266 — Automatic discovery of cell types and microcircuitry from neural connectomics

So let’s imagine we have an anatomical connectome: Fine. We’ve found a way to convert every synapse in the brain into an edge, and every neuron into a vertex. Fine.

But this still doesn’t tell us much about how the neuronal network actually works, because a graph of 100% excitatory cells will function very differently than a graph of a reasonable mixture of inhibitory and excitatory cells. Separately, the distances between synapses and the cell body can also fundamentally shift how the network behaves.

Determining cell type is difficult using only anatomical data, but it’s very possible to determine using cell morphology and neurite topology. In this paper, Jonas and Kording propose a predictive algorithm that generates probablistic neuron types for cells in the Helmstaedter 2013 mouse retina dataset and the C. elegans connectome. This unsupervised model incorporates statistical priors from the field of neuroscience — for example, retinal ganglion cells tend to synapse on nearer amacrine cells — and graph theory.

Of the 950 cells across 71 ground-truth cell types in the 2013 Helmstaedter mouse retina dataset, this technique found larger, less-specific clusters of cell types, with greater numbers of cells in each, but that also agreed with the GT data. Additionally, on the C. elegans dataset, the technique (automatically) roughly classified motor, sensory, and interneuron cell types.

The datasets and code are available here.