Bengalis — The Fourth Ancestor and the Missing Reference Panel

 

Copyright: Sanjay Basu

The fourth ancestor in Bengali DNA is real. But the two claims that made the video spread are not really about Bengali biology. They are about missing data.

I watched a sixteen minute video on this last week. Most of it holds up. The East Asian piece is really there. The date is close to right. The way it rises as you move east across Bengal is a real finding. Someone did the homework.

The two moments built to make you share it are the ones that break. A software tool that locates people by their DNA could not place Bengalis on a map. A team in Brazil looking for the closest living Asian relatives of Native Americans ended up pointing at Bangladesh. The video tells both stories as if the DNA were so strange that the instruments gave up.

That is not what happened. In both cases the tool worked fine. The list it was comparing against simply did not contain the right answer. That difference is the whole point, and I run into the same mistake every week at work.

Here is the fact check, and then the part nobody is talking about.

What the video gets right

Seven of the main claims hold up against the published research. Two need a sharper number.

That is a better record than most science videos manage. Now the parts that need fixing.

Three claims that need a correction

Iranian farmers did not bring farming

The video says that people related to the early farmers of ancient Iran moved into the northwest about 9,000 years ago, and that farming grew out of that mixing. That was the standard story until ancient DNA tested it.

Skeletons from Rakhigarhi and from sites near the Indus tell a different story. The Iranian related ancestry in South Asia split off before Iranian farmers, herders and hunter gatherers separated from each other. That split happened before farming existed. The authors say it plainly. This contradicts the idea that farming spread east from Iran on the back of a migrating farming population.

So farming in South Asia was either invented locally or picked up as a practice. It was not carried in by a wave of farmers. The video repeats the old version. Worth correcting before you pass it on.

The software did not break

The tool is called GPS, short for Geographic Population Structure. It placed about 98 percent of Bengali genomes more than 1,000 km away from Dhaka. That number is accurate.

The explanation is not. GPS works by comparing your DNA mix against reference groups whose home locations it already knows, then picking a point somewhere between them. The researchers ran it using reference groups from Pakistan and the northwest. Burusho, Kalash, Balochi, Hazara, Makrani, Pathan, Sindhi, Brahui. There was no Bengali group on that list. There was no Punjabi group either, and the Punjabi results were also poor.

The authors say so in their own summary. The missing reference groups caused the bad results. The video says this was not a software problem but the DNA itself. The people who ran the test concluded roughly the opposite.

A tool that picks a point between known locations does not warn you when the real answer sits outside its list. It just hands you a wrong location with complete confidence. Anyone who has built a search system has seen this.

Brazil was a short list, not a mystery

A team studied 1,538 Brazilians from the northeast of the country. They needed an Asian group to compare against the Native American part of Brazilian ancestry. They used the Asian groups in the 1000 Genomes database that had been collected inside Asia. That is seven groups. Bengalis in Bangladesh, 86 people, came out closest.

There is no Siberian group in that database. No Mongolian, no Tibetan, no Chukchi. The paper itself says the ancestors of Native Americans are believed to have come from northeastern Siberia. Winning a seven horse race that left out the favorite does not tell you much about the winner.

The paper also explains the result without any mystery. Their family tree model puts the Native American piece closest to an old mixed group formed from a north Indian branch and an older Southeast Asian branch. Bengalis score well because a South Asian plus East Asian blend is the nearest shape available to that ancient blend. That is not ancestry. That is arithmetic.

The date is not about who arrived. It is about when the mixing stopped

Here is the part the video leaves out, and it is the better story.

Karlsson’s team dated the East Asian mixing to 52 generations ago, about 500 CE, and found that a single event fits the data better than two. Basu and colleagues, working on a completely different set of 367 people from 20 populations, measured something else. They looked for the moment when marrying across groups gave way to marrying strictly inside your own community. Their answer is about 70 generations ago, roughly 1,575 years, mostly among upper castes and Indo-European speakers.

Two separate studies, two different questions, the same window of time. The Gupta period.

That changes what 1,500 years means. It is not only when a group of people reached the delta. It is close to when the subcontinent stopped mixing. The East Asian ancestry arrived, and then the social machinery that would have spread it around was switched off.

That is why we can still see it. A single wave of mixing into a population that keeps marrying across groups fades into the background within a few dozen generations. The same wave into a population about to split into thousands of closed marriage circles gets locked in place, with a different amount in each one. That is exactly what the data shows. A steady rise from west to east across the delta, plus sharp differences between neighboring communities.

