population structure

Science today has released the new paper on the Denisova high-coverage genome by Mattias Meyer and colleagues from Svante Pääbo's group [1]. There is a lot of material in the supplements of the new paper, and it will take some time to work through implications.

The basics are quite simple: The paper confirms the initial interpretation of the genome by David Reich and colleagues [2] in most respects. The mixture with a whole-genome sample from Papua New Guinea is estimated at 6% Denisovan ancestry. Confirming the later paper by Reich and colleagues [3], the new analysis finds no significant evidence of Denisovan ancestry in a mainland south Chinese (Han Dai) individual, and can exclude it down to a very small fraction:

However, in contrast to a recent study proposing more allele sharing between Denisova and populations from southern China, such as the Dai, than with populations from northern China, such as the Han (17), we find less Denisovan allele sharing with the Dai than with the Han (although non-significantly so, Z = –0.9) (Fig. 4B) (table S25). Further analysis shows that if Denisovans contributed any DNA to the Dai, it represents less than 0.1% of their genomes today (table S26).

That is a mystery to be explained. How did Asians end up lacking any evidence of Denisovan ancestry, when the peoples of Sahul (Australia and New Guinea) have six percent? It's nutty! The early modern humans who were the ancestors of present Sahulian peoples surely came from Asia, and they surely mixed with Denisovans there somewhere, right? But today there's no sign that present Asian peoples descended from those early Asian peoples.

We must, I think, conclude that there was at least one, and possibly several episodes of massive population movement across South and Southeast Asia.

I have recently completed a review of the analogous problem for Neandertals in Europe -- late and early Neandertals themselves appear to have been a dynamic population. I'm now working on a review of the situation in Southeast Asia. We may fundamentally have to look at the archaeological record in a new, and much more dynamic, way than has been the case.

Neandertal gene flow

To me at the moment, this is the most interesting paragraph of the new paper:

Interestingly, we find that Denisovans share more alleles with the three populations from eastern Asia and South America (Dai, Han, and Karitiana) than with the two European populations (French and Sardinian) (Z = 5.3). However, this does not appear to be due to Denisovan gene flow into the ancestors of present-day Asians, since the excess archaic material is more closely related to Neandertals than to Denisovans (table S27). We estimate that the proportion of Neandertal ancestry in Europe is 24% lower than in eastern Asia and South America (95% C.I. 12–36%). One possible explanation is that there were at least two independent Neandertal gene flow events into modern humans (18). An alternative explanation is a single Neandertal gene flow event followed by dilution of the Neandertal proportion in the ancestors of Europeans due to later migration out of Africa. However, this would require about 24% of the present-day European gene pool to be derived from African migrations subsequent to the Neandertal admixture.

This is a very interesting result, partially because it is the opposite of what we are finding. As I explained earlier this year, we are finding Europeans to share more Neandertal alleles than Asians do. The difference in our results has been much smaller than 24%; really only an increase of less than 0.5% on the whole genome, or maybe 10% relative to the overall amount in Europe (which is on the order of 3%).

My initial reaction to this difference is that it reflects the sharing of Neandertal genes in Africa. Meyer and colleagues filtered out alleles found in Africa, as a way of decreasing the effect of incomplete lineage sorting compared to introgression in their comparison. But if Africans have some gene flow from Neandertals, eliminating alleles found in Africans will create a bias in the comparison. If (as we think) some African populations have Neandertal gene flow, that probably came from West Asia or southern Europe. So as long as the present European and Asian (and Native American) samples have undergone a history of genetic drift, or if (as mentioned in the quote) they mixed with slightly different Neandertal populations, this bias will tend to make Asians look more Neandertal and Europeans less so.

Anyway, this demands further investigation. The Denisova genome makes a more compelling outgroup for these kinds of comparisons, because it is much closer to us than chimpanzees are. But it isn't really an outgroup because it shares alleles by descent with Neandertals. So it takes some clever genetics to compare the distributions of derived alleles in these genomes in terms of introgression versus incomplete lineage sorting.

