FOCUS4-D and Adavosertib in RAS-TP53-Mutant Disease

Why FOCUS4-D’s PFS Result Needs Careful Reading

A randomized comparison can show whether an assigned treatment changes outcomes for a defined group, but the result still needs to be read in context. FOCUS4-D studied adavosertib, a WEE1 inhibitor, against active monitoring in people with RAS- and TP53-mutant metastatic colorectal cancer whose disease was controlled after first-line treatment. Its principal result concerns progression-free survival (PFS): the time from randomization until disease progression or death, whichever occurred first.

The comparison was not adavosertib against no care from diagnosis. Participants entered this maintenance-setting comparison after initial treatment and were randomized to adavosertib or active monitoring. The question answered by the PFS analysis is therefore about the effect of assignment to adavosertib versus active monitoring in this selected population and at that point in their disease course.

Definition: PFS is the time from the defined starting point, here randomization, to the first recorded disease progression or death from any cause. A participant without a recorded PFS event is censored at the relevant last assessment under the analysis rules. A Kaplan–Meier estimate of PFS, written \(\hat{S}(t)\), estimates the probability of remaining alive without a recorded progression beyond time \(t\), subject to those rules.

The Main FOCUS4-D Estimate

FOCUS4-D reported median PFS of 3.61 months with adavosertib and 1.87 months with active monitoring. The reported hazard ratio (HR) was 0.35, with a 95% confidence interval of 0.18 to 0.68 and a reported \(p\)-value of 0.0022. These values provide related but different summaries: the medians describe points on the estimated PFS curves, while the HR compares event rates over follow-up under a time-to-event model.

Interpretation: The reported HR of 0.35 indicates that, under the model used to estimate it, the instantaneous rate of a PFS event in the adavosertib group was estimated to be 0.35 times the rate in the active-monitoring group at a given time. Equivalently, the estimated relative hazard was 65% lower. This is a relative rate comparison, not a statement that 65% fewer participants progressed, nor that PFS time was multiplied by a particular factor.

The 95% confidence interval, 0.18 to 0.68, is below 1 throughout, so it is consistent with a lower PFS-event hazard under adavosertib than active monitoring. In the repeated-sampling interpretation of a confidence interval, the method would capture the underlying HR in 95% of intervals constructed in this way under its assumptions. The interval does not assign a 95% probability to this particular interval containing the HR.

The reported \(p=0.0022\) indicates that, if the null hypothesis of no treatment-group difference held and the test assumptions were satisfied, a result at least as incompatible with that null as the observed one would be unusual. It is not the probability that the null hypothesis is true. Taken together, the HR, interval, and p-value provide evidence of longer PFS with adavosertib in this trial population. They do not establish benefit for every patient or every molecular subgroup.

Worked Example: Reading the Reported Hazard Ratio

Use the FOCUS4-D report: HR \(=0.35\), 95% CI \(0.18\) to \(0.68\), and reported \(p=0.0022\). State the comparison and interpretation.

The reference group is active monitoring, so the ratio is the estimated hazard in the adavosertib group divided by the estimated hazard in the active-monitoring group. Under the fitted time-to-event model, the estimate is \(0.35\), which is below the no-difference value of 1.

$$ 1-0.35=0.65 $$

Thus the point estimate corresponds to a 65% lower PFS-event hazard for adavosertib relative to active monitoring, conditional on the model interpretation. The confidence interval also lies below 1: \(0.18<1\) and \(0.68<1\). The result is statistically inconsistent with equal hazards at conventional two-sided significance levels, as also reflected by the reported \(p=0.0022\).

This does not mean the adavosertib group had 65% fewer events by the end of follow-up. Nor does it mean every participant's PFS increased by 65%. Those claims concern different quantities and cannot be inferred from this HR alone.

What the Median PFS Does and Does Not Mean

The median PFS is read from the Kaplan–Meier estimate, a step function that changes at event times. It is the earliest time at which the estimated PFS probability is at most 0.5, if the curve reaches that level. A step can move from above 0.5 to below 0.5 without taking the value exactly 0.5. If the estimated curve never reaches 0.5 during observed follow-up, the median is not estimable from that curve.

The reported medians of 3.61 and 1.87 months summarize when each estimated curve reached this threshold. Their difference is 1.74 months, and their ratio is about 1.93. These are descriptive comparisons of estimated medians. The HR of 0.35 is not the reciprocal of 1.93, and neither the difference nor the ratio of medians substitutes for the HR.

Worked Example: Finding a Kaplan–Meier Median

Suppose a hypothetical trial has an estimated PFS curve with the following steps: \(\hat{S}(2)=0.72\), \(\hat{S}(3)=0.54\), and, after an event at 4 months, \(\hat{S}(4)=0.46\). Assume the curve is at or below 0.46 immediately after that event and has not reached 0.5 earlier.

At 3 months the estimated PFS probability is \(0.54>0.5\). At 4 months it is \(0.46\leq0.5\). Therefore the earliest time at which the estimate is at most 0.5 is 4 months, so the estimated median PFS is 4 months. The curve need not equal exactly 0.5.

If instead the curve remained above 0.5 through the end of follow-up, the median would not be estimable from the observed curve. It would be incorrect to report the last follow-up time as the median merely because it was the last observed time.

What the Hazard-Ratio Summary Assumes

A hazard is the instantaneous event rate among participants who have not yet experienced the event. A commonly used proportional-hazards model expresses the treatment-group hazard as a constant multiple of the reference-group hazard over time. When proportional hazards is a reasonable approximation, one HR can summarize the relative event rate over the follow-up period. If the relative hazards vary substantially over time, a single HR may conceal that pattern and should not be treated as a constant effect at every time.

