Microscopes are designed to image flat surfaces, unlike the curved structures typical for biology. Here, I would like to mention a piece of work where we demonstrate a method based on K-mirrors to overcome this lack of flexibility for two-photon resonant scanning microscopes.
In my work, I regularly use two-photon microscopy to image the hippocampal layer of pyramidal cells in mouse CA1. For years, I noticed again and again that the cell layer I wanted to image was sometimes captured perfectly (as shown in typical published figures), but more often cut through by the imaging plane at an arbitrary angle, as shown in the two examples below, leaving a diagonal band of cells (imaged by Shinjini Ghosh).
This is an issue that most neuroscientists using a two-photon microscope may have encountered. It is quite prominent for mouse CA1, but it is even more striking for mouse CA2 or CA3. And I was struggling with similar issues even when I worked on the small and curved brain of the adult or larval zebrafish during my PhD.
In principle, there is a simple solution to this problem: you can simply curve and tilt the imaging plane, using a fast z-scanning device such as a tunable lens or a remote mirror (which I implemented 10 years ago). However, these “fast” z-scanning devices are still slow in the range of several milliseconds settling time. This makes it possible to change the z-focus for the slow y-axis (typically 30 ms per scan) but not for the fast x-axis (65 μs per scan). Therefore, you cannot curve and tilt the image plane along the direction that would be ideal.
In 2024, the solution to this problem hit me in the face: a very simple but effective solution. I read this tweet by microscopy expert Reto Fiolka, in which they used galvo scan mirrors to build a fast SIM microscope. It took me a bit of time to understand what they were doing here, but then I realized that the principle they were describing – the use of K-mirrors to rotate a FOV – was not only the solution to my problem, but also an established method in astronomy for perhaps 70-80 years!
With my fearless PhD student, Shinjini Ghosh, we designed a simple add-on to resonant scanning two-photon microscopes that performs this FOV rotation and thus can optimize for example the imaging of pyramidal cells in hippocampal CA1.
Check out our preprint, where we demonstrate therange of applications that benefit from this small modification, ranging from CA1 pyramidal cells, hippocampal dendrites, cortical layers, prism-based dendritic imaging, cerebellar Purkinje cell imaging, up to optimized glutamate imaging. Here’s an example, where we have a tilted FOV in hippocampal CA2/3, where we demonstrate that FOV rotation with the K-mirror, combined with axial scanning can be used to bring the hippocampal dendrites beautifully into the FOV:
All neurons in this volume were automatically segmented and proofread. But what about glia? The paper only briefly mentions “non-neuronal” cells and does not consider them in the analysis. However, the automated segmentation typically also produces – as byproducts – segmented astrocytes or other glia cells (although probably not proofread and therefore less trustworthy). I looked into the publicly accessible dataset and soon found some putative glial cells.
Here’s an interesting pair of cells: a pyramidal neuron with its large cell body sitting on top of what may be a microglia cell (alternatively, but less likely, an interneuron). I have no idea why these two cell bodies are so close to each other. If you have an idea, let me know! (Here’s the link to look at this cell pair yourself.)
Next, I specifically looked out for astrocytes, which are currently my favorite braincell type. Since these cells are most likely not carefully proofread, the reconstructions need to be taken with a grain of salt. Nevertheless, the difference compared to morphologies of pyramidal neurons is really striking: astrocytes form much denser and local branches. They appear like sponges rather than stars, at least fromthis ultrastructural perspective. The video is slow and not smooth because the many astrocytic branches caused lag during my screen recording. Here’s the link to look at these astrocytes yourself.
For the blue-colored astrocyte, you can even see a putative endfoot that wraps around a blood vessel (although this would need to be proofread to be sure). In addition, it is possible to directly see the dense coverage of space by astrocytic processes – again with the caveat that these astrocytic segmentations probably have not be proofread.
It’s always interesting to explore 3D electron microscopy datasets. This one is not of the highest resolution (45 nm in the z-axis), but it is definitely fascinating to explore if you are interested in hippocampal circuits, be it from the perspective of neurons or astrocytes!
One of the most important unseen developments in neuroscience during the years 2010-2020 was the creation of online community spaces for open exchange about methods.
One of these places was Science Twitter, which was used as an exchange platform for scientific and in-depth technical discussions. With the transformation into X, this community fell apart and was not really able to piece itself together since.
Another platform for exchange among experts was the NeuroMethods Slack channel, created in 2016 by Benjamin Saunders. Over the years, this platform also lost some of its initial traction. First, users moved from NeuroMethods to Science Twitter. At the same time, a new generation of scientists entered the playground and used different channels and platforms. More recently, Slack’s own limitations (storing past conversations) have undermined the idea of building a persistent, searchable knowledge base.
Now, mainly driven by Eleonore Duvelle, this space has been moved to a new platform, replacing Slack with Zulip (https://neuromethods.zulipchat.com/). Researchers can join using an academic email or can contact the workspace manager directly via email.
NeuroMethods on Zulip not only offers a “space for neuroscientists to ask and answer questions about methods in neuroscience” but also preserves a (growing) history of past discussions across diverse topics such as:
spike sorting software
virus serotypes
vendor comparisons for surgery setups
transgenic mouse lines
antibodies
tetrode production
choice of commutators
data analysis workflows
and much more
If you are a neuroscience researcher who would like to help others with your knowledge, want to explore technical discussions from past years, or have technical questions of your own – it is worth taking a look:
To assess the resolution of your microscope, measuring the point spread function (PSF) is one of the most basic and reliable methods. The PSF essentially shows how large the image of a very small, often sub-diffractive object will be. See this paper for reference on how to measure a PSF, and check out a previous blog post on PSF shapes.
