The data underlying our analysis is from the MICrONS consortium: a ~1mm³ volume of tissue from the mouse visual cortex, densely reconstructed using serial section electron microscopy, segmented into more than 54,000 individual neurons, among them ~30,000 excitatory ones. (2/12)

We employed GraphDINO, a self-supervised method for learning representations of neuronal morphologies without relying on manual annotations. The model outputs a vector embedding for each neuron that captures the morphological features of its dendritic tree. (3/12)

Our learned embeddings capture the essence of the 3D morphology of neurons and reflect known excitatory cell types from mouse V1. (4/12)

Dendritic morphologies form mostly a continuum, with distinct clusters only in deeper layers (e.g. layer 5 ET neurons). A quantitative test using synthethic surrogate data suggested that from layer 2 throughout upper layer 5 no density gaps exist and dendritic morphologies change continuously (5/12)

Dendritic morphologies vary with respect to three major axes: (1) the soma depth, (2) the total skeletal length of the apical dendrites and (3) the total skeletal length of the basal dendrites. (6/12)

Neurons in layer 2/3 show strong trends with increasing cortical depth: (1) decreasing width of their dendritic arbor and (2) smaller tufts. (7/12)

There morphological differences between primary visual cortex (V1) and higher visual areas (HVA): In layer 4, atufted neurons are primarily located in V1, while tufted neurons are more abundant in areas AL and RL. (8/12)

We found a novel morphological trait in layer 4: neurons that are primarily located in V1 in a narrow stripe around the L4-L5 boundary. These neurons are atufted and avoid reaching into L5 with their basal dendrites. (9/12)

Using the power of the MICrONS dataset, we could show that our functional digital twin of the neurons could predict the basal bias of cells in lower layer 4 without this model having access to any morphological information. (10/12) www.nature.com/articles/s41...

Thanks to co-authors Marissa Weis, Stelios Papadopoulos, @lhansel.bsky.social, Timo Lüddecke, Brendan Celii, Paul Fahey, Nuno da Costa, Forrest Collman, @csdashm.com , @clayreid.bsky.social, @sebastianseung.bsky.social, @viajake.bsky.social & Andreas Tolias. Funding: @erc.europa.eu & IARPA (12/12)