Piromelatine in mild Alzheimer's disease: a 2q12 responder signature that did not replicate

What was tried

NCT05267535 was a randomized, delayed-start, quadruple-masked, placebo-controlled multicenter study of piromelatine 20 mg once nightly in participants aged 60 to 85 with mild dementia due to Alzheimer's disease, sponsored by Neurim Pharmaceuticals with Syneos Health. The primary endpoint was ADAS-Cog 14 change at 26 weeks. A 26-week double-blind period was followed by a 26-week open-label extension on active drug. Participants had to meet NIA-AA criteria for probable Alzheimer's disease with documented decline over at least 12 months. The trial ran from 15 June 2022 to primary completion on 12 September 2024, enrolled 216 participants, and posted results.

One eligibility criterion set this trial apart. Participants who declined whole genome screening for the chromosome 2:107,510,000-107,540,000 polymorphism were excluded, as were carriers of it. The trial enrolled only 2q12 non-carriers.

Piromelatine (ChEMBL CHEMBL4297523, small molecule, maximum phase 2) is a melatonin MT1, MT2 and MT3 agonist with additional agonism at serotonin 5-HT1A and 5-HT1D receptors. The registry gives the reason for stopping directly: "Treatment with 20 mg piromelatine did not demonstrate efficacy over treatment with placebo based on analysis with the ADAS-Cog 14 subscale and with other cognitive and neuropsychiatric parameters."

The biological hypothesis

The melatonergic rationale rests on circadian and sleep disruption as an early and modifiable feature of Alzheimer's disease, and on preclinical work showing that melatonin receptor signaling supports synaptic function. Open Targets scores the MTNR1A association with Alzheimer's disease at 0.231, entirely from literature (0.327) and clinical (0.364) evidence. No genetic association, somatic mutation or animal model evidence contributes. The landscape is thin: Open Targets lists five clinical-stage molecules against MTNR1A, four of them (melatonin, ramelteon, tasimelteon, agomelatine) approved for sleep or mood rather than cognition.

The hypothesis tested here was narrower and came from a genetic post hoc analysis. In ReCognition (NCT02615002), a dose-ranging study of piromelatine 5, 20 and 50 mg in 371 participants with mild Alzheimer's dementia, there were no statistically significant differences from placebo on the primary computerized neuropsychological battery or on secondary outcomes including ADAS-Cog 14. A genome-wide association analysis was then run on the 107 participants who consented to genotyping, comparing responders with non-responders. Six single-nucleotide polymorphisms at 2q12 associated with response at p below 1x10-4 each. Post hoc, carriers of the cluster (27 percent of the genotyped sample) worsened on ADAS-Cog 14 while non-carriers improved, reported as a 2.91 point improvement on piromelatine 20 mg (n=23) against 1.65 on placebo (n=19), p=0.03 (Schneider et al., Journal of Prevention of Alzheimer's Disease 2022). The confirmatory hypothesis was that excluding 2q12 carriers would convert a negative trial into a positive one.

What actually happened

It did not. In the posted results, mean change in ADAS-Cog 14 at 26 weeks was -1.9 (SD 5.59) on piromelatine (n=105) and -2.0 (SD 5.20) on placebo (n=103). Both arms improved slightly, marginally more on placebo, and no statistical comparison was posted for the primary endpoint.

Computing the contrast from those arm-level values gives a difference of 0.1 ADAS-Cog points favouring placebo, a pooled standard deviation of 5.40, and an unadjusted 95 percent confidence interval running from 1.37 points favouring piromelatine to 1.57 favouring placebo. The standardized effect is 0.02. These figures are derived here from the posted means, standard deviations and arm sizes, not from the sponsor's own analysis.

Secondary endpoints ran the same way. ADCS-iADL change was 0.3 against 0.6, and ADCS-CGIC 3.8 (95% CI 3.684 to 3.989) against 3.9 (3.727 to 4.036).

Safety was unremarkable. Under blinding, serious adverse events occurred in 4 of 105 piromelatine participants (3.8 percent) and 0 of 103 on placebo, with no deaths. The four were single cases of coronary artery disease, sinus node dysfunction, hyperglycaemia and musculoskeletal chest pain. Retention was high (101 of 105 and 96 of 103 completed) and fell sharply in the open-label extension. FAERS returned no records for piromelatine, consistent with an agent never marketed.

