Checkpoint blockade accelerated renal medullary carcinoma

What was tried

NCT03274258 was an investigator-initiated, single-arm Phase 2 study at MD Anderson Cancer Center testing nivolumab plus ipilimumab in locally advanced or metastatic renal medullary carcinoma. Ten patients enrolled and all ten were treated. Patients received nivolumab intravenously over 60 minutes and ipilimumab intravenously over 90 minutes on day 1 of repeating cycles. The primary endpoint was objective response rate, with overall survival, progression-free survival and disease control rate also designated primary or co-primary measures in the registry record. The study opened on 29 May 2018 and reached actual primary completion on 21 November 2025. The registry status is TERMINATED and the posted reason is explicit: "The study met its prespecified futility criteria during the prespecified interim futility analysis after enrolling the first cohort of 10 patients." Results are posted on ClinicalTrials.gov and the trial was published in Nature Communications in 2025.

Renal medullary carcinoma is an aggressive kidney cancer that develops almost exclusively in people with sickle cell trait and is defined by loss of the tumour suppressor SMARCB1. A single-centre retrospective series of 135 patients reported that 78% presented with metastatic disease. Treatment options are limited, so a checkpoint combination with activity across other kidney cancers was a reasonable thing to test.

The biological hypothesis

Two lines of evidence pointed toward checkpoint blockade. Rhabdoid tumours, which share biallelic SMARCB1 inactivation with renal medullary carcinoma, were shown to carry T cell and myeloid infiltrates including clonally expanded tissue-resident memory CD8 populations, and checkpoint blockade in an experimental rhabdoid tumour model induced regression of established tumours and durable immune responses. The proposed mechanism was that SMARCB1 loss derepresses endogenous retroviruses and activates interferon signalling, generating immunogenicity that a checkpoint inhibitor can exploit. Separately, nivolumab plus ipilimumab had produced responses in non-clear-cell renal cell carcinoma cohorts, and renal medullary carcinoma is classified within that heterogeneous group.

The target-level support was weak and the trial record now shows it. Open Targets scores the association between CTLA4 and kidney medullary carcinoma at 0.0462, the only target association recorded for that disease term. The same disease term returns SMARCB1 at 0.4241, driven by somatic mutation evidence. The genetics pointed at the chromatin remodeller, not at the checkpoint.

What actually happened

There were no responses. Objective response rate was 0 of 10 and disease control rate was 0 of 10. Median progression-free survival was 1.38 months (95% CI 1.28 to 1.60). Median overall survival was 8.23 months (95% CI 3.45 to not estimable, because the required number of events had not occurred). Four of ten patients had a serious adverse event, with hypotension, abdominal pain, pericardial tamponade, pain and disease progression each reported once, and one death was recorded on study.

The published analysis went further than futility. All ten treated patients experienced rapid disease progression, and five of ten met radiological criteria for hyperprogression. A post-hoc single-cell RNA sequencing analysis of patient samples before and after treatment indicated that checkpoint therapy triggered an interferon-gamma response that induced a myeloid mimicry programme in the tumour cells themselves, regulated by a CEBPB and p300 axis and linked to proliferation. In an immunocompetent somatic mosaic genetically engineered mouse model of the disease, combination checkpoint therapy accelerated tumour growth while activating myeloid-affiliated transcriptional circuits, and selective pharmacologic inhibition of p300 suppressed that programme and restored sensitivity.

Failure mechanism, best guess

The registry archetype is efficacy failure, since the stop was a prespecified futility rule on response rate. The mechanism underneath it is an inversion rather than an absence. Checkpoint blockade was not inert in these tumours, it was actively counterproductive, and the same interferon-gamma signal that is supposed to mark a productive antitumour response instead drove a transcriptional programme that made the tumour grow faster.

The step that failed was the inference from rhabdoid tumour models to renal medullary carcinoma. Both are SMARCB1-null, but shared genotype did not deliver shared immune behaviour. Renal medullary carcinoma arises in a chronically hypoxic renal medulla under sickle cell trait, and SMARCB1 degradation is itself part of the hypoxic stress response in that setting. That is a different tissue context, a different selective history, and a different myeloid environment from a paediatric rhabdoid tumour. The trial design could not have detected the inversion early because a single-arm response-rate design treats acceleration and stable disease identically until progression is measured.

