Watch the companion video — narrated walkthrough of the meta-analysis
I want to say upfront that I am not a neuroscientist. I run a small company in Humboldt County, California, and most of my days are about shipping orders and writing code. But over the last few weeks I spent my nights pulling down four public single-cell RNA datasets and running the same statistical pass across all of them, and the result was clean enough that I think it is worth talking about.
The short version: when you pool 504,571 single-cell measurements of human midbrain tissue from four different research groups, one specific subtype of dopamine neuron — the cells that express the angiotensin II type 1 receptor, AGTR1 — is depleted in every single Parkinson's dataset compared to controls. Combined odds ratio 0.215. P-value less than 10 to the negative one hundredth power. That is not a marginal effect. That is one of the most depleted druggable cell types I have ever seen reported, and the receptor is already targetable by FDA-approved blood-pressure drugs that have been on the market for thirty years.
That is the headline. Here is the longer story of how I got there, what I think it means, and — more importantly — what I do not think it means, because the easiest way to embarrass yourself in a field you do not belong to is to overclaim.
Where the idea actually came from
I did not discover this. Tushar Kamath and his colleagues at the Broad Institute published the original observation in Nature Neuroscience in 2022. They ran single-nucleus RNA sequencing on postmortem human midbrain tissue from Parkinson's patients and controls, and they noticed that one subtype of dopamine neuron — defined by the markers SOX6 and AGTR1 — was preferentially lost in disease. It was a careful paper, peer-reviewed, well-cited. It did the science. What it did not do, and what no single study can ever do, is rule out the possibility that the effect they saw was an artifact of their particular cohort, their particular dissection protocol, or their particular sequencing chemistry.
That is what meta-analysis is for. You take the same biological question and you ask it of as many independent datasets as you can find. If the effect is real, it shows up across cohorts. If it is artifact, it washes out.
I am a hobbyist in this field, but I am not bad at this part of it. I have spent the last year teaching myself single-cell analysis on the side. I knew that since Kamath 2022, three more independent groups had released public single-cell Parkinson's datasets — Smajić et al. at DZNE in Germany, Wang et al. at Mount Sinai, and Martirosyan et al., an independent cohort released in 2024. None of them, as far as I could find, had run the AGTR1 question across each other's data. So I did.
What I actually ran
The pipeline is unsexy. Python, scanpy, statsmodels. Pull each dataset from GEO. Normalize. Run the same cell-typing logic on each one so the labels are consistent across studies. Count the fraction of dopamine neurons in each donor that fall into the AGTR1-positive subtype. Compare patients to controls. Pool the four resulting odds ratios using fixed-effect meta-analysis with effect-size weighting.
That is it. There is no fancy machine learning, no transformer, no proprietary model. The whole repository is a few thousand lines of straightforward bioinformatics that anyone with a laptop can rerun. I made sure of that on purpose. The code is on GitHub, the data is public, the figures are reproducible end to end with a single shell script.
The four-cohort breakdown looks like this:
| Dataset | Source | Year | Cells | Control AGTR1+ fraction | PD AGTR1+ fraction | Odds ratio |
|---|---|---|---|---|---|---|
| GSE184950 | Mount Sinai | 2022 | 12,778 | 3.31% | 1.00% | 0.295 |
| GSE178265 | Broad Institute | 2022 | 366,874 | 3.30% | 0.60% | 0.177 |
| GSE157783 | DZNE Germany | 2022 | 41,435 | 3.30% | 1.00% | 0.296 |
| GSE243639 | Independent | 2024 | 83,484 | 2.96% | 0.89% | 0.295 |
Every single one points in the same direction. Every single one. The combined odds ratio of 0.215 means roughly a 78 percent reduction in this cell type in Parkinson's brains across half a million cells from four independent cohorts run by four independent groups using slightly different protocols. That is the kind of consistency you almost never see in a noisy field like single-cell genomics, and it is the kind of consistency that — if I were a real neuroscientist running a real lab — would make me drop my other projects and chase this.
