Separating the genetics of disease, treatment and treatment response using graphical modeling and large-scale electronic health records.

This study introduces a novel graphical modeling framework applied to large-scale electronic health records to disentangle genetic effects on disease, medication selection, and treatment response, successfully identifying specific variants influencing blood pressure therapy and age-specific changes while controlling for confounding factors.

The Gist

Imagine you are trying to figure out why some people's blood pressure drops dramatically when they take a specific pill, while others see no change at all. Is it because of their genes? Is it because they were already very sick? Or is it just random chance?

For a long time, scientists have struggled to untangle these threads. It's like trying to hear a single instrument in a loud orchestra; the music of "disease," "medicine," and "response" all play at the same time, making it hard to tell who is playing what.

This paper introduces a new, clever way to listen to that orchestra. Here is a simple breakdown of what the researchers did and what they found.

The Problem: The "Before and After" Mess

Usually, when doctors study drugs, they look at a patient's health before the pill and after the pill. They try to calculate the "change."

• The Trap: If a patient has very high blood pressure to begin with, they are more likely to be given a strong drug. If their blood pressure drops, is it because the drug worked, or because they were just starting from a very high point?
• The Confusion: It's hard to separate the genetics of having high blood pressure from the genetics of responding to the medicine.

The Solution: A "Time-Traveling" Detective

The researchers built a new computer model (a "graphical model") that acts like a super-smart detective who understands time.

Imagine a timeline of a person's life:

1. Genes: You are born with these. They are the foundation.
2. Baseline Health: Your blood pressure before any treatment.
3. The Doctor's Decision: Based on your baseline, the doctor picks a drug.
4. The Result: Your blood pressure after the drug.

The new model looks at all these steps at once. It asks: "If we already know this person's genes and their starting blood pressure, does their gene still predict what happens after the drug?"

By doing this, the model can filter out the noise. It effectively says, "Okay, we know this gene causes high blood pressure. Let's ignore that part. Now, looking only at the change caused by the drug, is there a different gene at work?"

The Experiment: A Massive Digital Lab

To test this, the researchers used a massive digital library called the UK Biobank.

• The Crowd: They looked at over 211,000 people.
• The Data: They didn't just look at one snapshot; they looked at over 1.4 million blood pressure readings and 1.1 million prescription records over time.
• The Genes: They checked 8.4 million tiny genetic markers (SNPs) and thousands of "broken" genes (LoF variants) that stop a protein from working.

They ran this through their new "Time-Traveling Detective" model to see what it could find.

The Big Discoveries

1. The "Early Life" Blueprint

They found that your genetic blueprint for blood pressure is mostly set in stone before you turn 50.

• Analogy: Think of your blood pressure genetics like the foundation of a house. Most of the work is done before the house is even half-built (before age 50). While there are some small repairs made later in life, the main structure is already there.

2. The "Magic Bullet" Genes (Treatment Response)

This is the most exciting part. The model found four specific genetic spots that don't affect your blood pressure before you take a pill, but they do determine how well the pill works after you take it.

• The ADAMTSL1 Gene: One of these genes was already suspected to help with diuretics (water pills). The model confirmed it.
• The New Findings: They found three other genes (including PANO1 and GTSF1L) that seem to act like a "volume knob" for how much your blood pressure drops when treated. These genes are silent until the medicine is introduced, then they turn on.

3. The "Prescription Predictor" (Why you get this drug)

The model also found genes that predict which drug a doctor will choose for you, even after accounting for your health.

• The ARB Connection: They found a gene called KCNIP4 that makes it less likely a doctor will prescribe a specific class of drugs (ARBs) to a patient, even if that patient needs them. It's like a genetic "traffic light" that subtly influences the doctor's choice.
• The "Broken" Genes: They found that people with broken versions of genes like PKD1 (linked to kidney disease) and SLC35F2 are much more likely to be prescribed specific blood pressure blockers. This helps explain why some patients with kidney issues respond differently to standard treatments.

Why This Matters

Think of this new method as a high-tech filter.

• Old Way: "Does this gene affect blood pressure?" (Answer: Yes, but maybe it's just because the person was sick to begin with).
• New Way: "Does this gene affect the change in blood pressure caused by the drug, once we know the person was sick?" (Answer: Yes! This is the true drug-response gene).

The Bottom Line

This paper proves that we can now use massive computer models to separate the genetics of getting sick from the genetics of getting better.

Instead of a "one-size-fits-all" approach, this brings us closer to precision medicine. In the future, a doctor might look at your genetic "volume knobs" and say, "Based on your genes, this specific pill will work wonders for you, while that other one won't do much." It's a step toward treating the person, not just the disease.

Technical Summary

1. Problem Statement

Pharmacogenomic research faces significant challenges in disentangling three distinct genetic effects that are often confounded in observational data:

1. Genetic effects on baseline disease status: Genetic variants influencing the initial biomarker value (e.g., blood pressure).
2. Genetic effects on treatment selection: Genetic variants influencing whether a patient is prescribed a specific medication (often driven by baseline severity or side-effect profiles).
3. Genetic effects on treatment response: Genetic variants influencing the magnitude of change in the biomarker after treatment.

Traditional approaches, such as calculating "change scores" (post-treatment minus pre-treatment) or adjusting for baseline values in standard Genome-Wide Association Studies (GWAS), introduce statistical biases (e.g., endogeneity, scale-dependency) and fail to separate the genetic drivers of therapy choice from the genetic drivers of therapeutic efficacy. Furthermore, large-scale biobank data often suffer from complex missingness patterns and ascertainment biases that complicate longitudinal analysis.