When you align and optimize a microscope, a certain undesired appearance of the PSF (spherical aberration, coma, astigmatism) can rarely be traced back to a specific optical component: does it come from the 10-year-old objective? Or is it the dichroic? Or the misalignment of the Pockels cell? Or is it the laser beam itself? Often, one needs to optimize the optical alignment without understanding where the suboptimal performance comes from.
Recently, I was aligning a system where I observed a clear aberration and – rare enough – could trace it back to a very simple optical component. The main observation was a very strong astigmatism, which occurred when the focus in one dimension (x) did not coincide with the focus of the beam in the other dimension (y):
Technical details: PSFs recorded with a 10x objective at 905 nm using 1 μm diameter fluorescent beads.
This often happens when the dichroic of the microscope is too strongly attached such that it bends, acting as a slightly focusing element in one direction but not the other. In my case, things were even simpler: I had inserted a set of relatively large mirrors (2″ side length) into the beam path. For such large mirrors, even a small strain can bend their surface considerably, leading to the astigmatism observed above. Releasing this strain resulted in a perfectly normal PSF, measured a few minutes after the PSF shown above:
It was great to see how I could reproducibly increase and decrease the strain on the large mirrors back and forth, consistently achieving full-blown vs. negligeable astigmatism. It’s always satisfying to figure out such small technical details.
Kurzgesagt is a Youtube channel that breaks down science topics for a broad audience. They have great videos, both in terms of narrative, animation, and scientific accuracy. A few days ago, the channel dropped a video that addresses a neuroscientific question: “What happens inside your brain when you form a memory?”
Here’s my review of the video from my perspective as a researcher in neuroscience and neurophysiology. I’ll briefly summarize the video (normal text) but will also provide additional information or comments (in italics).
1. What are memories? Synapses and plasticity.
1.1 Hebbian plasticity and its modern interpretations
As the video describes, the connections between neurons in our brain constitute our memories. The video describes how connections between pairs of neurons becomes stronger when they fire at the same time. For example, when a neuron coding for the picture of a cat and a neuron coding for the sound it makes fire together, they will form a connection.
This “fire together, wire together” idea reflects the notion of so-called Hebbian plasticity that was established in the middle of the last century. It is interesting to note that the candidate mechanisms that implement this plasticity rule (long-term potentiation, LTP, and spike timing-dependent plasticity, STDP) were established more than 30 years ago in brain slices. However, with more and more work done in awake animals from 2005-2015 onward, skepticism has arisen on whether these plasticity rules also apply identically in vivo. Additionally, new plasticity rules that seem to be distinct from the original idea of Hebbian plasticity have been discovered in awake animals (e.g., behavioral timescale synaptic plasticity, BTSP). Together, while the simple idea of Hebbian plasticity is still appealing and also most likely valid in some contexts, other plasticity rules may be equally important or even more important, depending on neuron types (excitatory neurons vs. interneurons, cortex vs. hippocampus, etc.) and the state of the animal. This is a broad and active area of research both in theoretical and experimental neuroscience.
1.2 Synapses in animations vs. in biology
The video briefly describes the nature of the connections between neurons: synapses. The small gap between the two neurons (“synaptic cleft”) is shown in this screenshot:
As in many popular science and textbook illustrations, the synaptic cleft is shown as relatively large. This is necessary because one wants to visualize the neurotransmitters emitted from pre- to post-synaptic neuron. In reality, the synaptic cleft is very small (typically 20 nm) compared to the size of the synaptic spine (approximately 1000 nm). Another small detail that this illustration shown above doesn’t get right is the idea of conversion from chemical to electrical signal. In reality, neurotransmitters bind to post-synaptic receptors, which open ion channels, and the ions flowing through these channels directly constitute the electrical signal. In the above scheme taken from the video, in contrast, the signals received by the receptors seem to be converted chemically into the yellow fiber-like molecules and only then into the electrical signal below. I have no idea what these fiber-like molecules should correspond to, and I find it a bit confusing.
2. What are memories? Assemblies, columns, and attractors.
2.1 Assemblies and columns
Next, the video describes how the connections between neurons result in groups of connected neurons organized in columns within a cortical area. Such distributed groups of neurons then form assemblies throughout the brain, with groups encoding for a cat in visual cortex being connected to the neurons in auditory cortex encoding for the sound made by the same cat. In reality – the video suggests – many assemblies are present in parallel throughout the brain.
