Why Continental Population Differences Will Be Studied
For the last quarter century, medical researchers have been grappling with evidence that what’s true about a disease for one ancestral population isn’t necessarily true for another. It’s called population stratification.
The problem was first suspected in the mid-1990s. Medical researchers looking for candidate genes routinely compared the genetics of a group with the disease being studied with a comparison sample of people without the disease. In choosing those samples, ancestry was initially not a consideration. But as researchers got access to more genetic information, they began to worry that their results were being contaminated because the ancestral populations in the samples had different genetic profiles.3 A back-and-forth debate ensued in the technical literature. By 2004, the weight of the evidence had become clear. As one of the early DNA-based studies concluded, “Even small amounts of population admixture can undermine an association study and lead to false positive results. These adverse effects increase markedly with sample size. For the size of study required for many complex diseases, relatively modest levels of structure within a population can have serious consequences.”4
POLYGENIC SCORES
Polygenic scores will be discussed in detail in chapter 14. For now, think of them as analogous to test scores, but based on combined allele frequencies instead of combined answers to test questions. A polygenic score for schizophrenia (for example) measures the genetic risk of schizophrenia.
A decade later, the first studies using polygenic scores verified an explanation for population stratification that generalizes far beyond the study of diseases: Polygenic scores for one continental population don’t work as well for other continental populations no matter what the trait may be. In technical terms, the predictive validity of a polygenic score deteriorates as the genetic distance between the test population and the comparison population increases, consistent with population genetics theory.5 For example, a polygenic score based on a test population of English and Italians usually generalizes accurately for French and Germans, not so accurately for Chinese and Indians, and least accurately for the genetically most distant populations from sub-Saharan Africa.[6]
Population differences in predictive validity could reflect natural selection, genetic drift, or gene × environment interactions. Population geneticists have had strong scientific motivation to learn more about those differences but have been frustrated because the artifacts produced by population stratification are so common.7 Statistical analysis can correct for some of population stratification’s effects, but the only full solution is to have large samples from all the ancestral populations that are being compared. The problem is that genomic data have typically been collected from people who lived in the nations where geneticists worked, dominated by Europe and the United States, which in turn meant that large genomic databases were overwhelmingly people of European ancestry.
The collection of large samples from non-European populations was on the back burner through the first half of the 2010s. It’s understandable—samples in the hundreds of thousands are logistically demanding, and the foundations and government agencies with deep enough pockets to fund such samples have not (until recently) put them on their agendas. There also has been a lack of urgency: Geneticists have been kept busy with an ample supply of GWA research projects that can be done with European samples.
Then murmurings about underrepresentation of non-Europeans in genomic databases began appearing. They reached a broad audience in 2018 when British geneticist David Curtis charged that by using European samples, “UK medical science stands at risk of being institutionally racist.”8 In 2019, an article by a team of American geneticists in Cell, “The Missing Diversity in Human Genetic Studies,” widely picked up by the media, detailed the many ways in which the bias toward European samples “effectively translates into poorer disease prediction and treatment for individuals of under-represented ancestries.”9
It now appears likely that large samples from underrepresented populations—notably Africans and South Asians—will be available soon (China and Japan have been building such databases on their own). When they come online, ancestral population differences related to disease are going to be studied minutely.
Those same databases will potentially allow researchers to study genetic differences in personality traits, abilities, and social behavior across continental populations. That potential is likely to generate cross-cutting pressures. For highly charged topics such as IQ, many people will continue to urge that studying population differences does more harm than good. But what happens if findings from European samples about cognitive-related traits such as depression, autism, or schizophrenia lead to more effective treatments for Europeans but not for other populations? It will be ethically imperative to study the genetics of mental disorders in other populations as well, which means studying the ways in which they differ from Europeans. The idea that geneticists could ignore ancestral population differences indefinitely was always implausible. It is now out of the question.
Differences in Allele Frequencies Within and Across Continental Population
When SNPs cause differences in phenotypic traits, evidence for that role surfaces first in differences in target allele frequencies. The target allele is usually defined as one that is associated with an increase in the magnitude or intensity of a trait. If the topic is diabetes, the target alleles are the ones associated with an increase in the risk or severity of diabetes. If the topic is IQ, the target alleles are those associated with increases in IQ scores. Other labels used in the literature include risk allele, effect allele, and increasing allele. As in chapter 8, I express target allele frequencies exclusively as proportions ranging from 0 to 1 rather than as percentages of chromosomes.
My purpose in this discussion is limited to the wording of Proposition #7 as it applies to common SNPs: Continental population differences in target allele frequencies associated with personality, abilities, and social behavior are common. I am not presenting proof that those differences cause phenotypic differences, but showing you how different the situation actually facing geneticists is from the impression you may have when you hear that “race is a social construct.” Virtually all traits, whether physiological, related to disease, or related to cognitive repertoires, exhibit many large differences in target allele frequencies across continental populations.
