Why FOCUS4 Used Molecularly Stratified Recruitment
A maintenance trial in metastatic colorectal cancer faces two linked questions: which patients are well enough to enter a comparison after first-line treatment, and which treatment strategies are relevant to their tumors? FOCUS4 addressed both through a shared screening process followed by molecularly directed recruitment. The design made it possible to evaluate more than one maintenance strategy within a platform, while keeping each randomized comparison tied to a defined eligible patient group.
The central statistical distinction is between classification and randomization. A tumor’s molecular profile helped identify which comparison a patient could enter; it did not itself assign the patient to a treatment. Randomization then allocated eligible, consenting patients between the treatment strategies in that comparison. Keeping these steps separate is essential when interpreting FOCUS4: molecular groups were not interchangeable populations, and comparisons between groups were not randomized treatment comparisons.
From Screening to a Randomized Comparison
FOCUS4 was a UK platform trial for patients with metastatic colorectal cancer. Its screening pathway assessed tumor molecular features and clinical status during first-line treatment. The platform focused on patients whose disease had not progressed after an initial course of chemotherapy, commonly assessed at about 16 weeks. Patients with progression at that point were not candidates for a maintenance randomization designed for disease that was controlled.
Molecular testing was used to place tumors into protocol-defined groups, informed by features such as RAS, BRAF, PIK3CA, and TP53 status. The precise route through the platform depended on the protocol’s eligibility rules and on which comparisons were open at the time. Patients with a profile relevant to an available molecularly defined comparison could be considered for that comparison. Patients not eligible for a molecularly directed comparison could be considered for a non-stratified maintenance comparison, if they met its criteria. FOCUS4-N, for example, evaluated capecitabine maintenance against active monitoring in a non-stratified population.
A patient’s route therefore depended on more than a laboratory result. Clinical eligibility, disease control after induction treatment, consent, and the availability of an appropriate open comparison all mattered. Screening results could identify a molecular group, but a group label alone did not guarantee enrollment or randomization.
Obtain protocol-required tumor information and assess the patient’s clinical course during first-line treatment.
Determine whether the disease is controlled and whether the patient meets the clinical and protocol requirements for a maintenance comparison.
Use the molecular classification and current protocol options to identify an eligible randomized comparison, including a non-stratified option where applicable.
Allocate the consenting eligible patient to one of that comparison’s maintenance strategies, using its specified allocation procedure.
This structure differs from assigning every screened patient to one common treatment comparison. A molecularly directed comparison might enroll only a subset of the screened population, while a separate comparison could enroll patients with a different profile or eligibility route. The platform shared screening infrastructure and a disease setting; it did not make all participants members of one pooled randomized population.
Worked Example: Routing a Hypothetical Patient
Suppose a hypothetical patient has metastatic colorectal cancer, a tumor profile that meets the criteria for an open molecularly defined comparison, and no disease progression after the protocol’s induction-treatment assessment. The patient satisfies the comparison’s other clinical criteria and consents to participate. What determines the patient’s route?
First, the molecular result makes the patient eligible to be considered for that comparison. It does not assign the experimental treatment. Second, disease control and the remaining clinical criteria establish whether the patient can enter a maintenance randomization. Third, after consent and confirmation that the comparison is open, randomization assigns one of its treatment strategies.
If this comparison uses a 1:1 allocation, the probability of assignment to either strategy is \(1/2\) under the specified randomization procedure. This is a probability of treatment assignment among patients randomized in that comparison, not a claim that half of all screened patients will enter it. Screening failures, patients with progression, patients who decline, and patients without an applicable open comparison are outside that randomization denominator.
What the Randomized Comparison Estimates
The primary treatment contrast is defined within the comparison’s eligible population. For a time-to-event outcome such as progression-free survival (PFS), the comparison asks whether outcomes differ between randomized strategies for patients who met the criteria for that particular comparison. Randomization supports a causal interpretation of the treatment contrast within that population, subject to the usual conditions: treatment allocation is implemented as planned, outcome measurement and follow-up are adequate, and the analysis handles censoring and other protocol features appropriately.
Randomization does not make the molecular groups identical to one another. If one group has a different prognosis, different clinical characteristics, or different eligibility rules, its outcomes cannot be directly compared with those of another group to estimate a treatment effect. The valid randomized contrast is between arms within a trial comparison. Estimating whether a treatment works differently by biomarker status would require an appropriate design and analysis that directly tests treatment-effect variation; observing different results in separate groups is not, by itself, such a test.
Worked Analysis Within One Comparison
A common analysis for PFS uses a Cox proportional-hazards model. This model compares the instantaneous event rates between randomized arms over follow-up, conditional on the model’s covariates. For a simple two-arm comparison, the hazard ratio (HR) is the estimated hazard in the experimental arm divided by the estimated hazard in the control arm. An HR below 1 favors the experimental arm for the modeled event, but its interpretation depends on the model and the proportional-hazards assumption.
Worked Example: Interpreting a Hypothetical PFS Estimate
Consider entirely hypothetical results from one molecularly defined FOCUS4-style comparison. A Cox model estimates an HR of 0.72 for the experimental maintenance strategy versus control, with \(SE(\log(HR))=0.20\). Assume the proportional-hazards model is reasonable, censoring is handled appropriately, and the reported standard error is valid.
The estimated log hazard ratio is \(\log(0.72)=-0.3285\). A Wald 95% confidence interval on the log scale is:
Exponentiating the endpoints gives an approximate 95% confidence interval for the HR:
The point estimate corresponds to an estimated 28% lower instantaneous event hazard in the experimental arm than in control under the model, since \(1-0.72=0.28\). The confidence interval includes 1, so this example does not provide conventional two-sided 5% statistical evidence against an HR of 1 using this Wald calculation. It does not establish that the strategies are equivalent, and it does not mean there is a 95% probability that the true HR lies in this particular interval. These are invented values for illustrating interpretation, not reported FOCUS4 results.
