Common pitfalls in human genomics and bioinformatics: ADMIXTURE, PCA, and the ‘Yamnaya’ ancestral component

invasion-from-the-steppe-yamnaya

Good timing for the publication of two interesting papers, that a lot of people should read very carefully:

ADMIXTURE

Open access A tutorial on how not to over-interpret STRUCTURE and ADMIXTURE bar plots, by Daniel J. Lawson, Lucy van Dorp & Daniel Falush, Nature Communications (2018).

Interesting excerpts (emphasis mine):

Experienced researchers, particularly those interested in population structure and historical inference, typically present STRUCTURE results alongside other methods that make different modelling assumptions. These include TreeMix, ADMIXTUREGRAPH, fineSTRUCTURE, GLOBETROTTER, f3 and D statistics, amongst many others. These models can be used both to probe whether assumptions of the model are likely to hold and to validate specific features of the results. Each also comes with its own pitfalls and difficulties of interpretation. It is not obvious that any single approach represents a direct replacement as a data summary tool. Here we build more directly on the results of STRUCTURE/ADMIXTURE by developing a new approach, badMIXTURE, to examine which features of the data are poorly fit by the model. Rather than intending to replace more specific or sophisticated analyses, we hope to encourage their use by making the limitations of the initial analysis clearer.

The default interpretation protocol

Most researchers are cautious but literal in their interpretation of STRUCTURE and ADMIXTURE results, as caricatured in Fig. 1, as it is difficult to interpret the results at all without making several of these assumptions. Here we use simulated and real data to illustrate how following this protocol can lead to inference of false histories, and how badMIXTURE can be used to examine model fit and avoid common pitfalls.

admixture-protocol
A protocol for interpreting admixture estimates, based on the assumption that the model underlying the inference is correct. If these assumptions are not validated, there is substantial danger of over-interpretation. The “Core protocol” describes the assumptions that are made by the admixture model itself (Protocol 1, 3, 4), and inference for estimating K (Protocol 2). The “Algorithm input” protocol describes choices that can further bias results, while the “Interpretation” protocol describes assumptions that can be made in interpreting the output that are not directly supported by model inference

Discussion

STRUCTURE and ADMIXTURE are popular because they give the user a broad-brush view of variation in genetic data, while allowing the possibility of zooming down on details about specific individuals or labelled groups. Unfortunately it is rarely the case that sampled data follows a simple history comprising a differentiation phase followed by a mixture phase, as assumed in an ADMIXTURE model and highlighted by case study 1. Naïve inferences based on this model (the Protocol of Fig. 1) can be misleading if sampling strategy or the inferred value of the number of populations K is inappropriate, or if recent bottlenecks or unobserved ancient structure appear in the data. It is therefore useful when interpreting the results obtained from real data to think of STRUCTURE and ADMIXTURE as algorithms that parsimoniously explain variation between individuals rather than as parametric models of divergence and admixture.

For example, if admixture events or genetic drift affect all members of the sample equally, then there is no variation between individuals for the model to explain. Non-African humans have a few percent Neanderthal ancestry, but this is invisible to STRUCTURE or ADMIXTURE since it does not result in differences in ancestry profiles between individuals. The same reasoning helps to explain why for most data sets—even in species such as humans where mixing is commonplace—each of the K populations is inferred by STRUCTURE/ADMIXTURE to have non-admixed representatives in the sample. If every individual in a group is in fact admixed, then (with some exceptions) the model simply shifts the allele frequencies of the inferred ancestral population to reflect the fraction of admixture that is shared by all individuals.

Several methods have been developed to estimate K, but for real data, the assumption that there is a true value is always incorrect; the question rather being whether the model is a good enough approximation to be practically useful. First, there may be close relatives in the sample which violates model assumptions. Second, there might be “isolation by distance”, meaning that there are no discrete populations at all. Third, population structure may be hierarchical, with subtle subdivisions nested within diverged groups. This kind of structure can be hard for the algorithms to detect and can lead to underestimation of K. Fourth, population structure may be fluid between historical epochs, with multiple events and structures leaving signals in the data. Many users examine the results of multiple K simultaneously but this makes interpretation more complex, especially because it makes it easier for users to find support for preconceptions about the data somewhere in the results.

