Leveraging large-scale biobank EHRs to enhance pharmacogenetics of cardiometabolic disease medications

Marie C. Sadler, Alexander M Apostolov, Caterina Cevallos, Chiara Auwerx, Diogo M. Ribeiro, Russ B. Altman, Zoltán Kutalik

TL;DR

Researchers leveraged Electronic Health Records (EHRs) from over 50,000 participants in the UK Biobank and the All of Us program to identify genetic variants that influence how patients respond to common cardiometabolic medications. The study successfully recovered known pharmacogenetic markers, discovered novel associations near LDLR and ZNF800 for statin response, and provided a critical methodological warning: adjusting for baseline biomarker levels in longitudinal studies can create "statistical mirages" of genetic effects.

What Problem Does This Paper Address?

Pharmacogenetics (PGx)—the study of how genes affect a person's response to drugs—holds the promise of personalized medicine. However, most PGx studies to date have been small, often with median sample sizes of just over 1,000 participants. This lack of scale makes it difficult to find reliable genetic predictors of drug efficacy.

While large-scale biobanks like the UK Biobank exist, their EHR data is "noisy" and longitudinal. It’s not a controlled clinical trial; patients start and stop medications at different times, and their biomarkers (like cholesterol or blood sugar) are measured sporadically. The authors set out to prove that we can "clean" this EHR data to conduct robust genome-wide association studies (GWAS) and rare variant analyses, ultimately finding the genetic "knobs" that tune drug response.

Methods

The team developed a pipeline to emulate clinical trials using EHR data. They focused on ten "medication-phenotype pairs," such as statins and LDL cholesterol, or metformin and HbA1c (blood sugar).

  1. Cohort Emulation: They identified the first time a patient was prescribed a drug and looked for biomarker measurements in a "pretreatment window" (baseline) and a "post-treatment window."
  2. Genetic Analysis: They performed GWAS for common variants and "burden tests" for rare variants (using Exome Sequencing).
  3. Outcome Modeling: They tested two ways to measure response: absolute change (e.g., mmol/L drop) and relative change (logarithmic difference).
  4. Specificity Testing: Crucially, they compared their results against "medication-naive" individuals—people with the same disease who weren't on the drug. This ensured that a genetic variant was actually affecting the drug response and not just the natural progression of the disease.

Datasets

  • UK Biobank (UKBB): Used for the discovery phase. The researchers extracted data for approximately 50,000 participants across various drug cohorts.
  • All of Us (AoU) Research Program: Used for replication. This is a diverse US-based biobank, providing a vital check on whether the findings hold true across different populations and healthcare systems.

Key Insights and Ideas

  • The "Baseline Adjustment" Trap: One of the most important takeaways is methodological. Many researchers adjust for a patient's baseline level when measuring drug response. The authors demonstrate that if a gene is associated with the baseline level (e.g., a gene that makes your cholesterol naturally high), adjusting for it in a longitudinal model can create a spurious association. They showed that several "known" PGx hits might actually be statistical artifacts of this adjustment.
  • Novel Statin Response Loci: The study identified and replicated novel associations for statin response near the LDLR (Low-Density Lipoprotein Receptor) and ZNF800 genes. While LDLR is a well-known "cholesterol gene," this study specifically links it to how much a statin can further lower that cholesterol.
  • Absolute vs. Relative Gains: The authors found that patients with a high Polygenic Risk Score (PRS) for high cholesterol actually saw a larger absolute reduction in cholesterol when taking statins, but a smaller relative (percentage) reduction. This highlights the importance of how we define "efficacy" in clinical settings.
  • Rare Variant Impact: While common variants stole the show, the study found a significant association between rare variants in the HLA-B gene and relative LDL-C reduction, though this did not replicate in the smaller All of Us cohort, suggesting more data is needed for rare variant PGx.

Key Results

  • Recovery: The study successfully recovered major known signals like APOE and LPA for statin response.
  • Statin GWAS: In the TC (total cholesterol) response GWAS (N = 26,669), the novel ZNF800 signal reached genome-wide significance ($p = 4.09 \times 10^{-8}$).
  • Replication: Four out of seven primary signals discovered in the UKBB were nominally replicated in the All of Us cohort, despite differences in how prescriptions are recorded in the US vs. the UK.
  • PRS Utility: Baseline trait PRSs were significant predictors of drug response across almost all categories, proving that your overall genetic "burden" for a disease heavily influences how well a drug will work for you.

Limitations and Future Work

The authors acknowledge several hurdles:

  • Surrogate Endpoints: The study looks at biomarkers (cholesterol, SBP) rather than "hard" clinical events like heart attacks or strokes. While biomarkers are good proxies, they don't tell the whole story of patient health.
  • Polypharmacy: In the real world, patients are often on multiple drugs. This study struggled to account for the "cocktail effect" of different medications interacting, especially for blood pressure treatments.
  • Sample Size for Rare Variants: Even with 50,000 people, the "N" for specific drugs (like metformin) was too low to find many rare variant associations.

Future Direction: The authors suggest that as biobanks grow to millions of participants, we will finally have the power to move beyond common variants and map the rare genetic architecture of drug response, leading to truly personalized prescribing.

Summary updated

Abstract

Electronic health records (EHRs) coupled with large-scale biobanks offer great promises to unravel the genetic underpinnings of treatment efficacy. However, medication-induced biomarker trajectories stemming from such records remain poorly studied. Here, we extract clinical and medication prescription data from EHRs and conduct GWAS and rare variant burden tests in the UK Biobank (discovery) and the All of Us program (replication) on ten cardiometabolic drug response outcomes including lipid response to statins, HbA1c response to metformin and blood pressure response to antihypertensives (N = 932-28,880). Our discovery analyses in participants of European ancestry recover previously reported pharmacogenetic signals at genome-wide significance level (APOE, LPA and SLCO1B1) and a novel rare variant association in GIMAP5 with HbA1c response to metformin. Importantly, these associations are treatment-specific and not associated with biomarker progression in medication-naive individuals. We also found polygenic risk scores to predict drug response, though they explained less than 2% of the variance. In summary, we present an EHR-based framework to study the genetics of drug response and systematically investigated the common and rare pharmacogenetic contribution to cardiometabolic drug response phenotypes in 41,732 UK Biobank and 14,277 All of Us participants.
BibTeX
@inproceedings{sadler2025leveraging,
  author = {Marie C. Sadler and Alexander M Apostolov and Caterina Cevallos and Chiara Auwerx and Diogo M. Ribeiro and Russ B. Altman and Zoltán Kutalik},
  title = {{Leveraging large-scale biobank EHRs to enhance pharmacogenetics of cardiometabolic disease medications}},
  booktitle = {Nature Communications},
  year = {2025},
  doi = {10.1038/s41467-025-58152-3},
}

Loading related content...

Figure