4 ms·

Scientists generate first single-cell atlas of the primate brain

72 points by wglb 3y ago

seeknotfind 3y ago
From the title, I was expecting this to be a map of neural connections, but this is a map of gene expressions in 4.2 million cells from 30 brain regions.
frozenport 3y ago
Atlas but not spatial? Kinda false advertising.
- 331c8c71 3y ago
  Atlas of cell types. It's the standard terminology in the field by now. Spatial transcriptomics is also feasible now but at the expense of losing single-cell resolution.
twothreeone 3y ago
I think the article got the very first number wrong.. 100 million neurons only makes 1/1000th of a human brain.
stenl 3y ago
That’s not a terrible summary of the one paper, but there are 20 more papers. An overview by the Science editor: https://www.science.org/doi/full/10.1126/science.adl0913 https://www.science.org/doi/full/10.1126/science.adl0913 The full collection of papers is linked here: https://news.ycombinator.com/item?id=37878935 https://news.ycombinator.com/item?id=37878935 They are paywalled, but most are available as preprints on bioRxiv, e.g. https://www.biorxiv.org/content/10.1101/2022.10.12.511898v1 https://www.biorxiv.org/content/10.1101/2022.10.12.511898v1
scotty79 3y ago
What are the technical obstacles to just slicing preserved brain thinly and scanning each slice with microscopic detail?
- stenl 3y ago
  None, that was done 100 years ago, e.g. by Ramon y Cajal (Nobel prize 1906). But microscopic detail does not give molecular detail. What these current studies add is data on gene expression (mRNA molecules), chromatin accessibility (related to gene regulation), electrophysiology (in some cases), etc. We need such detail to connect disease genes inferred from genetics to specific brain cell types, for example.
  - scotty79 3y ago
    So basically just mapping each individual neuron and all of their connections is almost useless?
    - adr1an 3y ago
      Useless? No, but by today standards it doesn't amount to much. It's like having a low res pic in black and white. It worked alright in the past. That's the point.
    - stenl 3y ago
      No that would be amazing. But we don’t have the technology to map all the connections in large mammalian brains. It was done in the fruitfly just this year: https://www.science.org/doi/10.1126/science.add9330 https://www.science.org/doi/10.1126/science.add9330
      - agumonkey 3y ago
        that is a pretty amazing work
- clnq 3y ago
  I think: - Practically, we don't have microscopes to scan the brain at a resolution of every neuron and synapse quick or cheap enough. With infinite money, this might be overcome, but more on money later. - We wouldn't want to store this data as images, we'd want to store it as a graph, probably. It costs a lot to process this amount of data. That's a technical obstacle, not a technical impossibility. - There is risk of distortion in slicing, so some amount of reconstruction is necessary. We don't have the error-correction for this, which is where machine learning could help if we had enough samples of how things should be. But this field is not that developed. This ties in with another technical problem - validation. We probably don't generally know if we distorted things or not. - Cost and time. - End result might not be what we want. Slicing would let us produce something like a graph. But not necessarily understand what happens in the soma of a neuron when its dendrites receive a signal, and how that produces a signal in the axon. Although dendrites are the principal input zone, and axons are the principal output zone, all inputs and outputs of a neuron are not so clearly delineated or enumerated. There may be other reasons for particular output. In short - we don't see the dynamic changes with this method, so we can't even infer about the actual "weights", "heuristics", and other factors in the "decision making" of each neuron or vertex on the graph. For a fully functional graph, we'd need its vertices/tensors to have these weights, heuristics, and account for other factors in the signal processing that happens in the brain. Otherwise, the image would be a non-functional approximation, and I don't know if this is the end result we want. - Financing can be seen as a technical obstacle. In the recent decade, some interest has shifted from studying natural systems to ML as the more elegant solution for intelligence. Scanning the human brain could be in one of those technological freezes, like computing in ancient societies, where breakthroughs could have been made if there was interest. But interest was elsewhere. Those are my thoughts. It would be better to ask someone who works with this or close to this, but I think there are very, very few people in the world who work directly in brain imaging with slicing.

72 points by wglb 3y ago

From the title, I was expecting this to be a map of neural connections, but this is a map of gene expressions in 4.2 million cells from 30 brain regions.

frozenport 3y ago

Atlas but not spatial? Kinda false advertising.

331c8c71 3y ago
Atlas of cell types. It's the standard terminology in the field by now. Spatial transcriptomics is also feasible now but at the expense of losing single-cell resolution.

twothreeone 3y ago

I think the article got the very first number wrong.. 100 million neurons only makes 1/1000th of a human brain.

stenl 3y ago

That’s not a terrible summary of the one paper, but there are 20 more papers. An overview by the Science editor: https://www.science.org/doi/full/10.1126/science.adl0913 https://www.science.org/doi/full/10.1126/science.adl0913 The full collection of papers is linked here: https://news.ycombinator.com/item?id=37878935 https://news.ycombinator.com/item?id=37878935 They are paywalled, but most are available as preprints on bioRxiv, e.g. https://www.biorxiv.org/content/10.1101/2022.10.12.511898v1 https://www.biorxiv.org/content/10.1101/2022.10.12.511898v1

scotty79 3y ago

What are the technical obstacles to just slicing preserved brain thinly and scanning each slice with microscopic detail?

stenl 3y ago
None, that was done 100 years ago, e.g. by Ramon y Cajal (Nobel prize 1906). But microscopic detail does not give molecular detail. What these current studies add is data on gene expression (mRNA molecules), chromatin accessibility (related to gene regulation), electrophysiology (in some cases), etc. We need such detail to connect disease genes inferred from genetics to specific brain cell types, for example.
- scotty79 3y ago
  So basically just mapping each individual neuron and all of their connections is almost useless?
  - adr1an 3y ago
    Useless? No, but by today standards it doesn't amount to much. It's like having a low res pic in black and white. It worked alright in the past. That's the point.
  - stenl 3y ago
    No that would be amazing. But we don’t have the technology to map all the connections in large mammalian brains. It was done in the fruitfly just this year: https://www.science.org/doi/10.1126/science.add9330 https://www.science.org/doi/10.1126/science.add9330
    - agumonkey 3y ago
      that is a pretty amazing work
clnq 3y ago
I think: - Practically, we don't have microscopes to scan the brain at a resolution of every neuron and synapse quick or cheap enough. With infinite money, this might be overcome, but more on money later. - We wouldn't want to store this data as images, we'd want to store it as a graph, probably. It costs a lot to process this amount of data. That's a technical obstacle, not a technical impossibility. - There is risk of distortion in slicing, so some amount of reconstruction is necessary. We don't have the error-correction for this, which is where machine learning could help if we had enough samples of how things should be. But this field is not that developed. This ties in with another technical problem - validation. We probably don't generally know if we distorted things or not. - Cost and time. - End result might not be what we want. Slicing would let us produce something like a graph. But not necessarily understand what happens in the soma of a neuron when its dendrites receive a signal, and how that produces a signal in the axon. Although dendrites are the principal input zone, and axons are the principal output zone, all inputs and outputs of a neuron are not so clearly delineated or enumerated. There may be other reasons for particular output. In short - we don't see the dynamic changes with this method, so we can't even infer about the actual "weights", "heuristics", and other factors in the "decision making" of each neuron or vertex on the graph. For a fully functional graph, we'd need its vertices/tensors to have these weights, heuristics, and account for other factors in the signal processing that happens in the brain. Otherwise, the image would be a non-functional approximation, and I don't know if this is the end result we want. - Financing can be seen as a technical obstacle. In the recent decade, some interest has shifted from studying natural systems to ML as the more elegant solution for intelligence. Scanning the human brain could be in one of those technological freezes, like computing in ancient societies, where breakthroughs could have been made if there was interest. But interest was elsewhere. Those are my thoughts. It would be better to ask someone who works with this or close to this, but I think there are very, very few people in the world who work directly in brain imaging with slicing.