In practice, the best that can be expected is that the algorithms choose the smallest number of ancestral populations that can explain the most salient variation in the data. Unless the demographic history of the sample is particularly simple, the value of K inferred according to any statistically sensible criterion is likely to be smaller than the number of distinct drift events that have practically impacted the sample. The algorithm uses variation in admixture proportions between individuals to approximately mimic the effect of more than K distinct drift events without estimating ancestral populations corresponding to each one. In other words, an admixture model is almost always “wrong” (Assumption 2 of the Core protocol, Fig. 1) and should not be interpreted without examining whether this lack of fit matters for a given question.

admixture-pitfalls
Three scenarios that give indistinguishable ADMIXTURE results. a Simplified schematic of each simulation scenario. b Inferred ADMIXTURE plots at K= 11. c CHROMOPAINTER inferred painting palettes.

Because STRUCTURE/ADMIXTURE accounts for the most salient variation, results are greatly affected by sample size in common with other methods. Specifically, groups that contain fewer samples or have undergone little population-specific drift of their own are likely to be fit as mixes of multiple drifted groups, rather than assigned to their own ancestral population. Indeed, if an ancient sample is put into a data set of modern individuals, the ancient sample is typically represented as an admixture of the modern populations (e.g., ref. 28,29), which can happen even if the individual sample is older than the split date of the modern populations and thus cannot be admixed.

This paper was already available as a preprint in bioRxiv (first published in 2016) and it is incredible that it needed to wait all this time to be published. I found it weird how reviewers focused on the “tone” of the paper. I think it is great to see files from the peer review process published, but we need to know who these reviewers were, to understand their whiny remarks… A lot of geneticists out there need to develop a thick skin, or else we are going to see more and more delays based on a perceived incorrect tone towards the field, which seems a rather subjective reason to force researchers to correct a paper.

PCA of SNP data

Open access Effective principal components analysis of SNP data, by Gauch, Qian, Piepho, Zhou, & Chen, bioRxiv (2018).

Interesting excerpts:

A potential hindrance to our advice to upgrade from PCA graphs to PCA biplots is that the SNPs are often so numerous that they would obscure the Items if both were graphed together. One way to reduce clutter, which is used in several figures in this article, is to present a biplot in two side-by-side panels, one for Items and one for SNPs. Another stratagem is to focus on a manageable subset of SNPs of particular interest and show only them in a biplot in order to avoid obscuring the Items. A later section on causal exploration by current methods mentions several procedures for identifying particularly relevant SNPs.

One of several data transformations is ordinarily applied to SNP data prior to PCA computations, such as centering by SNPs. These transformations make a huge difference in the appearance of PCA graphs or biplots. A SNPs-by-Items data matrix constitutes a two-way factorial design, so analysis of variance (ANOVA) recognizes three sources of variation: SNP main effects, Item main effects, and SNP-by-Item (S×I) interaction effects. Double-Centered PCA (DC-PCA) removes both main effects in order to focus on the remaining S×I interaction effects. The resulting PCs are called interaction principal components (IPCs), and are denoted by IPC1, IPC2, and so on. By way of preview, a later section on PCA variants argues that DC-PCA is best for SNP data. Surprisingly, our literature survey did not encounter even a single analysis identified as DC-PCA.

The axes in PCA graphs or biplots are often scaled to obtain a convenient shape, but actually the axes should have the same scale for many reasons emphasized recently by Malik and Piepho [3]. However, our literature survey found a correct ratio of 1 in only 10% of the articles, a slightly faulty ratio of the larger scale over the shorter scale within 1.1 in 12%, and a substantially faulty ratio above 2 in 16% with the worst cases being ratios of 31 and 44. Especially when the scale along one PCA axis is stretched by a factor of 2 or more relative to the other axis, the relationships among various points or clusters of points are distorted and easily misinterpreted. Also, 7% of the articles failed to show the scale on one or both PCA axes, which leaves readers with an impressionistic graph that cannot be reproduced without effort. The contemporary literature on PCA of SNP data mostly violates the prohibition against stretching axes.

pca-how-to
DC-PCA biplot for oat data. The gradient in the CA-arranged matrix in Fig 13 is shown here for both lines and SNPs by the color scheme red, pink, black, light green, dark green.

