RELATIVITY-123: stopped for futility, with the enriched co-primary population faring worse

What was tried

Bristol-Myers Squibb ran RELATIVITY-123 (NCT05328908), a Phase 3 randomised, open-label trial of the relatlimab-nivolumab fixed-dose combination (BMS-986213) against investigator's choice of regorafenib or trifluridine-tipiracil in later-line metastatic colorectal cancer. It opened on 28 April 2022, recorded a primary completion date of 14 July 2025, and randomised 769 participants, 385 to Arm A and 384 to Arm B, treating 383 and 358. Arm A received relatlimab 480 mg plus nivolumab 480 mg every four weeks. Arm B received regorafenib 160 mg daily for 21 days of a 28 day cycle, or trifluridine-tipiracil 35 mg per square metre twice daily. Overall survival was assessed in two primary populations, all randomised participants and those with a PD-L1 combined positive score of at least 1. There was no masking.

Eligibility required colorectal adenocarcinoma, metastatic or unresectable, progressing during or within roughly three months of the last approved standard therapy, after one to four prior metastatic lines including a fluoropyrimidine, oxaliplatin, irinotecan, an anti-VEGF agent, and an anti-EGFR agent where RAS was wild type. Microsatellite instability high or mismatch repair deficient tumours were excluded, so this was a microsatellite stable population by design. Bristol Myers Squibb announced in December 2023 that an independent data monitoring committee had recommended stopping the trial at a planned futility analysis, and said the recommendation was not driven by safety.

The biological hypothesis

LAG-3 is an inhibitory receptor on antigen-activated T cells that delivers negative signals on binding MHC class II and FGL1 (Open Targets function annotation for ENSG00000089692, with FGL1 characterised as a major LAG-3 ligand in). Co-blockade with PD-1 relieves CD8 T cell exhaustion in a way neither antibody achieves alone (DOI 10.1016/j.cell.2024.07.016), and RELATIVITY-047 converted that mechanism into a registrational win in untreated advanced melanoma (DOI 10.1056/NEJMoa2109970).

The colorectal extension rested on descriptive immunology rather than genetics. Microsatellite-unstable colon cancers carry several counter-inhibitory checkpoints alongside PD-1, LAG-3 among them (DOI 10.1158/2159-8290.CD-14-0863), which invited the argument that LAG-3 blockade could widen checkpoint benefit into the microsatellite stable majority. RELATIVITY-123 excluded the deficient group outright and tested that argument directly. The microenvironment of most colorectal tumours is not exhausted-but-recoverable, it is poorly infiltrated (DOI 10.1038/s41568-025-00872-1), and a receptor that matters only on T cells already inside the tumour has little to act on.

Open Targets scores the LAG3 to colorectal carcinoma association at 0.1032, and the only contributing datatype is literature at 0.849. No genetic, somatic or pathway datatype supports the pair, and the hypothesis still reached a 769-patient Phase 3.

What actually happened

Median overall survival was 7.72 months in Arm A (95% CI 6.83 to 9.20) against 8.57 months in Arm B (95% CI 7.33 to 9.79), a stratified hazard ratio of 0.98 (95% CI 0.83 to 1.16). The PD-L1 CPS at least 1 population, a co-primary rather than a post hoc subgroup, did worse: 7.46 against 10.35 months, hazard ratio 1.09 (95% CI 0.87 to 1.38).

Progression-free survival by blinded independent central review was 1.87 months against 2.10 months, hazard ratio 1.63 (95% CI 1.39 to 1.90). Confirmed objective response rate was 3.1 percent (95% CI 1.5 to 5.4) against 1.7 percent (95% CI 0.6 to 3.6), a stratum-adjusted difference of 1.4 points (95% CI -0.8 to 3.6). Eleven of 359 evaluable participants responded in Arm A and six of 360 in Arm B, with median duration of response 11.10 against 5.55 months.

Safety behaved as the class predicts. Serious adverse events occurred in 227 of 383 treated participants in Arm A and 195 of 358 in Arm B, or 59.3 percent against 54.5 percent. Discontinuation for study drug toxicity was 18 of 383 against 10 of 358, or 4.7 percent against 2.8 percent. Immune-mediated events clustered in Arm A: hypothyroidism or thyroiditis in 36, hepatitis in 16 and adrenal insufficiency in nine, against zero or one each in Arm B.