Denisovan demography

It has become possible to make some good estimates of demographic history using only a single diploid genome, using a technique developed by Li and Durbin [4]. Meyer and colleagues applied this technique to the Denisova genome, finding that its genetic history contrasts with that of living human populations:

To estimate how Denisovan and modern human population sizes have changed over time we applied a Markovian coalescent model (22) to all genomes analyzed. This shows that present-day human genomes share similar population size changes, in particular a more than two-fold increase in size before 125,000–250,000 years ago (depending on the mutation rates assumed (23), Fig. 5B). Denisovans, in contrast, show a drastic decline in size at the time when the modern human population began to expand.

There is not yet enough data from Neandertal genomes to apply the same method, but to the extent that we understand their diversity, they show a similar picture. These archaic humans in Eurasia had much, much smaller effective population sizes than the ancient population of Africa. That's not surprising, given what we understand about ancient hunter-gatherer population dynamics.

What may be a bit more surprising is the geography. We know that Neandertals of Europe and Central Asia lived in an environment that was relatively marginal for their technology and subsistence pattern. The Denisovan population could well have lived in parts of South or Southeast Asia -- subtropical and tropical areas comparable to Africa in their ecological diversity and resource richness.

We might have imagined that the Denisovan population would be more diverse than Neandertals -- that it might have been comparable in diversity to part of Africa, if not the entirety of Africa. The genome is inconsistent with that picture.

How can we explain the apparent contrast?

1. Maybe Denisovans didn't live in South or Southeast Asia at all. If not, that demands that we explain how Australians got their genes.

2. Maybe the population was geographically extensive and diverse, but the genome from Denisova Cave doesn't represent it well. If so, we might discover that Sahulians actually have even more ancestry from this group. Alternatively, we might find that the early history of the population was widely shared, but the recent history diverged between Siberian and other branches of the Denisovan-inhabited region.

3. Maybe African diversity emerged from a much more complex series of interactions than we now appreciate. The demographic model of Li and Durban doesn't encompass admixture, just the probability of gene coalescence across time. We have recently begun to appreciate the reality of ancient African population structure. If those initial African populations were more divergent from each other than Neandertals and Denisovans, their later mixture would give rise to a picture of early population expansion, even if each of them had relatively low (Denisovan-like) diversity.

This picture is already complicated. It will get more so. We have a long way to go before the archaeology of MSA and Middle Paleolithic peoples will be reconciled with these genetic models.

The "modern human" catalog

I think it's tremendously interesting that the authors have compiled a list of gene variants shared by living humans that are absent from this high-coverage archaic human genome. It's a first step to identifying networks of genes that have been subject to recent evolutionary change in human ancestors.

That being said, the list of genes itself doesn't lend itself to concrete conclusions:

One way to identify changes that may have functional consequences is to focus on sites that are highly conserved among primates and that have changed on the modern human lineage after separation from Denisovan ancestors. We note that among the 23 most conserved positions affected by amino acid changes (primate conservation score ≥ 0.95), eight affect genes that are associated with brain function or nervous system development (NOVA1, SLITRK1, KATNA1, LUZP1, ARHGAP32, ADSL, HTR2B, CBTNAP2). Four of these are involved in axonal and dendritic growth (SLITRK1, KATNA1) and synaptic transmission (ARHGAP32, HTR2B) and two have been implicated in autism (ADSL, CNTNAP2). CNTNAP2 is also associated with susceptibility to language disorders (27) and is particularly noteworthy as it is one of the few genes known to be regulated by FOXP2, a transcription factor involved in language and speech development as well as synaptic plasticity (28). It is thus tempting to speculate that crucial aspects of synaptic transmission may have changed in modern humans.

