Neurotransmitter prediction confidence for 126,419 FlyWire neurons, corrected for the spatial clustering of synapses

A connectome describes which neuron connects to which, but not whether a connection isexcitatory or inhibitory. That depends on the neurotransmitter, which is predicted fromelectron microscopy images by a classifier that also publishes a confidence score for eachneuron. Whole-brain Drosophila models built on these predictions convert them into a fixed+1 or -1. Four published models were checked, and none of them carries the confidencescore in its input tables. This dataset restores that information for 126,419 FlyWire neurons, covering 120,588,207synapses, or 92.7% of the synapses in the released data. It also adjusts it. The publishedconfidence score is the fraction of a neuron's synapses that voted for the winningtransmitter, which assumes the synapses are independent observations. The classifier readsa 640 nm cube around each synapse, and a neuron's synapses are closer together than that:the median distance to the nearest synapse on the same neuron is 84 nm. After groupingsynapses that fall inside one cube, the median neuron has about ten times fewerindependent observations than its confidence score assumes. Between 2.2% and 6.7% of the covered synapses belong to neurons whose transmitter is notdetermined by the data at the 95% level. The lower figure grants the independenceassumption in full; the upper figure treats each 640 nm cube as a single observation. Thesame analysis on the hemibrain dataset gives 7.83%, against 7.80% for FlyWire on a matchedfilter. The record also reports an invariance property. Writing the signed connectome asW = A diag(s), with every sign either +1 or -1, gives W W^T = A A^T. The singular valuesand left singular vectors of a signed connectome therefore do not depend on the signs atall, and results built on those quantities are unaffected by the uncertainty measuredhere. Results built on the eigenvalues of W, on the trace of W squared, or on anexcitatory and inhibitory partition are affected. This was checked across four publishedsign conventions: changing the sign of glutamate across 19,604 neurons, about one neuronin seven, leaves the Frobenius norm unchanged to machine precision while moving thespectral radius of the excitatory block by 56%. Column definitions, usage examples and limitations are in README.md.