Racial Genetic Similarity and Difference: The Witherspoon et al. Study

One scientific topic that has been often previously discussed here and at other similar sites is the biological validity of the race concept.  This, unfortunately, has become necessary, because some people, perhaps with political motivations, assert, contrary to the evidence, that “race does not exist” and that race is a “social construct” with “no biological foundation.”

One popular and misinterpreted finding that has been eagerly grasped at by those who preach that “race is not real” is derived from the work of Richard Lewontin, which demonstrated that more genetic variation exits within than between groups.  In a previous article in this journal, I have explained how Lewontin’s finding in no way discredits the race concept.  However, there are “anti-racist” activists who still claim, based on their misinterpretations of population genetics, that it is possible for individual Europeans (“Whites”) to be more genetically similar to sub-Saharan Africans (“Blacks”) than to other Europeans.  Until now, there has been no formal proof that this assertion is incorrect.  I am now pleased to say that a recent scientific paper has delved into this very topic and that the findings of this paper clearly demonstrate that the race deniers are wrong.  First, let me give a brief introduction for the sake of clarity.

A number of scientific studies have shown that it is possible to genetically cluster individuals to their self-identified race with near 100% accuracy.  Further, racial categories can be determined by the genetic data even without any a priori information about the groups involved.  In other words, racial groups can be empirically observed through genetic analysis without any prior assumptions about these groups by the researchers.

However, does that imply that individual members of these races will always be more genetically similar to members of their own racial group compared to members of other groups?  Or, are genetic clustering and individual genetic similarity so different that this may not be always so?  Can individuals share more genetic similarity to members of other groups rather than to members of their own group, even if everyone is properly clustered with their self-identified race? In other words, can there be significant genetic overlap between individuals on the fringes of, say, the European and African clusters?

These are the questions asked, and answered, in the paper “Genetic Similarities Within and Between Populations” by Witherspoon et al. (online free).  I will simplify the authors’ statements and analogies so as to make the work more understandable to the broad readership; although this may mean that certain detailed specifics are glossed over, the main “take home” points and essential interpretations remain intact.  And, since the paper is available online at no cost, any reader interested in delving into the scientific details can do so at their leisure.

The authors introduced the metric “w”, which they defined as “…the frequency with which a pair of individuals from different populations is genetically more similar than a pair from the same population.”  In other words, what is being determined with “w” is the frequency with which, for example, individual Whites and individual Blacks may be more similar to each other than to members of their own race. This measurement, which is based upon gene by gene comparisons between individuals, is different from the two measurements of clustering that the authors compare to “w.”  Unlike “w”, the clustering measurements incorporate population-level genetic information, and thus consider the “aggregate” qualities of the population’s genetic information.  To put it simply, and bypassing many details, “w” compares individuals to each other, while clustering is, essentially, comparisons of individuals to the “genetic average” (or “centroid”) of different populations.  By crude analogy, we could consider physical traits. “W” would analogous to how similar two individuals are to each other in height, weight, eye color, skin color, hair color, facial features, etc.  Clustering, in contrast, is more analogous to how similar each individual is to the average measurements of height, weight, eye color, etc. for any group.  Thus “w” can tell us how similar individuals are to each other, while clustering tells us whether an individual is more similar to one group or another.  Clustering allows us to “bin” (or “cluster”) individuals as belonging to one group or another.

Is it possible for individuals from different groups to be more genetically similar to each other than to members of their own group?  More importantly, can this occur even if all of these individuals are correctly “binned” by genetic cluster analysis to their correct racial group?  In other words, is it possible to correctly cluster everyone to their self-identified race, even though members of different groups are more similar to each other than to some members of their own group?  In theory, yes, and the authors provide an example of how this may occur.  For the sake of understanding, I will simplify their explanation and calculations.

Assume that the measurement “q” represents the averaged gene frequencies for groups or for individuals.  The African genetic average (or “centroid”) of “q” may be 0.46; the European “q”, 0.61.  This “q” measures the average frequency of different gene types at various parts of the genome. Assume three individuals, two Africans and one European, with their own individual “q” measurements of 0.4, 0.52, and 0.55 respectively.  Consider the African with q = 0.52.  He is closer to the African average of 0.46 than to the European average of 0.61.  Thus, he clusters with Africans; in fact all three individuals would cluster with their identified group. Yet, at the individual level, the African at 0.52 is closer to the European’s 0.55 value than to the other African’s 0.4 value.  Thus, it would seem that individual racial overlap can be possible even though clustering is absolutely correct.  Does this actually occur in reality?

