Andy Garcia · freelance dev
FRENES
Back to the blog
Science & Decisions

An NAD Booster Did 2.72 Points Worse Than Placebo in 410 Parkinson's Patients. Blood NAD Rose.

October 19, 2026 10 min read

On October 8, JAMA published the result of NOPARK, a phase 3 trial run at 11 centers in Norway. 410 people with early Parkinson disease were randomized to nicotinamide riboside (NR), a precursor of the molecule NAD, at 500 mg twice a day, or to a matching placebo, for 52 weeks, double-blind. On ClinicalTrials.gov, the intervention is listed as a dietary supplement.

After a year, the symptom score of the treated group had risen by 2.45 points. The placebo group moved by minus 0.27, essentially flat. The adjusted difference, 2.72 points, favored placebo, with P = .02. The authors' conclusion: the findings "do not support nicotinamide riboside as a disease-modifying treatment for Parkinson disease."

The one-line version, "a supplement made Parkinson's worse", is close to what the abstract says and misleading at the same time. I am CTO of a LegalTech built on Claude, and I built BidForge, an AI that reads French public tender files for small businesses. Many product decisions start from a small, encouraging signal. This trial is a clean, large-scale example of what happens when a small signal meets a large test, and of how to read the word "slightly".

What I read: the full paper is not open access, and the PubMed Central copy is embargoed until April 8, 2027. So this rests on the abstract and key points, the article information PubMed Central does show (funding, disclosures, contributions), the trial registry, the University of Bergen's release, and a detailed summary of the full paper published by News-Medical on October 12, which I name each time I use it. I did not see the methods or the tables myself. I am not a clinician, and nothing here is advice about what anyone should take.

What the trial measured

Participants were 35 or older, diagnosed within the previous two years, with an abnormal dopamine transporter scan confirming dopaminergic denervation, and on stable dopaminergic treatment. Of 537 people screened, 410 were randomized between May 6, 2020, and June 20, 2024, and 393 made it into the primary analysis. Mean age 65.6, 34 percent women. The last follow-up ended on June 19, 2025.

The primary outcome was the change in the MDS-UPDRS, parts I to III: a standard rating of motor and non-motor symptoms that runs from 0 to 236, higher meaning worse. Next to it, the paper states two thresholds: an improvement of 6.7 points or a worsening of 5.2 points counts as clinically important.

What "slightly worse" means, in numbers

The difference is 2.72 points, with a 95 percent confidence interval of 0.47 to 4.98. Three readings follow.

The whole interval sits above zero. Even the most favorable value inside it leaves the treated group 0.47 points worse. A clinically important benefit, 6.7 points of improvement, is far outside it. On its main question, the trial is not inconclusive. It is negative.

The whole interval also sits below 5.2. The average harm is about half the threshold for a clinically important worsening (2.72 divided by 5.2 is 0.52), and even the top of the interval stays under it. The university's release says it plainly: statistically significant, "but modest and below the commonly used threshold for a clinically important difference."

And P = .02 means that if NR had no effect at all, a gap at least this large, in either direction, would turn up by chance in about 2 percent of trials like this one. It says nothing about size or importance. The interval does that job.

So "slightly worse" has a precise meaning here: worse on average, measurably, and smaller than what the field counts as clinically important.

The secondary outcomes point the same way, with less force. Of five prespecified key secondary outcomes, four showed no significant difference, including the dopamine transporter imaging. The fifth, the Non-Motor Symptoms Scale (0 to 360), worsened more on NR: 4.22 points, interval 1.07 to 7.37, P = .009. One hit out of five comparisons deserves caution, and I could not read how the paper handles multiplicity.

Serious adverse events went the other way: 17 of 205 people (8.3 percent) on NR against 29 of 205 (14.1 percent) on placebo. The abstract gives no test for that gap; News-Medical's summary of the full paper says it was not statistically significant, and the release says overall adverse events were similar. The harm signal is in symptom scores, not in serious adverse events.

NAD rose in the blood. The patients did not improve.

The sentence that stays with me comes from the study director, Professor Charalampos Tzoulis, in the release: "NR clearly increased NAD metabolism, but this did not translate into clinical improvement."

News-Medical's summary of the full paper adds the detail. In exploratory analyses, total blood NAD rose markedly on NR, by 30 μM on average against placebo. But among participants who had brain spectroscopy, the change in the ratio of cerebral NAD to ATP showed no statistically significant difference between the groups after 52 weeks. The compound did what the mechanism said it would do in the blood. In the brain, where the hypothesis needs it, the trial found no significant difference. And the outcome people care about did not follow. Per the release, brain imaging showed no evidence that NR either slowed or accelerated changes in the dopamine system.

