Research preview
Simulating biological neural circuits as computational systems, so their activity and behaviour can be studied directly.
Modelled mushroom body activity, Drosophila melanogaster
NeuroSim constructs computational models of biological neural circuits and runs them, so their activity can be observed as it develops rather than described after the fact.
What a nervous system does emerges from how its neurons are wired, how strongly and with what sign they interact, and how that activity is organised in time. Those relationships are hard to reason about on paper and straightforward to observe in a model that runs.
Sensory input
Activity arriving from receptors at the edge of the system.
Neural circuit
Populations of cells and the signed, weighted connections between them.
Population dynamics
Excitation, inhibition, delay and feedback, organised in time.
Output
The signal the circuit hands to whatever comes next.
The fruit fly is the only animal whose brain has been reconstructed at synapse resolution and characterised behaviourally in depth.
Optic lobe
Visual processing
Mushroom body
Olfactory learning and valence
Central complex
Heading and spatial representation
Antennal lobe
Olfactory input
Descending pathways
Motor output, not modelled
About 139,000 neurons, five orders of magnitude below a human brain, with a published wiring diagram and a literature that describes what many of its circuits do. That combination is rare, and it is what makes circuit models constrained by real structure possible at all.
The current research preview contains circuit models drawn from published Drosophila neuroscience. They are generated from described architecture, cell counts and projection rules, at a scale that runs interactively.
Olfactory learning and valence
Odour transduction
Odour identity
Sparse expansion
Global inhibition
Behavioural readout
Sparsity of the representation, the effect of removing inhibition, how odour identity survives noise.
Caron 2013, Lin 2014, Aso 2014
Heading and spatial representation
Localised visual cue
Heading bump
Rotate the bump
Keeps one peak
Whether activity stays localised, how the bump moves under asymmetric drive, what breaks the representation.
Turner-Evans 2020, Hulse 2021
Visual motion detection
Retinotopic input
Fast and slow arms
Direction selective
Pooled output
Direction selectivity, response to moving edges, and what the pooled output does when it drives a body.
Takemura 2013, Shinomiya 2019
Unstructured comparison
Drive
No imposed structure
Twenty per cent
Measured output
Whether an effect requires biological wiring, or whether the neuron model alone produces it.
Each modelled cell holds a membrane potential and a synaptic drive. Between events both decay. An arriving spike adds a step proportional to the number of synapses in that connection, positive or negative depending on the transmitter of the cell that sent it.
When the membrane reaches threshold the cell fires, resets, and cannot fire again for a fixed interval. The spike arrives at its targets after a propagation delay. Everything the network does is that rule, applied across thousands of cells.
Published values for the whole-brain Drosophila model of Shiu, Sterne and colleagues, 2024.
Every figure below was recorded from a run of the model. None of the numbers is illustrative.
A few hundred input channels expand onto thousands of Kenyon cells, held near silence by one inhibitory neuron. Only a small fraction respond to any odour.
73.9Hz
Receptor neurons
Odour transduction
34.6Hz
Projection neurons
Carry odour identity
2.8Hz
Kenyon cells
Sparse representation
71.0Hz
Feedback neuron
Global inhibition
4.0Hz
Output neurons
Behavioural readout
Silencing the interneurons that gate odour input, with the wiring, the stimulus and the seed all held.
Mushroom body · seed 20260820 · 1000 ms per condition
Observation and measurement
| Measurement | Control | Ablation | Change |
|---|---|---|---|
| Output firing rate | 4.0Hz | 33.4Hz | +726% |
| Network engaged | 67.1% | 90.7% | +35% |
| Response latency | 108.9ms | 58.7ms | -46% |
Caveat. A dependency shown here is a dependency of this computational model. It is not a measurement of an animal, and one seed is one sample.
The visual field of a simple agent drives the modelled optic lobe, and its wide-field output cells set the agent’s two wheel speeds.
This demonstrates a closed sensory to motor loop in the model. It is not a claim about how a fly navigates, and the body is a simple kinematic model rather than a physical one.
Simulate a neural circuit and watch its activity develop.
Examine individual cells, their state and their connections.
Change the input, the wiring or the conditions.
Quantify how the network responded.
No account, no upload, no install. The model runs on your own machine.
Engineering software and intelligent systems across simulation, artificial intelligence, robotics and digital environments.
NeuroSim is a software product developed by Alsadaany Industries. The company builds digital twin platforms and simulation engineering for industrial operations, and applies the same discipline here.
