David Reich – Why the Bronze Age was an inflection point in human evolution

David Reich is a geneticist at Harvard who works on human history — how ancient people related to each other, and to people living today.

Dwarkesh notes that their interview two years earlier became one of the most popular he has ever done, and that people found it compelling precisely because so much of human history is unknown and is only now being learned through the techniques Reich's lab uses.

The new preprint is the occasion for this conversation, and Dwarkesh asks for context.

Reich explains the original dream. When ancient DNA began, sixteen or seventeen years ago, the hope was that by extracting DNA from ancient human remains and tracking changes over time, we would learn a great deal about biology — about how people's biology changed.

That dream has not been realised. The field has been an enormous success for history: surprising findings about migrations, people not being descended from those who lived in the same place centuries or millennia before, mixture being common, sex-biased processes being common, and results archaeology did not predict. But learning about biology and biological change has not succeeded.

The main reason was sample size. A single person's DNA carries a tremendous amount of information about history, because one person's genome is not one person — it is two parents, four grandparents, eight great-grandparents, and so on. Thousands, tens of thousands, even hundreds of thousands of ancestors contribute to people today.

So a single genome, or a single Neanderthal genome, effectively represents tens of thousands of ancestors, and lets you position that individual exquisitely relative to everyone else you have data for. But if you want to know how one genetic variant — one affecting skin pigmentation, or the ability to digest milk into adulthood, or a behavioural trait — changes over time, a single person gives you one sample, or two: the one in their mother and the one in their father.

To get a high-resolution picture of frequency change you need genuinely large numbers of people, and we simply did not have them until the last few years. What motivates this study, and the work several groups will presumably do in the coming years, is that those numbers finally exist.

Dwarkesh warns that he will be asking a lot of naive questions, and starts with one: why are frequency changes especially interesting?

The point is to use the experiment of nature that has already run over the last tens of thousands of years to find out what is biologically significant in our DNA. If a population's environment changed — shifting to agriculture, living close to domesticated animals, moving from cold to warm, from lowland to highland — there was pressure to adapt to new stresses and new needs.

You detect that by seeing the frequency of a variant push systematically in one direction: a variant that lets you live at altitude, or that nudges you toward a behavioural pattern advantageous in the new situation. Slight shifts of a few percent or ten percent are very hard to detect without large samples. What the study hunts for are frequency changes too extreme to be chance, which would show that biology was pushed by a change in environment.

Dwarkesh asks what they found. Seven years ago Ali Akbari, then a postdoc in Reich's lab and later a permanent staff scientist, set out to use the data the lab was producing to study biological change over time. He was drawn to this lab because it had made a focus of generating truly large amounts of ancient human data — industrialising the process, making it cheap and high quality.

That volume of data made it possible to ask again whether frequencies have changed over time. The mainstream view in human evolution for several decades had been that natural selection was fairly quiescent over the last several hundred thousand years.

Several lines of evidence were deployed for this. One: compare diverse populations from different continents, say Europeans and East Asians, and look at mutations that differ in frequency. Every mutation differs a little, sometimes a lot, but there are almost no genetic changes that are 100 percent different in frequency between the two groups.

Europeans and East Asians descend from a common ancestral population that left Africa and the Middle East 40,000 or 50,000 years ago. That population had a set of gene frequencies, and the variants bopped around randomly — genetic drift — or perhaps under selection. The time since is short on an evolutionary timescale, so on average there is just not much differentiation.

But if there had been selection — to digest alcohol better, or milk better — you would expect some mutation to have rocketed to very high frequency. Forty or fifty thousand years is 1,500 to 2,000 generations, easily enough to produce a 100 percent difference. Yet you see nothing beyond what chance predicts.

Taken together, this made selection look quiescent: perhaps a few hundred thousand years ago the ancestral human population reached some optimum, and there has not been much genetic change since.

There has been some selection, including selection to remove bad mutations constantly raining down on the genome, but not directional selection — newly arising mutations, or mutations pushed systematically, to move a population to a different adaptive set point better suited to its conditions.

The study partitions the frequency changes across roughly ten million variable positions into directional selection versus everything else. Ninety-eight percent is everything else, especially genetic drift. It is overwhelmingly migration and population structure causing fluctuations.

Which is why adaptive signals are so hard to detect: they are a tiny fraction of total frequency change, while the vast majority is migration and mixture. And yet, as the study shows, there is so much natural selection that it has been rampant in the genome.

Dwarkesh asks a clarifying question: why discount admixture or population replacement as selection? At the group level, if one population replaces another, is that not selection?

He recalls the earlier episode: enormous changes in who lived in a given area, one population arriving and replacing the previous one, then another replacing that. To the extent genetics is relevant to why one replaced the other, why should that not count as selection over the last ten thousand years?

Reich grants that it could count, may count, and probably should in some respects. But the replacement could equally be a cultural phenomenon — technology held by one group and not another — with some genetic mutations contributing. Who knows? It is possible.

The statistical problem is that what you see is a whole-genome shift. What you want is one place in the DNA driving change differently from the rest of the genome. At times of migration there are huge fluctuations in frequency across the board, which makes those periods extremely uninformative for detecting selection. The best moments are the few hundred years when no migration or admixture is happening, because then you can watch a mutation slowly blowing in one direction.

