Neurotransmitter classification

Neurons use different neurotransmitters to communicate; some are excitatory, while others are inhibitory; others alter neural activities in more subtle ways. Previous efforts have used machine learning to predict what neurotransmitters a neuron uses based on electron microscopy images (Eckstein et al., 2020). First, I reproduce those results in this project. These previous results are from an electron microscopy volume, the "Full Adult Female Brain," (FAFB), which uses serial section transmission electron microscopy. More recent electron microscopy volumes use Focused Ion Beam Scanning Electron Microscopy, which generates superior resolution in the z dimension, at the cost of x-y resolution. I find that using FIB-SEM microscopy images results in significantly worse neurotransmitter predictions. Finally, I compare synapse-level neurotransmitter predictions versus synaptic bouton-level predictions. Synaptic boutons contain many synapses; Eckstein et al., 2020 predicted neurotransmitters at the synapse-level. However, this may be problematic because the synapse is at the edge of the neuron, so by definition, these images only contain 50% of the neuron of interest. Furthermore, these images may be too small to contain relevant information to predict neurotransmitters accurately. I find that predicting neurotransmitters using the entire synaptic bouton dramatically increases accuracy, even when the same total area of image data is used for training.