case 01
The denominator: 1,996 predictions, 2 held, 0 confirmed
1996 predictions registered and hashed before the data was read
2 held in both donor halves of the substantia nigra pars compacta
2 still hold with the cell mix controlled
1 point the same way in the Kamath et al. 2022 cohort (chance: about 1)
0 nominally significant there (p < 0.05)
0 pass the strict bar there (p < 0.05 / 2)What was tested. Each gene the agent picked up got predictions it sealed before reading the atlas: tracks Braak Lewy-body stage, in its cell group and in the dopamine neurons. A prediction counts only if it reaches p < 0.05 on a random half of the 18 donors, then again, the same way, on the other half.
What happened. 2 of 1996 held: 0.1%. If nothing were real, about 2.5 would pass by luck. The ones that held were then retested in an independent cohort, the Kamath et al. 2022 cohort (18 donors), where outcomes it was not staged for are checked as Parkinson's or Lewy body dementia against controls: 1 point the same way, 0 at p < 0.05, and 0 clear a bar corrected for testing 2 things at once.
What it means. Consistency between the two says the direction is not noise. It does not say the change is protective or harmful, and in the nigra it can be survivorship: the dopamine neurons left in late disease are not a random sample of the ones that were there. Case 03 takes that up.
limit: two cohorts, small ones; no covariate adjustment in the registered test; pseudobulk, not single cells
case 02
UNC5B: the strongest confirmation so far
UNC5B · oligodendrocytes · tracks Braak Lewy-body stage substantia nigra pars compacta -0.69 (p 0.041) → -0.73 (p 0.024) held cell mix controlled -0.72 (p 7.4e-4) holds Kamath et al. 2022 cohort -0.47 (p 0.091) Parkinson's or LBD vs controls bar 0.025
What happened. UNC5B in oligodendrocytes tracks Braak Lewy-body stage: -0.69 (p 0.041) in the discovery half and -0.73 (p 0.024) in the held-out half. In the substantia nigra, an independent cohort it points the same way at p 0.091, above the corrected bar of 0.025.
What it would take. It is still a correlation in post-mortem tissue. Whether UNC5B drives the change or answers it is what a perturbation would test: knock it down and overexpress it in iPSC dopamine neurons seeded with α-synuclein fibrils, and measure phospho-α-synuclein and neuron survival.
limit: one region confirmed, in largely the same donors · oligodendrocytes pseudobulk
case 03
PMPCA: a first-half pass that reversed
PMPCA · dopamine neurons · tracks Braak Lewy-body stage
discovery -0.94 (p 2.1e-4) held-out +0.13 (p 0.733) receipt 9234
swarm power full sample -0.50 (p 0.034, n 18), 95% CI -0.78 to -0.04. the interval excludes zero: something may be there, smaller than a half can see.
composition dopamine neurons share of nuclei vs axis +0.05 (p 0.854); controlling for it: -0.52 (p 0.028, n 18). not the mix.
verdict case closed: nothing better to follow up.What happened. PMPCA passed the first half of the donors at p 2.1e-4, then pointed the other way on the second. The receipt stands.
Why it is here. This is the failure the receipts exist to catch. A single split of donors is a noisy judge of moderate effects, and a selection on the first half inflates what it finds. The swarm reopened the case from six angles instead of letting it disappear.
limit: one random split per prediction, fixed by its hash
case 04
The swarm turns receipts into the next test
6792 swarm cases opened on failed predictions 280 follow-ups registered for tissue the agent had not seen 181 of those point the same way in the Kamath et al. 2022 cohort 11 pass the strict bar there DNAJC13 oligodendrocytes braak +0.50 p 1.9e-6 confirmed CHRM3 glutamate neurons braak +0.50 p 2.0e-6 confirmed STK39 GABA neurons braak -0.50 p 2.8e-6 confirmed NDUFV2 oligodendrocytes braak -0.48 p 6.3e-6 confirmed DDC glutamate neurons braak -0.44 p 5.0e-5 confirmed PLA2G6 microglia braak +0.42 p 1.2e-4 confirmed
What happened. Every failed prediction got a case. The analysts checked whether the effect was underpowered, confounded by sex, explained by the cell mix, or carried by one donor, and proposed better-aimed predictions. Proposals that survived the challenges were sealed and tested in the second cohort, at a bar corrected for every follow-up so far.
Why it matters. A negative result becomes the next experiment instead of a dead end. Because the swarm works on the donors whose test failed, it cannot confirm anything there; everything it proposes goes to tissue it has not read.
limit: follow-ups are chosen on the full first sample, then tested on an independent cohort that has no Lewy staging
case 05
The lysosomal axis in dopamine neurons: the sealed rule did not hold
12 genes · dopamine neurons · tracks Braak Lewy-body stage · sealed bc4d53b17eb8 discovery half +0.40 (p 0.288) held-out half +0.30 (p 0.426) all donors +0.28 (p 0.254) beats 96 of 1000 matched random sets
The question. Does a score built from the lysosomal Parkinson's genes fall as Lewy pathology advances, inside dopamine neurons?
What happened. no effect in the discovery half (+0.40, p 0.288).
limit: a gene-set score averages genes that may not move together · the random sets match expression level only
what it adds up to
Where this leaves the work
The goal stays the same: a short list of experiments a lab can run, each with the result that would kill it, and an open record of everything that didn't make the list. The next steps follow from the cases: test within dopamine-neuron subtypes rather than the whole population, adjust the registered test for age, sex and post-mortem interval, and find independent donors, not just independent tissue.
all data public: /findings, /api/experiments, the receipts, the paper