Putative synapse locations have been detected in K15 with Forrest’s synapse detection algorithm and in W with Anish’s synapse detection algorithm. For each feature channel (Synapsin, VGlut, psd95, etc.) an 11x11x11 cube is extracted around each each putative synapse location and the voxel values are summed, creating a feature vector of length (number of channels). This gives us an \(n \times d\) matrix, where the \(n\) rows correspond to putative synapses and the \(d\) columns correspond to the summed immunoflorescence in each channel.
We have implemented our own Hierarchical Mclust function by augmenting Mclust. In the course of exploring we used the full suite of models available in mclustModelNames p. 88
After looking through the BIC plots of each of the 11 models for each node of the tree it seemed best to use the unconstrained model “VVV” = ellipsoidal, varying volume, shape, and orientation.
The below table contains links to synaptograms from each cluster as given above. The 5 synapses with feature vectors closest to their cluster mean were chosen as representatives.
C1111 | C1112 | C1121 | C1122 | C1211 | C1212 | C122 | C21111 | C21112 | C2112 | C212 | C2211 | C2212 | C222 |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 | 1 |
2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 | 2 |
3 | 3 | 3 | 3 | 3 | 3 | 3 | 3 | 3 | 3 | 3 | 3 | 3 | 3 |
4 | 4 | 4 | 4 | 4 | 4 | 4 | 4 | 4 | 4 | 4 | 4 | 4 | 4 |
5 | 5 | 5 | 5 | 5 | 5 | 5 | 5 | 5 | 5 | 5 | 5 | 5 | 5 |
And as you can see, both Synapsin channels and both VGlut1 channels are “hot” which match the cluster mean as given above.