Last summer, Stanford University researchers delivered patterns of light to a mouse’s brain that generated a hallucination intense enough to influence the animal’s behavior. The experiment was possible due to a recent technology, known as sculpted light microscopy or all-optical interrogation of neural circuits, that uses advances in optics and optogenetics to simultaneously stimulate and record from dozens of neurons with near single-cell precision.
The unprecedented level of control offered by this new approach makes researchers well positioned to test theories about how the brain is wired and how it drives behavior. If you can turn groups of cells on and off in specific patterns, you can explore exactly what roles those cells play in neural computation. But significant challenges remain — with millions to billions of neurons in the mammalian brain organized into exquisitely complex circuits, scientists don’t yet know which types of neurons to target or the ideal sequences in which to stimulate them. As a result, most sculpted light experiments to date have targeted neurons through trial and error, and with little regard to timing. Without a way to make these experiments more efficient, the technique may fail to reach its full potential.
“We’re at the point of being able to do the kind of experiments that people have dreamed of,” says Chris Harvey, a neuroscientist at Harvard University. “Now, we have to do the hard work of trying to figure out exactly how to design these experiments to get the most out of them.”
For more than a decade, scientists have been using optogenetics to control neural activity with the flick of a light switch. An enormous amount of work has gone into discovering, and in some cases molecularly engineering, light-sensitive proteins that can excite or inhibit neurons with finer and finer temporal precision. Advances in optics over the past few years are adding another level of control. In traditional optogenetics, scientists flood the brain with light to activate wide swaths of neurons. With newer, more targeted lasers and opsins that are expressed only in the cell body, scientists are very close to being able to target individual neurons with precise, three-dimensional patterns of light from superficial to deep layers of cortex at kilohertz speed, many times faster than they can at present.
The ability to precisely control neural activity in this way finally makes it possible to rigorously test theoretical models of neural activity. “Just recording activity by itself is typically insufficient to test any of these models,” says Kayvon Daie, a postdoctoral researcher at the Howard Hughes Medical Institute’s Janelia Research Campus. “With sculpted light techniques, you can actually design neural perturbations that can best discriminate between two competing network models.”
For example, Harvey, along with his former graduate student Selmaan Chettih, recently published a study in which they stimulated individual neurons in layer 2/3 of the primary visual cortex to determine how they process stimulus-related inputs. Some previous findings have predicted, based on measures of connectivity, that layer 2/3 neurons amplify visual features — the activity in one neuron preferentially enhances the activity of similarly tuned neurons — in order to increase the chance that the brain will respond to a visual stimulus. Other experiments and theoretical work suggest that an opposing mechanism, feature competition, is important for suppressing redundant information and ensuring that fewer resources are needed to encode a stimulus. Using sculpted light to activate specific subsets of neurons, Harvey and Chettih showed that feature competition was much more prevalent than amplification in layer 2/3.
In a study posted to bioRxiv.org in April, Daie and collaborators at Janelia and Stanford used sculpted light to explore a different computation — memory. They wanted to figure out what sorts of networks represent short-term memory in the anterior-lateral motor (ALM) cortex. The researchers first identified neurons that were active during the memory phase of a task and then photostimulated groups of neurons randomly chosen from this set. Targeting certain neurons produced persistent changes in neural activity that far outlasted the photostimulation and influenced the animal’s behavior — properties that are consistent with working memory. Surprisingly, the researchers found that even within seemingly similar functional classes of neurons, ALM contains multiple groups, or “modules,” of recurrently coupled neurons that can be excited independently and maintain persistent activity during a delayed memory task.
A central challenge in using sculpted light microscopy to better understand neural circuits, including the memory-related circuits Daie and his colleagues discovered, is the small number of hypotheses that can be tested in each behavioral session. Reliable statistics can require about 40 to 50 trials, and animals can only do about 200 to 300 trials in a session before they tire out. “You very quickly find yourself in a position where you need to make really hard choices about what you intend to do,” says Shaul Druckmann, a computational neuroscientist and SCGB investigator and co-author on the study. “Unlike with most other kinds of experiments where you collect a dataset and then try to relate it to a theory, here theory comes in much earlier. The problem is that theories tend to be pretty abstract and broad, but the decisions you need to make about what neurons to stimulate, and how many, are very granular, and they can be pretty important.”
Druckmann and others say that bridging that gap — identifying the ideal neurons to stimulate to create meaningful perturbations — will require theories that incorporate the enormous functional diversity of neurons. For example, some cell types are much more excitable than others and require far less optical stimulation to elicit a desired level of activation. Cell types also vary in their selectivity. “There is this potentiating effect: A downstream neuron is more effectively activated if its selectivity is more similar to the neuron you’re photostimulating,” says Karel Svoboda, a neuroscientist at Janelia and an SCGB investigator and co-author on the study.
There is also emerging evidence that different cell types — distinguished by morphology, the numbers and types of neurotransmitters and receptors they express, and other factors — have different connectivity patterns. “If you ignore this heterogeneity, it will be at your peril because you won’t actually be doing the exact experiment you want,” says Hillel Adesnik, a neuroscientist at the University of California, Berkeley.
