FlyBoxLab

FlyBoxLab is an open-source interactive environment for exploring a simplified fruit-fly connectome simulation.

It places a FlyBrain neural simulation inside a closed-loop virtual world where sensory events are encoded into neural inputs, processed through the connectome simulation, read through descending neurons, and decoded into the movement of a simulated body.

Rather than presenting the connectome as a static graph, FlyBoxLab turns it into an experimental environment. Users can construct sensory scenarios, inspect live neural activity, stimulate or silence populations, introduce reproducible synaptic lesions, fork neural state, compare agents, rewind experiments, and export experiment data.

The project is designed as a playground, experiment interface, and connectome debugging environment while explicitly separating biological source data from simulated neural dynamics and experimental mappings.

The connectome and core simulation engine come from published research and the open-source flybrain project. My work is the infrastructure that makes the system usable interactively: the real-time sandbox, experiment tooling, SDKs, visualization, backend, and deployment around it.

Modern connectome datasets describe enormous neural graphs, but understanding how structure might relate to behaviour remains difficult when the data is explored only through static visualizations or offline analysis.

Working directly with millions of synapses also creates a significant accessibility barrier. Researchers, students, developers, and curious users need a way to interact with the nervous system as a dynamic system: apply controlled interventions, observe resulting neural activity, compare experimental conditions, and reproduce those experiments.

FlyBoxLab explores what that interface can look like for the fruit-fly nervous system.

FlyBoxLab wraps a FlyBrain connectome simulation in an interactive closed-loop environment.

Events in the simulated world are converted into neural stimulation through experimental sensory encoders. Activity propagates through the neural simulation, selected descending-neuron populations are read out, and an experimental motor decoder translates those signals into changes in the simulated body's movement.

This loop makes it possible to observe how interventions at the neural level relate to changes inside the virtual environment.

The platform treats experimentation as a first-class feature. Neural populations can be stimulated or silenced, deterministic synaptic lesions can be introduced, sensory gain can be modified, exact simulation states can be forked or restored, and multiple agents can be compared under controlled conditions.