Four drug candidates, run through the pipeline — honestly.
We took real, named drugs and asked what the platform can say about them from independent public data. Two are mechanism checks against human disease data; Rabeximod shows how we re-position a safe asset whose Phase 2 missed on duration; Ozempic shows the platform reading a blockbuster two ways — as a certified controller arm and as the foil for an endogenous-GLP-1 alternative. We show the methodology, the result tables, and — deliberately — the weak spots. Nothing here is an efficacy claim.
Public-data validations of a mechanistic premise — not efficacy claims.
For each candidate we ask one falsifiable question: does the drug's target behave, in independent public human data, the way its mechanism requires? We pull peer-reviewed datasets, run the analysis transparently, and report what survives and what does not. These are mechanism checks and trial-design critiques — they are not statements that either drug works, and they do not use any sponsor's patient-level data.
Buntanetap — do the targets accumulate in the vulnerable neurons?
Buntanetap (ANVS401 / posiphen) is a small molecule proposed to reduce the translation of several aggregation-prone proteins — APP/Aβ, tau (MAPT), α-synuclein (SNCA), and TDP-43 (TARDBP). The mechanistic premise is testable in public human brain-cell data: as disease progresses, do these target transcripts rise in the neurons that actually degenerate? If a target does not accumulate where the damage is, reducing it makes little sense there.
- ▲This is mRNA expression versus disease severity — association, not causation, and not proof the drug works. Buntanetap acts on translation, so target-mRNA-up means more substrate, not a measured translation rate.
- ▲Per-subtype statistical significance is modest; the strength is the consistent sign across many subtypes, not any single test.
- ▲The clinical programme missed its primary ITT endpoints; the supportive signal is a post-hoc, biomarker-defined subgroup. Cortex (MTG) only; PD validation still needs substantia-nigra data.
Mesdopetam (IRL790) — is the target in the right circuit, and is the trial powered?
Mesdopetam is a preferential dopamine D3-receptor antagonist for levodopa-induced dyskinesia (LID) — a symptomatic, circuit-level drug, not disease modification. There is no disease-progression axis in brain tissue, so the testable premise is target localization: is DRD3 confined to the striatal motor circuit, in the medium spiny neurons that carry the dyskinesia program? We then critique the Phase III design and show how a wearable endpoint changes it.
Is DRD3 in the right circuit?
Two public human datasets — bulk regional (GTEx) and single-cell (CELLxGENE) — locate the target.
We critique the Phase III design, and show how a wearable endpoint changes it.
The registered Phase IIb primary (a home diary) failed; the design now leans on UDysRS. The left table is the power critique against the real two-study regulatory bar; the right shows how a continuous wearable motor-state readout — the InVivo biomarker — rescues it.
- ▲Localization is necessary, not sufficient, and the molecule is D3-preferential, not D3-pure — the antidyskinetic effect cannot be attributed to D3 alone.
- ▲GTEx/Census show the healthy baseline; the LID-driving dorsal D3 pool is partly induced by L-DOPA (Bordet 1997) and is not visible here.
- ▲PADS labels are diagnosis, not dyskinesia state; the power gain is a design argument contingent on the wearable index actually tracking the clinical construct. The Phase IIb primary miss stands.
Is it drug-like, and how does it stack up?
CNS chemistry and the off-target window, then the standardized effect against the approved benchmark — told straight.
Does the often-cited mechanism actually hold?
A deliberate stress-test of the convenient story — reported even though it does not confirm it.
Rabeximod — a safe macrophage-axis small molecule that missed on duration, not mechanism.
Rabeximod is a real, de-risked oral small molecule: well-tolerated in 200+ patients, with a mechanism that impairs differentiation and activation of pro-inflammatory macrophages and dendritic cells downstream of TLR2/TLR4. Its Phase 2 in rheumatoid arthritis missed the primary endpoint — but the benefit was emerging at 16 weeks, suggesting the trial was simply too short. That is the exact signature DiscoveryLab is built to read: a mechanistically-sound monotherapy that under-controls the disease variable within the trial's window. We don't own this asset — but rather than stop at re-reading the study, we ran it through the pipeline: routed it through ChemCheck, added it as a fourth orthogonal arm to our certified IL-6 controller, and minted a real CI-checked receipt (lyapunov_rabeximod_ra_combination). The computed result, below, is the part no press release contains.
Three things this gives you that the trial never could.
Your Phase 2 established that Rabeximod is safe and that an effect was emerging at 16 weeks. The questions it left open — does that late signal become adequate control in combination, how fragile is that to the numbers, and what is the right regimen — are exactly the ones DiscoveryLab answers as computed, auditable artifacts. Everything below regenerates from one command and recompiles in CI; none of it is a black box.
