Hopefully as you see in a bit more depth how the paper unfolds, it will become more apparent that Chiang and friends make few real findings or even predictions. However, aside from the transparency and respect for freedom of knowledge that I mentioned in the Meat and Potatoes, the strongest analytical value of the paper is the predictive power of the “FlyCircuit” database they are introducing.
As I mentioned earlier, the FlyCircuit project took 3D images of single neurons and placed them as accurately as possible into the space of an average brain (they ran parallel male and female projects, with many fewer male than female neurons, so most of the paper deals with the female model). The average brains they constructed took into account the anatomically segregated neuropils of the Drosophila brain. The technique the authors used for labelling small groups of non-overlapping neurons or individual ones was predictably MARCM, which you can (eventually) read more about in the glossary. Then, through a method beyond my understanding and apparently beyond the scope of the paper, individual neurons were transformed into the space of the appropriate template brain, using a global coordinate system with its origin at the centre of the ellipsoid body. They did this for ~16,000 neurons, covering more than a tenth of the female drosophila brain’s neurons.
To check the accuracy of the registration of the neurons inside the template brain the group compared the size and a alignment of neuropil regions between their model brain and individual sample brains, finding 93% accuracy. However, when checking the registration of individual neurons – the parts that actually matter – accuracy was lower. In the one example they give of this, looking at intrahemispheric projection of specific neurons located in the antennal lobe, the accuracy was 69%. They could increas this to 95% accuracy, but only when they changed from their global registration algorithm to a local one that only considered the brain region containing that neuron’s cell body. They aren’t really clear on this, but I think that to construct a model of the entire brain they need to use the global algorithm, meaning that the overall accuracy of neuronal registration in their model might be as low as 69% or lower. Chiang and friends could have presented better summary data by giving the neuronal registration accuracy for the entire brain, or even better, for each neuropil pair, but they don’t try to make any firm conclusions, so they save themselves from major criticism. In the future though, if they want to make stronger assertions about how the brain is organized, registration accuracy will need to be documented more appropriately.
After describing their system, Chiang and friends define a “Local Processing Unit(LPU),” which seems to be what FlyCircuit is most appropriate for detecting. According to them, an LPU is “equal or smaller than the boundary of an anatomically demarcated neuropil” and has a segregated set of local neurons whose processes are confined to the the LPU. Later though, after applying a search algorithm to identify LPUs throughout the brain and offering the obscure tid bit that the ventrolateral protocerebrum may actually contain 2 LPUs, they offer 3 criteria for defining an LPU: 1.) An LPU has its own set of interneurons whose cell bodies are restricted to the LPU; 2.) nerve bundles projecting out of an LPU are segregated; and 3.) each LPU has its own long range tracts connecting it to and receiving input from distant partners. Then, directly following this list, they claim that, therefore, an LPU has its own ensemble of interneurons whose processes (and not only cell bodies as they mentioned before) are restricted to that area and each LPU receives input from at least one neural tract. This evolution of the definition of an LPU in the confines of a single paragraph is not explained, nor is it cited. Essentially they start off by defining an LPU with their search algorithm, but then it looks like once that algorithm pulls out a bunch of putative LPUs, they make a new definition. I am left wondering whether their LPU search algorithm creates the definition, or whether it is the traits of the discovered LPUs that came to define their concept of the LPU. So the major question I would like to ask the the group is, what the hell is an LPU? It seems to me that if they have an LPU search algorithm they must have a concrete set of criteria, or at least a set of very convoluted if-then terms, and a single definition shouldn’t be so hard to come up with.
Regardless of what the criteria are and what the algorithms look like, they applied them to the +16,000 female neurons they had squeezed into their model brain. They found 58 anatomically segregated brain regions, composed of 19 paired LPUs and 3 central, unpaired LPUs. The VLP was the only region that contained two separate LPUs, so maybe keep an eye out for a follow-up study looking into this. Four pairs of neuropils that couldn’t be defined as LPUs seem to be interregional tracts or communication hubs. I got a little hung up on this last part. That Chiang and friends’ definition of an LPU neglects hubs might suggest that there is no processing done at a communication hub, but I think a travelling signal could be modified drastically at a hub where information is being transmitted 1 to 1, converged or diverged – potentially very powerful forms of computation in their own right. Thus an LPU shouldn’t be seen as the be all and end all of info processing.
After sorting all their neurons into groups (LPUs or not) the group used graph theory (which I will neither attempt to explain or understand on this pass) to assess the extent of interconnectivity between the LPUs. From this analysis they determine that there are 4 segregated sets of densely connected LPUs, each of which tends to use different neurotransmitters. The specific implication of that tendency is lost on me, though. However, one interesting role for this wiring diagram is its ability to predict informaiton flow in the brain. Chiang and friends give a single example of a potential higher-level processing center based on this idea, but take it no further than that. This however may be the most important role for FlyCircuit to play in the future; it’s odd to think about, but we really know much more about the function of human brain areas than about flies’. Lesion studies, and now fMRI studies in humans have allowed us to peg a lot of our convoluted mess of a cortex with at least one, and more likely too many, functions. On the other hand, flies, which have taught us so much about genetics, molecular biology and development don’t yet have a functional brain map. FlyCircuit has hopefully opened the door for that.
Finally the group assessed the extent to which their model covers the volume of the brain, taking this to be a measure of completeness of their model. They find that 93% of voxels have at least 3 neurons intersecting them, while only 0.79% of voxels are without neurons. I could not, however, find any mention of the volume of their voxels, which is an important piece of info: the larger the voxel, the higher the measurement of completeness is going to be. Regardless, they estimate that “the basic pattern of the brain’s connectivity is revealed with fewer that 5,000 neurons,” and they note that their 16,000 probably does a pretty good job. Despite this they offer that “further mapping of the remaining 90% of the brain is continuing,” so presumably only more realistic pictures are to come.
So, in the end Chiang and friends have amassed images of more than 16,000 individual neurons, fit them into a set of brain regions that they defined as LPUs or otherwise, and assessed the connectivity of those regions. The paper is a testament to a long and arduous journey. In fact, the NY Times quoted Chiang as saying “Painfully, we had to throw all the old data away,” when they developed a better way to process the raw data partway through the project. With that in mind you can’t help but be impressed with the work. However, all that work has yet to produce any concrete results. Chiang and friends do find that the VLP is the only anatomical brain region that contains 2 distinct LPUs, but the implications of this are currently unknown. They also show that there are four main groups of LPUs, and although this is potentially very interesting, it is also left untouched. Most importantly, the authors give a representative example of the ability of their database and algorithms to predict the direction of information flow in the fly brain. As Chiang and friends say themselves, the great strength of the system lies in the predictions it can make. So the paper’s worth shouldn’t be gauged by it’s tantalizing glimpses of real conclusions, but in its heralding of a powerful tool. Aside from their transparency and making the as of yet unexploited fruit of their labor available to drosophilists everywhere, the strongest feature of this project is its future.