Amgen built a BiTE that only fires where two antigens meet, then stopped it at primary completion

What was tried

Amgen opened NCT05800964 in June 2023 as a first-in-human Phase 1 of AMG 305 in adults with advanced solid tumours. The design was non-randomised, sequential and open label, with Part A escalating doses by short intravenous infusion and Part B expanding selected tumour types at the recommended Phase 2 dose. Primary endpoints were dose-limiting toxicities over days 1 to 28 and treatment-emergent and treatment-related adverse events over two years. Secondary endpoints covered Cmax, Cmin, AUC and response rate by RECIST 1.1.

Eligibility required ECOG performance status 0 or 1, documented selected solid tumours, exhaustion of available standard therapy and life expectancy beyond three months. Expansion cohorts also required a measurable lesion of 10 mm or more. Untreated central nervous system metastases were excluded.

Actual enrolment reached 37 participants. Primary completion is recorded as 22 January 2026 and the record was updated on 19 October 2026 with an overall status of terminated. The reason given is two words: sponsor decision. No results have been posted.

The biological hypothesis

T cell engagers work in haematological malignancy because lineage antigens such as CD19 sit on a dispensable compartment. Solid tumours lack that luxury. Most candidate surface antigens also appear on normal epithelium, so a high-affinity engager tends to produce on-target off-tumour toxicity before it produces an antitumor effect.

AMG 305 was built to escape that constraint through conditional binding. Amgen described it at AACR in 2023 as a first-in-class dual-targeting BiTE molecule with selective activity for solid tumours co-expressing CDH3, the gene for P-cadherin, and MSLN, the gene for mesothelin. The logic is avidity. Each arm binds weakly enough that a single antigen does not sustain a synapse, while a cell displaying both crosses the threshold for T cell activation. Selectivity comes from the intersection of two expression patterns rather than the specificity of either one.

Independent work supports the principle. A 2025 avidity-driven bispecific antibody-drug conjugate against P-cadherin and CDH17 showed activity in dual-flank mice only where both antigens were present.

The evidence under each arm is uneven. Against the neoplasm term in Open Targets, CDH3 scores 0.475, from literature at 0.993, genetic association at 0.592, animal model at 0.461 and somatic mutation at 0.228. MSLN scores 0.151, from literature at 0.997 and clinical evidence at 0.196, with genetic association at zero. Neither target carries a recorded safety liability in Open Targets.

What actually happened

The study ran for roughly two and a half years, enrolled 37 of an unstated target, reached its recorded primary completion date of 22 January 2026, and was then marked terminated with sponsor decision as the stated reason. No results section exists, so there are no dose-limiting toxicity counts, no adverse event tables, no pharmacokinetic parameters and no response data. AMG 305 has no ChEMBL registration, as expected for an antibody construct, and openFDA returns no adverse event records for a molecule never marketed.

One structural observation needs no speculation. A Phase 1 that reaches primary completion and is then labelled terminated rather than completed usually indicates planned expansion work was not carried out. That distinction is visible in the registry. Everything beyond it is not.

Failure mechanism, best guess

The data needed to assign a mechanism were not posted, so the useful contribution is to name the constraints the molecule had to satisfy and the evidence that would separate them.

Therapeutic index is the first candidate. P-cadherin sits in basal layers of normal stratified epithelium, mesothelin on normal pleura, peritoneum and pericardium. Avidity narrows the window of normal tissue that can trigger a synapse rather than closing it, and any normal cell carrying both remains a target.

Population prevalence is the second and may be the more decisive. A conditional mechanism only works in patients whose tumours satisfy the condition. The posted eligibility criteria name selected solid tumour diseases but do not require testing for CDH3 or MSLN expression, and nothing in the registry indicates dual positivity was an entry requirement. Expansion run without that filter dilutes any true signal by whatever fraction of enrolled tumours express only one antigen.

Precedent supplies the third constraint. A systems pharmacology analysis of PF-06671008, an earlier P-cadherin targeting CD3 bispecific, predicted a clinical terminal half-life near one day and named P-cadherin expression and intratumoural T cell number as the parameters driving efficacy, with soluble P-cadherin acting as a sink. Those sensitivities transfer to any P-cadherin engager.

The archetype recorded is sponsor decision, taken directly from the posted text. No safety signal is asserted, because none is documented.

How to prevent this next time

Endpoint-level data are absent here, so a Bayesian or power calculation would be arithmetic performed on nothing. The levers the evidence does support are qualitative.

Biomarker enrichment is the largest. A molecule whose selectivity argument rests entirely on antigen co-expression should recruit on antigen co-expression. Prospective dual-antigen immunohistochemistry at screening, with a prespecified positivity threshold for each arm, turns the expansion cohort from a mixed population into a test of the actual hypothesis and makes a null result interpretable.

Base-rate adjustment is the second. The evidence profile above is literature-weighted on both arms, with no genetic association for MSLN against neoplasm. A programme in that position should enter the clinic with a lowered prior and heavier early translational readouts.

Red-teaming the population is the third. The design question is what fraction of the planned expansion cohort will carry both antigens above threshold, answered from expression datasets before the first patient is dosed.

The single highest leverage change would have been requiring prospective dual-antigen expression testing as an entry criterion for the expansion cohorts, so a negative result would separate a failed mechanism from an unenriched population.

What this means for similar programs

CDH3 is not a crowded target. ClinicalTrials.gov returns 28 interventional studies matching CDH3 or P-cadherin terms, and agents directed at the antigen itself number about five, including PF-06671008, FF21101, BC3195, a polyepitope plasmid vaccine containing CDH3, and AMG 305. Mesothelin is the opposite, with 116 interventional studies.

The transferable point is that conditional mechanisms need conditional trials. Avidity-gated bispecifics, logic-gated cell therapies and masked prodrug biologics all buy selectivity by making activity depend on a tumour property that varies between patients. Running them in unselected populations discards the property that justified the design.

Open questions

Why was the study terminated after reaching primary completion rather than marked complete? The registry does not say and Amgen has not published.

What fraction of enrolled tumours co-expressed CDH3 and MSLN above a functional threshold? Without that number the 37-patient experience cannot be read either way, and whether avidity delivered in humans the separation it gave in dual-flank models stays open.

Sources

1. • ClinicalTrials.gov, NCT05800964, protocol record, status terminated. https://clinicaltrials.gov/study/NCT05800964 - AACR 2023 abstract ND06, AMG 305, a dual targeting BiTE molecule with selective activity for solid tumors that co-express CDH3 and MSLN. Cancer Res 2023, 83(7 Suppl):ND06. https://- Open Targets Platform, CDH3 and MSLN associations against neoplasm (MONDO_0005070), tractability and safety, queried 26 October 2026. https://platform.opentargets.org/ - Synan A et al. MAbs 2025, 17:2441411. https://- Betts A et al. AAPS J 2019, 21:66. https://- openFDA drug event endpoint, queried 26 October 2026, no AMG 305 records. Article metadata from PubMed.