Scientists generate first single-cell atlas of the primate brain

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.

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.

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.