Action Potentials for June
1: A new strategy to increase the multiplexing capacity of the protein labeling method DNA-PAINT, called Secondary label-based Unlimited Multiplexed (SUM)-PAINT. Beautiful results in profiling 30 proteins at a single-protein resolution in formaldehyde-fixed neurons:
Here are zoomed-in 3d images of the boxes above at an excitatory (C) and inhibitory (D) synapse:
There’s been a little bit (not a lot, granted, but a little bit) of debate over the years about what level of structural detail would be needed to capture the information content of long-term memories. Some people think that unlabelled electron microscopy images alone would be sufficient, whereas others think that biomolecular labeling would be needed. My point has been that at least in the long run, it doesn’t really matter, because:
(a) biomolecular labeling technologies like this DNA-PAINT variant are getting so much better, and
(b) there will be some amount of damage in nearly every contemporary brain preservation procedure anyway. And single biomolecule labeling seems like the best way to infer the original morphological state in the case of damage.
2: A new study finds that connectivity information of the Drosophila optic lobe motion pathways derived from microscopy images alone is sufficient to predict the electrophysiologic responses of individual neurons to visual stimuli. For example, their model was able to predict the separation of neurons into light increment (ON) and light decrement (OFF) channels:
3: New study combining serial section EM and volumetric fluorescence microscopy to perform molecular-annotated connectomics of the mouse cerebellum. Much more scalable than DNA-PAINT.
4: Logan Collins on the feasibility of mapping the brain by expanding it and then using synchrotron x-ray microscopy. Argues that a whole human brain could be imaged at sub-30 nm voxel size for a price of roughly $10 million.
5: New cost estimates for scaling up electron microscopy connectomics by the Wellcome Trust. Estimates that a whole mouse brain would cost $7.5 - 21.7 billion, mostly due to the cost of proofreading ($7-21 billion), followed by registration and segmentation ($300 - 400 million) and imaging ($200- 300 million). So it basically won’t happen anytime soon unless registration, segmentation, and especially human proofreading are somehow automated by machine learning.
6: How much computation is the brain doing? We can count neurons and synapses, but a key question is the extent to which computation is occurring at the level of dendrites. Dendrites have nonlinear excitations, such as sodium spikes, which might be performing computations.
A new study in acute brain slices finds that rapid dendritic integration — i.e. the process by which neurons combine the electrical signals they receive from other neurons via synapses located on their dendrites — is dominated by passive cable properties, such that these nonlinear excitations can be ignored. On the other hand, nonlinear dendritic spikes played an important role in slower back-propagating action potentials, which mediate changes in synaptic strength.
7: Study finds that the psychological properties of risky and safe options may be different because they are encoded by separate sets of neurons. These distinct neural codes still seem to be comparable because they are linearly transformable.
8: There is a specific group of neurons in the visual cortex that mediate illusory contour perception, enhancing the brain's ability to infer objects from incomplete or ambiguous sensory information. When these neurons were selectively stimulated using holographic optogenetics, it led to the recreation of illusory contour representations across the visual cortex, even without any visual stimulus being presented.
9: Injecting a mutant version of the protein GNAQ that is constitutively active into the hippocampi of 24-month-old mice rescues their age-related impairments in long-term memory. My kind of gain of function experiments.
10: Another article points out that the whole brain death criteria is problematic because it doesn’t take into account the preserved hypothalamic function present in up to 50% of people declared legally dead by neurologic criteria. Makes the case that medical professionals and societies ignore, deny, or minimize this fact because doing so serves their interests.
11: Thomas Reilly reviews the evidence for early intervention in psychosis. The duration of untreated psychosis has an obvious association with poor outcomes, for example often by damaging a person’s support network. But this might just be a measurement artifact. And either way, it’s unclear what the best way to address this is.
12: In a new trial of the effects of running therapy and antidepressants (escitalopram or sertraline) in patients with depression and/or anxiety disorders, both interventions exhibited similar efficacy in mental health improvement at 16 weeks (remission rates of around 44%). However, running therapy (2-3x/week) showed superior results in other health outcomes, including weight, waist circumference, blood pressure, and heart rate parameters. A key limitation is that only 22 participants were willing to be randomized, while 119 chose their preferred option: antidepressant medication (n = 36) or running therapy (n = 83), so there’s an obvious selection bias.