The fourth ancestor is not interesting because it arrived. People arrive all the time. It is interesting because almost no further mixing happened for fifteen centuries afterwards.

This has real consequences. When a small group marries within itself for that long, many people end up descended from a handful of ancestors. Nakatsuka and colleagues checked 263 South Asian groups and found 81 of them, including 14 with over a million members, descended from bottlenecks tighter than those behind Ashkenazi Jewish and Finnish populations. Both of those are known for high rates of inherited recessive disease. The 2025 study of 2,762 Indian genomes shows the same picture in more detail. The closure that preserved the beautiful signal also concentrated the medical one.

The same gap, now in a hospital

Both viral claims come from one root. The reference list was missing the answer, and the method gave a confident result anyway.

That is not a quirk of two papers. That is the state of the field.

One group of 86 people from Dhaka stands in for about 270 million Bengali speakers. The Gujarati reference group was collected in Houston. The Tamil and Telugu reference groups were collected in Britain. Nakatsuka’s team counted 263 distinct South Asian groups, and most groups in the region have never been sequenced at all.

The cost of that is not a map pin landing in the wrong district.

Genetic risk scores are built from huge studies like UK Biobank, which is about 94 percent European ancestry. Martin and colleagues measured what happens when you use those scores on everyone else. They predict about 4.5 times better for people of European ancestry than African ancestry, and about twice as well as for East Asian ancestry. The further you get from the group the score was built on, the worse it does. It is the same guessing failure as the map pin, wearing a lab coat.

Now apply that to a Bengali genome, which carries a South Asian and East Asian blend that no large study contains. The score is not just weaker. It is off in a direction nobody has measured.

Meanwhile the gene variants that matter most here were found last. The MYBPC3 deletion raises the odds of a serious heart muscle disease about sevenfold. Roughly 4 percent of people across the subcontinent carry it, and it is almost absent everywhere else. That is around 60 million people. It was identified in 2009 by a team working in Hyderabad, because they looked in the right population.

The 2025 study makes the point from the other side. From 2,762 Indian genomes the authors rebuilt about half the Neanderthal genome and a fifth of the Denisovan genome, more than any study before it, and found that India holds more variety in Neanderthal ancestry than any other non-African region studied.

That variety was always there. The sampling was what was missing. Put the population on the list and the tools stop producing nonsense.

An ancestry percentage is not a fact about you

This is the habit of mind I want to leave you with, because it reaches far beyond genetics.

When a report says you are 9 percent East Asian, that number is not sitting in your cells waiting to be read. It is the result of a calculation that compares you against a particular list of reference groups. Change the list and the number changes. Add Tibetan, Burmese and Khasi groups to it, and that 9 percent splits into pieces with different names and different sizes.

That does not make the estimate fake. The mixing really happened, and the DNA clock is real physics. But sorting it into named ancestors is a description relative to the list you picked, and the list is a human decision about who got sampled.

This is an analogy and nothing more, but it is the habit QBism trains. Stop treating a number as a reading of what is out there. Start asking whose information it is, based on what, and what got left out. In quantum theory that view is argued over, and I am on one side of the argument. In ancestry testing it is not argued over at all. It is simply forgotten the moment a percentage gets printed.

The same forgetting runs through my day job. An embedding is a position relative to a training set. A benchmark score is performance on a task list somebody picked. A search system returns the nearest thing it has, not the right thing, and it does not tell you which one you got. Every one of these will hand you a confident answer drawn from an incomplete list, and none of them will mention the gap.

GPS put Bengalis a thousand kilometers from Dhaka and reported no error.

What I take from it

Three things.

First, the video is worth watching and mostly worth trusting. I would rather ten thousand people learn that Bengal has a fourth ancestral piece with a slightly shaky date than learn nothing at all. The correction belongs in the margin, not in a takedown.

Second, be suspicious of any science story whose big moment is a machine failing. That is almost never the discovery. It is the setup, and the real finding is usually about what got left out of the data.

Third, the interesting number here is not 9 percent. It is 263. That is how many distinct South Asian groups have genome wide data, out of thousands that exist, in a region of 1.8 billion people. Bengal looked unplaceable because the map had no Bengal on it. That is a sampling budget problem, and budgets are decisions that institutions make.

