Saturday, June 7, 2014

Wisdom Of The Week

The STRUCTURE method is actually fairly sensitive to sample size and normally this is what users want; people are less interested in populations with small numbers of individuals, even if they are actually differentiated. But this means that its results (like those of PCA) depend critically on what individuals are put in. If large numbers of individuals from Native American or Oceanian populations, which are often also quite drifted, are put in then they also will form additional subclusters in global analysis. If the Native Americans had survived the last 500 years intact, for example, this would seem quite reasonable.

As well as being dependent on sample size, STRUCTURE is dependent on the degree of genetic drift. The method is more likely to identify populations that have experienced founder effects than populations that have been evolving distinctly but with large effective population sizes for longer periods of time. Some of the oldest population subdivisions are likely to have been in Africa, but these tend to get de-emphasized by the method.

Wade goes on to say:

There is not one story of recent human evolution but at least five different stories, given that the populations on each continent have evolved largely independently of one another since the dispersal from Africa some 50,000 years ago.

But the Rosenberg et al. paper does not say this. And its not true. The STRUCTURE method is a powerful approach for detecting recent admixture, but it becomes progressively more difficult as the time since admixture increases. There is some discussion of this for example in our 2003 paper.

In practice, we know based on subsequent work that STRUCTURE certainly misses older admixture events. See our genetic atlas, for several examples of admixture events between individuals from these five groups Wade mentions within the last 5000 years some of which were not detected by Rosenberg 2002 but are visible with our methods e.g. African admixture into the Druze.

The Kalash themselves have had an admixture event from the West in the first millenium BCE that has been elusive to detect, e.g. using Admixture. Detecting ancient admixture is intrinsically hard and for admixture events older than 5000 years old the GLOBETROTTER approach as currently constructed is unlikely to work and we do not have many other tools to detect them. Moreover, there are substantial and interesting admixture events in the Indians and Japanese, for example, that we know are not well captured by our recent paper and do not obviously fit into the 5 continent scheme.

In the absence of disturbance, individuals within a continent are likely to become more homogenous due to genetic exchange over time. But it certainly does not mean that they were always similar to each other and there is evidence for multiple and distinct sources of ancestry for several regional groups (including of course from archaic hominids).

The Rosenberg et al. paper was a landmark because of the elegant visualization of ancestry and because it is one in a series showing what is possible using individual-based, rather than population-based analysis. But STRUCTURE barplots are in several ways over-interpreted, alas, by Nicholas Wade and by others.


No comments: