Three Different Claims
A pharmacogenomic result can be statistically associated with a drug response, fit a plausible biological mechanism, and still fail to support a prescribing change. These are not three ways of saying the same thing. Each claim answers a different question, and evidence for one does not automatically establish the next.
The first claim is about a relationship in data. The second concerns a possible explanation for that relationship. The third is a decision claim: whether acting on the genotype is supported in a defined clinical setting. A result may be compelling at one level and uncertain at another. For example, a variant may affect an enzyme involved in drug metabolism, yet the association with a clinical outcome may not replicate. Or the association may replicate, while the size of the dose adjustment needed to improve outcomes remains uncertain.
A useful translation chain is therefore not an automatic sequence. Association needs sound measurement, appropriate analysis, and credible replication. Mechanistic evidence can strengthen an interpretation, but a plausible pathway is not proof of causation. Clinical actionability additionally requires evidence about the consequences of genotype-guided decisions, including relevant benefits and treatment harms.
Worked Example: A Discovery Association That Replicates Weakly
Worked Example: A Discovery Association That Replicates Weakly
Suppose a hypothetical discovery study examines a genotype and a binary adverse drug outcome. The outcome is recorded for 200 carriers and 200 noncarriers. The data are:
| Group | Outcome | No outcome | Total |
|---|---|---|---|
| Carriers | 54 | 146 | 200 |
| Noncarriers | 36 | 164 | 200 |
For this illustration, treat participants as independent and each outcome as a binary observation. The crude odds ratio compares the outcome odds in carriers with those in noncarriers. It describes an association; it is not, by itself, a causal effect. The usual large-sample log-odds-ratio interval also relies on sufficiently large cell counts. Here all four counts are at least 36.
The estimated outcome odds are about 1.685 times as high among carriers as among noncarriers in this discovery sample. To quantify sampling uncertainty, use the large-sample standard error of the log odds ratio:
A 95% Wald interval on the log scale is \(\log(1.685)\pm1.96(0.2434)\), or approximately \(0.0447\) to \(0.9988\). Exponentiating the endpoints gives an odds-ratio interval of approximately 1.046 to 2.715. The two-sided Wald statistic is \(0.5218/0.2434=2.144\), corresponding to an approximate \(p\)-value of 0.032. Under the model, the discovery data provide evidence of an association at the 0.05 level. They do not show that the genotype causes the outcome or that changing treatment based on genotype will prevent it.
Now consider a separate hypothetical replication sample, with the same group sizes:
| Group | Outcome | No outcome | Total |
|---|---|---|---|
| Carriers | 44 | 156 | 200 |
| Noncarriers | 38 | 162 | 200 |
The replication estimate is \(44\times162/(156\times38)=1.202\). Its log-scale standard error is \(\sqrt{1/44+1/156+1/38+1/162}=0.2482\). Thus the approximate 95% odds-ratio interval is \(\exp\{\log(1.202)\pm1.96(0.2482)\}=(0.739,1.956)\), and the approximate two-sided \(p\)-value is 0.458. The replication estimate is in the same direction but is smaller and imprecise; its interval includes 1.
This replication does not establish that the true association is absent. Its interval also includes effects that could matter, as well as little or no association. But the discovery result should not be treated as firmly confirmed: the replication data provide weak evidence about the size and even the existence of the association. Differences in phenotype definitions, treatment context, population composition, or study methods could also affect comparability. A plausible mechanism cannot substitute for resolving these uncertainties.
Biological Plausibility Does Not Specify a Dose
A mechanistic story can be valuable. If a genotype changes the activity of a drug-metabolizing enzyme, it is plausible that drug exposure could differ by genotype. That observation can motivate studies of exposure, response, or toxicity. It does not, on its own, give the size or direction of a safe dose adjustment for an individual patient. Other pathways, co-medications, clinical characteristics, adherence, and variation within genotype groups can all matter.
Even a measured difference in dose or exposure is not automatically a dosing instruction. A dose recommendation must be linked to a relevant treatment target or outcome, and the uncertainty around the estimated difference matters. The following example illustrates why a point estimate alone can be misleading.
Worked Example: An Uncertain Difference in Stable Dose
In a hypothetical observational sample, investigators summarize the stable daily dose among patients receiving the same drug. The carrier group has \(n=20\), a mean dose of 4.8 mg/day, and a standard deviation of 1.6 mg/day. The noncarrier group has \(n=20\), a mean of 5.2 mg/day, and a standard deviation of 1.5 mg/day. Assume independent samples and use a Welch two-sample \(t\) interval, which does not require equal population variances. The dose distributions should be reasonably well behaved for this small-sample interval to be reliable.
Define the difference as carrier mean minus noncarrier mean. The estimated difference is \(4.8-5.2=-0.4\) mg/day. Its standard error is:
The Welch degrees of freedom are approximately 37.8, giving a 95% \(t\)-critical value of about 2.0245. The interval is therefore:
The estimated mean dose is 0.4 mg/day lower among carriers, but this interval is compatible with a meaningfully lower dose, little difference, or a higher dose in carriers. It does not justify instructing every carrier to reduce the dose by 0.4 mg/day. Moreover, these are observational stable-dose summaries: they do not establish that genotype caused the dose difference, that the difference improves a treatment outcome, or that a particular adjustment is safe. A mechanistically plausible genotype may still lack sufficiently precise, outcome-linked dosing evidence.
