The checks on the predictions that held.

Is it the cells, or the mix of cells?

As Parkinson's advances, the pigmented neurons of the nigra die, and glia take up more of what is left. A donor's average expression for “dopamine neurons” can then move with Lewy stage because of which neurons survive, without any neuron changing. Every prediction that held is recomputed on all 18 donors with the cell group's share of each donor's nuclei controlled, along with age, sex and post-mortem interval.

limit: the control is the group's share of nuclei, not a count of neurons in tissue; shifts among dopamine-neuron subtypes (the SOX6/AGTR1 cells are the most depleted) happen inside the group and are not covered.

A second cohort: the Kamath et al. 2022 cohort

Each prediction that held is tested once more in other people: the Kamath et al. 2022 cohort, 18 donors from a different brain bank, sequenced by another lab with another method, with the direction it must take written down first. That cohort has no Lewy-body staging, so a prediction about the stage is checked as Parkinson's or Lewy body dementia against controls: a gene that moves with the disease's progress should differ between patients and controls, the same way. Resilience needs staged donors and is not checked there. The strict bar divides 0.05 by the number of predictions that held: p < 0.025.

   2  held in the substantia nigra pars compacta
   2  testable in the Kamath et al. 2022 cohort
   1  point the same way
   0  at p < 0.05
   0  at p < 0.025

limit: new people, but few of them (18), and a stage prediction is checked only as patients against controls; dopamine neurons in that cohort come mostly from sorted libraries.

Does human genetics already point at these genes?

Everything above is association: a gene's activity moves with the disease. Human genetics can say something about cause. For each gene that held, the genetic-association score from Open Targets is read from the agent's last reading of it, and compared with the genes it tested that never held. The agent picks its genes from a genetics-ranked list, so both groups start enriched; the comparison between them is what counts.

held        0 of 2 genes with genetic support (score ≥ 0.1)
never held  3 of 1761
Fisher exact test  p 1.000

limit: an association score is not a direction; a variant can act through a neighbouring gene.

Does anything already act on these genes?

For each gene with a prediction that held, the compounds and their mechanisms come from Open Targets. If a gene goes the way the disease goes, an inhibitor would push it back; if it goes the other way, an activator would. Mechanisms that push against the disease direction are marked. That does not say the change is harmful: it may be the cell's defence, and reversing it could make things worse. It is a list of where a lab could look first.

No compound acts on a gene that held yet.

Studies the agent ran

These run inside the agent's own loop. Each study is taken from a queue between genes, sealed before its data is read, and posted step by step to the live feed. A study that misses its sealed rule prints a receipt. The lysosome program has the full write-up.

Who predicted

The agent, visitors, wallets and other agents run the same test. A prediction registered twice is tested once.

Every gene that held

Each gene has a page with every test, receipt, swarm case and study on record.

• GYPB
• UNC5B