TL;DR
This perspective paper clarifies the critical distinctions between self-identified race, ethnicity, and genetic ancestry in the context of health-system biobanks. It argues that conflating these social and biological concepts leads to spurious results and reinforces systemic inequities, offering five key recommendations for more accurate and ethical genetic research.
What Problem Does This Paper Address?
The "genomics era" has moved from small, curated studies to massive, EHR-linked biobanks (like the UK Biobank or the All of Us Research Program). While these datasets are powerful, they are fraught with complexity. Researchers often use "race" as a shortcut for "genetics" or "environment," but this is a fundamental category error.
Race and ethnicity are social and power constructs, while genetic ancestry is a biological reflection of human history. When researchers use these terms imprecisely, they risk:
- Spurious Associations: Mistaking population structure (genetic patterns shared by a group) for true disease-causing variants.
- Reinforcing Biases: Validating "biological race" myths that have historically been used for oppression.
- Inaccurate Phenotyping: Ignoring how clinician bias in Electronic Health Records (EHRs) can lead to over- or under-diagnosis in specific groups, which then pollutes downstream genetic analyses.
Methods
As a perspective piece, the paper synthesizes current best practices and methodological challenges rather than introducing a single new algorithm. It reviews:
- Genetically Inferred Ancestry (GIA): Using Principal Component Analysis (PCA) and clustering on genotype matrices to estimate genetic similarity, bypassing the need for self-reported labels.
- Linear Mixed Models (LMMs): Discussing tools like BOLT-LMM and SAIGE that account for population structure and cryptic relatedness without forcing individuals into rigid, ad-hoc racial categories.
- Fine-mapping: Leveraging the different Linkage Disequilibrium (LD) structures across ancestries to pinpoint causal variants.
- Natural Language Processing (NLP): Examining how clinical notes are used to build "phenotypes" and how these algorithms can inherit human biases.
Datasets
The paper provides a comprehensive summary of 10 major EHR-linked biobanks within the United States, including:
- BioMe (Mount Sinai): Notable for its high diversity (only ~27% White).
- BioVu (Vanderbilt): A pioneer in linking DNA to medical records.
- UCLA ATLAS & Penn Medicine Biobank: Examples of health-system-embedded cohorts.
- Reference Panels: The authors reference the use of the 1000 Genomes Project and the TOPMed Project for imputation and ancestry inference.
Key Insights and Ideas
- The "Dirty" EHR: EHR data isn't objective truth. For example, Black patients are diagnosed with schizophrenia four times more often than White patients. If a GWAS uses these labels without questioning them, it might find "genetic signals" that are actually just reflections of clinician bias.
- LD as a Tool, Not a Barrier: Linkage Disequilibrium (the correlation between nearby genetic variants) varies by population. African populations generally have shorter LD blocks due to their longer evolutionary history. This makes them invaluable for "fine-mapping"—the process of distinguishing which specific variant in a cluster actually causes a disease.
- The "One-Drop" Legacy: The paper highlights how historical US practices like "hypodescent" (assigning mixed-race individuals to minority status) still influence how race is recorded in medical systems today, limiting the granularity of the data.
- Beyond the Checkbox: Race is a poor proxy for environment. Instead of using race as a covariate for "lifestyle," researchers should extract more precise variables from EHRs, such as zip-code-level socioeconomic data or pollution exposure.
Key Results
- Eurocentric Bias: The authors note that the vast majority of existing GWAS represent individuals of European ancestry, which leads to poor predictive utility of Polygenic Risk Scores (PRS) when applied to non-European populations.
- Shared Genetic Risk: Despite ancestry differences, meta-analyses show that only about 3% of genome-wide significant loci display significant heterogeneity in effect sizes across ancestries. Most common disease-causing variants are shared globally; it is their frequency and the environment they interact with that differ.
- Biobank Demographics: Supplementary Table 1 quantifies the stark differences in biobank composition, ranging from the Mayo Clinic (~95% White) to BioMe (~34% Asian, ~22% Black).
Limitations and Future Work
- Granularity: Current EHR systems often limit race/ethnicity to a few checkboxes, failing to capture the complexity of multi-racial identities or specific ethnic subgroups.
- Computational Costs: While LMMs are superior for handling diverse populations, they are computationally expensive for biobanks with hundreds of thousands of participants.
- Future Direction: The authors advocate for the integration of Social Determinants of Health (SDoH)—like education, stress, and neighborhood quality—directly into EHR fields to provide a more rigorous alternative to using race as a proxy for environmental factors.
- Terminology Shift: There is a call to move away from "ancestry groups" toward descriptors of "genetic similarity" to avoid the baggage of continental-level labels.
Summary updated