Specific, cumulative and reversible odour learning in a whole-brain connectome model of Drosophila

Whole-brain models of the fruit fly built from the FlyWire connectome can predict how neurons respond to a stimulus, but they keep every synapse fixed and so cannot learn, while models that do learn simulate only the mushroom body and feed it with artificial odour signals. In this paper, I made one type of synapse plastic in a whole-brain leaky integrate-and-fire model of the FlyWire connectome: the synapses from Kenyon cells (KCs), those that encode odours, to mushroom body output neurons (MBONs), those that carry the fly's response to them. Under the learning rule, dopamine weakens each of these synapses in proportion to how strongly its KC fired, matching the weakening measured in living flies. Odours reached the KCs through the connectome's mapped wiring from the olfactory receptor neurons, but because the model has no sensors for pain, punishment was instead delivered by driving the 12 PPL1 dopamine neurons with random spikes at 100 Hz during the odour. Pairing an odour with punishment lowered the response of approach MBONs to that odour from 11.3 to 0.3 Hz in all five noise replicates of one simulated brain, while an unpaired odour, a never-paired odour and control runs with the rule switched off or scrambled changed little. When a second odour was paired on the brain already trained on a previous odour, it was learned at 85.5% of the strength reached by a naive brain, and at the end of training the approach MBONs still responded weakly to both odours, at 1.6 and 1.1 Hz against about 11 and 12 Hz before training. With a second rule, in which dopamine arriving before the odour strengthens a synapse instead of weakening it, a simulated fly body driven by the model unlearned and relearned in four of four cycles, whereas the first rule got stuck after one; the same potentiation, tested with backward pairing alone rather than as part of a full cycle, was too small to reach behaviour on its own. All results are for aversive, punishment-driven learning; reward learning could not be tested because sugar input does not recruit the reward dopamine neurons in this model. Learning depended on an authored plasticity rule, an injected punishment signal, and several other authored model choices, detailed in Methods; given these choices, the odour-specificity of learning followed from the structure of the connectome.