The percentage of variation captured by each PC is often included in the axis labels of PCA graphs or biplots. In general this information is worth including, but there are two qualifications. First, these percentages need to be interpreted relative to the size of the data matrix because large datasets can capture a small percentage and yet still be effective. For example, for a large dataset with over 107,000 SNPs for over 6,000 persons, the first two components capture only 0.3693% and 0.117% of the variation, and yet the PCA graph shows clear structure (Fig 1A in [4]). Contrariwise, a PCA graph could capture a large percentage of the total variation, even 50% or more, but that would not guarantee that it will show evident structure in the data. Second, the interpretation of these percentages depends on exactly how the PCA analysis was conducted, as explained in a later section on PCA variants. Readers cannot meaningfully interpret the percentages of variation captured by PCA axes when authors fail to communicate which variant of PCA was used.

Conclusion

Five simple recommendations for effective PCA analysis of SNP data emerge from this investigation.

  1. Use the SNP coding 1 for the rare or minor allele and 0 for the common or major allele.
  2. Use DC-PCA; for any other PCA variant, examine its augmented ANOVA table.
  3. Report which SNP coding and PCA variant were selected, as required by contemporary standards in science for transparency and reproducibility, so that readers can interpret PCA results properly and reproduce PCA analyses reliably.
  4. Produce PCA biplots of both Items and SNPs, rather than merely PCA graphs of only Items, in order to display the joint structure of Items and SNPs and thereby to facilitate causal explanations. Be aware of the arch distortion when interpreting PCA graphs or biplots.
  5. Produce PCA biplots and graphs that have the same scale on every axis.

I read the referenced paper Biplots: Do Not Stretch Them!, by Malik and Piepho (2018), and even though it is not directly applicable to the most commonly available PCA graphs out there, it is a good reminder of the distorting effects of stretching. So for example quite recently in Krause-Kyora et al. (2018), where you can see Corded Ware and BBC samples from Central Europe clustering with samples from Yamna:

NOTE. This is related to a vertical distorsion (i.e. horizontal stretching), but possibly also to the addition of some distant outlier sample/s.

pca-cwc-yamna-bbc
Principal Component Analysis (PCA) of the human Karsdorf and Sorsum samples together with previously published ancient populations projected on 27 modern day West Eurasian populations (not shown) based on a set of 1.23 million SNPs (Mathieson et al., 2015). https://doi.org/10.7554/eLife.36666.006

The so-called ‘Yamnaya’ ancestry

Every time I read papers like these, I remember commenters who kept swearing that genetics was the ultimate science that would solve anthropological problems, where unscientific archaeology and linguistics could not. Well, it seems that, like radiocarbon analysis, these promising developing methods need still a lot of refinement to achieve something meaningful, and that they mean nothing without traditional linguistics and archaeology… But we already knew that.

Also, if this is happening in most peer-reviewed publications, made by professional geneticists, in journals of high impact factor, you can only wonder how many more errors and misinterpretations can be found in the obscure market of so many amateur geneticists out there. Because amateur geneticist is a commonly used misnomer for people who are not geneticists (since they don’t have the most basic education in genetics), and some of them are not even ‘amateurs’ (because they are selling the outputs of bioinformatic tools)… It’s like calling healers ‘amateur doctors’.

NOTE. While everyone involved in population genetics is interested in knowing the truth, and we all have our confirmation (and other kinds of) biases, for those who get paid to tell people what they want to hear, and who have sold lots of wrong interpretations already, the incentives of ‘being right’ – and thus getting involved in crooked and paranoid behaviour regarding different interpretations – are as strong as the money they can win or loose by promoting themselves and selling more ‘product’.

As a reminder of how badly these wrong interpretations of genetic results – and the influence of the so-called ‘amateurs’ – can reflect on research groups, yet another turn of the screw by the Copenhagen group, in the oral presentations at Languages and migrations in pre-historic Europe (7-12 Aug 2018), organized by the Copenhagen University. The common theme seems to be that Bell Beaker and thus R1b-L23 subclades do represent a direct expansion from Yamna now, as opposed to being derived from Corded Ware migrants, as they supported before.

