Molecular Screening and Accrual in FOCUS4

From Molecular Screening to a Randomized Cohort

A molecularly stratified trial has two linked tasks: identify a patient's tumour characteristics and recruit patients who can enter a relevant treatment comparison. These tasks do not happen at the same moment. A patient may be registered, have tissue tested, receive a molecular classification, and still not be randomized. Disease progression, clinical fitness, consent, assay failure, and whether a suitable comparison is open can all affect the path from screening to randomization.

FOCUS4 was a UK platform trial for patients with metastatic colorectal cancer. Patients entered the platform around the start of first-line treatment, had tumour tissue profiled, and could be considered for randomization after an initial period of chemotherapy if their disease was controlled and they remained eligible. Molecular classification helped direct eligible patients to an available comparison. The trial reported 1,434 patients registered and 656 randomized. The overall randomized-to-registered fraction was \(656/1434 \approx 0.458\), or about 45.8%. This is a description of the overall accrual funnel, not the probability that any one registered patient would be randomized under every circumstance.

The introductory material on FOCUS4's molecularly stratified architecture establishes why treatment comparisons were organized around molecular groups. Here the emphasis is on the population that screening and recruitment produced: who entered the funnel, which patients could be classified, and which of those ultimately contributed to randomized comparisons.

What Each Stage Selects

In FOCUS4, the practical sequence joined molecular and clinical information. Tissue profiling was needed to assign a molecular category, while clinical status after initial treatment mattered for entry into a maintenance comparison. Consequently, the randomized population reflected more than the prevalence of molecular features in colorectal cancer. It also reflected whether suitable tissue was available and informative, whether a patient reached the post-induction point without disqualifying progression or other barriers, and whether an appropriate comparison was recruiting.

A useful accounting table keeps these denominators separate. Let \(N_0\) be the number registered, \(N_T\) the number with a usable tissue result, \(N_E\) the number meeting the later clinical eligibility criteria, and \(N_R\) the number randomized. These counts need not be nested in precisely the same way if the trial permits different routes through screening, so the protocol's definitions should be checked. Where they are sequential and nested, the overall fraction can be decomposed:

This decomposition is useful because a low final accrual fraction can arise from quite different causes: tissue not yielding a result, loss of clinical eligibility, or a mismatch between eligible subgroup and currently available comparison. Those causes call for different operational responses.

Screening Changes the Molecular Composition

The molecular composition among randomized patients can differ from the composition among all registered patients, even when the assays themselves classify tumours correctly. The reason is that randomization is conditional on several events after registration. Let \(G\) denote membership in a particular molecular group and \(R\) denote eventual randomization. The probability of belonging to that group among randomized patients is

This identity is Bayes' rule. It shows that the randomized group's molecular composition depends both on the group's prevalence at registration, \(P(G)\), and on its probability of reaching randomization, \(P(R\mid G)\). The latter can vary because of clinical eligibility, assay success, consent, or access to an open comparison. Thus randomized composition is not simply a readout of population prevalence.

Accrual Is Also a Time-Dependent Process

A platform trial's screening yield can vary over calendar time. An eligible patient may have a molecular category for which a comparison is open at one point but unavailable after recruitment closes or before a new comparison begins. Site activation, laboratory turnaround, and referral patterns can also change. The overall \(N_R/N_0\) fraction can therefore conceal periods or subgroups with poor access to randomization.

For FOCUS4, this matters when interpreting recruitment records: a patient not randomized was not necessarily clinically ineligible, and a molecular subgroup with fewer randomized participants need not have been rare at registration. To distinguish these explanations, a screening log should record mutually understandable reasons for non-randomization and, where possible, the dates of registration, result availability, clinical reassessment, and decision. Useful summaries include:

• the number registered and the proportion with usable tissue results, overall and by site or period;
• time from tissue submission to result, including the distribution rather than only the mean;
• the number clinically eligible at reassessment, separated from the number with an open suitable comparison;
• randomization fractions by molecular group and calendar period, with denominators shown;
• reasons for non-randomization, including assay failure, clinical progression or deterioration, patient choice, and lack of an available comparison.

These are diagnostics, not automatic evidence of a design flaw. For example, a longer assay delay may coincide with changes in case mix or site participation, and a small subgroup may produce unstable percentages. Counts, denominators, and uncertainty should be reported together. When a category has few patients, a one- or two-patient change can shift its percentage substantially; interpretation should not rely on percentages alone.

What Screening Does—and Does Not—Mean for Comparisons

Within a randomized comparison, random allocation supports a causal comparison of assigned treatments among the patients who were eligible, consented, and randomized to that comparison, subject to the usual assumptions about follow-up, outcome measurement, and analysis. Molecular screening does not undo randomization within that enrolled group. However, randomization cannot make the enrolled patients representative of all patients initially registered or of all people with metastatic colorectal cancer. Generalizing beyond the randomized population requires considering the selection pathway and whether treatment effects may vary across patients who did not enter the comparison.

This distinction also clarifies the alternatives. A biomarker-enriched trial recruits only patients meeting a specified molecular criterion; it can focus recruitment on that group, but does not directly provide randomized comparisons for excluded groups. A platform using molecular stratification can support several group-specific comparisons while screening a broader registered population, but it requires a reliable testing and assignment pathway and can face uneven accrual across groups. Neither approach automatically solves representativeness: eligibility, access, consent, and timing still shape the randomized sample.

The limitations of an accrual summary are important. A single overall randomized fraction cannot reveal whether missing tissue was non-random, whether a molecular group had limited access to an open comparison, or whether patients who progressed differed systematically from those with disease control. Likewise, comparing the observed molecular proportions with an external population requires compatible definitions, testing methods, and sampling frames. Such comparisons are descriptive unless their assumptions and uncertainty are addressed.

Common Mistakes and Practical Checks

• Calling the registered-to-randomized fraction a test success rate. It combines assay completion with clinical eligibility, consent, and comparison availability. Report those stages separately where possible.
• Assuming randomized subgroup proportions equal disease prevalence. They reflect both initial prevalence and the probability of reaching randomization. Compare counts at multiple stages before attributing differences to biology.
• Interpreting non-randomization as ineligibility. Some patients may be clinically eligible but have no suitable comparison open, an unresolved assay result, or another reason for not proceeding. Use the recorded reason rather than infer it.
• Assuming a representative sample is required for internal validity. Randomization supports the treatment comparison within the randomized population; representativeness matters chiefly when extending conclusions to other populations.
• Reporting percentages without denominators or dates. A percentage can be misleading when groups differ in size or recruitment exposure. Provide counts, denominators, and relevant time periods.