When Your Genome Overrides Your Prescription

The Hook

Two patients sit in the same cardiology clinic. Same age. Same LDL-C on their last panel. Both prescribed the same polypill—40 mg atorvastatin, ramipril, and aspirin in a single tablet.

One of them carries a PCSK9 loss-of-function variant. A quiet genetic trait that has been lowering their LDL cholesterol since before they were born. The other doesn't.

Here's the question most cardiovascular trials never ask: does the statin component of that polypill do the same thing in both patients?

The polypill is one of the most celebrated ideas in cardiovascular prevention. The SECURE trial showed it reduces recurrent events after myocardial infarction, largely by boosting adherence. The 2024 PolyPars trial cut major cardiovascular events by 50% over 4.6 years. The WHO added fixed-dose combinations for ASCVD prevention to the Essential Medicines List in 2023. These are real, population-level victories.

But the polypill is a population-level intervention. It was designed for a world where we treat risk factors, not genomes. And this is where causal inference forces us to think harder—because the effect of the statin component is not a constant. It depends on what the genome was already doing before the prescription was ever written.

The Biology: A Feedback Loop That Genetics Can Rewrite

To see the causal structure, you need to see the molecular feedback loop:

Step 1. Statins inhibit HMG-CoA reductase → intracellular cholesterol falls → the hepatocyte upregulates LDL receptors on its surface → more LDL is cleared from the blood.

Step 2. But statins also upregulate PCSK9 transcription. PCSK9 protein binds to the very LDL receptors the statin just created—and sends them to the lysosome for degradation. This is the built-in brake on statin efficacy. Every statin partially undermines itself.

Step 3. PCSK9 loss-of-function variants—R46L in Europeans, Y142X and C679X in African Americans—reduce PCSK9 activity. Fewer LDL receptors are degraded. More survive on the hepatocyte surface.

The net result: when you give a statin to a PCSK9 LoF carrier, you're upregulating receptors and fewer are being destroyed. The brake is partially released. The statin works harder.

This is not speculation. The pharmacogenomic data is striking:

The Cohen et al. (2006) NEJM landmark study established the principle: PCSK9 LoF variants in the ARIC cohort produced a 28% reduction in LDL-C among Black participants—but an 88% reduction in coronary heart disease. The cardiovascular benefit was wildly disproportionate to the lipid change. This is because lifelong exposure to lower LDL-C, starting from birth, produces far greater atherosclerotic protection than the same absolute reduction achieved pharmacologically in midlife.

This has a direct implication for the statin inside a polypill: the marginal benefit of the statin component is not uniform across genotypes.

The Causal Structure: Effect Modification, Not Confounding

This is not a confounding problem. This is an effect modification problem—and the distinction matters enormously for study design.

Confounding asks: is the treatment-outcome relationship distorted by a common cause? Effect modification asks: does the treatment effect itself change across levels of a third variable?

In this case, the PCSK9 genotype (G) modifies the causal pathway from statin treatment (A) through LDL-C (M) to cardiovascular outcome (Y). The DAG encodes this:

G → M: Genotype directly lowers LDL-C via lifelong LDLR preservation.

A → M: Statin lowers LDL-C via LDLR upregulation.

G modifies A → M: The statin's LDL-lowering effect is amplified in LoF carriers because the PCSK9 brake is weaker. This is a mechanistic interaction, not a statistical artifact.

M → Y: Lower LDL-C reduces atherosclerosis and cardiovascular events.

G → Y (direct?): The exclusion restriction question. If PCSK9 affects CVD only through LDL-C, the variant is a clean MR instrument. If it has pleiotropic effects on inflammation, Lp(a), or vascular biology—the direct path is open, and the MR estimate is biased.

The critical insight: in a standard polypill trial, PCSK9 genotype is typically unmeasured. The trial estimates an average treatment effect across all genotypes. But if effect modification is substantial, the ATE masks clinically meaningful heterogeneity.

What the Trials Missed—and What Observational Causal Inference Can Recover

The major polypill trials—TIPS-3, SECURE, PolyPars, NEPTUNO—were not designed to detect pharmacogenomic effect modification. They randomized at the patient level (or cluster level), measured composite MACE endpoints, and reported average effects. None genotyped participants for PCSK9 variants.

This is not a criticism. These trials answered the question they were designed to answer: does a polypill strategy improve adherence and reduce cardiovascular events in the average patient? The answer is yes.

But the next question is different: for whom does the polypill's statin component provide marginal benefit beyond what their genome already delivers?

This is a question that observational causal inference, combined with genomic data, is uniquely positioned to answer. Large-scale biobanks that link whole-genome sequencing to electronic health records and pharmacy dispensing data now exist—and they include the very populations where PCSK9 LoF variant frequencies are highest.

The Analytical Challenge: Time-Varying Confounding Meets Genetic Interaction

The analytical challenge is that LDL-C is simultaneously:

A mediator of the statin → CVD pathway (the mechanism by which the statin works).

A time-varying confounder of future treatment decisions (doctors adjust statin doses based on LDL-C levels).

Modified by genotype (PCSK9 LoF carriers have lower LDL-C at every timepoint).

Standard regression cannot handle this. Adjusting for LDL-C blocks the mediating pathway. Not adjusting for it leaves time-varying confounding unaddressed. This is precisely the kind of problem that modern causal inference methods were designed to solve—but they need to be extended with genotype-by-treatment interaction terms to answer the pharmacogenomic question.

The Polypill Paradox

This creates what I call the polypill paradox.

The polypill's greatest strength is its simplicity. One pill. One strategy. No titration. The SECURE trial showed adherence rates of 74% in the polypill arm versus 63% in usual care at 24 months. That adherence gap drives much of the clinical benefit.

But the polypill's greatest strength is also its blind spot. If PCSK9 LoF carriers already have lifelong low LDL-C, the marginal benefit of adding a statin is smaller. Conversely, PCSK9 GoF carriers may need augmentation beyond the fixed polypill dose.

The Mendelian Randomization Angle—and Its Limits

PCSK9 LoF variants are among the most celebrated instruments in Mendelian randomization. The Ference et al. (2016) NEJM study used genetic scores for both PCSK9 and HMGCR to compare LDL-C lowering pathways. Both reduced CVD risk proportionally, but PCSK9 variants produced less diabetes risk.

But MR estimates the lifelong causal effect of genetically lower LDL-C, not the incremental effect of adding a statin to someone who already carries the variant. These are fundamentally different estimands.

The question for polypill strategy is: "Given that this patient already has genetically low LDL-C, how much additional CVD protection does the statin provide?" This is a conditional treatment effect question. MR alone cannot answer it. It requires causal methods designed for observational data with genetic stratification.

"The polypill was born from a beautiful idea: that simplicity saves lives. And it does. But simplicity is not the same as uniformity."

What Needs to Happen Next

The Causal Thread

The polypill was born from a beautiful idea: that simplicity saves lives. And it does. But simplicity is not the same as uniformity. Your genome was writing prescriptions long before your doctor was.

PCSK9 loss-of-function variants don't just lower your cholesterol. They rewrite the causal architecture of how your body responds to statins. They modify the very pathway the polypill is designed to exploit. And until we measure this—until we design studies that respect the causal structure of gene-drug interactions—we'll keep estimating average effects in a world where the average patient doesn't exist.

The thread that connects your genome to your prescription is a causal one. It's time we traced it.