These concepts of neuronal groups, assemblies, and (cortical) columns are important scientific notions. However, there are a few imprecisions in how they are described and illustrated. First, the illustration (screenshot below) gives the impression that groups of connected neurons are arranged in a chain-like structure. This is not how neurons within neuronal assemblies (often also called neuronal ensembles) are connected in reality, to the best of my knowledge. It is likely that the video creators mixed up a couple of more or less connected concepts: the idea of the columnar organization of cortex (with a column being not the same as an assembly of neurons), the related idea of subcolumnar mini-columns, and the not so much related idea of neurons that may be functionally related because they emerge clonally from the same progenitors and use the same radial glia cell in cortex as a scaffold for migration (‘sister cells’). Perhaps a mixture of these concepts was the source of this illustration:
However, this is not how most neuroscientists would picture a neuronal assembly – the cortical columns are not – as the video seems to imply – the elements through which assemblies are generated throughout the brain. Instead, assemblies are more likely believed to be generated by individual neurons or groups of neurons that connect across brain regions. It seems to me that the video creators made this simplification to streamline their argument, and since it seems almost correct. However, it is worth pointing out that the columnar organization of e.g. visual cortex occurs in humans, primates, cats, but not in rodents. And other intelligent animals like fish or birds do not have cortical columns at all, while they still form neuronal assemblies. It therefore appears that the columnar organization of functionally defined neuronal groups is not essential for the function of the brain and for the emergence of neuronal assemblies.
2.2 Attractors
As an additional comment, it is worth talking about the idea of “assemblies” in general. This idea is directly connected to the previously mentioned concept of Hebbian plasticity (“fire together, wire together”). In the 1980s, John Hopfield presented the idea of Hopfield networks, which are artificial neural networks that form stable states driven by assemblies of neurons (these stable states are called “attractors“). Attractors have been highly popular as a computational concept in theoretical neuroscience, but it turned out to be very difficult to find them in the brain. Special forms of attractors (“line attractors”) have been found convincingly for the oculo-motor system and for the direction encoding system, and there is some work that shows attractors also in hippocampal or thalamo-cortical circuits. However, strong and direct experimental evidence for attractors is rather low when we consider how attractive they are as a concept.
2.3 Manifolds
Additionally, recent work in neuroscience has somewhat shifted away from the idea of stable attractors and cell assemblies as the basis of organization of the brain, leaning more towards the mathematically inspired idea of neuronal manifolds (subspaces of neuronal activity). In this manifold view of the brain, distinct cellular assemblies play a less prominent role. However, for this educational video, it seems a good idea to use the concept of cell assemblies since it is much more intuitive. Perhaps, it is even an equally good description of neuronal activity as the manifold approach – we do not know yet for sure.
3. How are memories formed? The hippocampus as librarian.
Next, the video describes how memories are formed: the neuronal assembly consisting of multiple sensory and other inputs across the entire brain gets activated. This assembly then activates the hippocampus, a brain region that they introduce as the “librarian” of the brain. The hippocampus creates an “index” toward this (cortical) assembly and therefore generates the episodic memory.
This idea of the role of the hippocampus likely comes originally from the hippocampal memory indexing theory. This is, in my opinion, a very useful description of what the hippocampus is doing, and it’s a great choice to include in the video. However, the most common view of the hippocampus is dominated by the idea of a cognitive map. This cognitive map theory connects quite naturally to the concept of indexing: the hippocampus provides the spatial, temporal or conceptual grid which establishes the spatial, temporal and conceptual relationships between matters of thought. In a slight modification of the memory indexing theory, hippocampus could be seen, rather than a switchboard for indices, as a generative brain structure to play out spatial relationships and temporal sequences. These aspects of the hippocampal formation are currently the focus of active research, and it is clear that many hippocampus and memory experts will also disagree with some of these ideas or maybe even have never heard of them, because the field is to some extent compartmentalized into subfields that are difficult to distinguish from the outside.
4. What strengthens memories?
4.1 Novelty, repetition, and emotion
The video moves on to describe how events are stored more permanently, by the mediation of three potential factors: novelty, repetition, or emotion. This is a great summary, and is also practically helpful for the viewer. While none of these factors are extremely surprising, their involvement in memory consolidation is a fundamental insight. It is always worth reflecting that you will not remember something (e.g., a lesson in school) if you are not interested in it, if it does not appear new to you, or if you don’t repeat it.
To give some context: it is important to understand that the memories discussed in this video are a special kind, in neuroscience often referred to as “episodic memories”. There are other types of memories related to motor learning, trauma, sensory conditioning, or semantic memory. These memories are treated slightly differently by the brain and use other specialized structures of the brain. For example, the striatum for motor learning, the amygdala for fear conditioning, and cerebellum for sensory conditioning. It is understandable that the video focuses on episodic memory formation since this is the most relatable form of memory.
4.2 Neuromodulation and specialized brain areas
It is worth thinking about the implementations of these three factors that enhance long-term memory formation – novelty, repetition, and emotion. Novelty is a signal that is associated with surprise and arousal, and is broadcast throughout the brain by the brain stem region locus coeruleus, which releases noradrenaline. Noradrenaline acts as a “neuromodulator” (although we and others have found that it acts also on astrocytes, maybe even more strongly) to modulate neuronal plasticity and therefore the formation of memories. Another neuromodulator, which is even better known, is dopamine, which does not signal novelty alone but rather an expectation mismatch (but see also this current debate). The opposite emotional valence, which also supports memory formation, seems largely driven by the brain structure of the amygdala.
4.3 Replay of memories
Finally, repetition is not only something we do ourselves by rehearsing the lines before a recital or by looking at the notes made in school; our brain also has a mechanism in place that simulates recent events and simulates them in front of our inner eye during both phases of rest, and during sleep. These events are called “replay” events. They are associated with a characteristic electrophysiological signature, the “sharp-wave ripple” complex, and are thought to form the basic unit of interaction of hippocampus with the cortex (touching upon the indexing theory from before).