Bamshad et al. (“Deconstructing The Relationship Between Genetics and Race”, Bamshad et al., Nat. Rev. Genet. 5, 598-609, 2004.) , using 377 DNA markers in 1,056 individuals, found that in 38% of the cases, individual Europeans were more similar to individual Asians than to other Europeans.  So it would seem that significant genetic overlap across broad racial lines exists, even if everyone is correctly binned to their own racial group.  But, is this really true?  Will that hold true when more markers are used?

These are the questions that the Witherspoon et al. paper attempted to address. What were their basic findings?  The authors first examined the amount of genetic overlap between individual Europeans and sub-Saharan Africans using 175 markers, comparing the “w” metric with two measurements of clustering. Since clustering is a less stringent measurement than is genetic similarity (“w”), it is not surprising that, with a given number of genetic markers, there is less overlap with clustering than with “w.”  For example, in the case of Africans vs. Europeans and using 175 markers, the two measures of clustering gave overlaps of 4.9% and 1.9%; in contrast, the “w” measure of similarity has an overlap of 23%.  This “w” means that, given these 175 markers, nearly one quarter of the time an individual European will be genetically more similar to an African than to another European.  This tracks fairly well with the findings of Bamshad, discussed above.  At the same time, 175 markers were sufficient to yield clustering at an accuracy of ~95–98%.   

Thus, given a moderate number of markers, accurate racial clustering of individuals may not coincide with individual members of a group always being more similar to members of their own compared to individuals of another group. Are the racial liberals then correct?  It is possible for a Dane to be more similar, genetically, to a Nigerian than to a fellow Dane, even if the error rate is less than 25% of the time?  The answer is, simply put, no.  This genetic overlap between individuals from the major racial groups is an artifact of not using sufficient numbers of markers.

As the authors used more and more markers to compare the three major racial groups (Europeans, East Asians, and sub-Saharan Africans), the less stringent clustering measurements rapidly fell to a 0% overlap, as expected from previous studies.  What about the more stringent measurement “w”, which looks at comparisons between individuals, and does not consider group data?  Once the authors reached 1,000 (or more) markers, the genetic overlap between these groups essentially reached zero. It is useful at this point to quote the authors about this fundamentally important finding:

This implies that, when enough loci are considered, individuals from these population groups will always be genetically more similar to members of their own group.

With respect to the question of whether individual members of one group may be genetically more similar to members of another group, they write:

However, if genetic similarity is measured over many thousands of loci, the answer becomes ‘never’ when individuals are sampled from geographically separated populations.

Thus, the naïve “anti-racist” view, actually stated at times (e.g., the NOVA program on race), that it is possible for individual Europeans and Africans to be more genetically similar to each other than to members of their own race, is simply false.  Any such “finding” is simply due to insufficient numbers of DNA markers being used.

With an adequate methodology, individual members of the major racial groups will always be more similar to members of their own group than to members of other groups.  Some may not like this and deem it “racist”, but these are the scientific facts, nonetheless.

For whatever reason, the authors were not satisfied with ending their study with these findings and decided to repeat their data analysis incorporating populations they term “intermediate” or “admixed.”  These included New Guineans, South Asians, Native Americans, African Americans and “Hispano-Latino” groups.  Not unexpectedly, it became somewhat more difficult to distinguish between groups, with a given number of markers, when these additional “intermediate/admixed” populations were added.  Even with more than 10,000 markers, the “w” measurement and the clustering measurements never quite reached zero with respect to overlap, although the numbers were low.  For example the authors state that with 1,000 or more markers the “w” measurement reached a value of 3.1%, meaning that even with the intermediate/admixed populations, genetic overlap was at a frequency of less than 5%.

Do these latter findings mean that there will always be genetic overlap between members of more closely related groups, especially when so-called “intermediate” and “admixed” populations are considered?  Although some people may fervently wish that 100% accurate classification will remain impossible, except for the most widely divergent groups, this may well not be the case.  We are entering an era in which reasonably affordable whole genome sequencing will be possible, and with the proper methodologies, it will be possible to compare a number of markers considerably larger than what is used in the current paper.  While 10,000 markers may not be sufficient to eliminate overlap between all groups completely – although it does reduce the overlap to very low levels – it is possible that larger numbers of markers, or even whole genome comparisons, could do so.  With more data, it may well be possible to distinguish, with near 100% accuracy, between groups that still demonstrate a low level of “w” with current data.