Failure mechanism, best guess

This was an efficacy failure, and the progression-free survival hazard ratio of 1.63 is the informative number. A progression hazard ratio that high, with a confidence interval well clear of 1, means that over the first two months the immunotherapy arm did measurably worse than cytotoxic and antiangiogenic control. Checkpoint blockade has no cytoreductive effect of its own, so removing chemotherapy from a fast-progressing population costs early disease control that LAG-3 and PD-1 blockade never recovered. The 3.1 percent response rate says any responsive subpopulation is smaller than this trial could resolve.

How to prevent this next time

No endpoint-level data were released for subgroups other than PD-L1 CPS and region, so a Bayesian recomputation is not supported. Three levers are.

Qualify the enrichment marker before making it a co-primary. PD-L1 CPS at least 1 was written in as a second primary population and returned a hazard ratio of 1.09 rather than enrichment. A randomised Phase 2 in that stratum would have shown this at a fraction of 769 patients.

Adjust the base rate for the tumour, not the mechanism. Checkpoint blockade in microsatellite stable colorectal cancer has a long record of single-digit response rates, and Claidex already holds a termination in that setting at [[evorpacept-cd47-mss-metastatic-colorectal-cancer-phase2-safety-termination]].

Do not drop chemotherapy in a fast-progressing population without an early-progression guardrail. The futility look that ended this trial ran on overall survival, and a progression-free survival guardrail would have fired earlier.

The single highest leverage change would have been to require a randomised Phase 2 in the PD-L1 CPS at least 1 stratum, powered on progression-free survival, before committing 769 patients to a Phase 3 whose only enrichment hypothesis turned out to select for worse outcomes.

What this means for similar programs

LAG3 now carries three recorded failures in the Claidex graph and a mechanistic risk score of 65, in the orange band. The other two are [[eftilagimod-alfa-lag3-her2-negative-metastatic-breast-cancer-aipac-003-phase2-3-sponsor-termination]] and [[favezelimab-lag3-hodgkin-lymphoma-mk-4280-003-phase1-2-strategic-termination]], and the second was itself closed after a colorectal readout. Open Targets lists six LAG-3 clinical candidates, one approved and three at Phase 3, so the saturation term is real. The score is driven by phase burden and the genetic deficit term, not by safety. LAG-3 blockade remains a real mechanism in melanoma. What RELATIVITY-123 constrains is its transfer into tumours whose problem is infiltration, not exhaustion.

Open questions

Was any tissue-based LAG-3 or MHC class II assay run on the archival samples the protocol required, and does it separate the 11 responders from the rest? Did they share a phenotype that PD-L1 CPS failed to capture? The futility analysis has not been published, so the interim estimates are not public.

Sources

1. • ClinicalTrials.gov, NCT05328908. https://clinicaltrials.gov/study/NCT05328908 - OncLive, 3 January 2024, RELATIVITY-123 futility analysis. https://www.onclive.com/view/futility-analysis-ends-nivolumab-and-relatlimab-evaluation-in-pretreated-mss-mcrc - Relatlimab and Nivolumab versus Nivolumab in Untreated Advanced Melanoma. N Engl J Med, 2022. https://- LAG-3 and PD-1 synergize on CD8+ T cells to drive T cell exhaustion. Cell, 2024. https://- Fibrinogen-like Protein 1 Is a Major Immune Inhibitory Ligand of LAG-3. Cell, 2019. https://- Counter-inhibitory checkpoints in microsatellite instable colon cancer. Cancer Discov, 2015. https://- Nature Reviews Cancer, 2025. The immune microenvironment of colorectal cancer. https://- Open Targets Platform, LAG3 and colorectal carcinoma. https://platform.opentargets.org/evidence/ENSG00000089692/MONDO_0024331 - ChEMBL, CHEMBL3990044. https://www.ebi.ac.uk/chembl/compound_report_card/CHEMBL3990044/ - openFDA FAERS drug event endpoint, 28 August 2026. https://api.fda.gov/drug/event.json.