Comparing Subpopulations from the Same Continent
I’ll use a specific example, schizophrenia, as an entry point to the topic. First, consider the landscape for subpopulations within the same continental population. The following graphs show what happens when the target allele frequencies for two populations are plotted against each other for three within-continent pairs: Kenyans and African Americans, British and Italians, and Chinese and Japanese.
Source: Author’s analysis, GWAS Catalog, and Phase 1 of the 1000 Genomes Project. A total of 962 SNPs in the GWAS Catalog are associated with schizophrenia. For this and the subsequent graphs, I chose 500 to plot (962 in a small graph would produce too many overlays, obscuring the pattern).10
The diagonal line identifies SNPs for which the target allele difference is zero. As you can see, the actual differences are closely bunched to either side of the diagonal on all three graphs.
Scatter plots like these imply extremely high correlations between the two sets of target allele frequencies, and indeed they are high: +.98 for the African and Asian pairs; +.97 for the European pair. These results are typical. Taking all of the unique SNPs for all traits that are part of both the GWAS Catalog and 1000 Genomes—a sample of 43,543 SNPs—the average correlations were +.98 for the three African pairs, +.98 for the six European pairs, and +.99 for the three Asian pairs.11
Comparing Continental Populations
Now look what happens when we repeat the exercise, but comparing Africans with Asians, Asians with Europeans, and Europeans with Africans.
Source: Author’s an
alysis, GWAS Catalog and Phase 1 of the 1000 Genomes Project.
The landscape is completely different. The cross-continent correlations are all high by the standards of social science, but even correlations of +.70, +.81, and +.74 (which are the ones represented in the figure) are associated with large differences between target allele frequencies. Why did I choose schizophrenia for the example? Because the three correlations for schizophrenia are nearly the same as the correlations for all SNPs related to cognitive repertoires in the GWAS Catalog: +.71 for Africans and Asians, +.76 for Africans and Europeans, and +.81 for Asians and Europeans. The schizophrenia example is typical, not extreme.
And that’s the nut of what I am trying to convey with Proposition #7. We don’t know what these differences mean yet (with a few exceptions to be taken up later), but the image fostered by “race is a social construct” does not apply. The raw material for investigating genetic sources of population differences in phenotypic traits consists of differences in target allele frequencies. For subpopulations within continents, the raw material is meager. For continental populations, the raw material is abundant.
An Operational Definition of “Large”
To demonstrate that abundance, I need a summary statistic for conveying how many SNPs fall far from the diagonal in the scatter plots. I settled on an operational definition of “large” that defines “large” relative to differences within continental subpopulations: A difference in target allele frequencies is called “large” if it is bigger than 99 percent of the target allele frequency differences found within continental subpopulations. To calculate that number, I used all 43,543 unique SNPs in the GWAS Catalog that are also found in Phase 1 of the 1000 Genomes Project. Combining all of the 12 pairs of within-continent subpopulations produced a sample of 522,516 pairs of target allele frequencies. Twenty percent of the absolute differences in target allele frequencies were less than .01, 63 percent were less than .05, and 88 percent were less than .10.12 Ninety-nine percent were less than .19—to be more precise, less than .186. Thus my operational definition says that the smallest between-continent difference that is “large” is anything greater than .186. For convenience, I will round up and use .20 as the criterion. It’s easier to remember.
In other words, if Asians have a target allele frequency of .45 for a certain SNP and Europeans have a target allele frequency of .65 or higher on the same SNP, that difference qualifies as “large.” If Europeans have a target allele frequency of .25 or less, that difference also qualifies as “large.” What’s important is the absolute difference between two populations.
How many SNPs show that large a difference? The following table shows the results for 112 phenotypic traits grouped into three types of noncognitive traits and three types of cognitive traits. The noncognitive traits are major diseases such as breast cancer and Parkinson’s disease, physiological biomarkers such as height and weight, and blood parameters such as red cell count and metabolite levels. The cognitive traits are cognitive disorders such as depression, cognitive ability (both IQ and neurocognitive functioning), and personality features such as risk-taking tolerance and life satisfaction. The note gives details.[13]
TARGET ALLELE DIFFERENCES QUALIFYING AS “LARGE” (.20+)
Physiological Traits
No. of Unique SNPs: 13,431
Total: 33%
African-Asian: 37%
European-African: 33%
Asian-European: 30%
Diseases
No. of Unique SNPs: 3,718
Total: 33%
African-Asian: 38%
European-African: 33%
Asian-European: 30%
Biomarkers
No. of Unique SNPs: 5,298
Total: 35%
African-Asian: 39%
European-African: 35%
Asian-European: 31%
Blood parameters
No. of Unique SNPs: 4,415
Total: 31%
African-Asian: 35%
European-African: 31%
Asian-European: 28%
Cognitive Traits
No. of Unique SNPs: 9,628
Total: 36%
African-Asian: 39%
European-African: 37%
Asian-European: 32%
Cognitive disorders
No. of Unique SNPs: 2,594
Total: 35%
African-Asian: 38%
European-African: 37%
Asian-European: 31%
Mental abilities
No. of Unique SNPs: 5,715
Total: 36%
African-Asian: 39%
European-African: 36%
Asian-European: 32%
Personality features
No. of Unique SNPs: 1,319
Total: 38%
African-Asian: 42%
European-African: 38%
Asian-European: 35%
Source: Author’s analysis, GWAS Catalog and Phase 1 of the 1000 Genomes Project.
