Featured
2024: A Review of the Year in Neuroscience
Feeling a bit wired
Eight years I’ve been writing these annual reviews. Eight. I now regret starting the tradition of opening each with a pithy, allegedly humorous take on the year’s political news. When I started these in 2016 the idea was that, surely, things couldn’t get much worse — many people of the UK had voted to leave the EU, many people of the US had voted to take leave of their senses — and so future news round-ups would show the arc of history bending back in a better, happier direction, as we returned to meaningful international cooperation and tackled crises like climate change and mass migration together. Bloody hell was I wrong. The arc of history bends slowly.
So stuff the politics, let’s do some brains. The arc of science tends forward, progressing, improving. Science is always, in the round, a source of hope. As individuals we scientists may moan about our lot as researchers, or about crap studies, or terrible statistics, or how our institutions treat us. Yet our individual grumpiness adds up to constant, yearning improvement. Here are four small yet significant ways we improved our understanding of the brain this year. And odds are good something here will be true now and forever. Isn’t that cool?
1. Vision quest
Here’s a thought: Might we consider studying vision in an animal that uses it? Don’t get me wrong, we’ve learnt a lot from studying the visual system of rodents, from the details of how cortex is wired to the limits of sensory coding. And all of our most powerful tools, from imaging the neural activity of specific neurons to perturbing that activity with light, are most easily and powerfully used in mice, with the consequence that they are the animal of choice for large-scale work in visual neuroscience, notably by the Allen Brain Institute and the International Brain Laboratory.
But mice and rats are nocturnal. (I confidently state ‘nocturnal’, like all writers on this topic, yet my mind can’t help but turn to the large rat that has recently taken to visiting our lawn, joyfully bounding across it pretending to be a squirrel, occasionally pausing to take the air, in broad daylight, while our cat sits on the dining room table staring at it through the window with a demeanour of “it’s not my problem, mate” when it very much is.) Mice and rats can navigate the world just fine without light, or eyes, through a mixture of their sense of smell, their ability to keep track of their self-motion better than us (not hard) and, especially, their whiskers, which probe and prod the ground in front and objects around. Crudely put, their eyes are for detecting potential threats (things above them) and food (things below them). Not for watching Bridgerton or football or clips of pandas falling off logs. Rodents are a poor model for how humans use their eyes.
A lovely paper this year showed us something of what we might be missing. It simply asked, how does primate and rodent navigation differ? Our intrepid authors turned to the (common) marmoset, that like us is a primate, lives its waking life in daylight, and uses it front-facing big eyes to guide its behaviour. It’s also a cute, small 20-cm high monkey.
By letting their marmosets forage freely in a 3D, split-level box rather like a massive hamster play arena (the arena is massive, not the hamsters), they found two basic differences from rodent vision. The first is that rats explore by moving, not looking; marmosets, by contrast, explore by looking, not moving. Not in itself a startling revelation, for that’s what we do: we scan our environments, noting salient features like the clock tower yonder or the small bronze statute of a urinating boy over there. But that’s the point. Rodents don’t do this.
The second difference is in the marmoset hippocampus. If you want to look at how animals are literally representing their place in the world, look there, for that is where are found the Nobel-prize winning place cells, the neurons that fire when a rodent is in a specific location of the world. So the team looked there in the marmosets. And found that no neuron was just a place cell. Sure, a majority of them had firing that related to where the marmoset was in the 3D, split-level box. But all of those neurons coded for something else too, and the majority of that was the view: what specific view the marmoset was looking at. And those place-related neurons did not cover much of the big box, whereas the view-related neurons did. Use an animal that has eyes to see and it turns out it uses different representations of the world than a rodent. What were the odds?
2. Fly in the ointment
As regular readers may recall, I’ve been sceptical about the worth of connectomes, of building the complete wiring diagram of the connections between each and every neuron in a brain. The source of the scepticism was simple: a wiring diagram is a terrific reference tool, just like the genome, making it easier to identify the parts of a brain and, when recording neurons, providing a way to better guess which neurons and what they connect to. But it is just a description, with little hypothesis-driven science possible without major further work. This scepticism has taken a beating recently. In 2021 Witvliet and co showed it was possible to derive scientific insight from the connectome alone — all you needed was more than one. They mapped the neural development of the tiny worm C Elegans by reconstructing connectomes at key stages of its life, providing, among other things, more strong evidence for the crucial role of randomness in the wiring of brains, even in a brain with just 302 neurons. And then, this year, FlyWire dropped.
