Representation changes what a study can resolve.
A more representative dataset can do more than widen participation. In the right study design, it can sharpen inference, reveal population-specific signals and show where a model may not travel unchanged.
Dr. Benjamin Harvey
Watch the 48-second teaching note
Representation is part of the study design
A dataset is never just a collection of rows. It defines which patterns a study can observe, which comparisons it can make and where its conclusions may travel. In genomics, representation is therefore not a footnote added after the analysis. It can change the resolution of the analysis itself.
A 2022 Nature Genetics paper that I coauthored examined the genetic regulation of 4,657 plasma proteins in 7,213 European American and 1,871 African American participants from the Atherosclerosis Risk in Communities study. The team also used data from 467 African American participants in the African American Study of Kidney Disease and Hypertension for replication. Those are the study’s population labels; they should not be treated as claims that socially defined categories describe simple biological boundaries.
The research question was narrower: how do genetic variants near a gene relate to the amount of a corresponding protein found in plasma, and what can analyses across these study populations reveal?
A pQTL connects variation with protein level
A protein quantitative trait locus, or pQTL, is a genomic region where genetic variation is statistically associated with the level of a protein. The paper focused on cis-pQTLs—associations involving variants near the gene that encodes the protein.
The study identified 2,004 proteins in the European American analysis and 1,618 in the African American analysis with associations to common variants in cis-regions, with most of those proteins overlapping. That overlap matters: it shows shared signal. The differences matter too: the analysis also identified population-specific cis-pQTLs.
This does not mean that one population is the reference and another is an exception. It means the evidence should be examined in the populations actually represented, with differences in sample size, genetic variation and correlation patterns kept visible rather than compressed into one universal story.
Broader representation can sharpen inference
One of the paper’s important findings was that the availability of African American samples led to smaller credible sets for some associations. A credible set is a bounded group of variants that statistical analysis considers plausible candidates for an observed signal. A smaller set does not prove which variant is causal, but it can narrow the field for follow-up work.
Why can this happen? Nearby variants are inherited together in different patterns across populations. When the same biological signal is observed against different correlation structures, the comparison can help separate variants that travel together in one dataset. That is an analytical benefit of representation: it can improve localization while also revealing where signals differ.
The published study further developed protein-prediction models in both populations and applied proteome-wide association methods to serum urate and gout. Those are peer-reviewed findings from the 2022 paper—not new measurements reported here.
Teach the boundary, not only the result
A useful classroom exercise would give students one simulated genetic-to-protein signal under two different correlation structures. First, ask each group to analyze one dataset alone. Then combine the evidence and compare the credible sets, overlapping signals and population-specific signals. Finally, ask what would have been invisible if only one dataset had been available.
This is a proposed teaching exercise, not a report of a completed classroom study. Its purpose is to make three boundaries visible: association is not causation, a narrower credible set is not a confirmed causal variant, and a prediction model should not be assumed to perform equally in populations it has not been evaluated in.
The lesson extends beyond genomics. When we ask whether a model is representative, we should also ask what the added representation lets the study resolve. Inclusion can change who benefits from the work. It can also change what the science is capable of seeing.