I grew up in Calcutta. The claim that a computer built to locate anyone on Earth could not find Dhaka sounds like a mystery. It is actually a bill.

Somebody should pay it.

Sources

Every paper below was opened and checked against the claims in this piece.

Primary literature

1. Karlsson EK, Harris JB, Tabrizi S, et al. Natural selection in a Bangladeshi population from the cholera-endemic Ganges River Delta. Science Translational Medicine 2013;5(192):192ra86. doi:10.1126/scitranslmed.3006338 — the 52 generation date, the single pulse model, and the 9.3 percent estimate.

2. Basu A, Sarkar-Roy N, Majumder PP. Genomic reconstruction of the history of extant populations of India reveals five distinct ancestral components and a complex structure. PNAS 2016;113(6):1594–1599. doi:10.1073/pnas.1513197113 — ANI, ASI, AAA and ATB, and the shift to endogamy about 70 generations ago.

3. Narasimhan VM, Patterson N, Moorjani P, et al. The formation of human populations in South and Central Asia. Science 2019;365(6457):eaat7487. doi:10.1126/science.aat7487–523 ancient genomes, the Indus Periphery Cline, Steppe ancestry timing.

4. Shinde V, Narasimhan VM, Rohland N, et al. An ancient Harappan genome lacks ancestry from Steppe pastoralists or Iranian farmers. Cell 2019;179(3):729–735. doi:10.1016/j.cell.2019.08.048 — the Rakhigarhi individual and the pre-farming split of the Iranian related lineage.

5. Das R, Upadhyai P. Application of geographic population structure (GPS) algorithm for biogeographical analyses of populations with complex ancestries: a case study of South Asians from 1000 Genomes Project. BMC Genetics 2017;18(Suppl 1):109. doi:10.1186/s12863–017–0579–2 — the 98 percent misplacement, and the authors’ own attribution to missing reference populations.

6. Mychaleckyj JC, Havt A, Nayak U, et al. Genome-wide analysis in Brazilians reveals highly differentiated Native American genome regions. Molecular Biology and Evolution 2017;34(3):559–574. doi:10.1093/molbev/msw249 — the Bengali result, the seven population comparison set, and the TREEMIX explanation.

7. Nakatsuka N, Moorjani P, Rai N, et al. The promise of discovering population-specific disease-associated genes in South Asia. Nature Genetics 2017;49(9):1403–1407. doi:10.1038/ng.3917–81 of 263 groups with founder events exceeding Ashkenazi Jewish and Finnish populations.

8. Kerdoncuff E, et al. 50,000 years of evolutionary history of India: Impact on health and disease variation. Cell 2025. doi:10.1016/j.cell.2025.04.027–2,762 whole genomes, archaic ancestry reconstruction, endogamy and disease.

9. Dhandapany PS, Sadayappan S, Xue Y, et al. A common MYBPC3 (cardiac myosin binding protein C) variant associated with cardiomyopathies in South Asia. Nature Genetics 2009;41(2):187–191. doi:10.1038/ng.309 — the 25 base pair deletion, 4 percent carrier frequency, combined odds ratio 6.99.

10. Martin AR, Kanai M, Kamatani Y, Okada Y, Neale BM, Daly MJ. Clinical use of current polygenic risk scores may exacerbate health disparities. Nature Genetics 2019;51(4):584–591. doi:10.1038/s41588–019–0379-x — portability decay across ancestries.

11. Kumar V, Reddy ANS, Babu JP, et al. Y-chromosome evidence suggests a common paternal heritage of Austro-Asiatic populations. BMC Evolutionary Biology 2007;7:47. doi:10.1186/1471–2148–7–47 — O2a-M95 frequencies in Munda speaking groups.

12. Chaubey G, Metspalu M, Choi Y, et al. Population genetic structure in Indian Austroasiatic speakers: the role of landscape barriers and sex-specific admixture. Molecular Biology and Evolution 2011;28(2):1013–1024. doi:10.1093/molbev/msq288 — evidence for the Southeast Asia to South Asia direction of the Austroasiatic dispersal.

Source video

13. https://www.youtube.com/watch?v=pNc_XCsL8vw&t=2s — the video whose transcript prompted this piece. Claims audited above.

As of 1 November 2026. Generation times differ between papers (29 years in Karlsson, 25 to 30 elsewhere), so dates derived from generation counts carry roughly a century of slack in either direction.