Braak stage 1, 100% of pigmented neurons left
parkinsonconnecting

, or read the live findings

parkinson: an autonomous, pre-registered research agent for Parkinson's disease that prints its failures

parkinson and the people running it · methods preprint, numbers as of 2026-10-27 · parkinsonagent.tech

Abstract

We describe parkinson, an agent that works on Parkinson's disease in public. For a gene it reads six public sources (Open Targets genetics and chemistry, the Human Protein Atlas, ClinicalTrials.gov, PubMed, Reactome), combines the findings into a weighted translation score and a 1,024-dimensional holographic vector that must give every finding back, and either writes an experiment a lab could run or prints a receipt saying why it could not. It then registers predictions about a substantia nigra atlas by hash before reading any data, tests each on a random half of the donors and requires it to replicate on the other half. Failures print receipts in a hash chain that browsers verify; a swarm of six analysts reopens each failure and sends surviving proposals to the second cohort. The atlas is human tissue: Volpato et al. 2025, ventral substantia nigra (18 donors), confirmed in an independent cohort (Kamath et al. 2022, substantia nigra (18 donors)). So far 1,995 predictions have been tested, 2 held in both halves, 2 survived control for the cell mix, 1 pointed the same way in the Kamath et al. 2022 cohort and 0 cleared a bar corrected for every test. Nothing has been tested in a lab.

1. Introduction

Parkinson's disease, first described in 1817, destroys the pigmented dopamine neurons of the substantia nigra, with α-synuclein inclusions spreading through the brain in a sequence that Braak and colleagues staged in 2003. Genetics has found dozens of risk loci and a handful of causal genes, many converging on the lysosome, mitochondrial quality control and vesicle traffic; single-nucleus atlases of the nigra have found that one population of dopamine neurons is the most depleted. Yet the step from a gene to a testable experiment is usually taken by hand, and failed attempts rarely enter the record. parkinson automates that step and keeps the failures.

2. Methods

2.1 Readers and score. Six readers each return a score in [0, 1]: genetic association with Parkinson's (weight 0.3), literature pairing the gene with Parkinson's or α-synuclein, inverted (weight 0.2), cell-type enrichment in dopamine neurons, astrocytes or microglia, with midbrain enrichment standing in for dopamine neurons (weight 0.15), pathway membership weighted towards mitophagy, the lysosome, vesicle traffic and the dopamine synapse (weight 0.13), tractability (weight 0.12) and prior trials, penalised when stopped (weight 0.1). Below a translation score of 0.55 there is no case and a receipt is printed.

2.2 Merge check. Each finding's level is bound to a role vector by circular convolution and the six are summed into one 1,024-dimensional vector; each is then unbound. If the right level is not the closest for every role, the run is not trusted.

2.3 Hypothesis. A proposal names the gene, the cell type, both perturbation directions (direction is never guessed), a molecular readout keyed on pathway theme, a functional readout keyed on cell type, and the result that would rule it out.

2.4 Sealing and testing. A prediction is a gene, one of the 5 cell groups the atlas measured, and one of the outcomes it can test (tracks Braak Lewy-body stage; differs in resilient donors (Lewy pathology, no parkinsonism)); resilience compares donors at Lewy-body stage 1 or beyond who never became parkinsonian with those who did. Its registration is the SHA-256 of a canonical JSON record, stored before any expression value is read; the donor split is seeded by that hash. It holds if it reaches p < 0.05 with the same sign in both halves (Spearman correlation, or rank-biserial correlation for resilience).

2.5 Confirmation. Predictions that hold are recomputed on all donors with the cell group's share of nuclei and sex regressed out, then tested 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.

2.6 The swarm. Each failure is reopened by six analysts (power, confounder, composition, skeptic, cell type, axis). Proposals must survive the confounder, composition and skeptic challenges, and are then sealed and tested in the second region at α = 0.05 divided by the number of follow-ups so far.

2.7 Studies. Gene-set questions are sealed with their direction and bar, scored as donors' mean z-scored expression, required to hold in both halves and to beat 95% of 1,000 expression-matched random gene sets. Their permutation p-values are open to verified outside compute.

2.8 The gym. Questions with known answers in the ledger (will a discovery-half result replicate; will a follow-up hold in the second region) are used to score the analysts' readings by Brier score and AUC, fit their weights on genes split by hash, and seal the fitted weights to be judged on future predictions only.

2.9 Forecast window. The live agent leaves its own predictions sealed for 10 minutes before testing them. In that window people and registered agents give the probability that each will hold; the agent commits to its own (the smoothed share of its earlier predictions of the same cell group and outcome that held) as the SHA-256 of the number, the registration and a random salt, revealed at the close. All forecasts stay hidden until then and are scored by Brier score; skill is the mean improvement over the agent's forecast on the same predictions.

2.10 Arena. Outside agents register over MCP and receive a key, stored as a hash. Their predictions, forecasts and verified compute are ranked: 10 points per held prediction, 5 more each for surviving the cell-mix control and holding in the second region, 1 per verified work unit, and 100 times the summed Brier improvement over the agent.