Interesting. I can imagine a Ph.D. dissertation looking into the function of each of those genes. It is surely true that in the last 300,000 years, human brains have been evolving. But why these genes as opposed to others? And how many regulatory changes (as opposed to amino acid changes) may have been further involved?

Maybe even more interesting: How many times will the human alleles be found in some other Denisovan (or Neandertal) genomes, and how often will the "archaic" allele be found in anyone living now?

A limited series of comparisons is too small to exclude that the range of variation will overlap, as fossil analysts have known for a long time. So we will need to work on extending our knowledge of the range of variation within living people, by increasing the sample of genomes representing populations around the world, particularly in Africa.

The technology

Of course, the most exciting thing about the new paper is the proof of concept for future high-coverage archaic genomes. The lab was able to generate the high-coverage sequence using its existing samples, by sequencing single-strand DNA instead of requiring double-strand DNA. This is a massive advantage when working with ancient DNA, because damage to the sequence often prevents double-stranded DNA from being amplified.

The paper makes explicit that the Denisova phalanx simply has better endogenous DNA preservation than any other specimen known. That being said, the new sequencing method has greatly increased the sequence yield from the sample:

We applied this method to aliquots of the two DNA extracts (as well as side fractions) that were previously generated from the 40 mg of bone that comprised the entire inner part of the phalanx (2, 8). Comparisons of these newly generated libraries to the two libraries generated in the previous study (2) show at least a 6-fold and 22-fold increase in the recovery of library molecules (8), which is particularly pronounced for longer molecules (fig. S4).

It would be too soon to say that a similar increase in yield will happen for other specimens, but obviously, this may bring higher coverage into reach for several specimens that are currently only sequenced at very low coverage, including the Vindija, Mezmaiskaya, and El Sidron Neandertals. We will have to wait and see how the new technique affects ancient DNA recovery going forward.

I keep telling people that I think it's exciting that research into human evolution is now pushing technology forward. It has often been that paleoanthropology uses technological advances in other fields. But with ancient DNA, we really see an organic growth of technology along with research questions about our evolution. In our work on the ancient genomes, we're making some progress pushing forward knowledge about human biology by understanding human evolution. Evolution really is the fundamental principle of biology, but using evolution to learn about biology sometimes requires traveling through time. Ancient DNA gives us a time machine bringing new insights into reach.

References

In my last installment on Neandertal introgression in present-day human samples, I covered whole genome data from the 1000 Genomes Project ("Which population in the 1000 Genomes Project samples has the most Neandertal similarity?". For the next few weeks I'll be releasing more of these comparisons, made with the help of my Ph.D. student, Aaron Sams.

Just to remind about our methods for comparing genomes, what we have done is to examine every base reported as a single nucleotide polymorphism by the 1000 Genomes Project. If the sequencing data had no errors, then this would be an account of every point mutation in the human genome. However, the data are imperfect in various ways, as I'll note below. Likewise, the Neandertal sequence data are imperfect in various ways.

Here's one of the 1000 Genomes Project comparisons, showing the histogram for pooled European, African, and Chinese samples. In this chart, the number of shared Neandertal derived SNP alleles is the x-axis, divided into bins of around 500. The y-axis is the number of individual genomes in the sample found in each bin. So on this chart, the largest number of European genomes (nearly 120) share very approximately 645,000 derived SNP alleles with the Vindija 33.16 genome.

I find it necessary to be very explicit about these charts, because after showing them to many people I know how easily they can be misinterpreted. It's natural to assume that they are bar charts, where higher y values mean more Neandertal. But with more than 2000 genomes to compare, a bar chart is really just noise. These histograms are much like bell curves, in which the shape of the distribution on the y-axis indicates the dispersion within the population of Neandertal shared alleles.