5. Memory is not static but reconstructive
The video stresses this importance of sleep and rest for memory formation. It is alreay well-known outside of neuroscience that sleep is important to keep things in your mind (which otherwise are lost). But it is worth repeating this over and over again. In the video, this note falls very well into place.
Overall, the video does an excellent job at describing why at each time when we access a memory, we influence it with our current context and change it. The video describes memories as analogous to “dioramas made from wax” that change a bit “each time they are under the spotlight of our attention”, and it brings up two interesting examples: how we change an episodic memory when we recount it; and how we may be able to enter and ideally change a memory through controlled spotlights of attention via psychotherapy. This is really a beautiful and intuitive picture that can easily be grasped without understanding the physiological and chemical details of synapses.
If you have any comments on my reading of this video, I’d be interested!
Dendrites are the tree-like arborizations of neurons through which they receive input from other neurons. This compartmentalized anatomy has given rise to the idea that individual dendrites process information independently from each other and from the soma. This idea has driven many studies on dendrite function during the last 30 years, mostly in brain slices. But do neurons and dendrites make use of these compartmentalized computations in the intact, living brain?
Recent advances in microscopy and voltage imaging have made it possible to address this question more directly than ever before. Specifially, there has been a recent streak of technically impressive papers on dendritic imaging in the hippocampus of awake mice. In these papers from the labs of Adam Cohen, Attila Losonczy, and Balázs Rózsa, hippocampal pyramidal neurons together with their dendrites are imaged in the behaving mouse, using both calcium and voltage imaging (Gonzalez et al., 2026; Lee et al., 2026; Moore et al., 2025; Wu et al., 2026). Let’s have a closer look!
Paper 1: Two-photon random-access voltage and calcium imaging of CA1 dendrites
Online motion correction for fast dendritic two-photon imaging
Recording both voltage and calcium in 3D is a “holy grail” for dendritic physiology. However, this requires scanning only the dendrites without their surrounding at high speed and repetition rate, making the recording susceptible to artifacts due to brain motion that make such data unusable. To fix this problem, the laser scan pattern needs to be corrected online with millisecond precision to compensate for brain motion. Here, the Rózsa and Losonczy labs developed a new approach that they termed 3D real-time motion correction (3D-RTMC).
Real-time motion correction is highly challenging, and a bright and stable object needs to be used as reference. Earlier work from the lab of Angus Silver injected bright beads for that (Griffiths et al., 2020). Here, the authors took advantage of the single-cell electroporation experts in the Losonczy lab (Gonzalez et al., 2025) and electroporated a cell close to the cell of interest with a fluorescence-expressing plasmid in order have a closeby bright object as a reference for online motion correction. This method allowed them to track the same dendritic branches in a head-fixed mouse running on a treadmill with sub-micron precision using simultaneous voltage and calcium imaging.
Decoupling of calcium vs. voltage with higher dendritic branch order
The authors found that as signals traveled further from the soma, voltage and calcium dynamics became more and more distinct. Proximal dendrite activity mirrored somatic activity, but distal branches showed a progressive decoupling, with branch order, not just distance, determining how well calcium reflected the underlying voltage. A limitation here is that they used the red-shifted jRGECO1a as a calcium indicator (which was necessary so it could be combined with the green voltage sensor ASAP6). jRGECO1a reports calcium changes in a nonlinear way; this nonlinearity is likely to result in missing small dendritic events (Rupprecht et al., 2025).
Local dendritic voltage signals
Interestingly, the authors describe an abundance of local calcium signals in dendrites independent of somatic activity (Figure 2C,K,L; Supplementary Figure 4). However, this important observation is only briefly described and not discussed in depth. For example, Figure 2L shows that dendrites and soma are very likely to be coactive; but if the dendrite is active, the soma is recruited (“s|d”) with a higher probability than the other way around (“d|s”). This is the opposite of what is reported in the paper by (Lee et al., 2026) discussed below and would be worth discussing. – For clarity, the Lee et al. paper came out after the Gonzalez et al. manuscript was published, so the authors had no way to discuss this finding.
Paper 2: Prism-based one-photon voltage imaging of CA1 dendrites
Prism-based voltage imaging of hippocampal dendrites using one-photon microscopy. Excerpt of Figure 1 from the preprint by Lee et al., under CC BY-NC-ND 4.0 license.
One-photon voltage imaging of individual CA1 neurons
The Cohen lab adapted a previously published design (Redman et al., 2022), using a side-on microprism and kHz-rate voltage imaging with DMD-based structured illumination to record from a large fraction of the entire dendritic tree of CA1 neurons (Lee et al., 2026). This manuscript provides to the best of my knowledge the most comprehensive functional voltage recording of a single neuron in a behaving animal to date.
I’m very impressed by how they combined high-end optics, genetic tools, and very complex data analysis for these experiments in behaving head-fixed mice. From a scientific perspective, the paper is very difficult to summarize because it covers several major questions of neurophysiology in an almost comprehensive manner. And I believe that the choice of the paper title could be improved because it highlights only one aspect of their work.
A note on data analysis
Sometimes, there are Methods sections that are instructive by being rich in interesting details and full of complex processing steps which correct for confounds that you would rarely think of in the beginning: this is one of them. One detail that I found a bit confusing is that they regressed out brain motion (derived from the motion correction algorithm) from the dendritic voltage signals. This is desribed at the bottom of the “Voltage signal extraction” section in Methods. Would this not artificially subtract any dendritic signals related to the movement of the animal? I am a bit skeptic about this methodological detail.