NOTE. Yes, the “Yamna → Corded Ware → Únětice / Bell Beaker” migration model is still commonplace in the Copenhagen workgroup. Yes, in 2018. Guus Kroonen had already admitted they were wrong, and it was already changed in the graphic representation accompanying a recent interview to Willerslev. However, since there is still no official retraction by anyone, it seems that each member has to reject the previous model in their own way, and at their own pace. I don’t think we can expect anyone at this point to accept responsibility for their wrong statements.

So their lead archaeologist, Kristian Kristiansen, in The Indo-Europeanization of Europé (sic):

kristiansen-migrations
Kristiansen’s (2018) map of Indo-European migrations

I love the newly invented arrows of migration from Yamna to the north to distinguish among dialects attributed by them to CWC groups, and the intensive use of materials from Heyd’s publications in the presentation, which means they understand he was right – except for the fact that they are used to support a completely different theory, radically opposed to those defended in Heyd’s model

Now added to the Copenhagen’s unending proposals of language expansions, some pearls from the oral presentation:

  • Corded Ware north of the Carpathians of R1a lineages developed Germanic;
  • R1b borugh [?] Italo-Celtic;
  • the increase in steppe ancestry on north European Bell Beakers mean that they “were a continuation of the Yamnaya/Corded Ware expansion”;
  • Corded Ware groups [] stopped their expansion and took over the Bell Beaker package before migrating to England” [yep, it literally says that];
  • Italo-Celtic expanded to the UK and Iberia with Bell Beakers [I guess that included Lusitanian in Iberia, but not Messapian in Italy; or the opposite; or nothing like that, who knows];
  • 2nd millennium BC Bronze Age Atlantic trade systems expanded Proto-Celtic [yep, trade systems expanded the language]
  • 1st millennium BC expanded Gaulish with La Tène, including a “Gaulish version of Celtic to Ireland/UK” [hmmm, dat British Gaulish indeed].

You know, because, why the hell not? A logical, stable, consequential, no-nonsense approach to Indo-European migrations, as always.

Also, compare still more invented arrows of migrations, from Mikkel Nørtoft’s Introducing the Homeland Timeline Map, going against Kristiansen’s multiple arrows, and even against the own recent fantasy map series in showing Bell Beakers stem from Yamna instead of CWC (or not, you never truly know what arrows actually mean):

corded-ware-migrations
Nørtoft’s (2018) maps of Indo-European migrations.

I really, really loved that perennial arrow of migration from Volosovo, ca. 4000-800 BC (3000+ years, no less!), representing Uralic?, like that, without specifics – which is like saying, “somebody from the eastern forest zone, somehow, at some time, expanded something that was not Indo-European to Finland, and we couldn’t care less, except for the fact that they were certainly not R1a“.

This and Kristiansen’s arrows are the most comical invented migration routes of 2018; and that is saying something, given the dozens of similar maps that people publish in forums and blogs each week.

NOTE. You can read a more reasonable account of how haplogroup R1b-L51 and how R1-Z645 subclades expanded, and which dialects most likely expanded with them.

We don’t know where these scholars of the Danish workgroup stand at this moment, or if they ever had (or intended to have) a common position – beyond their persistent ideas of Yamnaya™ ancestral component = Indo-European and R1a must be Indo-European – , because each new publication changes some essential aspects without expressly stating so, and makes thus everything still messier.

It’s hard to accept that this is a series of presentations made by professional linguists, archaeologists, and geneticists, as stated by the official website, and still harder to imagine that they collaborate within the same professional workgroup, which includes experienced geneticists and academics.

I propose the following video to close future presentations introducing innovative ideas like those above, to help the audience find the appropriate mood:

Related

The uneasy relationship between Archaeology and Ancient Genomics

Allentoft Corded Ware

News feature Divided by DNA: The uneasy relationship between archaeology and ancient genomics, Two fields in the midst of a technological revolution are struggling to reconcile their views of the past, by Ewen Callaway, Nature (2018) 555:573-576.

Interesting excerpts (emphasis mine):

In duelling 2015 Nature papers6,7the teams arrived at broadly similar conclusions: an influx of herders from the grassland steppes of present-day Russia and Ukraine — linked to Yamnaya cultural artefacts and practices such as pit burial mounds — had replaced much of the gene pool of central and Western Europe around 4,500–5,000 years ago. This was coincident with the disappearance of Neolithic pottery, burial styles and other cultural expressions and the emergence of Corded Ware cultural artefacts, which are distributed throughout northern and central Europe. “These results were a shock to the archaeological community,” Kristiansen says.

