Command Palette

Search for a command to run...

A genotype-enriched Phase 3 that the enrichment did not rescue

MetabolicBiomarkerAugust 27th, 2026·6 min read·10.5281/zenodo.20479005

Diamyd's DIAGNODE-3 restricted enrolment to HLA DR3-DQ2 carriers, the haplotype in which pooled individual-level data had concentrated the GAD-alum C-peptide effect at ratios of 1.596 and 1.441 versus placebo. At an interim covering 174 evaluable participants, the company reported no clinically meaningful effect on C-peptide in the overall population or in any pre-specified subgroup, and the 321-patient trial stopped for futility. The enrichment marker was carried by everyone and the effect it was meant to concentrate never appeared.

Mechanism Risk Score

ComponentPoints
Phase-weighted failure burden11.8 / 30
Archetype severity8.5 / 25
Temporal recency4.3 / 15
Genetic evidence deficit2.1 / 15
Programmatic saturation2.7 / 15

For GAD2 in Recent-onset type 1 diabetes mellitus (HLA DR3-DQ2 carriers), the Mechanism Risk Score is 29/100 (yellow band). The score is a failure-burden index derived from Claidex post-mortems on this target–disease pair, not a probability of approval.

MRS 29/100 (YELLOW), computed target-level across the 1 GAD2 failure on file in the Claidex graph. Components: phase burden 11.8/30, archetype severity 8.52/25, recency 4.25/15, genetic deficit 2.1/15, saturation 2.72/15. The genetic term uses the Open Targets association of 0.8600 between GAD2 and type 1 diabetes mellitus, which is high and therefore contributes little penalty: this failure is not a case of weak biological anchoring. Saturation is low because Open Targets records no clinical-stage agent against GAD2 and Claidex records one. The dominant terms are phase burden, since the failure was a Phase 3, and archetype severity, since the failure was a biomarker failure in a prospectively genotype-enriched population. The MRS summarises the documented failure record and is not a prediction of future trial outcomes. Archetype counters are left at zero because the failure_graph schema has no biomarker_failure column; the archetype is recorded on the claim row and in the MRS archetype term.

This score does not predict whether the next trial will succeed. It flags how heavy the documented mechanistic failure record is before a new program is justified.

Primary figure supporting this claim (Diamyd (recombinant human GAD65, rhGAD65) in Alhydrogel / GAD2 / Recent-onset type 1 diabetes mellitus (HLA DR3-DQ2 carriers)): A genotype-enriched Phase 3 that the enrichment did not rescue

What was tried

Diamyd Medical AB ran DIAGNODE-3 (NCT05018585), a Phase 3 randomised, quadruple-masked, placebo-controlled trial of intralymphatic recombinant human glutamic acid decarboxylase 65 (rhGAD65, Diamyd) in recently diagnosed type 1 diabetes. It opened on 19 May 2022, ran at 60 sites and enrolled 321 participants. Treatment was three intralymphatic injections of 4 micrograms of rhGAD65 adsorbed to Alhydrogel on days 0, 30 and 60, with oral vitamin D 2000 IU daily from day -30 through day 90. Placebo was Alhydrogel alone plus the same vitamin D.

Eligibility was narrow by design. Participants were aged 12 to 28, diagnosed within six months, GAD65 autoantibody positive, with fasting C-peptide of at least 0.12 nmol/L, and above all carriers of the HLA DR3-DQ2 haplotype, tested centrally because prior genotyping was not accepted. The dual primary endpoints were change from baseline to month 15 in C-peptide area under the curve during a two-hour mixed meal tolerance test, and change in HbA1c. Primary completion was 24 June 2026, and the record carries the terse note "Stopped for futility." No results are posted.

The biological hypothesis

GAD2 encodes the 65 kDa isoform of glutamate decarboxylase, an enzyme whose day job is producing GABA. It is one of the dominant autoantigens in type 1 diabetes, and autoantibodies against it are a diagnostic pillar of the disease. Open Targets scores the GAD2 to type 1 diabetes association at 0.860, with a clinical component of 0.965 and a genetic association component of 0.926, placing it among the better anchored pairs in autoimmunity.

The logic was tolerance induction rather than inhibition. Delivering the autoantigen straight into a lymph node, adsorbed to alum, was meant to push the GAD65-reactive T cell compartment toward a regulatory phenotype and slow destruction of residual beta cells. Preserving even modest endogenous insulin secretion matters clinically, since residual C-peptide tracks with lower rates of severe hypoglycaemia and long-term complications.

