019 paper, which was based on formal f-statistics rather than Global25.
Two observations follow. First, Cretan Christians and Cretan Muslims are genetically almost identical, because most Cretan Muslims descended from local converts. So DNA cannot say whether these travellers were Orthodox Greeks or Muslim subjects of the Ottoman Empire, a question that matters when asking what brought them to India around 1800. Second, we checked for the Turkic and East Asian ancestry that most Anatolian Turks carry. Group B shows none: a model of Crete, the Peloponnese, Anatolian Greeks and a Turkmen or Han Chinese source gives 0 percent to the latter. A spike-in test shows that as little as 2 percent of either would have been detected. These people were Aegean, not Anatolian Turkish.
One community, or several?
The average hides structure. We placed each group B individual on a PCA built only from modern individuals from Crete (100 samples), the Peloponnese (22), Greek Macedonia (15), Cappadocian Greeks from central Anatolia (13) and Cyprus (21).
To put a number on it, we measured how far each individual lies from the modern Cretan average and compared that with how far real modern Cretans lie from their own average. Among 100 modern Cretans the median distance is 0.027, the 95th percentile 0.034 and the most distant Cretan in the sample 0.037.
- Six individuals (I2869, I3405, I6937, I6939, I3403, I3345) sit comfortably inside the Cretan range. Their nearest modern populations are Crete, eastern Crete, the Mani, the Cyclades and the Dodecanese.
- Two individuals (I3348, I3350) are at the edge of the Cretan range, closest to the Mani, the Cyclades and Greek-speaking southern Italy. They are still plausibly Aegean islanders.
- I3404 and I6936 lie beyond any modern Cretan in our sample and fall among mainland Greeks: I3404 closest to Arcadia, the Taygetos and the Argolid, I6936 closest to Greek Macedonia and the western Peloponnese.
- I6935 is the most distant from Crete of all (0.044) and matches Cappadocian Greeks from central Anatolia best (0.021), then Cyprus and Rhodes. The published dataset already labels this person as an outlier with Near Eastern affinity.
Coverage noise cannot explain this pattern. At 47 to 63 percent coverage these three samples are among the best in the group, and they deviate in consistent directions toward specific, known populations rather than scattering randomly. The Moriopoulos split into Aegean, Balkan and Anatolian profiles is therefore well supported. Group B looks like a party of mostly island Greeks, probably Cretan or from the nearby southern Aegean, accompanied by at least two mainland Greeks and one Greek from Anatolia or Cyprus. That is the profile of a group assembled from across the Greek-speaking Ottoman world rather than a single village.
Group A: pilgrims from all of India
The earlier group is a different puzzle. Its 21 individuals are scattered along the entire South Asian genetic cline, from people close to the tribal populations of southern India to people close to Brahmins and Gujars of the northwest. The average spread of group A individuals around their own mean is 0.064, and 0.074 if we keep only the nine samples above 35 percent coverage. By comparison, individuals within a single modern Indian community (Tharu, Tamil Iyer Brahmins, Chamars of Uttar Pradesh) spread around 0.02. Group A is three times more diverse than any single community.
To express this on one scale we modelled each A individual with three sources: the Irula, a south Indian tribal population; Jat Sikhs of Punjab, rich in northwestern ancestry; and Tibetans, as a proxy for the East Asian ancestry common in the Himalaya. The model is well conditioned (condition number 4.8).
Two results stand out.
The range is enormous. The south Indian tribal-related share runs from 11 percent (I6934) to 88 percent (I6946). Several of the most extreme individuals have excellent coverage, so this is not noise. Some of the dead resemble today's Brahmins of Gujarat and Uttar Pradesh, others the lower caste and tribal populations of central and southern India. Their dates also spread from the early 8th to the mid 10th century. This fits the idea of pilgrims drawn from many regions and social groups over several generations, rather than a single royal party.
They do not look local. Modern Garhwali Brahmins and Rajputs, the people who live around Roopkund today, carry 18 to 22 percent Tibetan-related ancestry in this model. Not a single group A individual does: the highest value is 2 percent. We spiked Tibetan ancestry into the group A average and into a high coverage individual: 2 percent is invisible, 5 percent comes back as about 2 to 3 percent, 8 percent as about 5 to 6 percent. The detection floor is therefore about 3 to 5 percent, and the Garhwali level of 18 to 22 percent would have been obvious. The dead of Roopkund A were not Himalayan hill people. They came up from the plains.
One caution is necessary. Our Garhwal references are small (two Brahmins and one Rajput), and we do not know how old the East Asian ancestry of modern Garhwalis is. If it arrived after the 10th century, local people of that time might have looked more like plains Indians. Even so, the full spread of group A, from Gujarat-like to Tamil tribal-like profiles, is hard to reconcile with a single local community.
