How accurate must a connectome be? Dose–response of a whole-brain Drosophila model to simulated reconstruction errors

Brain emulation assumes that a wiring diagram reconstructed from electron microscopy is accurate enough for a simulation built on it to behave like the original. How accurate is enough has not been measured for a whole brain. We took the published leaky integrate-and-fire model of the adult Drosophila brain (127,400 neurons, 52.8 million synapses), reproduced its published firing rates with an independent implementation, and corrupted its connectome with ten kinds of copying error at graded levels. Each of 568 corrupted brains was scored on seven predictions of the model that had been confirmed in the fly. Tolerance differed by more than an order of magnitude between error types. Randomly missing synapses acted like a uniform reduction of all weights: predictions failed between 10% and 20% loss, but after rescaling the model's single free parameter all seven predictions survived a loss of 30% in every copy and of 50% in four of five. Spurious synapses scattered at random had no effect at 50%. Errors that change the pattern of connections (rewiring, shuffled or noisy weights) were tolerated up to about 5–10%. Errors that hit whole neurons were the least forgiving: with a wrong transmitter sign in 5% of neurons, only a quarter of copies kept all seven predictions, as expected if about 24 of the 127,400 neurons are individually critical. A reconstruction can therefore lose most of its synapses and still support an emulation, while a few misclassified neurons cannot be afforded. This deposit contains the preprint (PDF) and an archive with the simulator, the error models, the output of the sweep for all 568 corrupted connectomes, and the scripts that produce the tables and the figure.