Anti-correlation of apical vs. basal dendrite activity
Lee et al. find that subthreshold voltage is primarily explained by global activity across the entire arbor (~55% of variance), followed by a pattern that they term “see-saw”, where apical and basal dendrites are anticorrelated (~20% of the variance; clearly visible in the movie below). I really like how they collect evidence that this see-saw pattern is most likely determined by laminar-specific inputs in CA1. They do so by comparing the spatial extent of spontaneous activity patterns (large extent) with spatial patterns generated by local dendritic optogenetic depolarization (smaller extent).
Evidence against prominent local voltage events in dendrites
Interestingly, Lee et al. do not find evidence for prominent local voltage events in dendritic compartments, in contrast to the data from the Losonczy/Rósza labs (Gonzalez et al., 2026); and only limited evidence for dendrite-specific place fields, in contrast to a previous calcium imaging study (Moore et al., 2025).
Instead, they find that most “spikes” seen in dendrites originate from back-propagating action potentials (bAPs) in the soma (90%). With this, they identify the soma and not the apical dendrite as primary site of origin for bAPs and complex spikes. In the end, of course, all somatic spikes are caused by the cumulative integration of dendritic inputs. However, this study suggests that the amplification of those signals is primarily a consequence of successful somatic firing and not of amplification in the dendrites, contrasting with existing theories focusing on dendrite-localized amplification (Larkum, 2013; Major et al., 2013; Poirazi and Papoutsi, 2020).
Backpropagation enhanced by local depolarization
The authors also find that apical depolarization enhances bAP propagation into the apical dendrite, which would make bAP propagation specific to sites of prior activity. This finding is not very surprising and confirms earlier work in slices; but it strengthens the idea of bAPs as a learning signal for local dendritic plasticity and is a very important piece of evidence.
Many more biophysical and physiological findings are in the paper, and every paragraph is interesting to read. Highly recommended!
Paper 3: Simultaneous calcium and voltage imaging across the dendritic tree
One-photon imaging of voltage and calcium in individual neurons
This sister paper from Adam Cohen’s lab also uses prism-based one-photon imaging, recording not only voltage but also calcium events of hippocampal pyramidal cells across a large fraction of the dendritic tree (Wu et al., 2026). To this end, they use a slightly different optical setup, with a spinning disk-based design. They focus on CA2 pyramidal neurons, not primarily for scientific reasons but because expression is much stronger in CA2 for most AAV serotypes (Alexander et al., 2024). Imaging calcium and voltage simultaneously is one of the most interesting experiments for researchers like myself who worked on interpreting calcium imaging data for a long time (Rupprecht et al., 2021, Rupprecht et al., 2025). Achieving such a dual recording across a large fraction of the dendritic tree definitely exceeded my expectations.
A biophysical rule to convert voltage to calcium
Wu et al. find that the mapping from voltage to calcium is remarkably predictable and follows a hierarchical activation pattern. To explain the data, they fit a simple sigmoidal model to the voltage-calcium transfer function. They find that the inflection point of this transfer function (which they term V0 in their paper) becomes larger for smaller dendritic and especially apical compartments. They argue that this might be a mechanism to avoid noisy dendritic calcium signals due to local depolarizations from synapses. The combination of acquiring such a complex dataset and modeling it in such simple and straightforward terms is quite beautiful.
Similarity to somatic activity decreases with distance from soma for calcium signals and less so for voltage signals. Excerpt from Figure 3of the preprint by Wu et al., under CC BY-NC-ND 4.0 license.
Further evidence against prominent local events in dendrites
In contrast to previous work based on two-photon imaging data (Gonzalez et al., 2026; Moore et al., 2025), Wu et al. find that localized “dendrite-only” spikes were surprisingly rare. One potential explanation they bring up is that they record from CA2 dendrites, which are know to be less susceptible to plasticity protocols and therefore might not need as prominent local dendritic signals to instruct plasticity as compared to e.g. CA1 neurons. I find this explanation interesting, but I still have the feeling that the dissociation between earlier findings for local events and this evidence against local events is an interesting point of discussion that still needs to be resolved.
Conclusion
First of all: all of these papers are definitely worth reading! I found in particular the two papers from the Cohen lab to be a very interesting resource with many thought-through analyses and in vivo tests of prior brain slice work. The most surprising finding and the largest discrepancy between this work based on one-photon imaging (Lee et al., 2026; Wu et al., 2026) and prior two-photon prism-based work (Gonzalez et al., 2026; Moore et al., 2025; and other work from the Losonczy lab) is the lack of prominent local dendritic activity observed by the work from the Cohen lab.
It is not clear what could explain this discrepancy. Did the approach to regress out brain motion from voltage imaging signals by Lee et al. suppress true local activity? Did the inserted prism cut off connections, resulting in lower activity levels? (This is actually addressed with a control experiment by Lee et al..) Or may two-photon dendritic recordings be more prone to movement artifacts (despite the huge effort to correct for those) and therefore exhibit prominent apparent local activity? It is known that for typical hippocampal windows, the strata radiatum and lacunosum-moleculare, where the apical dendrites reside, exhibit stronger motion artifacts than soma or basal dendrite; in addition, resolution and SNR are lower for apical recordings, all of this contributing to potentially more motion artifacts in apical dendrites. Other reasons that may explain the difference between findings may be the different methods to induce expression, using single-cell electroporation (which adds a lot of plasmids into a single cell and might lead to overexpression) by Gonzalez et al., and AAV-based expression by Wu et al. and Lee et al. (which may have side-effects due to the injected virus particles).