13: In stressful situations, the adrenal glands secrete higher levels of glucocorticoids. A new study finds that chronically high levels of glucocorticoids drive the enteric nervous system to produce an inflammatory type of glia that promotes immune-mediated inflammation. Deleting the glucocorticoid receptors in these glia protected mice from the colitis-exacerbating effects of stress.
14: In the AI discourse, here’s Matthew Barnett’s most likely scenario for the future. Predicts that AI development will progress without abrupt jumps, leading to regulated deployment and alignment with human values via safety measures. Eventually, AI will automate all jobs, improving overall material conditions for the vast majority of people, and transitioning the world into an era where AIs are in control and humans are retired. One of my thoughts is that I’m not sure the division between humans and AIs is going to be so obvious.
15: Ari Allyn-Feuer and Ted Sanders on why transformative AI — i.e. AI that can accomplish all valuable human tasks — is very unlikely in the next 20 years, with <1% probability under their model. Seems accurate to me.
16: In related news, I made a bet for charity with Daniel Kokotajlo about whether all aspects of programming jobs will be automatable via AI by January 2027. I took the side of no. Manifold market:
17: A big breakthrough in cryobiology this month as a study reports transplantation of rat kidneys after cryostorage for 100 days, using perfused iron oxide nanoparticles to aid with radiofrequency rewarming:
The key advance compared to this group's previous efforts seems to have been in changing the cryoprotectant (from VS55 to VMP) and the loading protocol to decrease renal toxicity. The field is very much in the era of finding a cryoprotectant solution optimized for each tissue. This makes it more difficult to preserve heterogeneous tissues like the brain, which has big differences between white and grey matter.
18: Max Marty presents survey results from cryonicists and cryonics-adjacent people. A notable result is that once again, respondents tend to have a highly computational background, with the top academic degrees being Computer Science (31%), Mathematics (15%), and Engineering (14.5%).
19: Brian Wowk’s tribute to Saul Kent, who was cryopreserved by Alcor in May. One of his main accomplishments was starting a company with Bill Faloon that allowed him to contribute tens of millions to cryobiology research.
20: Q&A with Emil Kendziorra from the cryonics organization Tomorrow Bio. One part was especially well stated (paraphrased slightly):
Question: "Just to understand the intent, are you aiming for some kind of hibernation with continuity of consciousness or is this just preserving neural architecture?"
Answer by Emil: "The logic is that as long as the connectome is there — the connections and the weights — then for all intents and purposes, you're still there. Of course, the caveat is that we don't exactly know how consciousness works, so it's possible that this wouldn't work. But I think it's a valid educated guess, given that the alternative is literally burial or cremation. However, this would not involve continuity of consciousness. Your consciousness would stop and then, just like in the past when people thought that when the heart stops it's the end, our understanding of death could change over time. CPR, for instance, has only been around since the 1960s. So, not in the next 10 years, but maybe in the next 50, 100, or even 150 years, this concept could change again, and we will understand that death is recoverable under certain circumstances."
21: Upon discovering an article by Dr. Sigm. Freud from 1884 about brain preservation and staining, my official position on Freud is that I like his old stuff better.
22: In personal news, my family and I are moving from NYC to Oregon. Goodbye, terrifyingly difficult to quantify risk of being in the first city nuked by Putin. Hello, terrifyingly difficult to quantify risk of being attacked by an elk.
Action Potentials for June
Good luck with the move!
Fantastic Notes! Such a useful breadth of coverage of topics: neurobiology, memories, preservation, emulation, AI, theory of mind.
Saul Kent was a true visionary, it was great to read Brian's tribute, and to hear he's been preserved. I look forward to playing basketball with him again one day!
You get to the heart of the issue around AI and human futures. They are massively converging today, a journey that began with Turing in the 1930s. All our best AI is becoming increasingly neuroinspired, and it will all become evo-devo in nature, in my view. We the living are in for a major set of human-machine convergences, most obviously in personal AIs. It is also clear to me that all of us reanimated people will be partly biological and partly technological in nature. Even those who want to stay entirely biological will be doing so via a deep technology stack. The division between the two will become less relevant the more alive our technology becomes. Our leading tech will have to become alive, if it wants to thrive in the coopetitive tech ecosystem to come. Life not only finds a way, it copies and varies itself in any substrate that offers a selective advantage.
I wish your family a safe journey to Oregon. Such a lovely state!