(…)

Still, not everyone was satisfied. In an essay8 titled ‘Kossinna’s Smile’, archaeologist Volker Heyd at the University of Bristol, UK, disagreed, not with the conclusion that people moved west from the steppe, but with how their genetic signatures were conflated with complex cultural expressions. Corded Ware and Yamnaya burials are more different than they are similar, and there is evidence of cultural exchange, at least, between the Russian steppe and regions west that predate Yamnaya culture, he says. None of these facts negates the conclusions of the genetics papers, but they underscore the insufficiency of the articles in addressing the questions that archaeologists are interested in, he argued. “While I have no doubt they are basically right, it is the complexity of the past that is not reflected,” Heyd wrote, before issuing a call to arms. “Instead of letting geneticists determine the agenda and set the message, we should teach them about complexity in past human actions.”

Many archaeologists are also trying to understand and engage with the inconvenient findings from genetics. (…)
[Carlin:] “I would characterize a lot of these papers as ‘map and describe’. They’re looking at the movement of genetic signatures, but in terms of how or why that’s happening, those things aren’t being explored,” says Carlin, who is no longer disturbed by the disconnect. “I am increasingly reconciling myself to the view that archaeology and ancient DNA are telling different stories.” The changes in cultural and social practices that he studies might coincide with the population shifts that Reich and his team are uncovering, but they don’t necessarily have to. And such biological insights will never fully explain the human experiences captured in the archaeological record.

Reich agrees that his field is in a “map-making phase”, and that genetics is only sketching out the rough contours of the past. Sweeping conclusions, such as those put forth in the 2015 steppe migration papers, will give way to regionally focused studies with more subtlety.

This is already starting to happen. Although the Bell Beaker study found a profound shift in the genetic make-up of Britain, it rejected the notion that the cultural phenomenon was associated with a single population. In Iberia, individuals buried with Bell Beaker goods were closely related to earlier local populations and shared little ancestry with Beaker-associated individuals from northern Europe (who were related to steppe groups such as the Yamnaya). The pots did the moving, not the people.

This final paragraph apparently sums up a view that Reich has of this field, since he repeats it:

Reich concedes that his field hasn’t always handled the past with the nuance or accuracy that archaeologists and historians would like. But he hopes they will eventually be swayed by the insights his field can bring. “We’re barbarians coming late to the study of the human past,” Reich says. “But it’s dangerous to ignore barbarians.”

I would say that the true barbarians didn’t have a habit or possibility to learn from the higher civilizations they attacked or invaded. Geneticists, on the other hand, only have to do what they expect archaeologists to do: study.

EDIT (30 MAR 2018): A new interesting editorial of Nature, On the use and abuse of ancient DNA.

See also:

Something is very wrong with models based on the so-called ‘steppe admixture’ – and archaeologists are catching up

steppe-admixture

Russian archaeologist Leo Klejn has published an article Discussion: Are the Origins of Indo-European Languages Explained by the Migration of the Yamnaya Culture to the West?, which includes the criticism received from Wolfgang Haak, Iosif Lazaridis, Nick Patterson, and David Reich (mainly on the genetic aspect), and from Kristian Kristiansen, Karl-Göran Sjögren, Morten Allentoft, Martin Sikora, and Eske Willerslev (mainly on the archaeological aspect).

I will not post details of Klejn’s model of North-South Proto-Indo-European expansion – which is explained in the article, and relies on the north-south cline of ‘steppe admixture’ in the modern European population -, since it is based on marginal anthropological methods and theories, including glottochronological dates, and archaeological theories from the Russian school (mainly Zalyzniak), which are obviously not mainstream in the field of Indo-European Studies, and (paradoxically) on the modern distribution of ‘steppe admixture’…

The most interesting aspects of the article are the reactions to the criticism, some of which can be used from the point of view of the Indo-European demic diffusion model, too. It is sad, however, that they didn’t choose to answer earlier to Heyd’s criticism (or to Heyd’s model, which is essentially also that of Mallory and Anthony), instead of just waiting for proponents of the least interesting models to react…

The answer by Haak et al.:

Klejn mischaracterizes our paper as claiming that practitioners of the Corded Ware culture spoke a language ancestral to all European Indo-European languages, including Greek and Celtic. This is incorrect: we never claim that the ancestor of Greek is the language spoken by people of the Corded Ware culture. In fact, we explicitly state that the expansion of steppe ancestry might account for only a subset of Indo-European languages in Europe. Klejn asserts that ‘a source in the north’ is a better candidate for the new ancestry manifested in the Corded Ware than the Yamnaya. While it is indeed the case that the present-day people with the greatest affinity to the Corded Ware are distributed in north-eastern Europe, a major part of the new ancestry of the Corded Ware derives from a population most closely related to Armenians (Haak et al., 2015) and hunter-gatherers from the Caucasus (Jones et al., 2015). This ancestry has not been detected in any European huntergatherers analysed to date (Lazaridis et al., 2014; Skoglund et al., 2014; Haak et al., 2015; Fu et al., 2016), but made up some fifty per cent of the ancestry of the Yamnaya. The fact that the Corded Ware traced some of its ancestry to the southern Caucasus makes a source in the north less parsimonious.

In our study, we did not speculate about the date of Proto-Indo-European and the locations of its speakers, as these questions are unresolved by our data, although we do think the genetic data impose constraints on what occurred. We are enthusiastic about the potential of genetics to contribute to a resolution of this longstanding issue, but this is likely to require DNA from multiple, as yet unsampled, ancient populations.

Klejn response to that:

Allegedly, I had accused the authors of tracing all Indo-European languages back to Yamnaya, whereas they did not trace all of them but only a portion! Well, I shall not reproach the authors for their ambiguous language: it remains the case that (beginning with the title of the first article) their qualifications are lost and their readers have understood them as presenting the solution to the whole question of the origins of Indo-European languages.

(…) they had in view not the Proto-Indo-European before the separation of the Hittites, but the language that was left after the separation. Yet, this was still the language ancestral to all the remaining Indo-European languages, and the followers of Sturtevan and Kluckhorst call only this language Proto-Indo-European (while they call the initial one Indo-Hittite). The majority of linguists (specialists in Indo-European languages) is now inclined to this view. True, the breakup of this younger language is several hundred years more recent (nearly a thousand years later according to some glottochronologies) than the separation of Anatolian languages, but it is still around a thousand years earlier than the birth of cultures derived from Yamnaya.
More than that, I analysed in my criticism both possibilities — the case for all Indo-European languages spreading from Yamnaya and the case for only some of them spreading from Yamnaya. In the latter case, it is argued that only the languages of the steppes, the Aryan (Indo- Iranian) are descended from Yamnaya, not the languages of northern Europe. Together with many scholars, I am in agreement with the last possibility. But, then, what sense can the proposed migration of the Yamnaya culture to the Baltic region have? It would bring the Indo-Iranian proto-language to that region! Yet, there are no traces of this language on the coasts of the Baltic!

My main concern is that, to my mind, one should not directly apply conclusions from genetics to events in the development of language because there is no direct and inevitable dependence between events in the life of languages, culture, and physical structure (both anthropological and genetic). They can coincide, but often they all follow divergent paths. In each case the supposed coincidence should be proved separately.

The authors’ third objection concerns the increase of the genetic similarity of European population with that of the Yamnaya culture. This increases in the north of Europe and is weak in the south, in the places adjacent to the Yamnaya area, i.e. in Hungary. This gradient is clearly expressed in the modern population, but was present already in the Bronze Age, and hence cannot be explained by shifts that occurred in the Early Iron Age and in medieval times. However, the supposed migration of the Yamnaya culture to the west and north should imply a gradient in just the opposite direction!

Regarding the arguments of Kristiansen and colleagues:

[They argue that] in two early burials of the Corded Ware culture (one in Germany, the other in Poland) some single attributes of Yamnaya origin have been found.

(…) if this is the full extent of Yamnaya infiltration into central Europe—two burials (one for each country) from several thousands (and from several hundreds of early burials)—then it hardly amounts to large-scale migration.

