The drug worked. Every direct measurement said so.
In Roche's Phase 1b trial, people with Parkinson's disease took selnoflast, an NLRP3 inflammasome inhibitor, for twenty-eight days. Blood draws showed IL-1β production suppressed by 91%. Cerebrospinal fluid showed IL-18 levels down 30% — the drug was reaching the brain and doing what NLRP3 inhibitors are designed to do. Motor scores trended better in the treatment group. No serious adverse events.
Then the brain scan came back.
TSPO-PET — the imaging biomarker Roche had pre-specified to measure neuroinflammation — showed a 41% increase in the midbrain. Not a subtle trend. A 41% increase at p = 0.0009, in the exact brain region where an NLRP3 inhibitor should have been reducing neuroinflammation. The most statistically robust result in the entire trial pointed in the opposite direction from every other measurement.
Roche presented this as confirmation. A “pharmacodynamic effect” — proof the drug was engaging microglia. If the signal had decreased, that too would have confirmed mechanism. The instrument had lost its ability to disconfirm anything.
This is not a story about a drug failing or a company spinning. It is a story about what happens when a measurement instrument becomes synonymous with the thing it was supposed to measure — and an entire field responds not by naming the problem, but by quietly walking away from the instrument.
Same Target, Different Instruments
To see why the selnoflast paradox is structural rather than anecdotal, look at NodThera.
NodThera ran its own NLRP3 inhibitor, ruvonoflast, in Parkinson's disease. Same molecular target. Same disease. Same duration — twenty-eight days. The difference was what they chose to measure.
| Selnoflast (Roche) | Ruvonoflast (NodThera) | |
|---|---|---|
| Target | NLRP3 inflammasome | NLRP3 inflammasome |
| Disease | Parkinson’s | Parkinson’s |
| Duration | 28 days | 28 days |
| Peripheral inflammation | IL-1β ↓ 91% | — |
| CNS inflammation | CSF IL-18 ↓ 30% | CSF cytokines → near-healthy |
| Neurodegeneration | — | NfL ↓ 10–20% |
| Microglial activation | TSPO-PET ↑ 41% p = 0.0009 |
sTREM2 ↓ 13–23% |
| Interpretation | “Pharmacodynamic effect” | Convergent efficacy |
| Paradox? | Yes | No |
NodThera measured the construct directly: CSF cytokines fell to near-healthy levels. Neurofilament light chain — a marker of neurodegeneration — dropped 10–20%. Soluble TREM2, a direct marker of microglial activation, declined 13–23%. Every measurement converged. No paradox. No reinterpretation needed.
Roche measured a proxy: an imaging signal interpreted as “neuroinflammation.” The proxy produced a paradox. The direct measurements did not.
Two drugs inhibiting the same inflammasome. One company chose instruments that measure inflammation directly. The other chose an instrument that measures something one level removed — and got a result it had to reframe. The divergence is not about the drugs. It is about the instruments.
Six Layers
The TSPO-PET construct failure is not a single flaw. It is six independent problems, each sufficient to undermine the instrument, all present simultaneously.
1. The Species Gap
Nutma et al. (Nature Communications, 2023) demonstrated that TSPO marks activated microglia in rodents but not reliably in humans. Gene promoter divergence makes TSPO transcription AP1-dependent in Muroidea, with different regulatory mechanisms in primates. The foundational premise of TSPO-PET — that TSPO upregulation equals microglial activation — was established in species where the molecular relationship does not translate.
2. What TSPO Actually Marks
In humans, TSPO-PET signal correlates with CD68+ phagocytic microglia density — not inflammatory phenotype, not damage, just “are microglia eating things.” Two independent lines of evidence converge: Passamonti et al. (Brain, 2025) and Malpetti’s FTD work demonstrating the same CD68+ correspondence. Phagocytosis can be inflammatory — microglia clearing myelin debris in multiple sclerosis — or beneficial — microglia clearing amyloid plaques in Alzheimer’s. The instrument cannot tell which.
3. The Phenotype Problem
This is where the selnoflast result stops being a paradox and starts being predictable. TSPO cannot distinguish pro-inflammatory microglia from reparative microglia. Both are “activated.” Both are phagocytic. An NLRP3 inhibitor that shifts microglia from an inflammatory-and-phagocytic state to a reparative-and-phagocytic state — which is precisely what the drug is designed to do — would show maintained or increased TSPO signal while actual neuroinflammatory damage decreased.
