In a landmark paper demonstrating the capabilities of modern connectomics approaches, Moritz Helmstaedter et al. present a reconstruction of nearly 1000 cells and 500,000 cell-cell contacts in the mouse Retina using data collected with serial block face microscopy. This is a high-throughput electron microscopy method championed by Moritz and two of the other co-authors, Kevin Briggman (now at NIH in Bethesda) and of course Winfried Denk (a Max Planck Director with joint appointment at Janelia Farm). Valuable in its own right, this paper represents in part a promise of more and bigger things to come as the automation of data collection and segmentation continue to improve. Among the accomplishments in the paper: identification of a new type of bipolar cell type (XBC) based on its connectivity… an impressive finding considering a large amount of effort has been invested in categorizing retinal cells into types using a variety of other methods.
One of their achievements is hidden in, of all places, the acknowledgements. In pursuit of this project they were able to recruit the help of over 200 undergrads. I can only imagine the logistical and managerial issues involved with coordinating this. As I understand it though, it was a relatively well-oiled machine and played a big part in making this work possible.
Also missing from the paper is the hidden story of technology development, both software and hardware, upon which this paper relies. The challenges of producing these data are difficult on all fronts and were made possible by not being afraid to push some boundaries. Some of this story can be recovered from previous review and method papers by the same authors. Several of these authors likely bet their careers on the success of this unconventional approach long before it was a certainty. I’d say they are winning the bet.
I found one bit relatively hidden in the paper terribly interesting…. somewhere around 10% of the cells in the system were “orphans”… cells for which no cell type could be assigned. Maybe I am spoiled being a nematode guy, where every cell has a very specific identity, but this number seemed to me very high. What are these cells doing? Do we know anything about them? I’m not a retina guy, so maybe someone else has some answers… put it in the comments if you do!
Glancing at the connectivity matrix, cell-cell contact with glial cells is pretty extensive. In fact, they aren’t really visibly different in their “connections” when compared to neurons. Much of it seems to represent real synaptic input (figure 4D). Another bit of information that is missing is the distribution of electrical connections. These can be more challenging to identify and annotate (they were not in my connectome paper for this reason). I guess in an ideal world you always want more than can be delivered.
There was a nice discussion in the paper where they tried to estimate how much error there was in their estimations of connectivity.. Their definition of a synapse was a region where two cells came very close in proximity to each other.. a relatively liberal definition but necessary for their data for two reasons: 1) most of their data is processed in such a way as to ease the automated identification of cell contacts, but this blows out the synaptic vesicles making it impossible to used other criteria… having seen the data, it is not data that I would have been very happy to be working with. 2) identifying other synaptic features automatically is a bit more challenging (though possible). I imagine that with improved staining techniques their future work will be more richly annotated. See for example a micrograph with better staining in figure 3D.. the vesicles are very clearly preserved and stained in a nice way, and contains much more information than the micrograph in 3A which is what most of the data for this study looks like. They did not very extensively examine error rates.. only examining 20 connections as I understood it.. In these 20 connections they found 1 false positive, which means there might have been as many as 25,000 false positives. False negatives were almost certainly extremely rare. The false positives shouldn’t have had too much influence over interpretation of the data, however, because most connection types (neuron A presynaptic to neuron B), particularly important ones, are repeated hundreds of times.
Should we be concerned that they only have contact areas and not a more rigorous definition of a connection(i.e. presence of vesicles and synaptic densities)? They address this in figure 3E (and in other spots of the paper), where they have a plot of 93 contact areas and a histogram showing how many were synapses vs. incidental connections for difference sizes of cell-cell contacts. Synapses are pretty highly overestimated for smaller contacts, but large contacts nearly always were associated with a synapse. Even with these limitations, our understanding is much greater than it was before. The fact that we can address the data so critically I think is a huge victory for connectomics, not a drawback, because we have enough detail in our descriptions to understand our sources of error better than for light microscopy based techniques. We are simply close enough to ground truth to make it a point of conversation, and future papers will almost certainly improve data quality based on our understanding of where errors are occurring. Remember, other methods to data have been missing an entire neuron cell type.
Some amount of effort in this paper was placed on assigning cells to types based on anatomy and connectivity. These are essentially questions of homology, which I have taken quite seriously in all of my work. Read their “type-classification” section in the methods and you will see that in the details this is not always an easy problem. Particularly because of the need here to have the classification scheme mesh well with existing published schemes based on other methods overlapping and complementary information. I think the average reader would find many of their classification criteria either arbitrary or indecipherable. For example:
“OFF bipolar cells were sorted using the seventy-fifth percentile of the cumulative skeleton density in depth (starting from 0%) then, the lower 58.2% (their prevalence; see Table 1 in ref. 28) of cells were placed in the CBC3A/3B/4 and the remainder into the CBC1/2 category. The former was then sorted by the twenty-fifth percentile. Because this distribution was not clearly separable (consistent with the CBC4 width being smaller and more variable than drawn previously28), we began to collect the CBC4 cells starting at the highest twenty- fifth percentile numbers, adding cells consistent with the mosaic until the required prevalence was reached. The same procedure, now using the axonal coverage area, was used to separate CBC3A from B, reported to be larger CBC3A (ref. 28), and CBC1/2 using the spread in depth of the axon (twentieth to eightieth percentile). Finally, all mosaics were inspected again, six cells were reassigned, and one cell (cell 927, Supplementary Data 6) was moved to the ‘orphan’ group as it did not fit into any of the mosaics. In the resulting grouping, types 60–62 show a ramification-free zone adjacent to the INL that is lacking in types 58 and 59. Type 59 dendrites, if anything, are closer to the INL than type 58 dendrites. Type 62 ramifies slightly more widely in depth than types 60 and 61. Type 61 tends to be smaller than type 60.”
Some of the complicated language no doubt results from trying to express a complex problem in a highly restrictive publication format (Nature Magazine..). It is somehow surprising to me that they didn’t do more to use the connectivity in some more sophisticated way to more objectively define what might be cell types. It is hard for me to tell if someone else might have classified things differently.. and this classification, as a statement of homology, is the bread upon which the rest of the butter in the paper is placed.
The analysis of the connectivity data itself was not sophisticated (though the generation of the data was). Though they have information for individual cells, they focus mostly on a matrix that is collapsed into cell types. To me, figures like 6f are problematic in that they generally show edges and nodes that the authors have chosen to demonstrate a point… their subjective nature can be deceptive. One can choose a different set of edges and nodes to make an alternative argument. Essentially, there is no more information there than can be gleaned by staring at a connectivity diagram. There are certainly more interesting things that can be done… the data are made available though (supplemental 4), if you have an interest download it and have at it! I would especially liked to have seen some comments about variability between members of a cell type, as they have the information.
Any of the criticisms I throw out there shouldn’t overshadow how excited I am to see the work finally in print. The data matrix in itself is a huge resource (one that I might try and get around to playing around with myself…)