Actionability Requires Benefits and Harms
Clinical actionability concerns the consequences of using genetic information, not just whether a genotype predicts an outcome. Evidence that genotype-guided treatment improves one efficacy endpoint may be incomplete if important adverse outcomes were not measured. In that case, the evidence cannot establish the overall balance of benefit and harm. This limitation is especially important when a proposed dose change could plausibly alter both effectiveness and toxicity.
A treatment comparison should be interpreted in the population, setting, and outcomes actually studied. Random assignment can support a causal comparison of assigned strategies under appropriate conduct and follow-up, but it does not repair an outcome that was never measured. Nor does an observed difference in one endpoint imply that every relevant outcome changes in the same direction.
Worked Example: Efficacy Is Measured but Harms Are Not
Suppose a hypothetical randomized study assigns 100 patients to genotype-guided treatment and 100 to usual treatment. By the end of follow-up, the efficacy event occurs in 18 patients in the guided group and 24 in the usual-care group. Assume each patient contributes one binary outcome, follow-up is complete, and the groups are independently randomized. The study did not collect serious treatment-harm outcomes.
The efficacy event risks are \(18/100=0.18\) and \(24/100=0.24\). For guided treatment minus usual treatment, the estimated risk difference is \(0.18-0.24=-0.06\), or 6 percentage points fewer efficacy events in the guided group. A large-sample unpooled standard error for this risk difference is:
The approximate 95% interval is \(-0.06\pm1.96(0.05745)=(-0.1726,0.0526)\). In this hypothetical study, the point estimate favors genotype-guided treatment on the measured efficacy endpoint, but the interval includes no difference and also includes effects in the opposite direction. It is not precise evidence of an efficacy advantage.
More importantly for translation, the study provides no estimate of the difference in serious treatment harms because those outcomes were not measured. The efficacy result alone cannot show whether any possible benefit is offset by increased toxicity, or whether the overall consequences favor genotype-guided care. The correct conclusion is limited to the measured endpoint and its uncertainty; a claim of established net clinical benefit would exceed the data.
A Practical Check Before Translating a Result
When reviewing a proposed pharmacogenomic action, keep the claims separate and ask what evidence supports each one. The following sequence is a compact way to expose gaps without assuming that a positive finding at one stage guarantees the next.
Name the genotype, population, treatment, and outcome. Check that the outcome and genotype were measured in a way that answers the stated question.
Consider the estimate and its uncertainty, not only whether a significance threshold was crossed. Examine whether independent data support a similar association and whether study differences limit comparison.
Ask whether mechanistic evidence is directly measured or merely compatible with the result. Treat plausibility as support for a hypothesis, not as confirmation of the clinical association or its cause.
Identify the actual genotype-guided action, the outcomes measured, and the uncertainty in benefits and harms. A prescribing claim needs evidence relevant to the decision, not just evidence that genotype and outcome are associated.
This check also clarifies what a study can and cannot conclude. A replicated association may strengthen confidence that a relationship is real, yet leave the appropriate dose or treatment choice unresolved. A plausible mechanism may make an association worth investigating, yet leave clinical outcomes uncertain. A treatment comparison may show an effect on one measured endpoint, yet fail to establish the balance of outcomes if important harms are missing.
Common Mistake
Why this is wrong: Mechanistic plausibility does not validate the association, a single significant result does not establish reproducibility, and neither result alone shows that acting on genotype improves patient outcomes. The treatment decision may also affect outcomes that were not measured.
Correction: State each claim at the level supported by its evidence. Report the association estimate and uncertainty, evaluate replication and the proposed mechanism separately, and reserve a clinical-actionability claim for evidence that addresses the relevant treatment decision and its benefits and harms.
The distinction is not a demand for perfect certainty before any clinical use. It is a requirement to match the strength and scope of a recommendation to the evidence available. An association can be useful for prioritizing further study; mechanistic evidence can help explain or refine a hypothesis; and clinical evidence can establish whether using the information improves decisions. Keeping these roles distinct prevents a promising biological story from being mistaken for a tested prescribing strategy.
Check Your Understanding
Use the distinctions in this tutorial to answer the following questions.
- A genotype is associated with an adverse reaction in one small study, and a known protein pathway makes the result biologically plausible. Which claims are supported, and which still need evidence?
- In the replication example, the estimated odds ratio is above 1 but its interval includes 1. Why is it incorrect to conclude that the association is absent?
- A genotype is associated with a lower stable dose in an observational sample. Name two reasons this does not by itself justify a dose reduction.
- A randomized genotype-guided trial measures efficacy but not serious treatment harms. What can it conclude about the measured endpoint, and what important claim remains unsupported?
- Why should a clinical recommendation be evaluated separately from the statistical significance of a genotype-outcome association?