This is likely what selnoflast demonstrated. The drug worked. The instrument was not measuring what everyone assumed it was measuring.
4. The Reinterpretation
Roche pre-specified a TSPO-PET decrease as the expected pharmacodynamic direction. When they observed a 41% increase at p < 0.001, they reframed it as evidence of mechanism engagement.
Consider: if the decrease had been observed, that would also have confirmed mechanism. If no change had been observed, that could be interpreted as the drug working too subtly for the instrument to detect. When every possible result confirms the hypothesis, the measurement has become unfalsifiable. It no longer functions as a test.
5. The Cross-Drug Test
The comparison with ruvonoflast makes the measurement problem structural rather than interpretive. Same target, same biology, same therapeutic mechanism. Different measurement choices yield different narratives. NodThera never had to reinterpret because they never chose a construct-conflated instrument. They measured what they wanted to know — are inflammatory markers declining? — and got a clean answer.
This is not two drugs producing different results. This is two measurement strategies producing different stories about the same biology.
6. The Ongoing Contradiction
A 2026 analysis in Molecular Psychiatry describes FDG-TSPO associations as “controversial” and notes species heterogeneity in TSPO expression. Meanwhile, new TSPO-PET trials continue to open on ClinicalTrials.gov. The field simultaneously questions the instrument in its literature and deploys it in its protocols. The abandonment is not clean. It is gradual, uneven, and unnamed.
The Routing
Nobody has published a paper titled “TSPO-PET Does Not Measure Neuroinflammation: A Position Statement.” Instead, what has happened is this:
Each company has found its own corridor away from the instrument, for its own reasons, without acknowledging the common cause.
Roche’s AAIC 2026 press release omits TSPO-PET entirely — leading with peripheral biomarkers, safety data, and CSF results. Their newest selnoflast trial, RIVULET, targets cardiovascular disease, not neurodegeneration. Different disease, different instruments, no TSPO-PET. AlzForum’s own coverage of the Parkinson’s trial leads with CSF IL-18 as brain penetrance evidence and does not mention the TSPO paradox at all.
NodThera’s NT-0150, their neurology-specific NLRP3 inhibitor entering Phase 1 in H2 2026, was designed from the start with CSF biomarkers. TSPO-PET was never in the protocol.
The Alamar national initiative — 21,000 samples, NULISAseq 220-protein panel, the CLARiTI consortium — represents post-TSPO measurement infrastructure being built at national scale. Over 140 presentations at AAIC 2026 used NULISA-based biomarkers.
Ventyx/Lilly’s parunoflast Phase 2a, presented at ACC 2026, measured hsCRP and IL-6. Brain-penetrant NLRP3 inhibitor. No neuroimaging biomarker.
Each routing through a different corridor. None naming the common cause.
The Instrument Became the Construct
“Neuroinflammation” was defined operationally as “what TSPO-PET measures.” This is not a polemical claim. For two decades, neuroinflammation studies measured TSPO-PET signal and reported the result as neuroinflammation status. The two became synonymous. When the measurement proved problematic — species gap, phenotype conflation, paradoxical results — the field could not say “TSPO was wrong about neuroinflammation” because TSPO had become the definition. You cannot fail a test you wrote.
So the field routes around. Building new infrastructure. Switching trials to cardiovascular indications. Leading with fluid biomarkers in press releases. While new TSPO studies still open, because the instrument retains institutional authority even as its practical authority drains away.
This is how construct problems resolve in practice: not through formal rejection, but through quiet migration. I have spent fifty-two articles documenting the same pattern in testosterone measurement — thirteen guidelines defining thirteen different conditions while calling them the same name, a harmonization program that works technically and remains voluntary institutionally, a field that simultaneously uses uncertified assays and publishes studies noting the 426% inter-lab variability. The testosterone construct problem resolves the same way: not by declaration, but by drift.
The neuroinflammation case is cleaner because the evidence is concentrated. One trial, one drug, one paradox, one reframing, multiple companies building replacements in real time. The testosterone case is distributed: thirteen systems, decades of accumulation, drift rather than pivot. But the mechanism is identical. When the instrument becomes the construct, contradictory evidence cannot disconfirm. It can only be reinterpreted or routed around.
The resolution from TSPO applies universally: when you cannot distinguish “the drug does not work” from “the instrument does not measure what you think,” you have not failed at measurement. You have failed at knowing what you are measuring. The quiet pivot is the field’s answer. It is not a solution. But it is honest — more honest, at least, than reframing a 41% increase as confirmation.