Preprint: a virtual fly with a nervous system built from a real fly’s wiring map found 95% of odor sources without training or tuning for the tests; with shuffled connections, it found 18%
Preprint: a virtual fly with a nervous system built from a real fly’s wiring map found 95% of odor sources without training or tuning for the tests; with shuffled connections, it found 18%
On October 24, researchers at the University of California, Santa Cruz, posted a preprint on arXiv. They converted the BANC connectome, a map of connections in the brain and nerve cord of one adult fruit fly, into a network of 172 433 cells and connected it to a model of a fly’s body. The nerve cord carries commands to the legs.
In March, Eon Systems announced that it had “uploaded” a fly into a computer by connecting the brain model developed by Philip Shiu and colleagues to a NeuroMechFly body. In the debate over how to test such claims, Bing Brunton of the University of Washington showed that even a network built from a worm connectome can control a fly’s body if a trained intermediary translates its signals into leg movements. Benjamin Cowley of Cold Spring Harbor Laboratory proposed the necessary test: “If I simply create a randomly connected connectome, could it behave the same way?”
The preprint’s authors used neuron parameters from Shiu’s model and the same body to run this test, without training an intermediary or tuning parameters for the tests: “Tuning removes the opportunity to ask how much the wiring alone explains.” Two fixed rules connect the network and body: odors and approaching objects determine the firing rates of sensory cells, and leg motor cells determine step amplitude.
The network was first tested without the body on 15 cell pathways whose roles had been established in experiments on flies. A signal reached the target cell when the pathway supplied about 2% or more of its incoming synapses. One failure involved the giant fiber, which triggers takeoff. In a fly, it also connects to the jump muscle neuron through electrical contacts that the connectome does not record. The signal reached its target only after one connection was added manually.
In the body model, 20 flies traversed nine arenas, nearly all with walls, containing a total of 32 odor sources. The original network found 30,25 out of 32 sources; the shuffled network found 5,8. Each cell in the shuffled network retained its original number of connections and its sign (excitatory or inhibitory), but its partners were randomized. Leg commands recorded during network runs were then replayed without sensory input. From a rotated starting orientation, these commands led the flies to 2% of sources, while the active network reached 94% from the same starting orientation.
Turning toward an odor is calculated using the authors’ formula: the difference in odor input between the left and right sides is multiplied by a “balance,” defined as the number of spikes from mushroom body output cells that promote attraction to the odor minus the number from those that promote avoidance. The mushroom body is a center for odor learning. The network supplies this balance: it is 1,41 for the original network and 0,11 for the shuffled network. Substituting 1,44 into the formula restored odor approach in flies with the shuffled network: they reached 2,55 out of 3 sources in an arena with a wall, compared with 0,1 previously.
The researchers then added learning based on the mechanism in living flies: a dopamine reinforcement signal weakens synapses from mushroom body input cells to output cells, mainly those that promote odor attraction. After six episodes pairing odor with reinforcement, the balance became negative. None of the 20 flies found the source, whereas all had found it before learning. However, reinforcement without odor produced the same result: two thirds of the input cells fired even without odor because of a self-sustaining excitatory loop in the primary olfactory center. When the synaptic strength of one class of local cells in that center was reduced to 0,15 of its original value, 6,8% of the input cells responded to a typical odor, comparable to the proportion in a fly (5–10%). Almost all remained silent without odor, and the flies turned away only if reinforcement had coincided with odor. The authors predict that, in a fly whose input cells are active without odor, reinforcement delivered without odor will reduce the odor’s value.