When comparing the three continental populations, about a third of all target allele differences are at least .20.14 Note that .20 is the smallest difference that qualifies. The mean difference among those that qualify is .33 for both the physiological and cognitive traits.
The results for this subset of traits generalizes to all 2,147 traits in the GWAS Catalog as of May 2019 that also had SNPs represented in Phase 1 of the 1000 Genome Project. For the combined noncognitive traits, 32 percent of target allele differences across continental populations qualified as large. For the combined cognitive traits, 34 percent qualified as large.
One other feature of the results generalizes: The three continental pairs are consistently ordered. Africans and Asians have the highest proportion of large differences, Asians and Europeans have the smallest proportion, and Africans and Europeans are in between. This is consistent with the theoretically expected relationship between geographic and genetic differences between populations discussed in chapter 7.
The Traits Related to Cognitive Repertoires
The table here presents information on 22 traits related to personality, abilities, and social behavior that have at least 100 SNPs associated with them. Unlike the previous table, this one combines SNPs that are given different labels in the GWAS Catalog but are associated with the same trait. For example, the trait labeled “well-being” in the table combines SNPs from studies in the GWAS Catalog that were for traits labeled “eudaimonic well-being” and “subjective well-being.” The note gives additional information about the traits.[15]
The table provides more information than most readers need. Its purpose is to enable skeptical readers to look at the results from a variety of perspectives. Suppose, for example you think that an absolute difference of .20 is not sufficiently big. The table also shows you the percentage of allele differences that met a threshold of .25, which exceeds 99.9 percent of the within-continent differences. It also shows you the between-continent correlations and the mean allele difference for those traits that met the .20 threshold. Proposition #7 claims that “Continental population differences in variants associated with personality, abilities, and social behavior are common.” In effect, the table says that the data confirm that proposition no matter how you look at them.
The Inevitability of Interesting Questions to Ask
Even though we don’t know what analyses of these data will show, the existence of so many differences in target allele frequencies will raise interesting questions for a simple reason: We already know that the target alleles for two populations seldom balance out. Look again at the Asian-European scatter plot for schizophrenia.
Source: Author’s analysis, GWAS Catalog and Phase 1 of the 1000 Genomes Project.
DIFFERENCES IN TARGET ALLELE FREQUENCIES FOR TRAITS RELATED TO COGNITIVE REPERTOIRES
If the investigator’s ambition is to identify a role for natural selection in creating population differences, there’s no telling whether anything interesting lies in that plot. Getting from raw differences in target
allele frequencies to evidence of natural selection is a long and torturous process, and even then the results should be treated provisionally.16 For that matter, proof of the role of natural selection for many genetic differences will remain unobtainable without methodological breakthroughs. Recall from chapter 8 that one of the most commonly used tools doesn’t work for adaptations that occurred more than 30,000 years ago.
But while proving natural selection is difficult, the differences in target allele frequencies across populations can be analyzed without knowing what caused the differences. Such analyses can’t be done now with any confidence because of the problems of population stratification, but they will become feasible within a few years when large databases from the major ancestral populations are available.
For purposes of illustration, let’s jump ahead to that time and suppose that the schizophrenia scatter plot for Asians and Europeans is free of contamination by population stratification and that target allele frequencies and the weights associated with them can be taken at face value (very big suppositions). The 500 SNPs shown in the scatter plot do not reveal an obvious imbalance between the target allele frequencies above and below the diagonal, but it turns out a modest one does exist. In the full sample of 962 SNPs associated with schizophrenia in the GWAS Catalog, Asians have the higher target allele frequency for 513 SNPs compared to 449 for Europeans. This opens the possibility—only a possibility—that Asians are genetically more susceptible to schizophrenia than Europeans. Whether it is true depends on the magnitude of the differences in target allele frequencies, the effect sizes associated with the SNPs, and a variety of other considerations. Even if population stratification is no longer a problem, the raw difference is useful only for deciding whether it is worthwhile to curate the sample of SNPs to cleanse it of contaminating factors and to analyze polygenic scores for Europeans and Asians. Perhaps the imbalance of 513 to 449 in the raw data will turn out to be meaningful; perhaps it won’t.
Human Diversity Page 23