Quite recently we have witnessed the success of a group of geneticists from Stanford University and elsewhere (Poznik et al., 2016). They succeeded in revealing varieties of Y-chromosome connected with demographic expansions in the Bronze Age. Such expansion can give rise to migration. Among the variants connected with this expansion is R1b, and this haplogroup is typical for the Yamnaya culture. But what bad luck! This haplogroup connected with expansion is indicated by the clade L11, while the Yamnaya burials are associated with a different clade, Z2103, that is not marked by expansion. It is now time to think about how else the remarkable results reached by both teams of experienced and bright geneticists may be interpreted.

Regarding the work of Heyd,

(…) with regard to the barrow burials of the third millennium BC in the basin of the Danube, although they have been assigned to the Yamnaya culture, I would consider them as also belonging to
another, separate culture, perhaps a mixed culture: its burial custom is typical of the Yamnaya, but its pottery is absolutely not Yamnaya, but local Balkan with imports of distinctive corded beakers (Schnurbecher). I would not be surprised if
Y-chromosome haplogroups of this population were somewhat similar to those of the Yamnaya, while mitochondrial groups were indigenous. As yet, geneticists deal with great blocks of populations and prefer to match them to very large and generalized cultural blocks, while archaeology now analyses more concrete and smaller cultures, each of which had its own fate.

Iosif Lazaridis shares more thoughts on the discussion in his Twitter account:

As we mentioned in Haak, Lazaridis et al. (2015), the Yamnaya are the best proximate source for the new ancestry that first appears with the Corded Ware in central Europe, as it has the right mix of both ANE (related to Native Americans, MA1, and EHG), but also Armenian/Caucasus/Iran-like southern component of ancestry. The Yamnaya is a westward expansive culture that bears exactly the two new ancestral components (EHG + Caucasus/Iran/Armenian-like).
As for the Y-chromosome, it was already noted in Haak, Lazaridis et al. (2015) that the Yamnaya from Samara had Y-chromosomes which belonged to R-M269 but did not belong to the clade common in Western Europe (p. 46 of supplement). Also, not a single R1a in Yamnaya unlike Corded Ware (R1a-dominated). But Yamnaya samples = elite burials from eastern part of the Yamnaya range. Both R1a/R1b found in Eneolithic Samara and EHG, so in conclusion Yamnaya expansion still the best proximate source for the post-3,000 BCE population change in central Europe. And since 2015 steppe expansion detected elsewhere (Cassidy et al. 16, Martiniano et al. 17, Mittnik et al. 17, Mathieson et al. 17, Lazaridis et al. 2016 (South Asia) and …?…

I love the smell of new wording in the morning… viz. Yamnaya best proximate source for Corded Ware, Corded Ware might account for only a subset of Indo-European languages, Corded Ware representing Aryan languages (probably Klejn misinterprets what the authors mean, i.e. some kind of Indo-Slavonic or Germano-Balto-Slavic group)…

We shall expect more and more ambiguous rewording and more adjustments of previous conclusions as new papers and new criticisms appear.

Related:

Featured image from the article: Distribution of the ‘Yamnaya’ genetic component in the populations of Europe (data taken from Haak et al., 2015). The intensity of the colour corresponds to the contribution of this component in various modern populations

Heyd, Mallory, and Prescott were right about Bell Beakers

yamna-migration

Sometimes it is fun to read certain “old” papers. I have recently re-read some important papers that predicted what we are seeing now in aDNA analysis with surprising accuracy:

Harrison & Heyd (2007): “We predict that future stable isotope and ancientDNA analyses of Beaker skeletal material will support our view that immigration played an important role in the Europe-wide Bell Beaker phenomenon”. – Duh, obvious, right? Wrong. Read the whole paper. It was already becoming a classic in the study of the Bell Beaker culture before the latest research on Bell Beaker aDNA, and it will be still more important from now on. There are different models for the Bell Beaker origin and expansion, and this was only one of them: we had the Dutch model, the radiocarbon date-based attempts to locate Bell Beakers in Iberia or North Africa,… I tried to highlight the best sentences from Heyd’s article to include them in my article, and I just couldn’t stop highlighting almost everything. It is surprising that 10 years ago Volker Heyd was predicting so much from such a limited amount of material, and with conflicting reports coming from everywhere, from palaeogenetics to radiocarbon dating. Not that today their chronology of Le Petit – Chasseur is accepted by all, but their general Bell Beaker and Yamna model has been clearly established as the most likely one with support from aDNA.