At its heart, FlyWire is the full connectome of the brain of the adult female fruit fly (or Drosophila melongaster if you insist). It’s a monster, in connectome terms: 140,000 neurons divided into a strangely-precise 8453 types of neuron and connected by about 50 million synapses, making up to 15 million connections between neurons (if we count every single synaptic connection). A lot of wire in a tiny brain of a tiny fly.
But what made FlyWire special was the not the admittedly impressive technical feat of tracing all that wire but the effort put in to then doing science with it, to show how it could be used to drive knowledge forward. Officially, FlyWire dropped 11 papers as a collection across the Nature journals, ranging from a simulation that literally placed a model neuron everywhere there was an actual neuron, to a detailed probe of how the fly stops walking, using the connectome to show at least 3 separate neural pathways that call a halt, two seemingly specialised in stopping the fly to consume food rather than wandering past it looking mournfully at what it could have eaten, the third holding the fly still while grooming, so it doesn’t try to rub its face and walk at the same time, fall flat on its proboscis and look a bit of a prat to its mates.
Officially 11 papers; unofficially, I’d say at least 14. Pretty much every experimental, theory, and modelling paper relied on a crucial Cell paper from earlier this year that had inferred the sign of each connection already, whether it excited or inhibited it target — without that, much of the theory and modelling work would have been impossible. And two of the best papers appeared earlier this year, studying deep neural network models of the fly brain that were strongly constrained by the connectome: Lappalainen et al showed us that training wiring-constrained deep networks to track the visual flow of the world as the fly moves recovered many known types of neuron responses and predicted new ones; Cowley & co showed us a beautifully tight model-experiment interplay by training a deep network model to match the behavioural effects of knocking-out each of 23 types of early visual system neuron in the fly.
We cannot of course abandon all scepticism, for science proceeds through reasonable doubt. Types of neurons remain a nebulous concept: it was telling that about a third of the neuron types found in the previously-published, partial half-brain connectome of the adult female fruit fly could not be found in the newly-constructed full brain. The favoured explanation is that the disagreement represents variation between individual flies. The figure of 8453 types was arrived at by taking a consensus across three half-brains: the previous half-brain one, and treating the new, complete one as two halves. But how can it be a neuron “type” if it is simply missing from a different individual of the same species (and one, I suspect, that is genetically-identical)?
And deeply impressive as FlyWire is, the problem of scale remains: if we could scale the technology to the whole mouse brain, would such a synaptic-level connectome hold any value? After all, the constraints of genetically-specified connections between neurons must get weaker as the brains get bigger, so random chance plays more and more of a role in initial development, and life experience plays more and more of a role in sculpting connections from that baseline. What we want instead are the rules for wiring, but for that we need many, many connectomes, in order to infer the general principles they share. This creates a connectome paradox: as the brains get bigger so, conversely, the value of individual connectomes gets smaller. This suggests only making the leap to doing many brains at once is worth it, with little pay-off in the short-term. In the past I would have questioned whether it was worth it; FlyWire has given me pause for thought.
3. Shining a light in the dark
Two papers especially piqued my interest this year as they were on topics close to my heart. The first paper gave a new take on dark neurons. I’ve written much about them before, including a dedicated chapter in my book The Spike. The short version is that cortical activity obeys a rough 80:20 rule — 80% of spikes come from just 20% of the neurons. That means the other 80% are mostly silent, firing much less than 1 spike every second. For any given moment in time, for any given behaviour, they are not active: they are dark. This raises the question of what they are for. Perhaps they are reserve neurons, awaiting recruitment to represent something new; perhaps they’re just very picky and we’ve just never shown them anything they’re interested in.
Work from Oliver Gauld and colleagues in Michael Hausser’s lab this year offered a new view, that dark neurons are not reluctant to respond, they are actively suppressed. They looked at neurons in the barrel cortex, the neurons whose job is supposed to be responding to deflections of a single, unique whisker on a rodent’s face. They looked there because many of them don’t. It’s their only job, yet many neurons in a single “barrel”, representing a single whisker, don’t respond to deflections of that whisker. Indeed some of our best early evidence of dark neurons came from the total lack of response from neurons in barrel cortex.