2. Methodology

The authors propose a novel time-aware graphical modeling framework applied to large-scale Electronic Health Records (EHR) and genomic data.

• Data Source: The study utilizes UK Biobank data ($N=211,845$ individuals) comprising:
  • 1,420,443 repeated blood pressure (BP) measurements.
  • 1,117,900 prescription records for six drug classes (ACEIs, ARBs, CCBs, diuretics, beta-blockers, statins).
  • 8,430,446 imputed SNPs and 17,852 whole-exome sequence loss-of-function (LoF) variants.
• Study Design: A "pre-post" design is constructed where clinical measurements are stratified into age groups. For each individual, data is categorized as pre-treatment, treatment indicator (binary), and post-treatment.
• Graphical Inference Algorithm:
  • The core innovation is a time-aware conditional independence (CI) test algorithm.
  • It learns a graphical model where edges represent partial correlations.
  • Time Indexing: The algorithm enforces temporal causality. When testing the association between a genetic variant and a post-treatment trait, it conditions on:
    1. All other genetic variants.
    2. Pre-treatment baseline measures.
    3. Treatment indicators.
    4. Only preceding variables (never future variables).
  • Correlation Handling: To handle binary (treatment) and continuous (BP) variables, the authors use a heterogeneous correlation matrix (Pearson for continuous, polyserial for continuous-binary, polychoric for binary-binary) to place all variables on a liability scale.
  • Statistical Control: The method employs a Benjamini-Yekutieli (B-Y) False Discovery Rate (FDR) correction to control for the massive number of conditional independence tests performed.
• Validation: The approach was validated through extensive simulations (1,800 datasets) comparing True Positive Rates (TPR) and FDR against standard GWAS and various change-score models. It was also benchmarked against REGENIE GWAS results.

3. Key Contributions

• Novel Statistical Framework: Development of a time-aware graphical model that simultaneously models disease, treatment selection, and treatment response, effectively separating these genetic effects without the bias inherent in change-score phenotypes.
• Handling of Complex Data: The method robustly handles repeated measures, missing data (non-random), and the integration of binary treatment indicators with continuous biomarkers.
• Scalability: The algorithm utilizes GPU parallelization and block-wise LD processing to handle millions of SNPs and hundreds of thousands of individuals efficiently.
• Replication Strategy: The study includes rigorous replication in an independent hold-out set of UK Biobank participants and cross-cohort replication in the All of Us dataset across diverse ancestry groups.

4. Key Results

A. Simulation Performance

• The graphical model maintained FDR well below the 5% threshold across all scenarios, including those with structured missing data.
• It demonstrated superior power (TPR) in detecting age-specific and treatment-response genetic effects compared to standard age-stratified GWAS or models using change scores.

B. Blood Pressure (BP) Genetics

• Baseline Genetics: Genetic variation for BP is predominantly shaped prior to age 50. The study identified hundreds of variants associated with pre-treatment BP.
• Age-Specific Effects: While most genetic architecture is stable, the study identified 127 independent loci associated with age-specific BP changes later in life.
• Treatment Response: After conditioning on baseline BP and therapy, the study identified four independent loci influencing BP treatment response:
  • rs191701026 (PANO1): Associated with post-treatment SBP.
  • rs117686634 (ADAMTSL1): Associated with post-treatment SBP in the <50 age group.
  • rs6065646 (GTSF1L): Associated with post-treatment SBP in the <50 age group.
  • rs11668180 (AC005944.2): Associated with post-treatment DBP in the 56–60 age group.
  • Crucial Finding: These variants were not associated with baseline BP, confirming they specifically influence the response to treatment.

C. Pharmacogenetic Variants (Treatment Selection)

The study identified variants associated with the choice of therapy (conditional on disease status):

• Statin Use: Replicated known variants (e.g., APOE rs7412, PCSK9 rs11591147, LDLR LoF), validating the approach.
• Angiotensin Receptor Blocker (ARB) Use:
  • KCNIP4 (rs12513069): A common intronic variant associated with ARB selection. Carriers were less likely to be prescribed ARBs and more likely to receive ACE inhibitors. This gene regulates Kv4 potassium channels and has been linked to ACEi-induced cough.
  • PKD1 LoF: Loss-of-function variants associated with ARB use. PKD1 mutations cause Polycystic Kidney Disease; these patients may have distinct responses to beta-blockers and diuretics, favoring ARBs.
  • SLC35F2 LoF: A novel candidate gene associated with ARB use.
• Replication: The KCNIP4 and PKD1 associations were successfully replicated in the All of Us cohort across European and American ancestry groups.

5. Significance and Conclusion

• Decoupling Confounders: This study provides a robust statistical solution to the long-standing problem of separating the genetics of disease severity, drug selection, and drug response in observational data.
• Clinical Utility: By identifying variants that specifically influence treatment response (like ADAMTSL1 and KCNIP4) independent of baseline disease, the findings offer potential targets for personalized medicine and pharmacogenomic screening.
• Methodological Advancement: The proposed graphical modeling approach is applicable to any large-scale biobank with longitudinal EHR data, offering a pathway to discover time- and treatment-specific genetic associations that standard GWAS miss.
• Limitations: The study notes that power for binary treatment indicators is lower than for continuous traits, and while the method controls for many confounders, unobserved confounding in observational data remains a potential limitation.

In summary, this paper presents a powerful new framework for pharmacogenomics that moves beyond simple association testing to model the causal flow of genetics through disease, treatment selection, and therapeutic response, yielding novel biological insights into hypertension management.