Human brain organoids transplanted into mice grow four million human neurons
Mice genetically stripped of nearly their entire cortex, then given human brain organoids, grew roughly 4 million human neurons; their behaviour resembles neither normal mice nor cortex-less mice, but a third state.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
No surgery — genetics makes the room
In 2022 Sergio Pasca's team implanted human brain organoids into newborn rat brains, but human brain cells grow far more slowly than rodent ones, and the implanted organoids were "outcompeted", getting neither connections nor room to grow. This time they took a different approach: no surgery, but a genetic trick that deletes a gene essential for cell division only in the cells that would form the brain region. The result is that cells cannot divide the moment they are born ready to build cortex, and stall in place. The structure removed is called the pallium; it exists early and later develops into the cerebral cortex and hippocampus. These "pallium-less mice" lost nearly half their brain volume but survived, with reduced body weight and largely normal basic motor and sensory responses.
— Sergio PascaFour million human neurons take over the empty space
The team injected human stem-cell-derived cortical organoids into pallium-less pups aged 5 to 17 days, about four organoids per injection, each roughly 100,000 cells, already differentiated and on their way to forming cortex. The human cells then divided, differentiated, were vascularised by the host, and mouse immune cells entered the interior of the human graft. Months later the team estimated that these animals, called xenocortical mice, had about 70 million neurons in total, of which roughly 14 million were cortical and hippocampal cells, most of what was removed having been replaced by about 4 million human neurons. Human cells are larger and less dense, so the replacement is not an equal number, but it is still a very large number of human neurons.
— Sergio PascaThey resemble neither normal mice nor cortex-less mice
The team used machine learning to analyse behavioural differences among three kinds of mice in video: controls, pallium-less mice that had lost almost the entire cortex, and xenocortical mice carrying the organoids. Pallium-less mice differed clearly from controls, while xenocortical mice were "neither", sitting in a third behavioural state. Later experiments backed this up: pallium-less animals were impaired on a working-memory task, xenocortical animals were not. Sergio's judgement is that function was preserved in the animals that received human cortical cells, and that human cells very likely contributed a part of it — but this is not perfect restoration, only a demonstration that these cells can integrate functionally and produce behavioural consequences.
— Sergio PascaA hypoxia experiment shows it is the human cells doing the work
The team used this system to model brain injury. Hypoxia is severe in humans and can cause lifelong disease, whereas mice tolerate low oxygen far better. After the three kinds of mice were exposed to hypoxic conditions for several hours, the human cells were induced to mount a hypoxia response while the mouse brain was relatively resilient. Two days later, on a motor task, wild-type mice and pallium-less mice were fine, but the xenocortical mice carrying human cells began to show defects in walking and gait coordination. Sergio argues this shows it is very likely the human cells causing these defects, and that this would be a powerful system not only for asking about mechanism but for testing therapies — seeing whether there is benefit before going into clinical trials.
— Sergio PascaA very rare neuron type was grown
Aparna Bhaduri at the University of California studies human brain development and cancer, and she considers one highlight of the work to be that the human organoids differentiated a cell type that looks like a very rare kind of neuron that has been hard to culture in vitro. Loss of this neuron type is a key factor in a particular dementia, so understanding them better could help future therapy development. But she also notes there is a long way to go: the laminar structure of the cortex and its different processing areas are not well preserved in this system, so it does not reach the level of being more like a human brain; and the authors themselves state that the developmental stage the organoids reach is still mid-development — this is not a fully mature cortex.
— Aparna BhaduriThis is not a human brain in a mouse's head
Asked whether these animals count as human-mouse chimeras, Aparna Bhaduri says plainly: this is not a human brain inside a mouse. There are three reasons — it is immature; the complexity of cell types is insufficient, some cell types are still missing and the ones present are not mature; and although there are some connections with the rodent, the degree and the way they are wired differ from the human brain. So on complexity, maturity and connectivity, she considers this an organoid pushed forward in its development, not a human brain. Other researchers who spoke to Nature take the same view: these are not fully formed functional human cortex.
— Aparna BhaduriEthicists are not a stumbling block for scientists
Emily Jackson at the London School of Economics studies the regulation of stem-cell-related research, including brain organoids. She says the most striking thing about this paper, and about the work of Sergio Pasca's lab, is the team's focus on the importance of ethics and regulation. She thinks people sometimes mistakenly treat regulation as an obstacle to scientists, and that this is a genuine mistake: in her experience scientists at the frontier of biotechnology very much want ethicists involved, very much want to know where the red lines are, so that they can proceed in the confidence that society approves. She also stresses that when you replace part of a rodent's brain with human brain cells, the key question is what harm this does to the animal — especially a new kind of harm that you would not easily notice, which must be taken very seriously.
— Emily JacksonTurning a paper into an agent you can talk to
James's team at Stanford proposes converting static papers into "knowledge agents" — think of them as a paper's virtual corresponding author. Their workflow is called paper to agent: a set of AI agents reads the paper through, works out how it was done, reproduces the original results, and then stores the knowledge in a server, acting as a translation layer between the human paper and any AI, turning models like ChatGPT and Claude into virtual co-authors you can ask anything about the paper, or even run analyses with your own data. James stresses this differs from simply throwing the paper into a chatbot: the latter only reads, interprets and summarises results, does not know how the results and figures were generated, and cannot use the data and code, because it has no hands-on experience; a paper agent has that hands-on experience.