We must also consider the issue of genetic structure, not directly addressed in this study.  Although structure can include such genetic phenomena as inversions, deletions, and copy-number variation, the major component of genetic structure is the co-inheritance of specific genes.  In other words, we must consider not only the frequencies of each gene taken in turn, but the frequencies of specific genes together.  For example, there are genes that code for eye color, skin color, hair color, etc.  One can examine the frequency of each gene on a one-by-one basis in an individual (or group) and do all the pairwise comparisons to another individual (or group) and determine “w.”  But what are the frequencies of particular combinations of gene types inherited together?  For example, what is the frequency of having genes for blue eyes and blonde hair and fair skin, etc. co-inherited, rather than measuring the frequencies of each of these genes in turn and averaging the results?  Genetic structure superimposes further genetic differences on top of one-by-one consideration of genes; therefore, differences between groups are going to be larger when structure is considered compared to when only frequency differences of individual genes are measured and averaged.

[adrotate group=”1″]

To further explain the difference between genetic similarity and genetic structure, I present an analogy using colored marbles.  Assume that individuals of different races each have a set of marbles, numbered from one to 100, with the marbles being of various colors.  Genetic similarity (the basis of the “w” metric) would be analogous to comparing the marbles of two individuals one-by-one; first comparing the color of marble #1, then #2, then #3, and so forth, on an individual basis and then counting the total number of matches.  Genetic structure, on the other hand, would be analogous to asking if the two individuals have similar, or even identical, combinations of colors for specific marbles.  For example, person A may have red marbles for #1, #6, and #15; blue marbles for # 3, #10, #33, and #95; green marbles for #7, #8, #22, and #84, and a yellow marble for #38.  If this particular, specific combination of colored marbles is of importance, we can then ask if person B has a similar combination.  What is important here is not the one-by-one counting of matches, but whether the whole pattern is replicated, or almost replicated, between two individuals (or groups).

What about the relation between genetic ancestry and individual phenotype? The authors state that: “Thus it may be possible to infer something about an individual’s phenotype from knowledge of his or her ancestry.” However, since phenotypic traits are coded for by a number of genes smaller than that required to yield low genetic overlap, the authors assert that there may be significant phenotypic overlap between people of different groups.  They give an example of a trait “determined by 12…loci”, which would yield a 36% overlap of phenotypes between individuals of different groups.  Yet, racial groups show markedly different phenotypes.  How is this so, if what the authors state is true?

There are two points that the authors neglect to emphasize.  First, many phenotypic traits, including racially relevant ones, have been selected for because of their adaptive value, or the populations commonly exhibiting these traits have been subject to genetic drift isolated from other populations.  Thus, it is not reasonable to assume that genes that code for a particular phenotype are going to have the same “worldwide distributions” as markers used in this study. For example, gene alleles coding for skin color show markedly higher frequency differences between populations than do the neutral markers used in population genetics.

A second point is that racial phenotypes are the result of genetic structure, of many types of traits co-inherited together. It is the sum total of all these differences that allow for racial distinction at the phenotypic level.  Looking at individual phenotypic traits, just like looking at individual gene frequencies, is going to provide a markedly incomplete picture of human racial variation.

These findings powerfully support Frank Salter’s concept of ethnic genetic interests.  After all, there is essentially zero genetic overlap between individual members of different major racial groups; a member of one of these groups is always going to be more similar to a member of their own group than to that of another.  Multiplying over the large numbers of people that constitute racial groups yields a very substantial genetic interest.

Even if we take at face value this paper’s findings concerning the intermediate/admixed populations, the ethnic genetic interest concept holds as well. In the vast majority of cases, individuals will be more similar to members of their own group; overlap, while not zero, is low.  When one multiples these differences over the large numbers of people involved, then there are very large and crucial differences of genetic interests regardless of which populations are considered.

But that is not all.  First, consider that with sufficient numbers of genes assayed, the small degree of overlap observed with the intermediate/admixed groups may disappear; it would almost certainly disappear if genetic structure is considered.

Second, and perhaps most important, the ethnic genetic interest concept is not based on overall genetic similarity/difference, but rather on differences in frequencies of distinctive genes, above and beyond random gene sharing.  After all, those genes that do not differ in frequency between groups do not contribute to differences in genetic interests, because their frequency stays unchanged regardless of the outcome of competition.  Even if an entire racial group were to die out, the frequency of these “shared genes” would remain unchanged.  Note that measurements of overall genetic similarity, such as “w”, will as a matter of course also include genes that do not differ in frequency between groups. Therefore, even when “w” shows a low degree of overlap, there may well be no overlap at all with respect to those genes that are distinctive, that vary in frequency between populations.

