Neurotransmitter Classification from Electron Microscopy Images at Synaptic Sites in Drosophila

Abstract

High-resolution electron microscopy of nervous systems enables the reconstruction of connectomes. A key piece of missing information from connectomes is the synaptic sign. We show that for D. melanogaster, artificial neural networks can predict the transmitter type released at synapses from electron micrographs and thus add putative signs to connections. Our network discriminates between six transmitters (acetylcholine, glutamate, GABA, serotonin, dopamine, octopamine) with an average accuracy of 87%/94% for synapses/entire neurons. We developed an explainability method to reveal which features our network is using and found significant ultrastructural differences between the classical transmitters. We predict transmitters in two connectomes and characterize morphological and connection properties of tens of thousands of neurons classed by predicted transmitter expression. We find that hemilineages in D. melanogaster largely express only one fastacting transmitter among their neurons. Furthermore, we show that neurons with different transmitters may differ in features like polarization and projection targets.