Why this could actually matter
AGTR1 is the receptor that angiotensin II binds to. It is the same receptor that gets blocked by the entire class of drugs called angiotensin receptor blockers — ARBs. Losartan. Candesartan. Telmisartan. These are some of the most widely prescribed blood-pressure drugs on Earth. They have been FDA-approved for decades. Their safety profile is extremely well characterized. Several of them cross the blood-brain barrier in measurable amounts.
There is already a small body of epidemiological literature suggesting that long-term ARB users have a lower risk of developing Parkinson's. There is animal-model work showing that ARBs are neuroprotective in MPTP and 6-OHDA mouse models of Parkinson's. There is a 2025 iPSC paper showing that pharmacological inhibition of AGTR1 is pro-survival in human dopamine neurons in a dish. None of this is brand new. What I think is new is the cleanly meta-analyzed cross-cohort confirmation that the cells that get lost in Parkinson's are exactly the cells that express the receptor those drugs hit.
If a clinical trial were ever run — and I am not in any position to run one — the hypothesis would be: in early-stage Parkinson's patients, does adding a brain-penetrant ARB to standard care slow the rate of dopaminergic decline? It is the kind of trial that, on paper, costs maybe a few million dollars instead of the half-billion that a brand-new drug would cost, because the drugs already exist and are already off-patent.
I am genuinely hopeful that someone with the credentials to do this picks it up.
What I want to be very careful not to claim
I am going to say this part plainly because I have watched too many people in adjacent fields ruin their credibility by overclaiming. So, what this analysis does not do:
It does not prove ARBs treat Parkinson's. A retrospective bioinformatics finding is a hypothesis-generation step. The wet lab is where biology actually gets tested. Until somebody runs a real trial in real patients, this is a target that looks promising on paper, nothing more.
It does not establish causality. AGTR1-positive neurons being depleted in Parkinson's brains tells us they are vulnerable. It does not tell us why. It could be that AGTR1 signaling itself drives their death, in which case blocking it should help. It could also be that something else is killing them and AGTR1 is just a marker of a particular cell type that happens to be vulnerable for some other reason entirely, in which case blocking AGTR1 might do nothing. Wet lab work is the only way to distinguish those two possibilities.
It does not mean I am right and the field has missed something. The field has not missed this. Kamath saw it in 2022. Several groups have followed up. What might be missing — and where I think a citizen-analysis approach actually adds value — is the rigorous cross-cohort confirmation in a single place that anyone can reproduce. That is the thing I can contribute as a hobbyist. The hard biology is somebody else's job.
Why I am publishing this
A few reasons.
Mainly, I think this is what open science should look like. The datasets are public. The code is public. The methods are standard. The result is checkable in an afternoon by anyone who downloads the repository and runs the shell script. If I made a mistake, I want someone to find it and tell me. If I did not make a mistake, I want the people who can move this forward — neuroscientists, immunologists, drug-discovery teams, clinical trialists — to know it is there.
Also, honestly, because the more I read about Parkinson's the more I understand that it is not a far-off problem. About a million people in the United States are living with it right now. The diagnosis rate is rising. The treatments are decent for the early symptoms and bad for the long-term ones. The next breakthrough is not going to come from one person. It is going to come from a lot of small, well-aimed contributions stacking up. I would like this to be one of them.
If you are a neuroscientist, a clinician, an immunologist, a drug-development researcher, or a Parkinson's advocate and any of this lines up with something you are working on — I would genuinely love to hear from you. The repository has a contact section.
The repository: https://github.com/nicedreamzapp/parkinsons-vulnerability-predictor
Companion video walking through the full meta-analysis in about a minute: https://youtu.be/bC4hgeHS9cg
I will keep working on this. The next pass is going to look at whether the AGTR1 effect tracks with disease severity within each cohort, and whether the same signal shows up in Lewy Body Dementia, which shares biology with Parkinson's. If anything interesting comes out, it will go up on the same repository, in the open, with the code attached.
Thanks for reading.
— Matt
Companion video on YouTube: https://youtu.be/bC4hgeHS9cg. Full reproducible repo: https://github.com/nicedreamzapp/parkinsons-vulnerability-predictor. I run Nice Dreamz LLC and consult on private/local AI for compliance-sensitive firms via AirGap AI.
Top comments (0)