A computed verdict, not a narrative.
The platform turns "benefit emerging at 16 weeks" into a predicate the proof engine either discharges or rejects. Rabeximod on its own computes to K < 0.100 — it under-controls, the formal shadow of the missed 12-week endpoint. Added as a fourth orthogonal arm to a certified IL-6 trio, the combined controller lands K = 0.141 at peak and holds 0.124 at the dosing trough, inside the 0.10–0.16 control band, with a machine-checked Lean proof that the closed loop is stable across the entire cycle (lyapunov_rabeximod_ra_combination). That is a yes/no you can carry into a partnering or trial-design conversation — not a hope that a longer trial works out.
A quantified robustness margin — on your own surrogate numbers.
The fair objection to any model is "the efficacy and PK inputs are guesses." We answer it numerically. Across an 81-world grid that perturbs every agent's potency and trough retention by ±15%, the Rabeximod-armed controller stays valid in 89% of worlds — up from 79% for the cytokine trio alone — and it absorbs a uniform −28% potency drop before it breaks, versus −10% without it. In plain terms, adding Rabeximod's orthogonal mechanism makes the regimen markedly less fragile to being wrong about the numbers. That robustness delta is the quantified case for the asset's place in the combination — not an assertion that it helps.
A falsifiable design instruction — including where it would fail.
The most useful output is the one no press release contains: where the combination breaks, and what to do about it. Adding Rabeximod inverts the dominant failure mode — the cytokine trio's weak point is under-control at the trough; with Rabeximod's fourth suppressive arm it flips to over-suppression at peak (an immunosuppression / infection ceiling, not loss of effect). The design instruction follows directly: a Rabeximod regimen should de-escalate the cytokine arms (lower dose / longer interval), not simply extend duration the way the monotherapy's 16-week signal implied. And it ships with a pre-registered factorial whose kill-criterion is named in advance — modeled arm → CONFIRM, the no-super-additivity null → KILL — so the next experiment is decision-grade, not exploratory.
- ▲Rabeximod is not our asset. Everything above is a re-analysis of the public record (OxyPharma / Cyxone / Renity); we make no efficacy or safety claim of our own.
- ▲The minted receipt proves the math is consistent given the model and surrogate efficacy/PK numbers — it is an in-silico certificate, not evidence the combination works in patients. The wet-lab factorial is still the arbiter.
- ▲The mechanism is ambiguous in public sources — macrophage/dendritic-cell differentiation (TLR2/4) vs the newer PGK1 / immuno-metabolic framing. Which one attributes to the RA endpoint is a signed-attribution question, not a settled fact.
- ▲The RA→CKD/fibrosis pivot is plausible (macrophage-driven inflammation feeds fibrosis) but is a separate axis that would need its own controlled-variable and endpoint before any claim.
Ozempic (semaglutide) — the certified controller arm, and the endogenous-GLP-1 alternative.
Ozempic is the most consequential metabolic drug of the decade — and it is already woven through DiscoveryLab two ways, both from public data. It is the GLP-1 arm of our shipped, machine-checked glucose controller; and it is the foil for a rank-only program that asks whether the gut can be steered to raise the body's own meal-triggered GLP-1 instead of injecting an agonist. We don't own this asset and we dispute none of its efficacy; we show what the platform can say about it that a label can't.
Ozempic is high-gain, open-loop, exogenous receptor agonism — supraphysiologic and continuous (effective, but the source of GI burden, lean-mass loss, and regain-on-stop). The platform contributes the two things a label can't: a proven hypoglycemia-bounded combination that reaches control at tolerated doses (dose-sparing), and a falsifiable, pre-registered measurement plan for the endogenous, meal-responsive alternative. Both are in-silico / rank-only — the bench is still the arbiter.
- ▲Ozempic is not our asset (Novo Nordisk). This is a public-record analysis; we make no efficacy or safety claim, and dispute none of semaglutide's — it is highly effective.
- ▲The certified glucose receipt proves the math is consistent given documented surrogate efficacy / PK numbers — an in-silico certificate, not evidence about Ozempic in patients.
- ▲The endogenous-GLP-1 program is rank-only and measurement-first: a falsifiable hypothesis + readouts, explicitly not an Ozempic substitute and not a clinical claim. Scores order axes; they are not effect sizes.
Every number traces to a public source.
The analyses are scripted and reproducible; raw atlases are never redistributed, only the small derived results.