How to prevent this next time

No control arm and ten patients means there is no endpoint-level data here that would support a quantitative recalculation, so the levers are qualitative.

First, treat shared driver genotype as a hypothesis about biology, not as evidence of it. The rhabdoid tumour immunogenicity result was strong in its own system. Extending it required at least one experiment in a renal medullary carcinoma model before opening enrolment, and the model that eventually produced the answer, an immunocompetent somatic mosaic mouse, existed as a technology well before this trial closed.

Second, build hyperprogression detection into single-arm immunotherapy trials in rare aggressive tumours. Pre-treatment growth rate can be estimated from routine staging scans, and comparing on-treatment growth rate against that baseline turns an ambiguous progression event into an interpretable signal. Five of ten meeting hyperprogression criteria was recoverable from the imaging already being collected.

Third, require serial on-treatment biopsy or liquid sampling in first-in-disease immunotherapy studies. The mechanism here was found in paired pre-treatment and post-treatment single-cell data. Collecting that prospectively rather than post hoc would have converted a negative trial into a mechanistic result years earlier.

The single highest leverage change would have been to test the combination in an immunocompetent renal medullary carcinoma model, rather than relying on rhabdoid tumour data, before enrolling the first patient.

What this means for similar programs

This is the first CTLA4 failure in the Claidex graph, and the target-level Mechanism Risk Score is 47 out of 100 (yellow band). The largest single component is the genetic deficit term at 14.31 of 15, which reflects the near-absent Open Targets association between CTLA4 and this disease. That is the signal worth carrying forward: the risk here was not that CTLA4 is a crowded or discredited target, it is that this particular target and disease pair had almost no supporting evidence outside an analogy. Programmes that justify a target on tumour-type analogy rather than on evidence in the tumour type itself should expect the analogy to be the thing that breaks. The finding that p300 inhibition restored sensitivity in the mouse model also converts this failure into a testable next step rather than a dead end.

Open questions

Does the myeloid mimicry programme appear in other SMARCB1-null tumours treated with checkpoint blockade, and if so does it explain heterogeneous responses in rhabdoid tumour cohorts? Is baseline CEBPB or p300 pathway activity measurable in pre-treatment tissue and does it predict which patients hyperprogress? Would sequential p300 inhibition followed by checkpoint therapy be tolerable in patients with sickle cell trait, whose baseline renal and haematological reserve is already compromised? And does anti-PD-1 monotherapy produce the same acceleration, or does the CTLA4 component drive it?

Sources

1. • ClinicalTrials.gov record and posted results for NCT03274258, retrieved 2026-08-26. https://clinicaltrials.gov/study/NCT03274258 - Trial publication with the hyperprogression and myeloid mimicry analysis: Nature Communications 2025.https://- Rhabtumour immunogenicity and checkpoint blockade in an experimental model: Cancer Cell 2019.https://- SMARCB1 and the hypoxic stress response in sickle cell trait: PNAS 2023.https://- Clinical characteristics and outcomes in 135 patients with renal medullary carcinoma: European Urology Oncology 2025.https://- Nivolumab plus ipilimumab in non-clear-cell renal cell carcinoma, CheckMate 920: Journal for ImmunoTherapy of Cancer 2022.https://- Open Targets Platform: CTLA4 (ENSG00000163599) association with kidney medullary carcinoma (MONDO_0006260), score 0.0462; SMARCB1 association with the same disease, score 0.4241; ipilimumab (CHEMBL1789844) mechanism of action. Retrieved 2026-08-26. https://platform.opentargets.org/disease/MONDO_0006260 - openFDA FAERS, ipilimumab generic-name query, retrieved 2026-08-26: 42,912 total reports, 38,136 serious, 9,140 with a fatal outcome. https://api.fda.gov/drug/event.json.