Researchers are further subdividing cell types, such as excitatory and inhibitory neurons, using RNA sequencing and other techniques. And toolmakers are developing opsins that would specifically stimulate functionally unique neural subtypes. “That’s beyond the reach of current technology, but I think it’s a goal that people are considering,” says Daie.
However, that still leaves the challenge of folding cell-type specificity into neural connectivity models. “We know that there’s value in adding cell types to models,” says Kanaka Rajan, a theoretical neuroscientist at the Icahn School of Medicine at Mount Sinai. “But to be honest, as theorists, we are only just scratching the surface of the full implications of incorporating cell types in our models and mathematical theories.”
Sequence and Timing
Assuming experimentalists figure out a more efficient way to identify the best neurons to stimulate to drive neural circuits, questions still remain regarding the optimal timing or sequence of stimulation needed to prompt particular behaviors or percepts. Sculpted light studies to date have generally stimulated neurons at the same time, but exploring parameters of sequence and timing — the order in which the neurons are stimulated and the rate (and timing) of action-potential induction — can unlock one of the holy grails of neuroscience: understanding the neural code.
More precise control over timing would also bring optogenetics closer to treating psychiatric conditions — one of the original goals of its co-inventor, Karl Deisseroth. In a 2019 Cell paper, Deisseroth, along with Rajan and others, used brain-wide imaging in zebrafish to observe that a passive coping behavior (with parallels to depression in humans) is driven by the encoding of stress via the progressive recruitment of habenular neurons. In a future experiment, the researchers plan to determine if optogenetics using sculpted light can be used to slow this recruitment, possibly staving off the stress-driven behavior.
So far, the number of neurons that can be targeted with the sculpted light approach has maxed out at a few dozen, though Adesnik says that his lab has developed a technique to increase this capacity fivefold. Michael Häusser, a neuroscientist at University College London and an investigator with the SCGB, says the next evolution of this technology will be the ability to target thousands of neurons in physiologically relevant patterns. Many scientists believe that this increased capacity for stimulation is necessary to unambiguously drive behavior, but stimulating too many neurons can muddy the waters. “The patterns of activity in the brain are like ripples in a pool,” Druckmann says. “You can drop a rock into the pool, but if you don’t take into account how the ripples that you’re generating interact with the existing ripples you are likely to miss or misattribute some really strong effects.” The solution, he says, is to strike a balance between stimulating enough neurons to generate an interesting behavioral effect and leaving an opportunity to clearly observe the rest of the network.
Druckmann believes that quality of stimulation, not just quantity, will be important in the next stage of sculpted light microscopy. Currently, it is difficult to control the expression of optogenetic proteins; neurons that express a lot of the protein are easily stimulated by a given amount of light, while neurons with low expression may not be stimulated at all. Moreover, it’s often infeasible to excite some neurons and inhibit others. But the types of manipulation needed to discriminate between computational models might require exactly that. “When we can very precisely perturb neurons, we can really have faith that we are nudging population activity in the particular way that is required to test or develop more refined models,” Druckmann says.
Researchers also need to address other technological hurdles. First, photostimulating a neuron requires moving the laser around a bit, usually with a spiral beam, to hit multiple parts of the neuron. As a result, cells that are not directly targeted by the laser can be activated. Adesnik says sculpted light experiments typically “activate as many as two off-targets for every one actual target.” One potential solution is ‘temporal focusing,’ whereby the laser is decomposed into separate colors, so that the different wavelengths propagate along separate light paths and only come into focus in a narrow 2D plane, minimizing off-target stimulation.
The other major concern is overlap in the excitation spectra for the opsin and the calcium sensor: The wavelengths used to excite the opsin and the calcium sensor peak at 1040 nm and 920 nm, respectively, but can be a few hundred nanometers wide. For experiments combining optogenetics and multiphoton imaging of individual neurons, this overlap has been relatively inconsequential. But multiphoton imaging of a population of neurons means that the 920 nm laser is scanned over a large field of view, which, because of the spectral overlap, will likely result in a considerable amount of unintended optogenetic stimulation.
Adesnik sees a few ways to get around this complication. His lab uses fast opsins, which are less sensitive to calcium imaging-induced photo-potentials. Another alternative, employed in a 2018 Cell Reports study led by the neuroscientist Tommaso Fellin of the Italian Institute of Technology, is to use blue-shifted opsins and red-shifted calcium sensors, an approach that almost completely eliminates spectral overlap. However, high-energy optogenetic lasers only work at 1040 nm, meaning blue-shifted opsins only work with low-power lasers, limiting the number of neurons that can be activated simultaneously. Finally, Adesnik says that in the future, calcium imaging could be replaced by voltage imaging, which could theoretically be performed at wavelengths near 635 nm using one-photon excitation, thereby preventing crosstalk altogether. The problem there is that red-shifted voltage sensors for two-photon imaging are not yet available.
“Optogenetics, CRISPR, these sorts of technologies became hugely successful not only because of their power, but because of their ease of use,” Adesnik says. “[Sculpted light microscopy] takes optogenetics to the next level. But it’s not plug and play; you can’t put a gene in, turn an LED on, and boom, you have the behavior you want. It’s going to require quite a lot more effort.”