Bock and friends have produced a formidable paper, pioneering the new combo of functional calcium imaging and serial recostruction electron microscopy, a technique that will undoubtedly revolutionize the way we study neural circuitry. Numerous sources have called their work (and that in Briggman and friends’ companion paper) a “technical tour de force.” That being said, their findings are rather weak, with low sample sizes leading to a flimsy conclusion. After getting through the somewhat roundabout logic of the experiments, I found myself wanting something more (specifically in the way of n’s) from such a high profile project.
Generating the Data
To generate their functional data, Bock and friends opened a live mouse’s skull and monitored the activity of a single plain of layer 2/3 pyramidal cells in the primary mouse visual cortex with a fluorescent calcium indicator. The specific information they extracted from this process was the orientation selectivity of the pyramidal cells they monitored. To do this they presented the mouse with visual stimuli in the form of lines moving across their field of view in a specific direction and watched for increases in calcium in the cell. Increases in calcium-dependent fluorescent indicated increased activity, so when the experimenters saw increased fluorescence in a cell in response to a line of specific orientation, they knew this was the preferred stimulus for that cell. Once they amassed their functional data, which contained info on which cells were selective for which orientation, the mouse given a substance that died its blood vessels fluorescent red and was then killed.
At this point the experimenters removed the area of the visual cortex that contained their functionally imaged pyramidal neurons, sliced it up into 1215 sections (~40 nm thick each – that’s 4 millionths of a centimeter) with a machine called a cryostat microtome, which is essentially a glorified, very accurate, difficult-to-operate deli-meat slicer. Then, by matching their fluorescent blood vessels between their original functional imaging session with those in their excised chunk of brain (now in 1200 thin slices) they found the population of cells they had functional data for. Next, they performed transmission electron microscopy on those 1200 slices of brain, a bout of microscopy that took them “several months,” and generated 36 terabytes of data. The authors have kindly made their data set freely available for download at the cleverly named Cell Centered Database, and if you look here you can have an online look at what their slices looked like. (If you scroll around these images you get an idea of the amazing amount of information this group gathered. If you’d like to learn how to analyze them yourself, Dr. Bock pointed me to Kristen Harris’ Synapse Web)
Although the authors have informed me in the comments section that these electron micrographs are not all that hard to figure out, for many of us, they initially appear unintelligible. And there were 3.2 million of them. So Bock and friends enlisted a small army of students for help: A. Arons, J. DiMartino, T. Law, M. Pershan, S. Reid, and V. Swanti. These were presumably undergrads – often the workhorses that fuel these ridiculously tedious projects, and who often get no recognition. Thankfully though, these students were mentioned briefly in the supplementary material that most people don’t bother to look at. The students looked through the pertinent images, manually tracing axons and dendrites, keeping track of synapses, until all the cells located in the volume of interest were found and their neurites classified. The result is an exhaustive connectivity map – microconnectome(?) – of the 1500 neurons whose cell bodies resided in the volume, plus axons and dendrites projecting into the volume from outside of it. But! Only 14 of these neurons had been functionally characterized with the calcium indicator experiment that I mentioned earlier.
Examining the Data
Of those 14 functionally characterized neurons, 13 were selective for lines moving in specific directions, one of them was not. Judging by the morphology of that lonely, unselective neuron, and by examining the synapses it made, Bock and friends determined it was an inhibitory interneuron, while the other 13 were excitatory pyramidal neurons.
Ten of the 14 functionally characterized neurons synapsed with other dendrites within the imaged volume. They focused on these 10 neurons from this point onward, and by assessing the characteristics of the rest of the neurons in the volume (the ones that they had no functional info for) Bock and friends found that, of the 245 synapses they made inside of their volume, 51% of them were onto inhibitory cells and 49% were onto excitatory cells. As for the cells that received those synapses, 38% were inhibitory and 62% were excitiatory. That means a good majority of synapses were onto inhibitory neurons, suggesting that local excitatory output in the visual cortex is generally more extensive onto inhibitory cells. This finding falls into line with the theory that inhibitory neurons are important for pooling excitatory input. The idea is that the pooled info can then feed back onto the excitatory cells to prevent runaway excitation and seizures and/or to refine the excitatory signal.
