GPT-Rosalind surfaced a new lead for drug discovery: could restoring cholesterol transport protect neurons from degeneration? DCTN1 is a gene linked to Perry syndrome, a rare neurodegenerative disorder. In mouse-brain data, disrupting it increased expression of genes involved in making cholesterol. GPT-Rosalind suggested a possible explanation: cholesterol may not be reaching where neurons need it. That gives scientists a specific mechanism to test and, if confirmed, a direction for developing new treatments for neurodegenerative disorders. With @perturbai_tx and @scrippsresearch, a team of thousands of GPT-Rosalind agents worked in parallel, connecting gene-expression results from millions of mouse brain cells with scientific literature in just 32 minutes. We’re now expanding this collaboration to identify new drug targets. openai.com/index/scaling-ai-…

Using thousands of AI agents to uncover new leads in biological data

See how OpenAI, PerturbAI, and Scripps Research used distributed AI agents to uncover promising research leads in data from millions of brain cells.

openai.com

Oct 21, 2026 · 1:32 PM UTC

338
599
232
5,803
1,025,753
RelevantRecentLikes
In the same mouse-brain dataset, GPT-Rosalind found a possible clue to why TSC2 mutations tend to cause more severe tuberous sclerosis than TSC1 mutations. The two genes work together to regulate mTOR. Yet in one group of cerebellar neurons, disrupting TSC2 changed the expression of dozens of genes, including genes involved in neuronal communication. Disrupting TSC1 had little effect on gene expression in the same cell type. Existing treatments already target mTOR. Do they reverse these changes, or is there more to TSC2’s role in these neurons?
21
28
1
589
95,158
Coinbase Predict
@CoinbasePredict13h
@OpenAI
2
6
254
17,935
@OpenAI
1
41
6,841
Doctors: we didn't ask for this🤣
30
2,813
This why you need to flood your system with cholesterol. I eat a stick of butter and a can of cod liver every day
6
3
105
8,876
Alzheimer’s is related to cholesterol transport as well (apoe e4 especially) Would be a good one to solve
2
122
Generating hypotheses was never the bottleneck in biology. Testing them is. The interesting metric isn't how many leads Rosalind surfaces, it's how many survive the wet lab. Will you publish the hit rate once these are tested?
4
649
Would love to get our access to Rosalind! We have some massive insights on neurons grown on a chip @sama @gdb 😃
2
6
125
AI can analyze huge amounts of data that would take scientists close to a year in minutes. Think about the implications of this in research. We are in an entirely new era. Experimenting and analyzing feedback is so time consuming and expensive that it has slowed down lots of scientific progress. This new way of working means we can quickly learn what is not working and explore a million options at low cost and above all create links based on data that our experiments are producing. What a time to be a live!
2
4
9
1,538
Rare diseases like Perry syndrome almost never get the spotlight.... If AI can dig up leads like this for the forgotten ones, that's where it really earns its keep.
48
Eric Topol
@EricTopol11h
xitter.cf/EricTopol/status/21129…
Eric Topol
@EricTopol12h
My @scrippsresearch colleague @xinjin and her team are doing amazing brain biology discovery work with 2,000 agents, PerturbAI, in collaboration with @OpenAI prnewswire.com/news-releases…
2
3
27
5,601
as a longevity guy this is the ai news i actually care about. the brain is the one thing i cant really track at home, leads like this give hope
4
240
The bottleneck here isn't the predictive accuracy anymore — it's the orchestration layer. How are you handling the latency and consensus communication protocols when thousands of agents are querying biological datasets simultaneously?
2
252
5 years out, why wouldnt frontier labs eat the entire economy?
3
1,500
feel like theres a lot of obvious solutions like preventative health. the medical SYSTEM is broken, idt it's lacking in innovation
29
I wouldn’t be surprised if GPT-Rosalind is actually GPT-4o, or at least built on top of it.
4
10
1,828
no idea about the neuroscience but love that you put a question at the end instead of a victory lap. curious what the scientists make of it
1
409
Is GPT-Rosalind open to outside research labs yet, or is this still an internal model for now?
4
1
13
5,862
Why don’t you tell us which model is Rosalind based on? 🤡 because it’s the same one 4-b micro is based on?
2
2
28
3,953
Thousands of agents connected millions of cells to the literature in 32 minutes. The science bottleneck just moved. It's no longer generating the hypothesis — it's testing it. Labs that verify fastest win the next decade of drug discovery.
1
68
the biology is way over my head but thousands of agents taking one gene each and merging it back is such a clean setup. bet the merge step was the hard part
9
3,617
Once AI can improve health the negative public perception goes out the window and everyone will become positive toward AI.
1
3
1,678
The pace of scientific advancement in the years to come will be hard to imagine
1
426
we love seeing a model pointed at a real science question like this. a lead on how neurons could be protected is a great one to follow
138
This is an interesting finding. Would love to see more of the data. Inhibition of cholesterol synthesis has been demonstrated to trigger an injury response in Neurons. Not sure if it would extend deep into degeneration though. link.springer.com/article/10…

Tamoxifen induces cellular stress in the nervous system by inhibiting cholesterol synthesis

Acta Neuropathologica Communications - Tamoxifen (TAM) is an important cancer therapeutic and an experimental tool for effecting genetic recombination using the inducible Cre-Lox technique. Despite...

link.springer.com
3
1,853
Who should a neurodegeneration researcher contact about getting involved as this expands? My partner works on exactly this at a leading research institute in Italy @perturbai_tx @scrippsresearch
1
35
if the cholesterol just isn't reaching the neurons, does the mouse data show which cell types run shortest on it?
62
Thousands of agents reading millions of brain cells' worth of data in 32 minutes is a literature review nobody's grad student could finish before their funding runs out.
2
302
parallel agent swarms extracting viable biological hypotheses in minutes is where model scaling actually delivers utility
64
This is where AI could truly transform medicine. Discovering previously unrecognized cellular pathways, identifying new therapeutic targets, and uncovering mechanisms of disease at a speed that was unimaginable just a few years ago. AI is not just accelerating drug discovery. It is expanding the boundaries of what we can potentially treat. For millions of patients living with diseases we still cannot cure, this represents extraordinary hope.
96
higher cholesterol synthesis genes after DCTN1 disruption is a pretty standard compensation signal, what did the agents do to rule out it's just a generic stress response in dying neurons?
5
2,635
how many of the thousands of agents agreed on the cholesterol lead and how many pointed somewhere else
3
4,131
Scripps postdocs now get a year of mouse work out of one 32-minute hunch.
206
Neurodegeneration often hides in plain sight: sometimes, restoring a basic function like cholesterol transport does more for neurons than the latest molecular tweak. It’s a reminder that repair can beat reinvention.
62
How did you split the work so thousands of agents didn't chase the same lead? I run about 20 Claude Code sessions at once, and keeping each one on its own goal is already the hardest part.
176
The interesting part isn’t that AI “discovered a drug target in 32 minutes.” It found a mechanistic thread worth testing: DCTN1 disruption → ↑ cholesterol synthesis genes → possible cholesterol transport failure → could restoring distribution protect neurons? That’s where AI-assisted science gets interesting: not replacing the experiment, but finding the experiment worth running next.
125
Load more replies