So the study treats the history of Europe and the Middle East as an archipelago of small populations in space and time, each fairly isolated between big events. A little population in Britain isolated for a few hundred years; a little population in Hungary isolated for a few hundred years. In each of those experiments of nature: does this mutation rise slightly? Does the same mutation rise slightly there too? If all the arrows point the same way, we win — that is selection.

For example, 4,500 years ago in Europe almost all mutations went through huge frequency changes, and not because of selection. It is the steppe migration from north of the Black and Caspian seas: 40 to 80 percent of the DNA becomes Yamnaya from steppe pastoralists. Their mutation frequencies differed not necessarily through selection but because they had evolved in different places for thousands and tens of thousands of years. In the descendant populations you see enormous frequency change, so what you have to do is ask whether selection explains a shift beyond chance.

Dwarkesh checks a number: 3,800 locations at 50 percent confidence of having been under selection in the last 10,000 years?

It is 7,200 at 50 percent confidence — and since only half of those are real, and we do not know which half, about 3,600 are real.

Does that mean everything outside those 7,200 is confidently not under selection?

No. At a 25 percent cutoff there will be tens of thousands, many of them real too. Multiple analyses suggest the genome is vibrating with natural selection.

There are all sorts of weaker effects that even larger studies will pick up. Almost every position in the DNA is correlated with another position being dragged one way or the other by selection. Rather than quiescent, selection is everywhere — even though it accounts for only 2 percent of frequency change, it is tugging positions in one direction or another throughout.

They then analysed the positions they had identified — the hundreds they were very confident about — to see whether they were randomly distributed or patterned. They looked at around a hundred traits with genome-wide association studies: immunity, autoimmunity, behaviour, metabolism and others.

For each, they asked whether the variants known to affect that trait carry an unusual number of selection signals. The answer was a vast enrichment, four- or five-fold, for immune traits — a super-concentration of selected signals there. There was also strong enrichment for metabolic traits, the kind that affect obesity, fat traits or Type 2 diabetes, and almost no detectable enrichment for behavioural or psychiatric traits.

Dwarkesh checks the interpretation: this does not mean behavioural, psychiatric or cognitive traits are not under selection — only that the individual sites controlling such traits are not especially likely to be among the locations identified.

Dwarkesh has it exactly right, and Reich says the wrong conclusion to draw is that immune traits are highly selected and behaviour has not been selected at all in this part of the world over the last 18,000 years. There is clear evidence of selection on behavioural traits too.

The reason behavioural signals look much weaker is a property of the traits, not of selection. Medical studies show behavioural traits are underpinned by far more genes than immune traits, which rest on relatively few genes of strong effect. Behaviour is shaped by a very large number of genes of weak effect, and the study simply lacks the statistical power to detect such weak signals.

An analysis restricted to the strongest signals is effectively querying immune traits well and behavioural traits badly. Immune traits may still be the most selected category — but it is not at all the case, and can be proved not to be the case, that behaviour is unselected.

Dwarkesh says that is interesting.

Reich says they can prove there are two ways to reconcile the old picture with the new one. The old picture: selection has been quiescent over hundreds of thousands of years, because you do not see 100 percent frequency differences between Europeans and East Asians.

The new picture: hundreds of positions rocketing in frequency, often at selection rates of 1 percent or more, which means rapid doubling over dozens of generations. Over the 1,500 or 2,000 generations separating Europeans and East Asians, should you not then see many variants that are 100 percent different in frequency?

The first explanation is that in Europe and the Middle East we are in a period of accelerated selection. One way to see it is the enrichment pattern: immune traits are unusually associated with these signals. Compare the last 5,000 years — the Bronze Age and after — with the previous 5,000, and the intensification around immune traits, and likewise around metabolic traits, has accelerated.

Selection has not run at the same rate everywhere and always; it is increasing across the period analysed, and plausibly the whole period is elevated relative to earlier ones. That is not implausible, because this is a population that went through a huge shock in how people lived. Almost everyone analysed is a farmer or food producer of some kind. Farming was invented for the first time anywhere in the Middle East 11,000 or 12,000 years ago; the people who invented it exploded into Europe after 8,500 years ago and expanded rapidly across the continent.

In the Bronze Age that intensifies further: much higher population densities, people living more and more beside their animals, catching their diseases and exchanging diseases with the animals and with each other. It is a period of rapid change in how people live, generating different biological needs. It is perhaps not surprising that under such dramatic change the population's biology would not be ideally adapted.

This is what some call evolutionary mismatch: take a genetic variant that evolved in hunter-gatherers, put it into farmers or pastoralists, and it is not quite right. What you are seeing is the DNA of a population that descended from hunter-gatherers only 10,000 years ago, reacting to the shock of being moved into an agricultural, Bronze Age, high-density, urban environment. The hypothesis is that what the data shows is the adaptation that occurs as a result.

Dwarkesh notes the paper has many examples of this intensification around the Bronze Age, and suggests going through some of them.

Reich explains that the work examines many of these positions closely. They built an internet tool, the AGES browser, with Ali and a co-author, which lets anyone query each of the ten million positions and see the trajectory and the evidence for selection at each one.