The Phase 3 hypothesis was narrower than the general one. Pooled individual-level data from three earlier placebo-controlled GAD-alum trials showed the C-peptide effect concentrated in HLA DR3-DQ2 carriers, with a treatment effect ratio of 1.596 (95% CI 1.132 to 2.249, adjusted p = 0.0035) in carriers lacking DR4-DQ8 and 1.441 (95% CI 1.188 to 1.749, adjusted p = 0.0007) across all DR3-DQ2 genotypes. DIAGNODE-2 then reported that among its DR3-DQ2 carriers, time in range declined by 5.1 percentage points on GAD-alum against 16.7 on placebo at month 15 (p = 0.0075). DIAGNODE-3 was built to confirm that enrichment prospectively.

What actually happened

On 10 April 2026 Diamyd announced that the trial had met pre-specified futility criteria and would stop. The interim analysis covered 174 evaluable participants followed to month 15. The company reported no clinically meaningful effect on C-peptide in the overall population or in any pre-specified subgroup, and no observable difference between arms in HbA1c or time in range. An independent external statistical validation reviewed data derivation, randomisation and statistical assumptions and found no factors that meaningfully affected interpretation. No new safety concerns emerged. Diamyd initiated a strategic review.

No effect estimates or confidence intervals from DIAGNODE-3 have been released, so none are reported here.

Failure mechanism, best guess

The most consistent reading is biomarker failure rather than collapse of the underlying autoantigen hypothesis. Every participant carried the enrichment marker, and the effect it was supposed to concentrate did not appear.

Three features of the supporting evidence are worth weighing. The DR3-DQ2 restriction came from post hoc subgroup analysis of trials that had already missed on their full populations, then pooled retrospectively. The Phase 2b signal that most directly matched the Phase 3 endpoints came from an exploratory continuous glucose monitoring analysis in 27 treated and 15 placebo DR3-DQ2 carriers, 42 people in total. And the pooled effect ratios were compatible with effects as small as 13 to 19 percent at their lower confidence bounds, a thin margin to carry into a 15-month Phase 3 primary endpoint.

HLA DR3-DQ2 selects for a class of antigen presentation. It does not measure whether a given patient's GAD65-reactive repertoire is still plastic enough to redirect, how much beta cell mass remains, or how fast that person is progressing. A genotype that shifts population-average response is not the same instrument as a marker that identifies responders. That distinction is the whole failure.

How to prevent this next time

No endpoint-level estimates were released, so the data cannot support a Bayesian recomputation. Three qualitative levers apply instead.

Separate the discovery subgroup from the confirmation subgroup. The DR3-DQ2 effect was found in the same body of data used to argue it forward. A prospective Phase 2b powered on the enriched population, using the eventual Phase 3 primary endpoint, would have tested the enrichment rather than assuming it.

Adjust the base rate before sizing. Antigen-specific tolerance in recent-onset type 1 diabetes has a long record of Phase 2 signals that do not replicate, and a subgroup-derived effect inherits that base rate rather than escaping it.

Demand a readout that tracks the mechanism, not the risk allele. Immune monitoring of GAD65-reactive T cell phenotype at months 3 and 6 would have answered whether tolerance was being induced at all, independent of whether C-peptide moved.

The single highest leverage change would have been to run a prospectively enriched, adequately powered Phase 2b in DR3-DQ2 carriers using the eventual Phase 3 primary endpoint, instead of promoting a retrospectively pooled subgroup straight into a 321-patient confirmatory trial.

What this means for similar programs

Any programme planning to rescue a marginal effect through genotype enrichment now has a closer comparator. The Claidex graph records a structurally similar case in MAGNET, where lithium was tested in UNC13A rs12608932 C/C carriers with amyotrophic lateral sclerosis and failed. The pattern is the same: a subgroup found after the fact, promoted to an inclusion criterion, never reproduced prospectively.

For antigen-specific immunotherapy the read is narrower than a verdict on the field. Route, dose, timing from diagnosis and booster schedule were fixed choices, and the trial answers only the combination it tested. It does not establish that GAD65 tolerance is unreachable.

Open questions

Will Diamyd publish the DIAGNODE-3 effect estimates, subgroup results and immune-monitoring data, or post results to ClinicalTrials.gov? Without them the field inherits a negative headline with no quantitative detail to calibrate against.

Did the DR3-DQ2 effect fail to replicate, or did it shrink? Those imply different next steps.

Sources

Related failure claims

Linked claims sharing target, indication, or failure mechanism.