A quantitative census of millions of postsynaptic structures in a large electron microscopy volume of mouse visual cortex

Abstract Neurons display remarkable sub-cellular specificity in their synaptic targeting, which varies by cell type—for example, excitatory neurons prefer to target the spines of other excitatory cells. Modern electron microscopy connectomes enable the study of this sub-cellular specificity and its context in a circuit at unprecedented scale and resolution. However, this scale has also made it challenging to create accurate and efficient methods for classifying and segmenting fine cell components (including spines) across entire volumes. Here, we present a cost-efficient computational pipeline for classifying postsynaptic targets and segmenting structures such as spines. Our method relies only on having a mesh representation of a neuron and avoids processing image or segmentation data directly. Instead, we leverage tools from geometry processing to create features capturing the local geometry of a neuron’s surface. We couple this core technique with computational and storage optimizations, enabling reliable deployment over hundreds of thousands of neurons for a few hundred dollars in cloud compute cost. We then show that a simple classifier trained on the MICrONS mouse visual cortex dataset can use these mesh-based features to accurately classify synapses as targeting somas, dendritic shafts, or spines (weighted F1 score 0.961). Using this pipeline, we create a map of the postsynaptic structures at over 207.3 million synapses in MICrONS. We present an overview of this census of postsynaptic targeting, finding expected patterns (e.g., excitatory neurons preferentially targeting excitatory spines) as well as unexpected exceptions (e.g., Layer 5 near-projecting and Layer 6 corticothalamic cells often connecting to excitatory neuron shafts). We also demonstrate that these tools can be used to detect spines receiving multiple synaptic inputs, revealing surprising variability in their frequency across cells even within a cell type. We make our postsynaptic target predictions available for study, as well as the code for the computational pipeline and cloud deployment. Beyond MICrONS, we find that the model generalizes well to the H01 connectome without retraining (weighted F1 score 0.949), indicating that these tools will be useful in future connectomics reconstructions. More generally, our work demonstrates that representations derived from neuronal meshes can be a scalable and generalizable primitive for describing morphologies. Competing Interest Statement The authors have declared no competing interest. Footnotes This version adds some information about running the proposed method on several other connectome datasets, including validation of model performance in one of these. It also includes various small changes to the text and wording.