Brain Function—What’s Next?
A Conversation with David Hildebrand, Harvard University
Aug. 17, 2017
The May 10, 2017, issue of the prestigious journal Nature featured an article by researchers at Harvard University and PSC that charted out the first “nano-scale” reconstruction of a vertebrate brain—that of the zebrafish larva. They used electron micrographs of ultra-thin brain slices, reconstructed in 3D thanks to the work of PSC’s Art Wetzel, to follow all the connections in a subset of neurons. (Read more about those findings here.) We took the opportunity to talk with the lead author of that paper, David G. C. Hildebrand, about what the work means in the larger field of brain function.
PSC: What is the next step from the Nature paper? If the connections in the brain are a “wiring diagram,” is the next step to put labels on the components? (To keep to the wiring metaphor, “this is a transistor, this is a capacitor”?)
Hildebrand: This is one possibility, but probably not a major next step. In general, as far as this analogy holds true, I think people tend to associate synapses with transistors and believe that other components are essentially a consequence of intrinsic properties of individual neurons (such as the ion channels they express) or fallout of sub-circuit motifs (such as activating a neuron that then inhibits the [neuron that activated it]). One big next step for the field is to figure out what motifs exist and what purpose they serve, the latter of which will require additional functional experiments. Some studies are already doing this. Wei-Chung Lee’s 2016 Nature paper, for example, looks at three-neuron connectivity motifs and their prevalence in their reconstructions. Mitya Chklovskii’s work and Andres Tolias’s work may contain more on this, too, but I haven’t read into it heavily enough. There are also in silico efforts constrained by experimental data that are trying to analyze what falls out of simulations. Most of the studies so far are biased by the size of the region they are sampling. Looking at this at the whole-brain scale would be very interesting, because one would also start to see motifs that are important in long-range communication between brain areas.
PSC: Is there further resolution work to figuring out the connections, or are the data already there? What percentage of the zebrafish larva brain do we have mapped at this point?
Hildebrand: Additional imaging would need to be done in order to identify all the connections (and follow all the wires) in this larva’s brain. The “data are there” only in that an “analog copy” exists in the tissue sections we have available for re-imaging. It is certainly possible to do, but the time would need to be taken to acquire additional images at higher resolution (with the beam scanning over the sample in smaller steps to finely resolve additional structures).
Note that one additional limitation of the main fish we presented in the paper is that it was sectioned into relatively thick partitions of ~60 nm. Ideally, one would want to “Nyquist section” the smallest structure of interest. Ultimately the goal would be for a plasma membrane to be the target structure here, but for physical sectioning they’re too thin, at something like 8 nm thickness (note that this sets the ~4 nm/pixel side target lateral resolution for “high resolution imaging” that most people use). For physical sectioning, most folks are targeting 30 nm-thick sections or less at this stage. Other approaches allow for thinner sectioning but there are other trade-offs.
A very small percentage of the larva’s axons and dendrites are mapped in [our] study. Very, very small. We expect to know more relatively soon on total numbers, but just to get a quick approximation: People have assumed in the past that there are ~100,000 neurons in the larval zebrafish at this age (we will know for sure for this fish soon). We reconstructed myelinated axons of ~830 of these. Total path lengths of the axons and dendrites for the different neurons will vary substantially, but this already points to a very tiny proportion. So, another “next step” is to get automatic mapping working. Lots of efforts are under way to do what is called “automatic image segmentation,” which basically subdivides each tissue section into different neuron labels. A group at Google Research has released an awesome improvement in this recently, so progress is definitely being made and investigators will definitely start using these approaches more in the near future. Not likely on this particular sample, but on new samples.
PSC: In the Nature paper, you shared an image that highlighted the neurons/axons underlying rheotaxis (orientation/swimming in response to a water current). Is rheotaxis going to be an early focus for correlating the structure you’ve found to behaviors?
Hildebrand: … there is one small section that draws on this possibility very briefly:
“The resulting projectome included 94 lateral line afferents that innervated 41 neuromasts (Fig. 3b). These reconstructions revealed striking bilateral symmetry in the lateral line system (Supplementary Video 8). Only one neuromast and its afferents lacked contralateral counterparts. This may be an important anatomical feature that facilitates comparisons of local velocity vector fields for detecting differential flow along the left and right sides, which is essential for rheotaxis in larval zebrafish …”
Another result in [our] lab that was just recently [published (see Nature, July 27, 2017, here)] looks at this behavior in detail, but not much at the anatomy underlying it. Further modeling of these lateral line afferent axons and their connectivity could definitely clarify whether or not their organization (and in particular apparent symmetry) is important for the behavior that is observed. We aren’t completely sure that we could answer the question with this dataset because the posterior neuromast sensory organs seem to be more important for rheotaxis, and those are outside the volume we captured.
In my view, there are a lot of parallels that can simultaneously start to be drawn between anatomy and behavior with this kind of dataset, but some of them just require looking at the circuitry and starting to analyze it more closely.
PSC: This is pretty big picture/future, but once we have a fully labeled wiring diagram of a brain in the computer, how far is that from modeling the brain’s activity? Is it basically a “blank slate” brain that just needs virtual learning to be “functional,” or is there more to it than that?
Hildebrand: Great question. In short, we don’t know. Different neuroscientists believe very differently about this. I’m not sure how I feel about it, but it is very clear that it is harder to model the complete system without a fully labeled wiring diagram. Once we have such a circuit diagram, we can start to model what is happening more precisely than just the connectivity permits, too. For example, we can impose in the model the particular properties of neurons in a certain area or those that have specific morphological features. To me, the pie-in-the-sky speculation of being able to simulate the entire organism is interesting, but not really the primary reason to do the work. The reason to do the work is that we learn so much about how the individual pieces—which we can study with other methods—can fit together.