To further explain the importance of distinctive genes vs. “w”, I will go back to my colored marbles analogy.  Imagine that the distribution of colors for marbles 1–80 was completely random, but the colors for marbles 81–100 were specific to a person’s race.  Overall similarity in marble color (analogous to “w”) would consider all 100 marbles.  However, if we were to ask how the color frequencies of the marbles were to change if people of one race were completely removed from the example, we would observe that only marbles 81-100 would be affected.  For marbles 1–80, since the color distribution is completely random with respect to race, it doesn’t matter if one race or another is eliminated from this marble counting exercise.  Only the “population-distinctive marbles” are at issue here.

Likewise, when considering competition and conflicting genetic interests between human groups, the gene frequencies that really matter are those that exhibit differences in frequency between the groups, not those that are randomly distributed between the groups.

Thus, while the Witherspoon et al. paper strongly supports the concept of ethnic genetic interests, we need to remember that ethnic genetic interests is a more stringent and specific concept than simply measuring the degree of genetic similarity.  If we are not careful, we may otherwise conclude that a group of mice constitute a greater genetic interest for a person than does another person, since the group of mice would contain more copies of the person’s gene sequences than would another single person! (By some measurements, mice and humans are ~90% genetically similar.)

But this is not the case: Genetic interests are determined by the gene frequencies that are distinctive between humans and mice (as well as differences in genetic structure between the two species). They are not determined by overall genetic similarity, and they are not determined by counting the numbers of gene sequences held in common.

In summary, this is a crucially important paper that demonstrates that individual members of the major racial groups will always be more genetically similar to members of their own group than to individuals of the other major races.  The paper demonstrates the importance of using sufficient numbers of markers in these studies, and the findings also underscore the differences between the concepts of clustering (“binning”) of individuals into groups vs. measurements of the genetic similarity between individual members of these groups (“w”).

Although the inclusion of “intermediate” and “admixed” populations prevented the genetic overlap of cross-racial individuals from reaching zero, with a sufficient number of markers the overlap was at a very low level.  Further, it is quite possible that when utilizing a greater number of markers, or even a whole genome analysis, this genetic overlap may vanish completely.

Another important point to consider when evaluating this (and any other) genetic study is that genetic structure is an important part of human genetic variation that has not yet been carefully examined, but which will likely amplify the differences in genetic variation between human population groups.  When considering the totality of genetic structure, individual overlap between racial population groups, including “intermediate” and “admixed” group, will almost certainly be nil.

Finally, the data from this paper support Frank Salter’s conception of ethnic genetic interests, although we must remember that genetic interests are properly thought of as derived from differences in the frequencies of distinctive genes, rather than counting total copies of genes shared in common.

In the final analysis, the primary findings of this paper are a devastating blow to politically motivated assertions of “no genetic differences between human races.”


With respect to the issue of clustering itself, there has been some controversy, which has been laid to rest with a recent article “Geography and genography: prediction of continental origin using randomly selected single nucleotide polymorphisms”, Allocco et al.BMC Genomics 8:68, 2007; online free.

Race deniers, as we know, claim that there are no genetic differences at all, of any significance, between even the major continental racial groups.  When confronted with the ease by which people can be “binned” (or “clustered”) into specific racial groups, the deniers bluster that such clustering requires an enormous number of markers and/or requires the choice of “biased” markers specifically picked because these markers are known, in advance, to sharply vary in frequency between groups.

These assertions and accusations are incorrect.  Allocco et al. have demonstrated that only 50 randomly chosen markers (with the emphasis on random) can cluster individuals into the major continental racial groups (Europeans, sub-Saharan Africans, and East Asians) with 95% accuracy.  The “misclassifications” resulting in the 5% “error” rate were of two African Americans, likely of admixed racial heritage, who were observed to be in between the European and African clusters.  The authors also demonstrated that as few as 5 completely random markers are sufficient to yield a 63% accuracy rate in clustering individuals into racial groups.  The authors state that “differences between continentally defined groups are sufficiently large that even a randomly selected, minute fraction of the genetic variation in the human genome can be used to characterize ancestral geographic origin in an accurate and reproducible manner”, and they conclude that their findings “argue strongly against the contention that genetic differences between groups are too small to have biomedical significance.”  The authors also assert that the clustering methodology can be “easily extended” for distinguishing more closely related groups and those with mixed origins, as long as more genetic data is obtained, sufficient to make these distinctions.

Much of this type of work is freely available to the public. It would seem that the race deniers are running out of excuses as to why they continue to promote what amounts to fraudulent pseudo-science to an unsuspecting public.

Ted Sallis (email him) writes on scientific issues.