Bock and friends actually try to address this issue more directly, leading them to their big quasi-finding. One major question about this mass convergence of excitatory input into inhibitory neurons is whether or not it confers orientation selectivity onto the inhibitory interneurons. The thinking is that if inhibitory neurons pool excitatory input regardless of the orientation selectivity of the presynaptic excitatory neurons, they probably aren’t selective. On the other hand, if inhibitory neurons pool excitatory information only for specific orientations, they would be orientation selective. Over the past 3 years or so, a number of groups have come up with contradictory findings pertaining to this argument, some finding that input into the inhibitory interneurons is selective and some not. However, none of these groups have used techniques that successfully combine functional data and very precise 3D mapping of networks the way that Bock and friends have done with their chunk of tissue. With that in mind Bock and friends undertook to solve the conundrum. They focused on a subcircuit of their network, consisting of the 10 functionally characterized excitiatory neurons that made synapses inside their volume and the postsynaptic targets that received input from more than one of these 10 excitatory neurons (recall, we know the orientation of these excitatory neurons). Bock and friends claim to have “found multiple examples of pyramidal cells with diverse preferred stimulus orientations that provided convergent input to inhibitory neurons.” But, when I look at their conveniently schematized network, I see their 10 fucntionally characterized excitatory neurons converging onto a total of only 14 inhibitory neurons. By my count, 10 of these inhibitory neurons receive the majority of their inputs from pyramidal cells of the same preferred stimulus orientation – hardly the diverse preferred stimulus orientations that Bock and friends claim. In fact, 13 of the inhibitory neurons receive at least half of their input from pyramidal cells of the same preferred stimulus orientation. In other words only 1 of their 14 inhibitory neurons pool inputs from pyramidal cells of diverse preferred stimulus. Thus, I would argue that if you were to take this data as a measure of the distribution of the preferred orientation of all the inputs into inhibitory interneurons, it suggests that these interneurons do indeed pool input of specific orientation selectivity. However, you should by no means take this data at face value because the inhibitory interneurons in question only receive between 2 and 9 inputs from the functionally characterized pyramidal cells. So, considering that these interneurons probably receive 1000s of excitatory inputs, I will concede that my analysis means next to nothing… as long as Bock and friends do the same. (In their defence they don’t make any formal conclusions from this data other than that “they found multiple examples or diverse preferred stimulus orientations” providing input the interneurons. Also, they do admit in the discussion that their sample size is pretty low, and the “News and Views” opinion piece on the paper thoroughly acknowledges this shortcoming.)
Regardless of their low sample size, they try to extract something meaningful. Again, looking at the small network of their functionally characterized excitatory pyramidal cells and their postsynapstic targets that received input from more than one of them, they ask what predicts whether a pair of excitatory neurons will converge onto a common interneuronal target. Coming up with the strange metric that they refer to as “cumulative synaptic proximity,” Bock and friends found that the closer two cell bodies are to one another, and the closer their synapses are to one another, the more likely they are to converge onto an inhibitory cell. This seems like it makes perfect sense; cells that are closer are simply more likely to run into the same things when putting out axons and making synaptic contacts. Well, yes – but only if there aren’t any other rules governing where synapses are formed. So essentially what Bock and friends are trying to say here, although never explicitly, is that, since what determines convergence of two pyramidal cells onto a common target is the most basic, random chance choice, there apparently isn’t an overriding rule that makes pyramidal cells of similar orientation selectivity converge. So they are using this as roundabout evidence that inhibitory cells don’t pool input from excitatory neurons with a specific orientation selectivity. It would be much more straight forward to actually test the question “Do inhibitory interneurons pool orientation selective input,” by looking at the inputs and their preferred stimulus, and I’m sure they would have done that had their sample size been big enough. Despite the shortcomings of this approach, Bock and friends make the tentative conclusion that inhibitory interneurons in the primary visual cortex do not pool excitatory inputs of a specific preferred orientation.
As you can tell, the sample size used to make the tentative conclusions in this study is my main concern. The data supporting the absence of direction selective input onto inhibitory cells relies on data from only 10 functionally characterized neurons, “with some inhibitory targets recieving input from three or four distinct cells with a range of orientations” (although like I said, in most cases the majority of inputs are of a specific preferred orientation). They present this as “striking” data, but it means nothing because inhibitory neurons receive 1000s of inputs, so while they see “three or four” inputs of varying selectivity, there may be an unidentified overall bias in the total population of synapses onto a single cell for one orientation or a small range of orientations. Bock and friends reflect the weakness of their findings in their discussion, mainly blowing the horn on the techniques they pioneered and new advances that will make it easier to scale up the size of the networks reconstructed by similar approaches. They emphasize that while we are stuck looking at populations of ~14 cells we aren’t going to learn much – as evidenced by their work. They tout improvements like semiautomated tracing (reducing dependence on cumbersome undergrads) and serial block face electron microscopy as tools to usher in the new age of functional connectomics. To hit home the significance of the technical improvements to come, they note that “the number of interconnections between [neurons] increases as a square of the sampling density” meaning that small gains in the amount of data acquired results in disproportionate increases in the amount of connections to study. Good news for the future, but its a shame they didn’t manage to acquire just a bit more data this time around.