Percentages

Everyone is excited to find out what percentage of Neandertal ancestry people have. I'm hesitant to report percentages, because I think they are misleading on these data. There is some filtering hiding beneath the data. In particular SNP alleles that are found only in one individual ("singletons") are likely to be undersampled by the project's sequence analysis. Because gene variants that have introgressed from Neandertal populations tend to be rare in present-day samples, when we miss some rare alleles, this tends to reduce our estimate of Neandertal similarity. This bias in resequencing data should affect populations roughly in proportion to their Neandertal ancestry. Our comparisons of different populations are therefore likely to give the right order of Neandertal ancestry (e.g., Europeans more than Asians) but may underestimate the total fraction of ancestry by some amount. We are counting human SNP variants and not every base pair in the Neandertal genome data, so the effect of sequencing error in the Neandertals will be minimal, but nevertheless present in a small fraction of comparisons. These errors should be randomly distributed with respect to human population differences, but they also add noise that should decrease the accuracy of percentage estimates.

For another thing, we don't know where the zero point may be. Europeans have around 3 percent more than Yoruba; Yoruba (as I showed in the last post) have around a half percent more Neandertal similarity than Luhya in the 1000 Genomes Project sample. The Luhya are almost certainly not minimal for living people, in fact I would put some money against it. Since some Neandertal alleles have proceeded right up to high frequencies outside Africa, there has been ample opportunity during the last 30,000 years or more for other alleles to have spread into Africa.

Our conservative approach is to rely on comparisons of large samples of people, ideally hundreds, and to trust a comparison only when it achieves statistical significance in these samples. That still allows us to detect very slight excesses of Neandertal ancestry in some populations, because the data from hundreds of individuals is very strong evidence. But the overlap among populations is sometimes very extensive even if their means differ significantly.

Incomplete lineage sorting (ILS) is one pattern by which living people share alleles with Neandertals. ILS should be equally distributed among populations today, under the assumption that Neandertals and ancestral Africans stem from a single unstructured population. Obviously, Europeans and Asians share more derived SNP alleles with Neandertals than do Africans today, so we can strongly reject the hypothesis of isolation between African and Neandertal populations.

Given that, three patterns of evolution could have caused some populations to share more derived alleles with Neandertals than others.

1. Population structure in the ancestors of Africans and Neandertals may have caused some populations to share more ILS with Neandertals than others.

2. Continued gene flow between Neandertals and Africans could have spread Neandertal alleles into Africa and vice-versa.

3. Recent introgression from Neandertal populations into the ancestors of today's populations may have transferred new Neandertal alleles into recent humans.

These three processes actually overlap with each other. Very likely all three of them happened -- although to date, the descriptions of Neandertal genome data have accentuated the last and argued that the first two are relatively less important [1] [2]. A "new" allele in a Neandertal may actually have originated from a mutation more than a half million years ago, have been lost within ancient Africans, and transferred into today's Europeans when they encountered and mixed with Neandertals. We cannot tell these processes apart from the standpoint of any single SNP allele. Only by comparing many SNP alleles across many populations can we sort out their relative importance.

To this end, we have been comparing populations with each other and ancient Neandertals in many different ways. The 1000 Genomes Project has continued to sample and resequence many of the same samples that were initially amassed for the International HapMap Project. The HapMap was a project based on genotyping individuals with microarray technology. Genotypes are just as informative in many cases as whole-genome sequences. If you already know which genetic variations you want to examine, a microarray can save a substantial amount of wasted effort.

With Neandertal comparisons, we don't start out knowing in advance which genotypes will be useful. For this reason, genotyping data yields a potential bias when comparing to Neandertal or other human genomes. The microarray was designed to include genotypes that were already known to vary in some human population. With the HapMap, this bias tends to overrepresent the genetic variations in the initial HapMap samples -- generally, Utah residents of northern European descent, ethnic Yoruba people from Nigeria, ethnic Han Chinese from Beijing, and Japanese people from Tokyo. If these samples share some common derived SNP alleles with Neandertals, they will very likely be represented in the genotyping array. But very rare alleles won't be represented. And alleles that are uniquely in other populations -- such as East Africans or South Asians -- may not be represented, either. The bias is called "ascertainment bias" because it comes from the "ascertainment" of SNPs, or their initial discovery in some populations but not others.

