Sotorasib versus consolidation durvalumab in ctDNA-positive stage III KRAS G12C lung cancer: a trial whose eligible population was defined out of existence

What was tried

Memorial Sloan Kettering Cancer Center opened NCT06333678 on 20 March 2024 as a randomised phase 2 study in unresectable stage III non-small cell lung cancer. Eligibility stacked three conditions. Patients had to carry a KRAS G12C mutation, had to be planned for or recently finished with definitive chemoradiation of 56 to 70 Gy alongside a platinum doublet, and had to show detectable circulating tumour DNA within eight weeks of completing it. Those who cleared all three were randomised open label to continue standard consolidation durvalumab for up to twelve months, or to switch to oral sotorasib at 960 mg daily until progression. The primary endpoint was progression-free survival, and ctDNA was measured again at cycle 3 alongside imaging.

The record was updated on 2 September 2026 with a status of terminated and an actual enrollment of 10. The posted reason reads "Study closed by Amgen due to lack of enrollment."

The biological hypothesis

Two lines of evidence converged on this design.

The first concerned the target. In a Caris cohort of 17,095 NSCLC specimens, 27.5 percent carried a KRAS mutation and G12C accounted for 40 percent of those. Open Targets scores the KRAS to non-small cell lung carcinoma association at 0.804, and its composition matters. It rests on somatic mutation evidence at 0.840, clinical precedence at 0.946 and literature at 0.997, with no germline genetic component. This is a target validated by somatic driver biology and by drug approvals, not by human genetics.

The second concerned the population. Consolidation durvalumab became standard on the strength of PACIFIC, where median progression-free survival was 16.8 months against 5.6 months with placebo, hazard ratio 0.52, and five-year median overall survival was 47.5 against 29.1 months. That average conceals a subgroup that does poorly. On the Big Ten LUN 16-081 trial, patients with ctDNA still detectable after chemoradiation had 24-month progression-free survival of 29 percent against 65 percent for those without, p equals 0.0048.

Together these produced a defensible question. If a patient has a druggable driver and molecular evidence that consolidation immunotherapy is not clearing residual disease, is switching to targeted therapy better than continuing a treatment that appears to be failing?

What actually happened

Ten patients enrolled over roughly 22 months of open accrual before Amgen, the drug supplier, closed the study. No results have been posted, and no analysis of the primary endpoint was possible at that sample size.

The failure was structural rather than scientific. The trial required the simultaneous intersection of a molecular subtype, a disease stage, a completed treatment sequence and a dynamic biomarker measured in a narrow window. Each filter is defensible alone. Multiplied together they left almost nobody.

Failure mechanism, best guess

This was an enrollment collapse produced by compounding eligibility filters, not by a wrong hypothesis.

The molecular arithmetic follows from sourced inputs. Applying the Caris frequencies, KRAS G12C represents 0.275 multiplied by 0.40, or about 11.0 percent of unselected NSCLC, consistent with an independent estimate of 10 to 13 percent of advanced non-squamous NSCLC. Roughly one NSCLC patient in nine passes the first gate.

The remaining gates cannot be quantified from the public record, and this analysis assigns them no numbers. Unresectable stage III is a minority of presentations, and the proportion of those patients with detectable ctDNA in the eight-week window depends on assay, sampling time and disease burden, with no rate specific to this trial published. What can be said is that the product of four filters, one of which already removes about 89 percent of the population, leaves a per-site eligible pool measured in single-digit patients per year at most centres.

A second contributor sat outside the eligibility criteria. The trial asked patients to be randomised away from an established standard of care with a documented survival benefit, on the basis of a biomarker not part of routine practice. Trials that ask a patient to leave a proven therapy accrue more slowly than trials that add to it, and this design combined that friction with a minute denominator.

How to prevent this next time

No endpoint-level data were generated, so no efficacy inference is available and no Bayesian update on the mechanism is warranted. The available levers are design levers, the correct toolkit for this class of failure.

First, compute a pre-activation feasibility count from real screening logs rather than prevalence literature. Multiplying published frequencies gives a national estimate. What determines accrual is how many patients per site per year clear every filter in sequence, which requires each site to count backwards through its own records.

Second, interrogate each filter independently. The persistent ctDNA requirement was the scientifically interesting element but also the narrowest and least standardised gate. A staged design that randomised all ctDNA-positive stage III patients to a switch strategy, with KRAS G12C as a stratification factor rather than an entry criterion, would have preserved most of the question at a fraction of the screening burden.

Third, price the cost of discontinuation. Where the control arm is an established standard with a survival benefit, an add-on design accrues where a switch design does not. Fourth, red-team the cascade before activation, with a reviewer who did not write the protocol computing the compounded pass rate and empowered to call the trial infeasible.

The single highest leverage change would have been to require a site-level, backwards-counted eligible patient census across every filter before activation, and to treat a projected accrual below one patient per site per year as a design failure rather than an operational risk to be managed later.

What this means for similar programs

Biomarker-enriched switch designs in the post-chemoradiation setting are multiplying, and they inherit this arithmetic. Any protocol intersecting an oncogenic driver with a dynamic residual disease marker in a stage-restricted population faces the same collapse regardless of how strong its biology is. The KRAS entry in the Claidex failure graph now carries two programs and a modelled risk score of 33 in the yellow band, driven by phase burden and target saturation rather than by weakness in the biology.

There is also a supply-side lesson. The sponsor of record was an academic centre, but the drug supplier closed the study. Investigator-initiated trials depending on one company for study drug carry a termination risk that appears nowhere in the protocol.

Open questions

The planned enrollment target is not retrievable from the public record as it currently stands, so the shortfall cannot be expressed as a percentage of target. The proportion of stage III patients with detectable ctDNA after chemoradiation using the assay specified here has not been reported. Whether any of the ten enrolled patients were randomised is not disclosed. It is also not established whether patients declined randomisation because of the switch design or whether sites never identified eligible candidates, and those explanations imply different fixes.

Sources

1. • ClinicalTrials.gov, NCT06333678, record as of 2 September 2026. https://clinicaltrials.gov/study/NCT06333678 - Nassar AH et al. Characterization of KRAS Mutation Subtypes in Non-small Cell Lung Cancer. Mol Cancer Ther 2022.- Veluswamy R et al. KRAS G12C in advanced NSCLC: prevalence, co-mutations, and testing. Lung Cancer 2023.- Antonia SJ et al. Durvalumab after Chemoradiotherapy in Stage III NSCLC (PACIFIC). N Engl J Med 2017.- Spigel DR et al. Five-Year Survival Outcomes From the PACIFIC Trial. J Clin Oncol 2022.- Analysis of Circulating Tumor DNA Predicts Outcomes of Short-Course Consolidation Immunotherapy in Unresectable Stage III NSCLC (Big Ten LUN 16-081). J Thorac Oncol 2024.- Open Targets Platform, KRAS (ENSG00000133703) and non-small cell lung carcinoma (MONDO_0005233), association score 0.804, retrieved 8 September 2026. - ChEMBL, sotorasib, CHEMBL4535757. - openFDA FAERS, sotorasib, retrieved 8 September 2026.