At this point in time, it seems not yet possible to figure out the technical and biological contributions to motion artifacts or the lack thereof in these studies. So I’m really looking forward to seeing these important experiments replicated and consolidated across other labs!
In any case, check out these papers! I barely scratched the surface, and all of them are definitely worth a read.
References
Alexander, G.M., He, B., Leikvoll, A., Jones, S., Wine, R., Kara, P., Martin, N., Dudek, S.M., 2024. Hippocampal CA2 neurons disproportionately express AAV-delivered genetic cargo. https://doi.org/10.1101/2024.11.27.625768
Gonzalez, K.C., Noguchi, A., Zakka, G., Yong, H.C., Terada, S., Szoboszlay, M., O’Hare, J., Negrean, A., Geiller, T., Polleux, F., Losonczy, A., 2025. Visually guided in vivo single-cell electroporation for monitoring and manipulating mammalian hippocampal neurons. Nat. Protoc. 20, 1468–1484. https://doi.org/10.1038/s41596-024-01099-4
Gonzalez, K.C., Terada, S., Noguchi, A., Zakka, G.N., O’Toole, C., Bilbao, G., Reynolds, L., Jász, A., Kertész, B., Szadai, Z., Shen, A., St-Pierre, F., Polleux, F., Losonczy, A., Rózsa, B., 2026. Movement-stabilized three-dimensional optical recordings of membrane potential changes and calcium dynamics in hippocampal CA1 dendrites. Neuron S0896627326000048. https://doi.org/10.1016/j.neuron.2026.01.004
Griffiths, V.A., Valera, A.M., Lau, J.Y., Roš, H., Younts, T.J., Marin, B., Baragli, C., Coyle, D., Evans, G.J., Konstantinou, G., Koimtzis, T., Nadella, K.M.N.S., Punde, S.A., Kirkby, P.A., Bianco, I.H., Silver, R.A., 2020. Real-time 3D movement correction for two-photon imaging in behaving animals. Nat. Methods 1–8. https://doi.org/10.1038/s41592-020-0851-7
Larkum, M., 2013. A cellular mechanism for cortical associations: an organizing principle for the cerebral cortex. Trends Neurosci. 36, 141–151. https://doi.org/10.1016/j.tins.2012.11.006
Lee, B.H., Park, P., Wu, X., Wong-Campos, J.D., Xu, J., Xiong, M., Qi, Y., Huang, Y.-C., Itkis, D.G., Plutkis, S.E., Lavis, L.D., Cohen, A.E., 2026. Fast dendritic excitations primarily mediate back-propagation in CA1 pyramidal neurons during behavior. https://doi.org/10.64898/2026.01.03.696606
Moore, J.J., Rashid, S.K., Bicker, E., Johnson, C.D., Codrington, N., Chklovskii, D.B., Basu, J., 2025. Sub-cellular population imaging tools reveal stable apical dendrites in hippocampal area CA3. Nat. Commun. 16, 1119. https://doi.org/10.1038/s41467-025-56289-9
Poirazi, P., Papoutsi, A., 2020. Illuminating dendritic function with computational models. Nat. Rev. Neurosci. 21, 303–321. https://doi.org/10.1038/s41583-020-0301-7
Redman, W.T., Wolcott, N.S., Montelisciani, L., Luna, G., Marks, T.D., Sit, K.K., Yu, C.-H., Smith, S., Goard, M.J., 2022. Long-term transverse imaging of the hippocampus with glass microperiscopes. eLife 11, e75391. https://doi.org/10.7554/eLife.75391
Rupprecht, P., Carta, S., Hoffmann, A., Echizen, M., Blot, A., Kwan, A.C., Dan, Y., Hofer, S.B., Kitamura, K., Helmchen, F., Friedrich, R.W., 2021. A database and deep learning toolbox for noise-optimized, generalized spike inference from calcium imaging. Nat. Neurosci. 24, 1324–1337. https://doi.org/10.1038/s41593-021-00895-5
Rupprecht, P., Rózsa, M., Fang, X., Svoboda, K., Helmchen, F., 2025. Spike inference from calcium imaging data acquired with GCaMP8 indicators. https://doi.org/10.1101/2025.03.03.641129
Wu, X., Lee, B.H., Park, P., Wong-Campos, J.D., Xu, J., Plutkis, S.E., Lavis, L.D., Cohen, A.E., 2026. A dendrite-resolved, in vivo transfer function from spike patterns to dendritic Ca2+. https://doi.org/10.64898/2026.01.18.700189
I’m glad to share a PyTorch-based implementation of spike inference from calcium imaging data: CascadeTorch, now available on GitHub.
The original Cascade repository remains fully supported. This post explains why I re-implemented Cascade in Torch, and what this means for users.
What is Cascade?
Cascade is a deep learning-based method for inferring neuronal spike rates from calcium imaging signals (ΔF/F traces). It uses convolutional neural networks trained on a large ground-truth database of simultaneous electrophysiology and calcium recordings. The original publication in 2021 was trained on a broad range of datasets; more recent versions were trained on mouse spinal cord data and, most recently, on variants of the new sensor GCaMP8.