– Mallory in Celtic from the West 2 (2013), as the last of many to propose Bell Beaker as the vector of spread of Late Indo-European languages, but the first to relate it to North-West Indo-European: “The spread of Indo-European languages from Alpine Europe may have begun with the Beaker culture, presuming here a non-Iberian Beaker homeland (Rhineland, Central European) for that part of the Beaker phenomenon that was associated with an Indo-European language. While it is possible that IE language(s) spread with the Beaker phenomenon, it is questionable that this was associated with Proto-Celtic rather than earlier forms of Late Indo-European, at least part of which might be subsumed under the heading NW Indo-European. This is because the time depth of the dispersal of the Beakers is so great and the earliest attested Celtic languages are so similar (…)”. You might think that it is related to the Atlantic Indo-European theory favoured by Cunnliffe and Koch in the book… Wrong, he specifically dismisses a Neolithic spread of Indo-European, and a Calcholithic spread of Celtic languages as too early. You might also think that to publish that in 2013 has no merit, given the data. Wrong again. Just look at the trend among renown archaeologists – like Anthony (with Haak) and Kristiansen (with Allentoft) – trying to hop on the bandwagon of Corded Ware-driven Indo-European dispersal based on the “steppe admixture” proportion of recent genetic papers, and you realize he is going against the grain here.

Prescott and Walderhaug 1995 (as referred to in Prescott 2012): “The Bell Beaker period is the most, perhaps the only, reasonable candidate for the spread and final entrenchment of a common Indo-European language throughout Scandinavia (and not just Corded Ware core areas of southern and eastern Scandinavia), and particularly Norway”. Duh again? Not so fast. While Bell Beaker had been proposed before as a vector of Indo-European languages in Europe, the association with Germanic was far more controversial. Only the unifying Dagger Period was more clearly established as of Pre-Germanic nature, but it could be interpreted as of Corded Ware, Úněticean, or even early Neolithic origin, or a mix of them. Bell Beaker groups were never good candidates, if only because of the desire by some researchers to offer a romanticized (either more unifying or ancient) picture of a Germanic Northern Scandinavian homeland, explained as a culturally and genetically homogeneous group.

Their papers seem to state the obvious now that the latest aDNA samples are proving them correct, but it was far from clear years ago: remember the native European Basque-R1b – Uralic-N1c harmony disrupted by invasive Eurasian Indo-European-speaking warriors carrying R1a lineages from Yamna to Corded Ware? Well that is still a thing for some. And even today the most popular interpretation of the spread of Indo-European-speakers in Europe is based on the defined “steppe ancestry” proportion found in Corded Ware individuals, and a supposedly Yamna community formed by R1b-R1a lineages, which is obviously reminiscent of the identification of R1a lineages with Proto-Indo-Europeans based on the initial analysis of haplogroups in modern populations.

It is sad to imagine how much we would have improved in our knowledge, had we read their work with interest when it was necessary, and not now that we have most of the aDNA clues. Still sadder is to see people rely on genetic studies alone to derive today what are likely the wrong conclusions. Again.

I will end with a mea culpa. I hadn’t read those works; but even if I had, I would have stayed with the simpler, R1a-Corded Ware model of Indo-European dispersion. That oversimplification will remain in the different editions of our Grammar of Modern Indo-European as a permanent reminder. Simpler seems always better, and Cavalli-Sforza had famously asserted that ancient population movements could be solved with the study of the structure of modern populations. I think he was right, that we can in fact ascertain ancient population movements by studying modern populations if we include anthropological disciplines, but it is such a complex task – and geneticists have not shown a good grasp in (or interest for) Anthropology -, that it is nowadays clearly wrong to rely on modern population samples to derive conclusions about ancient populations, and we are better off studying ancient DNA samples in their context.

We were Back-to-the-Future-wrong, overestimating our potential in some aspects – like the results of researching modern DNA -, and underestimating it in others – like the potential changes that ancient DNA investigation could bring for anthropological disciplines. Just as we are wrong today in trusting the potential of admixture analysis to be self-explanatory, without a need for wide anthropological investigation (or even able to revolutionize archaeological and linguistic theories).

I hope to keep a more critical view of publications – especially the most popular ones – from now on.