Interpretation also depends on how censoring and event ascertainment operate. Kaplan–Meier and standard time-to-event analyses require censoring to be suitably non-informative, often understood as the censored participants having the same subsequent event prospects as comparable participants remaining under observation, conditional on the analysis information. If participants leave observation for reasons related to unrecorded progression risk, the estimated curves can be biased. In a randomized trial, randomization supports a fair comparison of assigned groups at baseline, but it does not by itself eliminate bias from differential follow-up, outcome assessment, or departures from the analysis assumptions.

Worked Example: Checking the Log-HR Scale

A hazard ratio is often analyzed on the logarithmic scale. For the reported FOCUS4-D estimate, \(\log(0.35)\approx-1.050\). The confidence limits transform to \(\log(0.18)\approx-1.715\) and \(\log(0.68)\approx-0.386\).

SE(\log HR)\ \approx\ \frac{\log(0.68)-\log(0.18)}{2(1.96)} =\frac{-0.386-(-1.715)}{3.92} \approx 0.339

This is an approximate standard error reconstructed from rounded confidence limits, assuming a roughly symmetric interval on the log scale. The corresponding approximate test statistic is \(-1.050/0.339\approx-3.10\), which is consistent with strong evidence against a zero log-HR. It will not reproduce the reported \(p=0.0022\) exactly: the published HR and limits are rounded, and the reported test may use details not recoverable from these rounded values.

Why the Comparison Matters—and Its Limits

Active monitoring is a meaningful comparator in a maintenance setting: it represents monitoring rather than immediate assignment to the study drug, while allowing clinical care as appropriate. FOCUS4-D therefore addresses whether adavosertib delayed a PFS event relative to this strategy for the eligible molecularly defined group. It does not compare adavosertib with every possible maintenance treatment, and it does not isolate what would happen if treatment were given earlier in the disease course.

The randomized comparison gives stronger support for a treatment-effect interpretation than an uncontrolled comparison of outcomes before and after treatment. Still, the result pertains to patients who met the trial’s eligibility criteria and reached the point of randomization. Molecular selection is central to the question: the estimate should not be generalized automatically to patients without the RAS- and TP53-mutant profile.

PFS is clinically relevant, but it is not identical to overall survival, symptom improvement, or quality of life. A favorable PFS estimate alone does not establish effects on those outcomes. Nor does statistical evidence establish that the size of benefit is large enough to outweigh adverse effects, inconvenience, or other treatment burdens for an individual patient. Such judgments require the broader clinical evidence, not the HR in isolation.

Key takeaway: FOCUS4-D reported longer estimated PFS with adavosertib than with active monitoring in its RAS- and TP53-mutant metastatic colorectal cancer population. The HR of 0.35 summarizes a relative event-rate comparison under model assumptions; the medians describe where the estimated curves first reach or cross the 0.5 threshold. Neither summary alone answers every clinical question.

Common Mistakes and Diagnostics

Calling the HR a ratio of median times. An HR of 0.35 does not mean PFS lasts \(1/0.35\) times as long. It compares hazards, whereas medians are points on estimated survival curves.
Claiming the median must equal a time when the curve is exactly 0.5. The Kaplan–Meier curve is a step function. The median is the earliest time it is at most 0.5, if it reaches that level; otherwise it is not estimable.
Reading a 65% lower hazard as 65% fewer patients progressing. A hazard is an instantaneous event rate among those still event-free, not a cumulative event proportion.
Ignoring follow-up and curve shape. A single HR can obscure changing relative hazards. Inspect the estimated curves, numbers at risk, follow-up duration, and censoring patterns when available.
Overgeneralizing the result. The estimate concerns the trial-eligible, biomarker-defined population randomized in this treatment setting, not all colorectal cancer or all maintenance strategies.

As a practical diagnostic, check whether the estimated curves are compatible with a reasonably stable relative separation over time before treating one HR as a complete description. Also check how many participants remain at risk late in follow-up: tail estimates based on few people are especially uncertain. If curves cross or diverge sharply, report that a single proportional-hazards summary may be incomplete rather than forcing a constant-effect interpretation.

Check Your Understanding

Answer each question using the definitions and distinctions in this tutorial. Solutions follow the questions.

In FOCUS4-D, what are the event and the time origin for PFS?
What does an HR of 0.35 compare, and why is it not a claim of 65% fewer events?
A hypothetical Kaplan–Meier curve is 0.52 at 5 months and falls to 0.47 at 6 months, with no earlier value at or below 0.5. What is its median PFS?
What does the reported 95% CI of 0.18 to 0.68 indicate about the HR, and what does it not mean?
Name one limitation of concluding clinical benefit from the FOCUS4-D PFS result alone.

Solutions. (1) The event is the first disease progression or death, whichever occurs first; time starts at randomization. (2) It is the estimated instantaneous PFS-event hazard in the adavosertib group relative to active monitoring under the fitted model. A hazard ratio is not a cumulative event count, so it does not state that 65% fewer participants experienced events. (3) The median is 6 months: that is the earliest time the estimated probability is at most 0.5. (4) The interval is wholly below 1, supporting a lower estimated hazard with adavosertib under the model. It is not a 95% probability that this fixed interval contains the true HR, nor a probability that the null is true. (5) PFS is not overall survival, quality of life, or symptom benefit; the PFS result alone does not establish effects on those outcomes or settle the balance of benefit and treatment burden.