macOS desktop gains 3D fruit fly simulation driven by real FlyWire connectome neurons

The project bridges biological data and desktop software by embedding a real-time neural simulation of a fruit fly into macOS. The core of the system is a 668-neuron circuit derived from the FlyWire connectome, which includes escape command neurons (Giant Fibers), steering neurons, and grooming pathways. These neurons are modeled using a leaky-integrate-and-fire (LIF) simulation running at 1 kHz, ensuring that the fly’s behavior, such as fleeing from a cursor or grooming, emerges from the underlying neural activity rather than pre-programmed scripts. This approach provides a direct link between biological neural architecture and observable behavior, offering a tool for studying how connectome data translates into real-world actions.

For engineers, the implementation is notable for its minimal system requirements and lack of permissions. The application relies on macOS APIs that do not require user consent, such as cursor position, window frames, and thermal state, to create an interactive environment for the fly. The fly’s behavior adapts to these inputs in biologically plausible ways: fast cursor movements trigger escape responses via the looming-detector neurons, while window edges serve as ledges for the fly to walk along. The procedural body, though not derived from biological data, is controlled by the neural simulation, with gait, wing beats, and grooming all tied to the activity of specific neuron groups. This design demonstrates how complex, data-driven simulations can be integrated into consumer operating systems without friction.

The project’s transparency about its limitations is a strength. The connectome data provides wiring and neuron positions, but the simulation makes assumptions about synaptic weights and neurotransmitter predictions. The 668-neuron circuit is a subset of the full FlyWire connectome, focusing on pathways relevant to escape, grooming, and locomotion. While the simulation replicates key behaviors like the 4 ms escape latency, it does not model the entire fly brain. Engineers can regenerate the underlying data from raw FlyWire Codex dumps, ensuring reproducibility and allowing for experimentation with different subsets of the connectome. The inclusion of diagnostic tools, such as circuit invariants and behavior tests, further underscores the project’s utility as both a research tool and a technical demonstration.

From a practical standpoint, the application’s design choices highlight trade-offs between biological fidelity and computational efficiency. The 1 kHz LIF simulation balances real-time performance with neural accuracy, but the procedural body and simplified sensory inputs (e.g., cursor as looming threat) introduce abstractions. The fly’s circadian rhythm and temperature-dependent activity add realism, but these are driven by macOS system states rather than biological clocks. For engineers, the project serves as a case study in embedding scientific data into interactive applications, with potential applications in education, neuroscience research, and even the development of bio-inspired control systems for robotics or AI.