Shortest-path memory maps miss memory specificity and inhibition in the fly mushroom body

Complete connectomes now make it possible to ask where a memory is stored by analyzing changes in a wiring graph. Shortest-path measures such as Dijkstra’s algorithm are an attractive tool for this, because they are fast and need only the wiring. But, whether they reflect how memories actually behave in a circuit has not been tested. This study asked how accurately a shortest-path “memory map” represents an associative memory in the Drosophila mushroom body, and what such a map must be validated against. Memories were modeled as dopamine-gated depression of Kenyon cell (KC) to output neuron (MBON) synapses, or as a change in each MBON’s overall excitability. Models were built on a toy circuit and on the real hemibrain connectome. They were tested on the trained odor and on similar and unrelated untrained odors, and compared with 100 shuffled circuits and with a simulation of activity in the same networks. Three findings emerged. First, the two storage mechanisms produced identical shortest-path maps for the trained odor in all 220 paired trials, and only untrained odors revealed that the excitability rule generalizes to every odor. Second, the real wiring generalized memories like shuffled wiring, but gave a stronger innate ‘avoid’ bias than all 100 shuffles (p = 0.010). Third, the shortest-path map misjudged how far a memory spreads on both toy and real wiring, and lost most of the memory when even 10% of downstream neurons were inhibitory. Shortest-path memory maps can detect that a memory exists but not reliably where its boundaries lie. They should be validated against untrained stimuli and against activity simulations before being used to localize memories in connectomes.