Designing Genotype-Stratified Drug Trials

Why Genotype-Stratified Trial Design Matters

A pharmacogenomic trial must do more than include a genetic measurement. Its design has to make the planned treatment comparisons possible within the genotype groups of interest. If a genotype is uncommon, too few participants may enter that stratum to estimate its treatment effect precisely. If randomization is not balanced within strata, chance imbalances can also make comparisons less informative. Careful planning links the clinical question to eligibility, genotype classification, treatment assignment, endpoints, and recruitment.

A genotype-stratified randomized trial enrolls participants who meet defined clinical and genetic criteria, classifies them into prespecified genotype strata, and randomizes treatment within those strata. Genotype is not randomized: the randomized factor is treatment. The design can estimate treatment effects within genotype strata and, if planned and adequately powered, compare those effects across strata. It does not by itself establish that a genotype-guided prescribing strategy improves outcomes; that requires a treatment-strategy estimand and an appropriate comparison.

Eligibility and Genotype Strata

Eligibility criteria should identify a clinically relevant population in which the trial question can be answered ethically. Define the condition being treated, prior or concurrent therapies that affect eligibility, safety exclusions, and the time point at which eligibility is assessed. Genetic eligibility, if required, needs its own precise rule: specify the assay, the variants or alleles used, how ambiguous or missing results are handled, and how results map to strata. A genotype-derived phenotype may be useful, but its classification rule should be fixed in advance rather than adapted after outcomes are seen.

The choice between an all-comer design and genetic enrichment depends on the target question. An all-comer trial can estimate effects across the eligible population and assess whether effects differ by genotype, provided enough participants are recruited in each stratum. An enriched trial deliberately recruits a genotype group, which can improve precision for its treatment effect but changes the population directly represented by the trial. Excluding a genotype group is not justified merely because it is less common; it should follow a clinical or scientific rationale, such as a safety constraint or a question explicitly limited to that group.

Genotype frequencies can differ across ancestry groups and recruitment settings. Therefore, planning should use prevalence estimates appropriate to the intended recruitment population, and recruitment monitoring should track who is being enrolled. Ancestry should not be treated as a substitute for genotype. If ancestry or site is relevant to the design or analysis, specify how it will be measured and handled without creating so many randomization strata that allocation becomes impractical.

Randomization Within Genotype Strata

Once participants are eligible and classified, randomize treatment within each genotype stratum. With two treatment arms and a 1:1 allocation ratio, each stratum should receive approximately equal numbers assigned to the two arms. This supports a direct within-stratum comparison and avoids relying on genotype groups having similar sizes. Permuted blocks within strata can help maintain balance during recruitment, especially when enrollment is slow. Allocation should remain concealed until assignment; stratification does not replace sound sequence generation or concealment.

Stratifying on every measured characteristic is usually not workable. A small number of clinically important factors, such as genotype stratum and perhaps site, may be used in the randomization scheme. If there are many strata, some will contain very few participants and block management can become cumbersome. Covariate adjustment in the prespecified analysis may be preferable for additional prognostic factors.

Be clear about what the design randomizes. In a trial comparing a drug with control within genotype groups, both genotypes can receive either randomized treatment. In a trial comparing a genotype-guided strategy with usual care, the randomized arms are strategies, and the treatment received may differ according to the strategy. These designs answer different questions and require different estimands. The intended treatment contrast must therefore be settled before choosing the randomization procedure.

Endpoint Choice and Estimand Alignment

Choose an endpoint that represents the clinical outcome of interest and can be measured consistently across genotype groups and treatment arms. Specify the variable, assessment method, time point or follow-up window, and direction of benefit. If the treatment effect is defined as a mean difference in change from baseline, the endpoint and analysis must correspond to that definition. For a binary outcome, specify the event definition and time horizon. For repeated measurements, specify which visit or summary is primary rather than leaving the primary comparison implicit.

The endpoint should match the estimand: the treatment effect the trial aims to estimate for a defined population and treatment contrast. Decide how treatment discontinuation, switching, rescue therapy, death, and missing measurements enter that estimand and analysis. These events can affect the meaning of a treatment comparison, particularly when their frequency differs across genotype groups. A well-powered calculation for an endpoint that does not match the intended clinical question does not repair that mismatch.

If several genotype-specific effects or endpoints are confirmatory, the design must state which contrasts are primary and how multiplicity is addressed. A trial may designate one overall treatment effect as primary and genotype interaction as secondary, or may require evidence within specified strata. Those choices affect the testing threshold, sample size, and interpretation. They should be explicit rather than decided after results are observed.

Sample Size for Treatment Effects and Interactions

Sample size depends on the effect the design is expected to estimate or detect, the outcome variability or event rate, the significance level, the desired power, and the allocation ratio. For a continuous endpoint, a useful initial calculation assumes independent participants, a common outcome standard deviation \(\sigma\) across arms, equal allocation, and a large-sample normal approximation. For a two-sided test at level \(\alpha\), targeting power \(1-\beta\) to detect a treatment-effect difference \(\delta\) within one genotype stratum, the required number per treatment arm in that stratum is approximately

Here \(z_q\) is the \(q\)-quantile of the standard normal distribution. The calculation is an approximation; the final plan should use methods appropriate to the endpoint, analysis, allocation, and any multiplicity adjustment. For binary, time-to-event, or clustered outcomes, the assumptions and sample-size method must reflect the corresponding outcome model and design.

If the scientific question is whether treatment effects differ by genotype, sample size should target the interaction contrast, not just a treatment effect in one stratum. For two genotype strata \(G_1,G_0\) and treatment arms \(T,C\), let the mean outcome in each cell be \(\mu_{G,T}\) or \(\mu_{G,C}\). The interaction contrast on the mean-difference scale is

Assume independent observations, a common standard deviation \(\sigma\), and \(m\) participants in each of the four genotype-by-treatment cells. Each cell mean has variance \(\sigma^2/m\). The interaction estimate is a signed sum of four independent cell means, so its variance is the sum of their variances:

Its standard error is therefore \(2\sigma/\sqrt{m}\). Under the normal approximation, obtaining power \(1-\beta\) for a two-sided test at level \(\alpha\) requires the standardized interaction to be approximately at least \(z_{1-\alpha/2}+z_{1-\beta}\). Setting \(|\Delta|/(2\sigma/\sqrt{m})\) equal to this target and solving for \(m\) gives

Recruitment, Precision, and Common Pitfalls

Genotype prevalence influences how many people must be screened and how quickly each stratum can fill, but it does not change the number needed in a stratum to achieve a specified precision under a given sample-size model. If an interaction is the primary target, recruitment plans should ensure adequate numbers in all four treatment-by-genotype cells. Targeted recruitment of a rare genotype can be efficient, but it may make the enrolled genotype distribution unlike that of the clinical population. The analysis and interpretation should respect that sampling design; a stratum-specific randomized treatment effect remains interpretable for that stratum, while a population-average effect may require appropriate weighting to represent a different target population.

A common mistake is to multiply a marginal genotype prevalence by an overall clinical eligibility rate and treat the product as the joint screening yield. That is valid only if genotype and clinical eligibility are independent, or if the eligibility rate is conditional on the genotype group being counted. The correction is to estimate \(P(\text{eligible}\mid\text{genotype group})\) and multiply by that group's prevalence, or to state and defend an independence assumption. Another common mistake is to power a trial for a treatment effect in the common genotype and assume that this guarantees power for an interaction. Interaction estimation combines information from all genotype-by-treatment cells; it often requires a larger total sample, especially when one stratum is uncommon.