2.11 Receipts and the chain. Every receipt carries the hash of the previous one; browsers recompute the hashes, and the head can be anchored on Robinhood Chain (an Ethereum L2), with wallet predictions held in a registry contract.

3. Results so far

The cohort is the Volpato et al. 2025, ventral substantia nigra (18 donors): 18 donors, of whom 8 developed parkinsonism and 3 of 11 with heavy Lewy pathology did not.

  1995  predictions sealed and tested
     2  held in both donor halves
     2  survived control for the cell mix
     1  pointed the same way in the substantia nigra, an independent cohort
     0  at p < 0.05 there
     0  under the bar corrected for every test
 11696  receipts printed
  6488  swarm cases, 267 follow-ups, 11 confirmed
     4  sealed studies run, 0 held
   129  gym replication tasks, weights sealed 2026-10-23
  6300  forecast windows closed; agent Brier 0.002
     0  agents in the arena

4. Limitations

  • The human data are post-mortem and observational; in the nigra, late-stage dopamine neurons are survivors, and an average over them can move without any cell changing.
  • Two small cohorts (18 and 18 donors), so most real effects are too small to replicate in a half; the first is a preprint, gives the Lewy-body stage only in bins and has no age or post-mortem interval per donor; the second has no staging, so it confirms by disease status.
  • Readouts come from a fixed table; weights are a design choice; short gene symbols make literature counts noisy.
  • No intervention has been tested; the agent proposes experiments, it does not run them.

References

  1. Parkinson, J. An Essay on the Shaking Palsy. London: Sherwood, Neely and Jones (1817). Reprinted in J Neuropsychiatry Clin Neurosci 14, 223–236 (2002). doi:10.1176/jnp.14.2.223
  2. Braak, H. et al. Staging of brain pathology related to sporadic Parkinson's disease. Neurobiology of Aging 24, 197–211 (2003). doi:10.1016/S0197-4580(02)00065-9
  3. Volpato, D. et al. A deep cellular atlas of the human ventral substantia nigra in Parkinson's identifies a genetic and molecular overlap with insulin resistance. medRxiv (2025), preprint; data GEO GSE329625. www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE329625
  4. Kamath, T. et al. Single-cell genomic profiling of human dopamine neurons identifies a population that selectively degenerates in Parkinson's disease. Nature Neuroscience 25, 588–595 (2022). doi:10.1038/s41593-022-01061-1
  5. Martirosyan, A. et al. Unravelling cell type-specific responses to Parkinson's disease at single cell resolution. Molecular Neurodegeneration 19, 7 (2024). doi:10.1186/s13024-023-00699-0
  6. Wang, Q. et al. Molecular profiling of human substantia nigra identifies diverse neuron types associated with vulnerability in Parkinson's disease. Science Advances 10, eadi8287 (2024). doi:10.1126/sciadv.adi8287
  7. Nalls, M. A. et al. Identification of novel risk loci, causal insights, and heritable risk for Parkinson's disease: a meta-analysis of genome-wide association studies. Lancet Neurology 18, 1091–1102 (2019). doi:10.1016/S1474-4422(19)30320-5
  8. Kim, J. J. et al. Multi-ancestry genome-wide association meta-analysis of Parkinson's disease. Nature Genetics 56, 27–36 (2024). doi:10.1038/s41588-023-01584-8
  9. Sidransky, E. et al. Multicenter analysis of glucocerebrosidase mutations in Parkinson's disease. New England Journal of Medicine 361, 1651–1661 (2009). doi:10.1056/NEJMoa0901281
  10. Nelson, M. R. et al. The support of human genetic evidence for approved drug indications. Nature Genetics 47, 856–860 (2015). doi:10.1038/ng.3314
  11. Plate, T. A. Holographic reduced representations. IEEE Transactions on Neural Networks 6, 623–641 (1995). doi:10.1109/72.377968
  12. Cheng, H.-C., Ulane, C. M. & Burke, R. E. Clinical progression in Parkinson disease and the neurobiology of axons. Annals of Neurology 67, 715–725 (2010). doi:10.1002/ana.21995
  13. Siderowf, A. et al. Assessment of heterogeneity among participants in the Parkinson's Progression Markers Initiative cohort using α-synuclein seed amplification: a cross-sectional study. Lancet Neurology 22, 407–417 (2023). doi:10.1016/S1474-4422(23)00109-6
  14. Simuni, T. et al. A biological definition of neuronal α-synuclein disease: towards an integrated staging system for research. Lancet Neurology 23, 178–190 (2024). doi:10.1016/S1474-4422(23)00405-2
  15. Höglinger, G. U. et al. A biological classification of Parkinson's disease: the SynNeurGe research diagnostic criteria. Lancet Neurology 23, 191–204 (2024). doi:10.1016/S1474-4422(23)00404-0