Group C: one traveller from Southeast Asia
The lone individual I6940, dated around 1750, is closest to Javanese (0.038), then Balinese and several Austroasiatic-speaking groups of mainland Southeast Asia such as the Khmer, Kuy and Mlabri (about 0.047). He or she is far from Tibetans (0.26) and from the Tibeto-Burman populations of the Himalaya. With one sample we cannot say more than that this person's ancestry points to maritime or mainland Southeast Asia, not to the mountains around the lake.
So what happened at Roopkund?
The genetic evidence rules out the single catastrophe. The lake holds at least two separate disasters a thousand years apart.
The first, spread over the 8th to 10th centuries, killed people from across the Indian subcontinent and its social spectrum, but apparently not locals. This fits pilgrims on the Nanda Devi route, the kind of mixed crowd the legend of a royal party and its entourage would also describe. Whether a hailstorm killed them, or several accidents over generations, the DNA cannot tell.
The second, around 1800, killed a group of Greek-speaking Ottoman subjects, mostly islanders with a strong Cretan or southern Aegean signal, plus at least two mainland Greeks and one Anatolian or Cypriot Greek, together with one person of Southeast Asian ancestry. No known document records such a party in the Garhwal Himalaya. They could have been merchants, travellers attached to a larger caravan, soldiers of fortune, or people swept into a pilgrimage for reasons we cannot reconstruct. Their mixed Greek origins fit a group that formed on the road rather than one that set out from home together.
Two limits deserve emphasis. Ancestry is not identity: DNA shows where someone's ancestors lived, not their religion, language or citizenship, so "Greek" here means genetically Aegean. And the dates for group B carry wide uncertainty; the 2019 paper places them between the 17th and 20th centuries, centred around 1800.
Global25 coordinates used in this article
Group averages from the Moriopoulos 2026 collection and individual group B coordinates from Global25 (Davidski). Paste directly into the Vahaduo tool.
India_Uttarakhand_Early_Medieval_Roopkund_c.800CE_(South_Asian_Profile)_(n=21),0.045583,-0.062286,-0.145263,0.107775,-0.070123,0.057837,-0.002026,0.010109,0.035003,0.019421,-0.003619,-0.000528,-0.000736,0.000832,-0.003697,-0.003277,0.001018,0.001906,-0.000341,0.002817,0.001165,-0.001867,0.001925,0.000556,-0.002902
India_Uttarakhand_Late_Modern_Roopkund_(Aegean_Greek_Profile)_(n=8),0.106994,0.144459,-0.012068,-0.044453,0.008117,-0.015025,-0.001586,-0.000663,-0.002966,0.016811,0.001015,0.002810,-0.002527,0.002202,-0.014200,-0.000414,0.008426,0.000142,0.000786,-0.004080,-0.001045,0.002040,-0.002157,0.000181,-0.000434
India_Uttarakhand_Late_Modern_Roopkund_(Balkan_Greek_Profile)_(n=2),0.118945,0.142682,0.000377,-0.018249,0.021697,-0.005856,-0.001058,0.001038,-0.003988,0.009021,-0.001299,-0.000974,-0.003940,0.006331,-0.014386,-0.004508,0.008344,0.000507,0.002828,-0.006566,-0.009732,-0.001484,0.002403,-0.001566,-0.003054
India_Uttarakhand_Late_Modern_Roopkund_(Central_Anatolian_Greek_Profile)_(n=1),0.108132,0.144205,-0.044123,-0.061047,-0.008309,-0.025100,-0.000705,-0.007154,-0.018816,0.012392,0.011367,0.000150,-0.004608,0.004954,-0.011672,-0.005038,0.007041,0.003801,0.005656,-0.010880,-0.005116,0.000618,0.001849,0.001807,0.006227
India_Uttarakhand_Late_Modern_Roopkund_(Southeast_Asian_Profile)_(n=1),0.009106,-0.397072,-0.088246,-0.024225,0.154183,0.067770,-0.002585,-0.001846,-0.009408,-0.018224,0.096621,0.008692,-0.013379,0.013625,-0.000950,0.002917,0.002347,-0.006841,-0.002388,0.022261,-0.006613,0.007048,-0.013927,0.014219,0.040236
India_RoopkundB_oNearEast:I6935,0.109270,0.144205,-0.041860,-0.062339,-0.009232,-0.025100,-0.000470,-0.006923,-0.018816,0.012757,0.011367,-0.002398,-0.003717,0.004954,-0.013301,-0.003447,0.005867,0.003167,0.004902,-0.009755,-0.005241,0.000866,0.003821,0.002892,0.007784