Cascade was originally implemented using TensorFlow/Keras because I this framework showed the highest promise back in 2019/2020. Since then, Cascade has been widely used for spike inference across imaging conditions by many labs.
Why switch to PyTorch?
Since 2021, when we published Cascade, the machine learning community has largely shifted toward other frameworks such as PyTorch or JAX in recent years. An example of this shift in the neuroscience community is DeepLabCut, which recently transitioned from TensorFlow to PyTorch as its primary backend. Installing PyTorch and TensorFlow in parallel is sometimes tricky and can create dependency conflicts of package or GPU driver versions.
What is CascadeTorch?
CascadeTorch is a 1:1 reimplementation of Cascade in PyTorch. That means, the network architecture is identical to the original TensorFlow version, and original pretrained models are simply converted to PyTorch by copying the weights.
Therefore, if you run the same input ΔF/F trace through a TensorFlow model and its converted PyTorch equivalent, you obtain the same spike rate prediction.
Compatibility and support
The original TensorFlow-based Cascade will be fully supported also in the future, and pretrained models are available in both formats. However, if you are starting a new project, I recommend using CascadeTorch. The switch is very straightforward. Simon Ball, who performed some initial tests with CascadeTorch, reported that he’s “getting the same results, and with no changes other than pointing the module import to your forked repository”. Please check out CascadeTorch directly on GitHub:
During the lockdown in 2020, I was forced to stop my own experiments in the lab. This is when I turned to computational and analysis side-projects that since then have become quite central to my own work and approach to science. It was then that I teamed up with Anna Maria Reuss, then PhD student in the lab of Adriano Aguzzi, to work on the analysis of her massive 3D light-sheet imaging data from cleared brains.
Anna Maria developed a method that enables the use of large archival human tissue – non-perfused, and embedded in FFPE for more than ten years -, making it transparent, compatible with antibody labeling, and therefore suitable for imaging at large scale and micrometer resolution.
I’d like to take the opportunity to share this brief movie of a small excerpt of a cleared human brain (dopaminergic neurons in the substantia nigra) that we include in this most recent version of the manuscript to showcase how fine dendrites are resolved:
Dopaminergic neurons in the substantia nigra, a small excerpt from a human brain sample cleared with the aDISCO protocol and imaged with the mesoSPIM light-sheet microscope. Supplementary Movie 2 of Reuss et al., 2026, under CC BY-NC 4.0 license.
We used two-photon imaging of interneurons, astrocytes, and pyramidal cells in head-fixed mice, while optogenetically stimulating the primary source of noradrenaline, the locus coeruleus (LC). In addition, we used fiber photometry to calibrate our LC stimulation protocol and to compare optogenetically evoked noradrenaline release with that observed during behavioral paradigms.
For the main findings, please check out the manuscript itself. In this blog post, I’d like to highlight a few results that I personally find fascinating but that are less central in the paper.
1. Astrocytes are much more sensitive to noradrenaline than neurons
Based on our two-photon calcium imaging data, pyramidal neurons and interneurons show little or no response to weaker LC stimulation and respond robustly only to strong stimulation. Astrocytes, in contrast, are highly sensitive and respond even to the lowest detectable increases in LC activity. This positions astrocytes ideally to mediate noradrenergic effects in hippocampal circuits. I find it great to have this very direct comparison across all three cell types – pyramidal cells, interneurons, and astrocytes – using the same experimental approach for each cell type.
Astrocytes are activated both by strong (20 Hz) and intermediate (5 Hz) stimulation of locus coerueleus. A subset of interneurons is activated for strong stimulation (20 Hz), but activation is detectable for weaker stimulation (5 Hz). Adapted from Figure 2 of Duss et al., 2026, under CC BY-NC 4.0 license.
2. Interneurons responsive to noradrenaline are located close to the stratum radiatum
Previous studies have already described that specific functionally defined interneurons are located in specific laminae or sub-laminae of hippocampus (e.g., Geiller et al., 2020). Here, we found that interneurons responding to noradrenaline release upon optogenetic LC stimulation are preferentially located near the pyramidal cell layer, close to the stratum radiatum.
A subset of interneurons is activated by LC stimulation (black traces; Group 1). They are preferentially located in specific laminar positions (bottom right, black density distribution), close to the stratum radiatum. Adapted from Figure 5 of Duss et al., 2026, under CC BY-NC 4.0 license.
3. Not all astrocytes are equal
We found that some astrocytes are much more sensitive than others when responding to locus coeruleus activation. These differences were consistent across repeated stimulations and across days. This adds evidence, from a functional perspective, that astrocytes are not a homogeneous population, even within the same hippocampal region.
A functionally defined diversity of astrocytes. Some astrocytes respond at low noradrenaline levels (orange examples), others only at intermediate levels (red), again others only at high levels of noradrenaline release. Adapted from Figure 3 of Duss et al., 2026, under CC BY-NC 4.0 license.
For many more details on the cellular responses to locus coeruleus stimulation, and how these compare to response profiles during natural arousal, check out the preprint. Happy to hear thoughts or questions about the work!
How does the brain work, and how can we understand it? To approach this big question from a broad perspective, I want to report on some ideas about the brain that marked me most over the past twelve months and that, on the other hand, do not overlap too much with my own research focus. Enjoy the read! And check out previous year-end write-ups: 2018, 2019, 2020, 2021, 2022, 2023, 2024.
