Discrepancies in connectivity data between paper and NMC portal data

#1

Hello all,

We’ve been successfully using the connectivity data provided at the NMC portal for constructing neural networks of our own. After going through the paper and looking hard at the data, we however haven’t been able to answer some questions regarding differences between the paper and the data at the NMC portal (https://bbp.epfl.ch/nmc-portal/documents/10184/7288948/pathways_anatomy_factsheets_simplified.json). Specifically:

  • Issue 1: In the paper for the total number of synapses in the microcircuit, you provide the estimate ~37 synapses (36.7 ± 4.2 million synapses). However if you take the sum of total_synapse_count 's, you get a number close to 60 million (59 245 093).

  • Issue 2: When plotting the connection_probability variable as a heatmap similarly to Figure 7B in the paper, two spots stand out: L4_BTC:L4_ChC with 40% probability and L4_NGC:L4_ChC with 100% probability. These spots are blank in the paper. Why have they been weeded out from Fig 7B but left in the json data?

  • Issue 3: There are also some connections in Figure 7B that don’t have a connection_probability set in the json data, eg. L6_NGC:L6_SBC. Why are they missing from the json data?

We’d be really happy if someone took the time to clarify these issues for us.

Cheers,
Henri Hokkanen
PhD student, Univ.of Helsinki

#2

Hello Henri,

great to hear you are using our data in your projects.
About issue 1 I have to check and will come back to you later.

Issue 2: First of all, in the big 55 (morphological types) by 55 matrices we did not want to report values for connection types that might be just statistical fluctuation. After all, the connectome derivation has substantially stochastic components. So, we used a rather conservative approach: We show data in the 55 by 55 matrices only if there is at least one connected pair within 100 um* in seven statistical instantiations of the microcircuit. And the value we show is the average over all instantiations.

On the other hand in the portal data we give data for one statistical instantiation and without such filtering.

In your issue 2, what happened is that that the connection types were filtered out because they are not present in all instantiations, but they are present in the instantiations selected for the portal.
Take for example L4_NGC:L4_ChC, which indeed has a 100% connection probability. However, compare the total number of synapses to the mean number of synapses per connection: They are the same! In other words, there is only 1 connection. One potential pair, and it happens to be connected. Since this does not really mean much and might even be misleading, we applied the filter rules described above to the figures.

For issue 3: That is odd, they should be in the json data. I checked the connection type you mentioned manually and these are its values:
convergence 1.25 ± 0.43
divergence 0.59 ± 0.69
synapses per connection 11.7 ± 3.2
total synapse count 117
connection probability 0.1

I’ll look into why it was missing.

–Michael Reimann

*100 um in 3d space if both types are in the same layer, otherwise only in the x-z plane. Also it might be 125um, I’ll have to check, not sure right now.

#3

Thank you Michael,

Your explanation certainly clarifies issue 2.

Looking forward to hearing your comments on the other issues,
Henri, inspired by the HBP

#4

Okay, here I am again.
I confirmed the reason for issue 1:
In the paper, we report the numbers for one single microcircuit, i.e. the internal connectivity of that microcircuit only.

On the other hand, for the portal we first took one microcircuit and surrounded it on all sides with six additional microcircuits. The numbers we report are for considering pairs of neurons where the presynaptic neuron is in the central microcircuit, and the postsynaptic neuron in any of the seven (one central + six peripheral) microcircuits. In other words, the additional ~29 million synapses are from connections to neighboring microcircuits.

This has the advantage that the analyzed connectivity is more saturated, because the size of a single microcircuit is standardized with respect to the extent of a dendrite (Fig 3), but an axon can extent substantially further!
But, of course for your purpose of connecting neurons, the completely internal synapse numbers are what you need. I will talk to the people responsible and try to get them to update the portal such that it reports both numbers: internal and ‘saturated’.

In the meantime, I can already give you the completely internal numbers right now. Contact me under
the email address listed in my profile and tell me how you want the data (for example as a text file, or hdf5, etc) and I’ll send it to you.

–Michael

#5

Alright, that makes sense.

I think it’s a good idea to update the portal and add some details about the contents of the JSON files. There might be people other than us who are tinkering with the connectivity data.

Now one more thing: are most of the simulations in the paper run actually with the 1+6 architecture you described? Say, figure 12 for example (with high vs. low calcium comparison). So is it only neurons in the central microcircuit that are sending out axons?

Cheers,
Henri