Gauld and co flicked a whisker on each side of a mouse’s face at the same time and simply asked it to report which side had moved the most. Many neurons in layers 2 and 3 of that whisker’s “barrel” were dark to the stimulus, not responding differently whether the left or right whisker had be stimulated. Yet activating enough neurons using light could bias the mouse’s choice to the other side, as though, when forced to be active, these dark neurons were perfectly capable of transmitting information about the whisker. So why didn’t they normally? Because they were locally and actively suppressed: the more some neurons were stimulated, the more the unstimulated neurons around them reduced their activity. This gives us a new “sparse coding” hypothesis for why neurons are dark: not because they are hardwired to respond to just a few, select features of the world, but because they are actively suppressed from responding. Some light in the darkness.
4. How low can you go?
The other paper that especially piqued my interest was a thought-provoking take on the dimensionality question: how many dimensions does neural activity have? It’s now routine to take a recording of hundreds or thousands of neurons and ask if the dynamics so captured have fewer dimensions than there are neurons. One reason to do this is that it makes understanding the brain easier: we don’t have to worry about coding in thousands of neurons, but just a comparative handful of dimensions of activity. Recently, I’d pointed out that as we increase the number of recorded neurons we hope the number of additional dimensions doesn’t scale with them; rather, we hope the number of dimensions of neural activity has a hard upper limit that is a tiny percentage of the number of neurons. Consider what happens if it doesn’t. There are about half a million neurons in the primary visual cortex of a mouse and about 50 million neurons in the primary motor cortex of a macaque monkey. Even if the number of dimensions of activity is just 1 percent of the number of neurons then if the dimensions of activity scales all the way to the full extent of the region’s neurons that’s still 5,000 dimensions in mouse visual cortex and 500,000 dimensions in the macaque’s motor cortex. That’s still high-dimensional in anyone’s book. This year, the 1-million-neuron paper appeared in Neuron, and had a first crack at answering to this question.
I’d previously covered this technical marvel when it was a pre-print: the scaling of calcium imaging to cover vast chunks of a mouse’s cortex, with typical recordings in the low hundreds of thousands, and one exemplar obtaining the activity of 1 million neurons at (roughly) the same time. So this team were the best-placed on the planet to ask how dimensions of activity scale with the number of neurons.
They estimated the number of “reliable” dimensions in the activity, essentially by finding how many dimensions were needed to capture the variation in activity of two sets of neurons at the same time — the idea being that these dimensions must be common to both sets, and thus not noise. And while 16 of these dimensions captured about half the variation in activity, the number of reliable dimensions grew unbounded with the number of neurons.
On the one hand, this means that we can indeed describe a good chunk of activity with just a handful of dimensions. But on the other, it implies that recordings have to scale with the size of the targeted population in order to capture all that population’s shared dynamics. Which raises the obvious question: how do we know how big the targeted population is? But that’s a story for a different time.
Is this thing on?
There was so much more of course. Rafal Bogacz and crew kicked the year off by taking a different slant on the problem of how the brain could approximate backprop and hence learn like an artificial neural network. Their slant being: it doesn’t, that would be silly. Instead, they proposed that a network of neurons predicts what the neural activity at the end of learning should look like, and then change synaptic weights according to whether that prediction looks correct or not. We also had much more on how psychedelics alter brain function, including suggestions they desynchronise activity in the human brain. And if you like science cartography, there’s now a map of the whole of PubMed.
We also welcomed a new voice for neuroscience, The Transmitter, an ambitious online magazine offering us deep dives into neuroscience news and opinions, with that most unusual of things, a dedicated editorial team and a roster of journalists. Two examples of some great journalism this year: Shaena Montanari’s investigation into the whether simply flickering a light at 40Hz could clear Alzheimer’s brains of the amyloid plaques linked to the dementia, and Angie Voyles Askham’s deep dive into the controversy of whether its time to abandon one of the success stories of modern neuroscience, the reward prediction error theory of dopamine.
And, full disclosure, The Transmitter is also in my news this year: I’ve started a regular column there (apologies for the massive picture of my face that is at the end of the link — please scroll down quickly), kicking off with why averaging is holding back neuroscience and imagining the ultimate systems neuroscience paper.
You could sign up to follow me there for more in 2025. You could continue to mournfully scroll through your social media feed. Either way, all I can suggest is that, if you’re after good news in 2025, it’s probably best to stick to science.
Liked this? Then you might like to read my book, an almost complete guide to all we know about how neurons work
Bluesky: @markdhumphries.bsky.social
LinkedIn: Mark Humphries