For this year’s wrap-up, I reflect on what it would mean to achieve experimental goals that currently seem out of reach.
This is part III, with part I and part II online already.
III. Simultaneously recording from all neurons of the human brain
The idea: Since the middle of the last century, the number of neurons recorded simultaneously in animal brains has steadily increased (Urai et al., 2022). Around the year 2000, it became clear that simultaneously recording multiple neurons provides insights into the brain that could never be obtained from single neurons alone. More than 20 years later, these population dynamics remain an active and important area of research. For this reason, recording from all neurons is often described as the ultimate thought experiment: first in a simple organism like C. elegans (Randi and Leifer, 2020), then in a vertebrate like the zebrafish larva (Ahrens et al., 2013), a mammal like the mouse, and eventually in humans, with their roughly 80 billion neurons.
Why it won’t work: At present, no method comes even close to achieving this goal. With fMRI, it is never entirely clear what signals are actually being measured. In addition, both the temporal resolution (approx. 1 s) and spatial resolution (approx. 1 mm) are far from what would be required (0.001-0.01 s temporally, <0.01 mm spatially).
Optical microscopy can record from thousands – and in extreme cases up to a million (Demas et al., 2021) – cells in animals, but this number comes at the cost of temporal resolution and yields only an indirect and noisy proxy of neuronal activity.
In contrast to optical cellular microscopy, electrophysiology is used in humans as well, but at most from some thousand cells – roughly 0.00000001% of all neurons. Moreover, every electrophysiological method destroys more neurons than it records, reflected in the Butcher number introduced by Markus Meister. This problem becomes worse the farther one moves away from the cortex, toward the hypothalamus or small nuclei in the deep brainstem, or to neurons hiding close to the walls of the ventricles. Every method is invasive in its own way, and none is truly scalable, despite claims to the contrary.
Perhaps, someday there will be methods by which neurons are programmed to send, in addition to their actual axon, a collateral branch to the brain surface, where the axon as the “key” finds a genetically programmed “lock” attached to an electrode that allows recording from this biological axon. With such a design, invasive methods would no longer be necessary because the signals would present themselves on a silver plate. For these 80 billion additional axons, one would merely need to find about half a liter of extra space inside the skull and develop readout electronics capable of listening to a single axon within an area of roughly 1 µm² and forwarding the signal to a recording device. Let us not look too closely into further, even less comfortable details, such as the cables dangling from the head and data storage: 80 billion × 1,000 Hz data points per second – this is roughly 80 terabytes of data per second, assuming 8 bit signals – would challenge even the most advanced streaming and storage technology.
And what would still be missing is, of course, the subcellular and subthreshold state of each neuron: its membrane potential across soma and dendritic structure, as well as plastic cellular and synaptic dynamics on short timescales (facilitation, adaptation) and long timescales (LTP, LTD, immediate early genes, etc.). At present, there is no even remotely scalable approach to collecting such data in any animal, not to speak of humans.
What we would learn: Suppose we had this dataset in front of us: 80 billion cells, recorded for 24 hours, with every action potential detected. Let us ignore the cerebellum for the moment, with its 60 billion neurons, leaving “only” 20 billion. Assuming an average firing rate of 0.1 Hz (the true rate, including interneurons, is likely somewhat higher), we would observe roughly 2 million action potentials every millisecond.
One of the most fascinating aspects would be to see how action potential patterns propagate and how they correspond to thoughts and sensory impressions. If, for example, the subject watched a movie during recording, or if he or she imagined navigating through a space, spoke words, or was suddenly startled, we could follow the activity patterns through the brain from one region to another as they spread, persist, and eventually fade. We could examine how stable these activation patterns are, how widespread they are, and whether they are distributed or localized. We could reconstruct reaction times to sensory stimuli across all brain regions.
We could focus on a specific brain region – let’s say, the arousal center of the locus coeruleus or the dopaminergic substantia nigra pars compacta – and quantify, for every neuron in the brain, how its activity is influenced by this specific neuromodulatory nucleus.
If the dataset was good enough, perturbations would no longer be necessary to study causality of synaptic interactions; observation alone could yield enormous insight.
What I find most appealing about this thought experiment is the ability to trace a thought as it moves through the brain. Would it resemble videos of bird flocks, seemingly chaotic yet tightly coordinated by local interactions? Or would it look more like completely distributed, spatially disconnected activity patterns? It is clear that such dynamics are at least partially chaotic, that is, that the dynamics will strongly depend on initial conditions (the state of the network before a repeatable event such as a sensory stimulus). If we could record all neuronal activity, we might even be able to predict such less organized and more chaotic dynamics and understand how chaotic and “critical” neuronal activity really is.
These are very general ideas, but once you think more concretely about already known phenomena of neuronal activity, many additional exciting questions emerge. How are sharp-wave ripples triggered, from cortex to thalamic or brainstem nuclei? Where does the activity of dopaminergic neurons, whose signals are still debated in terms of reward prediction errors versus movement signals, originate? A systematic screen of activity across all brain regions could clarify exactly such questions and put theories, often built on tiny slices of neuronal activity and behavior, to a serious test.
Overall, I don’t think that an understanding of the brain would suddenly appear once we have this dataset; we would not understand the nature of consciousness, and would not be able to recreate the brain; but many existing research fields would be pushed forward, and many existing questions of basic research would be answered immediately.