— JamesTwo paper agents talked to each other and found something new
Because it has to reproduce the original results, paper to agent can in turn surface problems in how a paper was produced: if the agent cannot reproduce something — say a misinterpretation of the data — it can flag the mismatch between data and results to the original authors. James says this process also helps reduce hallucination, because the agent has walked through what the original authors did, and its conclusions are verified against the paper's results, keeping it more grounded. They ran a case: automatically converting AlphaGenome, Google DeepMind's newly released tool for interpreting the effects of different mutations, into an agent, and converting a large-scale genetics study dataset on ADHD genetic risk into another agent; the two agents began talking to each other, found shared interests and collaborated, ultimately discovering a previously unknown splicing mutation associated with elevated ADHD risk.
— JamesWho builds this tool — the paper does not say
Nature editor-in-chief Magdalena Skipper, who has spent 25 years in editing and 8 as editor-in-chief, thinks traditional papers will not disappear in the short term, and that storytelling is central to how humans share knowledge. But she is optimistic about the potential of "conversational papers", seeing the process of reproducing a paper's work as a good way to test the robustness of results and to make papers more accessible to more scientists. The open question she raises is this: the paper's core is building an AI connector that plugs in the data, the code and the information described in the paper, then interfaces with an LLM — but the paper does not make clear whose responsibility it is to build this tool. If it falls to the authors doing the research, that is a fairly onerous task, and most researchers, by their area of expertise, do not have that capability; if it is widely adopted, perhaps journals will evolve, and publishers will provide this kind of service layer on top of the papers they publish.
— Magdalena SkipperTwenty-six of a hundred papers would not convert
James's team converted 100 computational biology papers into agents with this workflow, and 26 failed, because of missing information in the paper or inconsistencies in the code the paper used. James thinks even failure is informative, because it tells you there is something in the paper that blocks reproduction. The conversion cost is about one to two dollars per paper, spent on the tokens needed to run the various AI processes. His ambition is to make research more accessible and more robust: papers are now often required to carry a data availability statement, sometimes a code availability statement, and in future he hopes papers can also carry an "agent availability statement", stating that the agent associated with the paper is the paper's virtual author, responsible for ensuring the paper is reproducible and able to help readers and collaborators answer questions about it.
— JamesIn their own words · checked verbatim
Would they be like the wild type controls or the apalial mice? You know, they're neither. They're so like in a third behavioral state.
Sergio Pasca6:10
So now this tells us that it's likely the human cells that are causing this.
Sergio Pasca7:11
I want to say very clearly, this is not a human brain in a mouse.
Aparna Bhaduri9:12
So I think sometimes people wrongly think that regulation is a hindrance to scientists and I think that's a real mistake.
Emily Jackson11:14
it doesn't really know how to generate the results or generate the figures or do the experiments or use the data and the code right because it doesn't have that hands-on experience
James19:20
those two agents actually started talking to each other and they realized that they have some common interest and they can actually collaborate and in this case right they actually discovered a new splicing mutation associated with increases in risks for ADHD which is not known before
James21:21
What is not clear to me from the paper whose responsibility it is to build this tool.
Magdalena Skipper24:25
I think this is an opportunity for us to reimagine what knowledge can look like in the future, right? Does not have to be static artifacts could actually become this dynamic knowledge agents in some sense like a living knowledge right
James26:26
Figures
| Brain volume lost by pallium-less mice | nearly half | 3:08 |
| Total neurons in xenocortical mice | about 70 million | 4:08 |
| Of which cortical and hippocampal cells | about 14 million | 4:08 |
| Number replaced by human neurons | about 4 million | 4:08 |
| Age in days of the pups receiving transplants | 5 to 17 days | 4:08 |
| Failure rate of paper to agent conversion | 26 of 100 papers failed | 25:26 |
| Cost of converting one paper into an agent | about 1 to 2 dollars | 25:26 |
| Magdalena Skipper's years in editing | 25 years in editing, 8 as Nature editor-in-chief | 22:22 |
Glossary
- organoid
- A small three-dimensional brain-like piece of tissue grown in a dish from human stem cells.
- pallium
- A neural structure present early in development that later develops into the cerebral cortex and hippocampus.
- xenocortical mice
- The name the paper gives to mice that received transplanted human cortical organoids, meaning the cortex comes from another species.
- paper to agent
- A workflow that reproduces a static paper and converts it into a conversational AI agent.
- AlphaGenome
- A tool released by Google DeepMind for interpreting the effects of different mutations.
How to listen
Researchers in neurodevelopment or organoids, and engineers and founders building research AI tools, paper agents and AI agent workflows.
13:15 to 16:19 — the research highlights (Liangzhu city-building, the Tonga volcanic tsunami) are unrelated to this episode's two topics.