Examining viral videos of a “connectome-based fly”: testable predictions distinguish a biological model
Examining viral videos of a “connectome-based fly”: testable predictions distinguish a biological model
On 14 September, Patrick Mineault, a researcher working at the intersection of neuroscience and AI, examined viral videos in which a “connectome-based fly” plays Beat Saber, Doom, or Minecraft. The movement on screen depends on how the map of neural connections was turned into a network and coupled to a controller or virtual body. A model’s biological relevance is assessed by whether its inputs and outputs are grounded in measurements and whether it produces testable predictions.
A connectome is a map of synaptic connections between neurons. MaleCNS contains a complete map of the central nervous system of a male fruit fly, including the ventral nerve cord, which connects the brain to the body. The map shows which neurons are connected. To simulate their activity, a model also needs firing thresholds, signal strengths, and delays within cells and at synapses.
One way to simulate activity using this map is with a LIF network: its neurons integrate incoming signals and fire when they reach a threshold. In one of the demos Mineault examined, the network receives random input, and signals from a few neurons are converted into button presses. In Eon’s March demo, a few output neurons sent the virtual body commands to “walk forward” and “turn,” while prebuilt motor controllers coordinated the legs.
Philip Shiu and colleagues used a LIF model of the entire adult fruit fly brain to address a more specific question: which neurons would become active in response to a taste or mechanosensory signal. They then used light to activate selected neurons in real flies and observed their behavior. In this way, the connectivity map yielded testable predictions about feeding and grooming circuits.
Bingni Brunton and colleagues conducted a control experiment by coupling the connectome of the roundworm C. elegans to a physical model of a fly’s body. A separate neural network, trained through trial and error, translated the worm’s neural activity into commands for 42 actuators controlling six legs. The simulation produced a realistic gait.
“Plausible behavior can be produced even without biological fidelity.”
Janne Lappalainen and colleagues used the structure of the fly’s visual system, fitted model parameters to estimate motion, and compared predicted neural activity with data from 26 studies. For a whole-fly model, Mineault proposes grounding visual inputs, movements, and sensory feedback from the body in measurements, then testing the model against natural behavior and physiology.