The road here: 30 patients, then 20, then 410

The case for the trial was not weak. In June 2018, a paper in Cell Reports found that NR improved mitochondrial function in neurons derived from patients with a genetic form of Parkinson's and prevented dopamine neuron loss in fly models. In 2022, the Bergen group published NADPARK, a phase 1 trial: 30 newly diagnosed patients, 1,000 mg a day or placebo for 30 days. Brain NAD rose by an amount the paper called "significant, but variable", and recipients whose brain NAD rose showed altered brain metabolism "associated with mild clinical improvement". In 2023, the same group published NR-SAFE, which had randomized 20 patients, half of them to 3,000 mg a day, for four weeks. The treated group improved on the same scale, but that change was "also associated with a shorter interval since the last levodopa dose."

Each clinical signal was small and short, and the papers said so: NADPARK called for larger trials, NR-SAFE flagged the levodopa timing. The phase 3 trial had more than 13 times the patients of NADPARK and ran a year instead of a month.

There was a quieter warning too. A 2024 mouse study led from the University of Helsinki, which NOPARK cites, reported that NR first partially rescued behavior and later lowered dopamine in the substantia nigra. Per News-Medical's summary, the authors point to preclinical evidence that NR may affect dopamine metabolism, as consistent with a possible interaction with dopaminergic treatment, without identifying a cause.

Who published the negative result

Tzoulis and a co-author, Christian Dölle, reported patents on the use of NR to treat Parkinson disease (one granted in Europe, the others pending), licensed to Niagen Bioscience. Tzoulis also reported personal fees from Niagen Bioscience. ChromaDex, a Niagen Bioscience company, supplied the NR and the placebo. The funding came from the Research Council of Norway, the K. G. Jebsen Foundation, the Western Norway Regional Health Authority and KlinBeForsk, which, per the paper, had no role in design, analysis or the decision to publish. The trial was registered in June 2018. Archived copies of its registry page from 2021 to 2023 already name the MDS-UPDRS at 52 weeks as the primary outcome, then described with a 0 to 260 range, which covers all four parts of the scale; the current page, last updated in September 2025, specifies parts I to III, the version JAMA reports.

People with patents on the hypothesis ran the trial that undercut it, and published it in JAMA under a conclusion that leaves no room for spin. That is how evidence is supposed to work.

What the evidence does not show

It does not speak to other diseases. The release says the findings do not rule out benefits of NAD-boosting treatments elsewhere, which would need their own controlled trials.

It tests one dose, 1,000 mg a day, for one year, in early Parkinson disease, on top of stable dopaminergic treatment, in Norway.

It cannot cleanly separate a symptomatic effect from a change in the course of the disease. Per News-Medical's summary, the authors list that limitation themselves: there was no washout phase, and assessing patients on medication may hide small differences.

It reports averages, not who worsened; the team says it will look at whether groups of patients responded differently.

It does not explain why the result tilted the wrong way. Per News-Medical, the authors offer hypotheses (sustained NAD augmentation may adversely affect biological processes in Parkinson's, or NR may interact with dopaminergic therapy) without establishing a cause; the team says it is still analyzing the mechanisms.

And it does not turn into a safety alarm: serious adverse events were fewer on NR, and overall adverse events were similar. The researchers' own guidance, in the release, is that they do not recommend NR or other NAD-boosting supplements for people with Parkinson disease outside controlled clinical studies.

What I take from it as a builder

Write the threshold before the result. NOPARK prints 6.7 and 5.2 next to its primary outcome. In an A/B test or a model evaluation, the equivalent is the smallest change worth shipping, written down before the data arrives. Without it, "statistically significant" gets read as a verdict it was never meant to be.

Treat the proxy as a proxy. Blood NAD rose; symptoms did not improve. More clicks, more fields extracted, a higher benchmark score: each is the molecule, not the patient. And measure the proxy as close to the outcome as you can: the easy one moved, the one in the brain did not move significantly. For BidForge, which reads French public tender files for small businesses, the outcome is a better decision about whether to bid, not the number of lines the model pulled out.

Size the claim to the pilot. Thirty patients over a month is a reason to run a bigger test, not a result. So are thirty beta users.

Name the primary metric in advance, and report it when it disappoints. A trial registry does that for medicine, with a version history that shows when the wording changed. A decision log does it for a product team.

The most useful number in this paper is a 2.72 that went the wrong way, measured carefully enough that the next team starts from an answer instead of a hope.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk