Action Potentials for November
1: New study on the neuronal networks controlling eye movements in the larval zebrafish brainstem.
The brain was initially preserved by cutting off the skin adjacent to the hindbrain and simply immersing it in a mixture of 2% paraformaldehyde and 2.25% glutaraldehyde. Turns out good ultrastructural preservation is easy when the brain is tiny and not stuck inside of a skull. The tissue was then embedded for electron microscopy, serial sectioned, scanned, and the brainstem network was reconstructed:
They identified two types of abducens (ABD) neurons that coordinate eye movements. The ABD motor neurons (ABDm) are the pool of output neurons that directly control the lateral rectus muscle, which moves the eye outward on the same side of the body. The ABD internuclear neurons (ABDi) first connect to motor neurons on the opposite side of the brain, which then control the medial rectus muscle that pulls the opposite eye inward:
Other neurons they identified clustered into two submodules, connected to one or the other type of ABD neurons (ModOM or ModOI):
Further analysis showed that each submodule contained a three-part cycle of neural connections — like a triangle where information flows from point 1 to 2 to 3 and back to 1. One cycle primarily controlled the same-side eye movements through the ABDm neurons, while the other controlled ABDi neurons. Before mapping the connectome, this organization pattern not known.
They built a model to predict the change in firing rates of each neuron for a given change in eye position. They used very simple rules for their electrophysiological simulation, including approximating the connection strength between neurons based on the number of synaptic connections between them. They then compared functional imaging results from real zebrafish to their simulation results. They found that their model had a good correspondence to the true data (d), better than you would expect just based on predicting “potential synapses” instead of having the actual connectome (e):
I haven’t assessed the accuracy of their results. I don’t believe they pre-registered their analysis plan, so it should be thought of as exploratory. However, I trust that they are good scientists and I have no reason to doubt them more than I doubt any other paper I read. It doesn’t really matter that much to me, because what is most interesting to me is the method: it’s not just predicting behavior, but predicting whether the connectome-based simulation has the same electrophysiology as is seen in vivo. To me, this is key to bridging the structure to function gap in neuroscience, and is part of what will eventually distinguish neural emulations from just behavior cloning.
Note that although it is being published now, it seems that much of this work was likely done years ago. The lead author left Princeton in May 2021. Just a reminder that when we read about science, especially in journal publications, it is often a few years behind the cutting edge.
2: CaMKIIα is one of the most famous molecules in neuroscience. It makes up ~1% of total protein content in the brain, its knockout was the first genetically modified mouse line to show impaired learning without gross morphological brain alterations, and it has long been thought of as a “memory molecule.” Complicated models have suggested that autophosphorylation at a specific site (Thr286) could maintain memories through self-perpetuation.
A new review describes the accumulating evidence challenging this belief. While CaMKIIα activity is clearly essential for certain types of learning, multiple lines of evidence suggest that it’s not required for memory storage. This suggests a need to explore other mechanisms of memory storage, including the possibility of memories being stored purely in higher-order structural changes.
On the meta level, I want to point out that this seems to me like an example of the brain being less complex than previously thought. Endless claims of “the brain is the most complex object in the universe” are not necessarily helping us get closer to the truth.
3: Study performs postmortem perfusion fixation of the whole body through the carotid artery in six cases and does MRI scans of the brains. They find that a loss of grey matter-white matter contrast on T1-weighted images is a proxy for inadequate fixation. However, the tissue preservation was still sufficient for light microscopy even when the fixation was not ideal.
4: The 2024 Nobel Prize in Chemistry was awarded to three people: Demis Hassabis and John Jumper (from Google DeepMind) for developing AlphaFold, and David Baker (University of Washington) for his work on computational protein design. This article points out how Helen Berman and others in the community helped create the Protein Data Bank, which now holds over 200,000 protein structures and was essential for training AlphaFold. To me, the decades of data representation and collection that made it possible is probably more impressive than the AI work, which is also important but more obvious as a thing to do.
5: H5N1 has been in the news and is obviously concerning. However, this paper is a reason for hope. It suggests that if H5N1 bird flu mutates to prefer upper respiratory tract receptors (SA α2,6) instead of lung receptors (SA α2,3), which would favor transmission, it would likely also become less deadly since it would cause mainly upper respiratory infections rather than severe pneumonia. This is supported by ferret studies where no ferrets died from transmissible H5N1 variants that had this receptor switch.
6: Study finds that tracking poses with AI can accurately monitor infant movement from video data in the NICU. It can also predict neurological changes like sedation and cerebral dysfunction.
7: Fine tuning the language model Llama 3.1 70B on “Psych-101,” a data set with 10,681,650 choices by 60,092 participants across 160 psychology experiments, creates a model (“Centaur”) that is reportedly better able to predict human choices than existing cognitive models. They suggest that we can use this approach to make general models of human cognition and create a new field of psychology research.
8: Study finds that higher levels of testosterone are associated with more somatic symptoms in men with depression, especially loss of appetite and weight loss. Also finds that men with lower pre-treatment levels of testosterone who are prescribed SSRIs are more likely to have decreased libido. The sample size is probably too low to trust these results very much, but interesting nonetheless.
9: Randomized trial finds that short-term supportive psychodynamic therapy is non-inferior to cognitive behavioral therapy for the treatment of major depressive disorder (n = 290, 16 sessions, 8 weeks).
10: Randomized trial of hyperbaric oxygen therapy finds that it is helpful for symptoms of PTSD in veterans (n = 56, 60 daily sessions).
11: Donanemab, an anti-amyloid antibody treatment for Alzheimer’s disease, has a new, slower dosing protocol. Starting at lower doses and gradually increasing over 4 months reduced the risk of brain swelling by 40%, while maintaining its effectiveness at clearing amyloid. Might be because the first reactions between amyloid plaques in the vessels and the antibodies are the most dangerous.
Interesting to see studies on the slow titration of any medication leading to improved results. This is one of the few times that pharmaceutical companies would do these expensive studies, because these medications are potential blockbusters. I wonder the extent to which this would hold true for other medications as well.
12: New data suggests that brain aging in dogs and cats more closely mimics that of humans than brain aging in mice. This suggests that studying the mechanisms of cognitive decline in dog and cat brains could yield findings relevant for human health as well.
13: The first amber is found in Antarctica. It is estimated to have been preserved for 83-92 million years. Suggests that Antarctica once had a temperature warm enough to have resin-producing trees.
14: Podcast with Jeff Alstott on US government AI policy.
15: The Allen Institute is putting on a free workshop in July 2025 on using neuronal network simulation software.
16: There was coverage of our survey on the neuroscience of memory in The Times. The article was mostly focused on Ariel Zeleznikow-Johnston’s new book, “The Future Loves You: How and Why We Should Abolish Death,” which is available in the UK on Nov 28th.
17: Ariel also has an interview with the Foresight Institute, in which he reports that the best advice he ever received was when he was kid, when he wanted to do risky things and the adults around him would say “You can't enjoy things if you're dead.” This made him think that we need to be safe and healthy, so that we can have healthy and flourishing lives long into the future.
18: I am planning on attending Foresight Institute Vision Weekend USA in SF on Dec 7th-8th: https://foresight.org/vw2024us. I will be giving a 10 minute “Lightning Talk” in 3:30-4:30 PM slot at Lighthaven on Sunday Dec 8, titled “The race against time in preserving the brain.” I would love to meet any readers of Neurobiology Notes there. Please reach out if you’re planning on going.
Fantastic job as always, Andy. You are like a super-AI that scans the literature and summarizes it for us.
"To me, this is key to bridging the structure to function gap in neuroscience, and is part of what will eventually distinguish neural emulations from just behavior cloning." Yes! Vital to do more than just emulate behavior. The Turing Test is worthless for those of us interested in surviving on a more reliable substrate.
Thanks for the shoutout Andy!
Have a great time at Lighthaven, that sounds like it'll be a blast.