Failure mechanism, best guess

This is a biomarker failure rather than a target or safety failure. The selection rule came from a subgroup of 42 participants nested inside a genotyped subset of 107, drawn from a trial of 371 that was itself negative on every prespecified endpoint. Every layer of that chain multiplies the chance that the genotype-by-treatment interaction was noise.

The prospective test was also underpowered for the effect it was built on. At the pooled standard deviation of 5.40 observed here and arm sizes of 105 and 103, the standard error of the difference is 0.75 points, and the smallest difference detectable with 80 percent power at two-sided alpha 0.05 is 2.10 points. Against the 1.26 point post hoc estimate from ReCognition, power was approximately 39 percent. Even had that effect been real and fully reproducible, this trial would more likely than not have missed it. The observed effect was instead near zero, which argues for absence rather than a power shortfall.

A secondary contributor is the endpoint switch. ReCognition's 2q12 signal was defined on a computerized battery, with ADAS-Cog 14 secondary. The confirmatory trial made ADAS-Cog 14 primary, so a genotype effect defined on one instrument was asked to reappear on another.

How to prevent this next time

Endpoint-level data exists on both sides of this comparison, so a quantitative check is available.

Shrink the post hoc estimate before powering. A 1.26 point difference estimated from 42 people, selected after the fact from a negative trial, should be discounted heavily. Powering on the unshrunk estimate produced a trial with roughly 39 percent power for the effect it was meant to confirm.

Match the replication endpoint to the discovery endpoint. The 2q12 cluster was discovered on a computerized battery, and confirming on ADAS-Cog 14 changes the question being asked.

Use an enrichment design that preserves the comparison. Excluding carriers removed the only group that could show whether the interaction was real. Randomizing both genotypes with a prespecified interaction test would have returned an interpretable answer.

The single highest leverage change would have been running the 2q12 hypothesis as a prespecified interaction in a genotype-stratified trial powered on a shrunk estimate, rather than as an exclusion criterion powered on the raw post hoc difference.

What this means for similar programs

MTNR1A now carries its first Claidex failure, and the target Mechanism Risk Score is 43 (yellow band). The dominant component is genetic deficit at 11.53 of 15, reflecting an Open Targets association of 0.231 with no genetic evidence behind it. Saturation is low at 9.48 of 15, so the mechanism is not crowded. Melatonergic agents remain reasonable candidates for sleep and circadian endpoints in dementia, but treating them as disease-modifying on cognition now has one clean negative against it in a genotype-selected population. The wider lesson applies to any pharmacogenomic responder signature found in the genotyped minority of a failed trial: low prior, high confirmation cost.

Open questions

Did the excluded 2q12 carriers differ in any measured way from enrollees?

Would a sleep or circadian primary endpoint, where melatonergic pharmacology has the strongest support, have separated in this population?

Was 20 mg the right dose, given that ReCognition tested 5, 20 and 50 mg and the post hoc signal was defined at 20 mg only?

Sources

1. ClinicalTrials.gov. NCT05267535, piromelatine 20 mg in participants with mild dementia due to Alzheimer's disease, including posted results. https://clinicaltrials.gov/study/NCT05267535 Schneider LS, Laudon M, Nir T, et al. A Polymorphism Cluster at the 2q12 locus May Predict Response to Piromelatine in Patients with Mild Alzheimer's Disease. Journal of Prevention of Alzheimer's Disease 2022;9(2):247-254. https://ClinicalTrials.gov. NCT02615002, ReCognition study of piromelatine in mild Alzheimer's disease. https://clinicaltrials.gov/study/NCT02615002 He P, Ouyang X, Zhou S, et al. A novel melatonin agonist Neu-P11 facilitates memory performance and improves cognitive impairment in a rat model of Alzheimer's disease. Hormones and Behavior 2013;64(1):1-7. https://McCleery J, Sharpley AL. Pharmacotherapies for sleep disturbances in dementia. Cochrane Database of Systematic Reviews 2020;11(11):CD009178. https://Open Targets Platform. MTNR1A (ENSG00000168412) and Alzheimer disease (MONDO_0004975). Association score 0.231. Queried 12 September 2026. https://platform.opentargets.org/evidence/ENSG00000168412/MONDO_0004975 ChEMBL. Piromelatine, CHEMBL4297523. https://www.ebi.ac.uk/chembl/compound_report_card/CHEMBL4297523/ openFDA FAERS drug event endpoint, queried 12 September 2026 for medicinalproduct "piromelatine". No records returned.