MODEL & DATA
Real wiring. Explicit assumptions.
This prototype tests how activity could spread through a measured fly CNS subnetwork. It is a connectome-constrained hypothesis generator, without physiological calibration. The arena couples an expanded subnetwork to a separate illustrative body adapter.
What is represented
Each computational unit pools neurons with the same annotated cell type, side, and, for nerve cord cells, soma neuromere. Individual differences and retinotopic positions within a population are lost. The dataset is one male specimen.
We use Traced neurons with a named type from the official v1.0 flat files, then extract a visual-to-motor subnetwork. The preparation script processes the full connection table. The browser simulates the extracted populations, not the entire CNS.
How the subnetwork was selected
Six rounds of unsigned propagation from LC4, LPLC2, T4a/b, and T5a/b rank connected populations. Anatomical group quotas retain visual, central, descending, cord, and motor populations. Strong directed path ancestors and existing bilateral counterparts are retained; DNp01 and DNg13 are explicitly included. All measured edges among the retained populations are used.
Incoming counts from the full raw graph remain in each target’s denominator. Cropping or silencing cannot artificially strengthen the surviving connections. Unrepresented inputs remain silent.
The computational unit
rᵢ(t + Δt) = rᵢ(t) + α [ max(0, tanh(Iᵢ(t) + g Σⱼ Wᵢⱼ rⱼ(t))) − rᵢ(t) ]
Wᵢⱼ = sⱼ Cⱼᵢ / total_inputᵢ, where C is the observed synapse count and s is the assumed presynaptic sign. r is normalized population activity. I is external feature drive. g is connection gain.
α = 1 − exp(−Δt/τ). All populations update simultaneously from the previous state, with Δt = 2 ms and τ = 20 ms. Initial activity and baseline input are zero. These times are illustrative model parameters, not measured circuit latencies. Activity is bounded between 0 and 1.
Why a rate model?
Pooling many neurons into one cell-type unit calls for a population-rate abstraction. A published whole-brain model used individual leaky integrate-and-fire neurons and transmitter predictions to study feeding and grooming. That work motivates this prototype; its experimental validation does not transfer to our different dataset, visual circuits, or equations.
Shiu et al., Nature (2024) ↗
What the stimulus means
Looming: direct drive into LC4 and LPLC2. During the 60–240 ms stimulus window, LC4 receives A(0.35 + 0.65q), and LPLC2 receives Aq², with q progressing from 0 to 1. These are illustrative velocity/size feature proxies, not a fitted optical response.
Motion A/B: a constant pulse into T4a/T5a or T4b/T5b. The display illustrates the selected input; no image is processed by a retina. The model does not infer direction selectivity from the wiring.
No stimulus: no external input. A resting model must remain silent.
LC4 and LPLC2 are established visual inputs to the giant fiber pathway. DNg13 is a descending steering neuron. Motor units are identified using the dataset’s cb_motor and vnc_motor annotations, independently of their names.
Ache et al., Current Biology (2019) ↗
Yang et al., Cell (2024) ↗
Limits that affect interpretation
- Synapse counts are structural evidence, not measured conductances.
- Acetylcholine is positive; GABA and histamine are negative. Glutamate is inhibitory by default, but its effect depends on receptors. The control tests alternative assumptions.
- Unknown and modulatory transmitter outputs have zero fast effect in signed mode. Unsigned mode sets every sign positive and shows only structural spread.
- Gap junctions are absent. This is especially relevant to the giant fiber escape circuit.
- Group-level paths can join different neurons within a type. They are not proof of a continuous single-neuron route.
- Silencing is a model intervention. A changed motor trace is not an experimentally established causal effect.
- The rate model omits spontaneous activity, adaptation, plasticity, neuromodulatory state, and muscle mechanics. The arena adds geometric visual feedback, without other sensory modalities.
What would make it a stronger twin?
Use individual neurons with retinotopy, receptor-aware or experimentally fitted synaptic effects, and recordings for parameter fitting and held-out validation. Add gap junctions, a retina, then a neuromechanical body and sensory feedback. Validate simple visual experiments before interpreting behavior.
The embodied arena
The arena simulates an expanded 526-population graph, representing 9,615 neurons and 14,713 directed links. It includes the original 414 populations plus DLM/DVM wing power and b2 steering motor units, their strongest presynaptic populations, strong input paths, and existing bilateral counterparts. Original incoming-synapse denominators are retained. The full-CNS comparison below belongs to the original circuit experiment.
A sphere’s position relative to the fly determines a geometric looming feature. Signed activity propagates through the real graph, then a declared heuristic maps motor activity to launch, lift, thrust, and yaw. Position and orientation feed back into the next sensory step. A moving target alone cannot push the fly. Closing vision or disconnecting motor output from rest prevents powered motion.
The displayed fly is decimated anatomy from Flybody (Vaxenburg et al., Nature 2025, Apache 2.0). This app uses its appearance, not its MuJoCo dynamics or trained controllers. A stronger next model would couple this neural interface to those dynamics and fit motor readouts against held-out behavior. NeuroMechFly / FlyGym is another research platform for embodied sensory feedback.
Download arena graph ↓ · Arena numerical checks ↓ · Equations and reproduction notes
We also ran the full typed CNS
The sparse Python model ran the same looming input on 25,850 populations, representing 162,517 neurons and 6,742,293 directed population links. Cropping changes both motor amplitudes and rankings. These two models must not be treated as quantitatively interchangeable.
| Population | Subnetwork peak | Full CNS peak | Full CNS: DNp01 silenced |
|---|
| TTMn · R / T2 | 0.03401 | 0.02491 | −99.1% |
| TTMn · L / T2 | 0.01589 | 0.01073 | −90.5% |
| hDVM MN · L / T3 | 0.02385 | 0.03854 | −3.2% |
| hDVM MN · R / T3 | 0.01999 | 0.03112 | −3.1% |
DNg13 is silent in the cropped looming experiment but weakly active in the full model (peaks below 0.0004). The wider recurrent network can change a qualitative conclusion as well as response strength. The full-model results are recorded experiments; changing controls above runs the 414-population subnetwork.
All numbers use the default signed model: bilateral looming, input 1, gain 3, glutamate inhibitory. These are predictions from this model only; gap junctions and receptor-specific effects are not included.
Download all full-CNS population results ↓ · Full-CNS experiment settings & comparison ↓
Reproduce and inspect
Directionality, zero-input silence, inhibitory effects, bounded activity, silencing, and structural reachability checks pass. Browser and Python subnetwork trajectories agree to within 1.2 × 10⁻¹⁶ absolute activity. The sparse Python implementation also matches the scalar reference on controlled circuits. These are numerical checks, not physiological validation. Browser visual testing was not performed.
Download code, graph & research notes ↓Download graph JSON ↓
Source files and SHA-256 checksums
Data: FlyEM at HHMI Janelia, Cambridge Drosophila Connectomics Group, MRC LMB, and Google Research. CC BY 4.0. This is an independent prototype, not an official Janelia simulator. Official downloads ↗