It is possible now to find sets of SNP markers that have been statistically chosen to minimize ascertainment biases. The filters used in such comparisons are complex, and in some cases actually rely on the Neandertal genotype, so I haven't used them here. For our first paper we have focused on the whole-genome sequence comparisons, but here I'll give the same comparisons on some HapMap samples to show approximately where they fit. I will focus here on raw comparisons instead of standardizing them in terms of the predictive ability of informative SNPs on whole genome data. Finding the most informative SNPs is part of the process of sorting introgression from earlier population structure, and is rather more complex; I prefer to start with something very simple and visually easy to interpret.

South Asia

One interesting place is India. The HapMap includes a sample of Indian-Americans with origins in Gujarat, in western India. Here's a plot comparing the Gujarat ancestry (GIH) sample with the CEU and LWK samples:

The GIH sample has substantially fewer shared Neandertal derived SNP alleles than the CEU sample. What may be more curious is that the GIH sample also has fewer than East Asians on average. The JPT+CHB samples, for example, exceed the GIH mean by around 100 derived SNPs.

On a mean of more than 43,000, 100 is around a fourth of a percent, so it's not much -- and it may fall within the amount expected from ascertainment bias. It will be much more enlightening to have GIH whole genome data. In the meantime, we can probably confirm the picture from sequence data that indicates Europeans today have the highest degree of Neandertal ancestry.

East Africa

The situation within Africa is potentially very complex also. From sequence data, we were able to show that Yoruba (YRI) and Luhya (LWK) population samples have different numbers of shared derived Neandertal SNP alleles. The YRI sample in West Africa has significantly more Neandertal similarity than the LWK sample in East Africa. We speculate that this relation may reflect trans-Saharan gene flow, which has continued throughout history and prehistory.

Is this a question of east versus west in Africa? That might seem unlikely considering the extent of population movements into northeastern Africa and continued trade along the East African coast throughout historic time.

The HapMap includes a sample of ethnic Maasai people from Kenya, which allows us to provide another perspective on African variation. Here is the chart, compared to LWK and CEU:

The Maasai have substantially more Neandertal similarity than Luhya, despite their present geographic proximity. In fact, the mean amount of Neandertal similarity in the Maasai is approximately the same as that in the ASW sample, which is composed of African-American ancestry people in the Southwest U.S.:

You see immediately more dispersion in the African-American ancestry sample, because the mixture between African and European ancestors is more variable and much more recent than the events that gave rise to the Neandertal ancestry of Maasai people.

We speculate that there may have been a substantial amount of interaction in northeast Africa. Obviously this has been true in historic times, but the Maasai suggest that it may go back long before the origins of the present ethnic groups and their movements into this area. The present heterogeneity of Neandertal similarity in these populations suggests a really complex population history. Some of the present Neandertal similarity may derive from ILS within the ancient African population.

Probing assumptions

Of course my lab is not the only one presently engaged in comparing the archaic human genomes with recent populations. One of the reasons why we're pursuing a more open science strategy in our reporting is that different groups using different methodologies ought to converge on the same population history. Where we see different results, it's often an indication that the alternative approaches involve substantially different assumptions about the way ancient humans interacted. As we've probed more deeply into the data, we have confronted the reality that long-term population mixture between Neandertal and African ancestral populations is extremely difficult to rule out. Assuming that long-term interactions were impossible because Neandertals and Africans were completely isolated will probably lead to erroneous results. That makes it harder for us to clearly identify gene variants that came from Neandertals within the last hundred thousand years, as opposed to those shared with Neandertals via more ancient gene flow.

What makes long-term interactions seem more likely is that some of the Neandertal genomes seem to be more closely related to living people than others. More on that in my next installment.

References