India_RoopkundB:I2869,0.103579,0.151314,-0.020742,-0.043928,0.011079,-0.011435,0.000940,-0.005307,-0.004500,0.015855,-0.000487,-0.000300,-0.000595,0.003716,-0.016965,0.000796,0.009127,0.004307,0.007039,-0.000625,-0.005366,0.007048,-0.005916,0.002289,0.000599
India_RoopkundB:I3345,0.103579,0.140143,-0.016970,-0.045220,0.009540,-0.020638,0.000470,-0.004846,-0.004295,0.012028,0.003248,0.001649,0.002081,-0.013074,-0.011672,0.010209,0.024251,-0.002534,-0.000503,0.002001,-0.003119,0.003339,-0.001602,0.005543,0.005987
India_RoopkundB:I3348,0.104717,0.143190,-0.001508,-0.042313,0.013541,-0.017012,-0.001410,0.001846,0.008795,0.013850,0.000974,0.015137,0.004162,0.007019,-0.019272,-0.014983,-0.003651,-0.003674,-0.002765,-0.016883,-0.006738,0.002844,-0.005176,-0.002048,-0.005628
India_RoopkundB:I3350,0.105855,0.149283,-0.008297,-0.037791,0.003385,-0.015618,-0.005640,-0.003000,-0.010635,0.013850,0.002923,-0.003297,-0.011001,0.002202,-0.003257,0.009546,0.026859,0.002407,-0.000880,0.007003,0.000749,-0.002226,-0.004683,0.004579,-0.003353
India_RoopkundB:I3403,0.106994,0.140143,-0.010559,-0.044251,0.013541,-0.012271,-0.005170,-0.007384,0.001023,0.029522,0.000162,-0.001049,-0.004608,0.012661,-0.017101,0.000398,0.006389,-0.003927,-0.004777,-0.010005,-0.001622,0.004328,0.003821,-0.010483,-0.000838
India_RoopkundB:I3404,0.111547,0.138112,-0.000377,-0.021641,0.022773,-0.003904,0.001410,0.002769,-0.007976,0.017130,-0.001299,0.001199,-0.002676,0.006468,-0.010179,-0.006762,0.001434,-0.000887,-0.000754,-0.005127,-0.013726,-0.004699,0.000493,-0.001084,-0.009101
India_RoopkundB:I3405,0.108132,0.143190,-0.011691,-0.049096,0.004001,-0.013108,-0.005875,-0.004384,-0.009204,0.020593,0.003085,0.006145,-0.013231,-0.000138,-0.013843,0.002917,0.007432,0.000760,0.015084,-0.007629,-0.001747,0.002349,0.005669,-0.003735,0.003592
India_RoopkundB:I6936,0.121791,0.145221,0.001131,-0.016473,0.018157,-0.004741,0.001410,0.004846,-0.004295,0.001093,0.003410,0.001798,-0.003865,0.011423,-0.014929,-0.002784,0.007562,0.000127,0.000126,-0.009505,-0.007861,0.001855,0.003081,-0.002169,0.002275
India_RoopkundB:I6937,0.102441,0.151314,-0.014331,-0.043282,0.012002,-0.015060,-0.005640,0.001154,-0.002045,0.017312,-0.000162,-0.001649,-0.001635,0.003991,-0.019001,-0.015380,-0.004955,0.009122,-0.007039,-0.008004,-0.002745,0.001978,-0.001109,0.006145,0.003592
India_RoopkundB:I6939,0.126344,0.147252,-0.011314,-0.052649,-0.002462,-0.015060,0.000940,-0.003692,-0.002659,0.017312,0.000162,0.002398,0.006244,0.002202,-0.012758,0.008353,0.001173,0.004814,-0.003645,-0.002126,-0.002121,0.004575,0.000739,-0.001084,-0.002036
Methods in brief
Global25 scaled coordinates were merged with first occurrence priority from Davidski's population averages, modern averages, the Moriopoulos 2026 collection and the dated individual ancients file; individual Roopkund coordinates come from the Global25 individual file. Distances are Euclidean over 25 dimensions. NNLS models use a sum to one constraint enforced by a heavily weighted extra equation, with results normalised; condition numbers are reported. Spike-in tests add a known fraction of a source to the target and re-fit to determine the detection floor. The Cretan range in Figure 3 uses the 100 modern Cretan individuals in the Global25 modern file. G25 is a proxy; the formal qpAdm and f-statistic results of the original paper take precedence where they differ.
References
- Genomics Harney E, Nayak A, Patterson N, Joglekar P, Mushrif-Tripathy V, et al. (2019). Ancient DNA from the skeletons of Roopkund Lake reveals Mediterranean migrants in India. Nature Communications 10: 3670. doi.org/10.1038/s41467-019-11357-9
- Press release Max Planck Institute for the Science of Human History (2019). Biomolecular analyses of Roopkund skeletons show Mediterranean migrants in Indian Himalaya. shh.mpg.de
- Commentary Hawks J (2019). Geneticists work to understand how skeletons wound up in a mysterious Himalayan lake. johnhawks.net
- Background Roopkund. Wikipedia. en.wikipedia.org/wiki/Roopkund
- Data Global25 PCA coordinates (Davidski) and the Moriopoulos 2026 collection of Global25 averages.