Why Stratification Matters—and Its Limits
Stratifying recruitment allowed FOCUS4 to investigate maintenance strategies in groups selected for molecular relevance, rather than requiring one universal treatment comparison for all screened patients. A shared platform could use common screening processes while opening distinct comparisons as the scientific questions and eligible groups required. This can make recruitment more focused and help avoid diluting a treatment signal by including patients for whom the treatment rationale does not apply.
The design also creates operational and inferential constraints. Biomarker testing must be timely and sufficiently reliable for decisions at the maintenance assessment. Delayed results can make a patient ineligible for an open comparison or complicate enrollment. Small molecular groups may accrue slowly, limiting precision. A comparison that closes or is not yet open cannot enroll otherwise suitable patients through that route. Finally, the population randomized after disease control is selected: results do not directly describe all patients starting first-line therapy.
Useful diagnostics include checking screening and enrollment counts by molecular group, reasons for non-enrollment, time from specimen collection to result, and the proportion of results that are missing or inconclusive. Investigators should also examine whether important protocol deviations or differential follow-up could undermine the randomized comparison. These checks help reveal whether the platform is reaching its intended populations and whether estimates are based on the groups the design claims to address.
Worked Example: Screening Yield and the Randomization Denominator
Suppose, purely hypothetically, that 240 patients are screened. Tumor results classify 60 as eligible for a particular open comparison. Of those 60, 48 have disease control at the maintenance assessment, meet the remaining criteria, and consent. Under 1:1 randomization, how many patients are expected in each arm?
The randomized denominator is 48, not 240 and not 60. With equal allocation, the expected number in each arm is:
The 12 patients among the 60 who did not proceed did not enter the randomized comparison, and neither did the other 180 screened patients. The actual arm counts can differ slightly from 24 if the randomization procedure permits imbalance, but the expected allocation under a 1:1 scheme is equal. Reporting only the 48 randomized patients would describe the comparison; reporting the full screening flow is also important for understanding how the comparison population arose.
Comparison With Other Recruitment Approaches
| Approach | How patients enter comparisons | Main implication |
|---|---|---|
| Single, broadly eligible randomized trial | All clinically eligible patients enter the same treatment comparison, regardless of molecular subgroup. | Provides a broad estimate for its enrolled population, but may not address biomarker-specific treatment questions efficiently. |
| Molecularly stratified platform such as FOCUS4 | Shared screening routes patients to an eligible comparison based on molecular and clinical criteria. | Supports distinct randomized contrasts in defined groups, but requires careful subgroup-specific interpretation and screening oversight. |
| Nonrandomized biomarker-matched treatment assignment | Biomarker status determines treatment without random assignment within the comparison. | Treatment effects can be confounded by prognosis and selection; molecular matching alone does not provide randomized evidence. |
FOCUS4’s platform structure should therefore be understood as a way to organize multiple targeted randomized questions, not as a single experiment that automatically pools all patients or all treatment effects. Each comparison has its own eligible population, allocation, follow-up, and estimand.
Common Mistakes and a Short Justification
- Wrong: “The biomarker assigned the patient to treatment.” Correction: The biomarker helped determine comparison eligibility; randomization assigned treatment within that comparison.
- Wrong: “Patients in two molecular groups can be compared directly to see which treatment works better.” Correction: A randomized treatment contrast is protected by randomization within its own comparison, not across separately recruited groups.
- Wrong: “The results apply to every patient with metastatic colorectal cancer.” Correction: The direct target population is the group meeting that comparison’s clinical and molecular eligibility criteria and entering randomization.
- Wrong: “A result from patients with controlled disease after induction applies equally to patients whose disease progressed.” Correction: Patients with progression were not part of that maintenance randomization, so the result does not directly estimate treatment effects for them.
The reason within-group randomization supports a treatment contrast can be stated simply. Let \(Y(1)\) and \(Y(0)\) denote the outcome a patient would have under the experimental and control strategies, respectively. For patients eligible for a given comparison, random assignment makes treatment independent of these potential outcomes in expectation. Thus, with valid follow-up and analysis, differences between assigned arms estimate the treatment contrast for that comparison’s randomized population. This argument does not establish exchangeability between different molecular groups: those groups were formed by tumor characteristics, not randomized to one another.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions.
- What is the difference between molecular classification and treatment randomization in FOCUS4?
- Why should a treatment effect from one molecular comparison not automatically be applied to another molecular group?
- In a hypothetical comparison, 36 eligible patients are randomized 1:1. What is the expected number per arm?
- A hypothetical estimated HR is 1.15. What does the point estimate indicate about the event hazard, and what further information is needed before drawing a statistical conclusion?
- Why does a maintenance comparison after disease control not directly answer a treatment question for patients whose disease progressed during induction?
Solutions. (1) Molecular classification determines which comparison a patient may enter; randomization assigns treatment among eligible participants in that comparison. (2) Groups may differ in prognosis and eligibility, and assignment across groups was not randomized; a within-group contrast does not establish effects elsewhere. (3) \(36 \times 1/2=18\) patients are expected per arm. (4) The point estimate is a 15% higher instantaneous event hazard in the experimental arm than in control under the model; a confidence interval or test, model assumptions, and the outcome definition are needed to assess uncertainty and evidence. (5) Those patients were excluded from the maintenance randomization, so its randomized contrast does not directly estimate their treatment effect.