Virtual Fly Lab v0.5 — taste circuit (see LEARNING-NOTES.txt for synaptic learning)

SOURCE AND PROVENANCE
Model: Shiu et al., A Drosophila computational brain model reveals sensorimotor
processing, Nature (2024), https://doi.org/10.1038/s41586-024-07763-9
Code and data: https://github.com/philshiu/Drosophila_brain_model
Pinned commit: 91bdd1e7dcf193f3e7ca5a8933497fcef63b7960
Anatomy: FlyWire Consortium, FAFB female adult brain, materialization v630.
This is the older dataset used in the authors' example, not current v783 or BANC.
Source file URLs and SHA-256 checksums are embedded in taste-circuit.json.
The authors' MIT license is retained in SHIU-LICENSE.txt. FlyWire's broader
public-data terms also apply; this educational derivative should be treated as
noncommercial. A code license does not replace anatomical-data terms.
FlyWire public-data guidance: https://flywire.ai/guidelines
CC BY-NC 4.0: https://creativecommons.org/licenses/by-nc/4.0/
Attribution: FlyWire Consortium; Dorkenwald et al. (brain reconstruction),
Schlegel et al. (annotation), and the synapse/transmitter contributors credited
by FlyWire. Selection and browser adaptation by Virtual Fly Lab; no endorsement
by the source authors is implied. The extracted data is modified by subsetting.

SELECTION
The 21 sugar-sensing input IDs and output MN9 (720575940660219265) are taken
directly from the authors' example.ipynb. Include every neuron on a directed
input-to-output path with at most three edges, plus all input and output cells.
Retain ALL edges between included neurons, including recurrent and inhibitory
edges, with original synapse counts and author-assigned signed counts.
Result: 447 neurons, 24,403 directed connections, 218,932 anatomical synapses.
No IDs or anatomical edges are invented. Other neuron roles are not annotated
here; "network" does not assert a cell type. Root IDs remain strings to avoid
JavaScript integer precision loss. External inputs/outputs are omitted.

DYNAMICS (ASSUMED, NOT MEASURED PER CELL)
Identical LIF cells with dv/dt=(Vrest-v+g)/20 ms and dg/dt=-g/5 ms.
Rest/reset -52 mV; threshold strictly above -45 mV; refractory 2.2 ms;
synaptic delay 1.8 ms; gain 0.275 mV per signed anatomical synapse.
These parameters come from the authors' model.py. Signed transmitter effects
are assumptions from that model, not a direct measurement of receptor effects.
The browser integrates the linear subthreshold equations analytically on a
0.1 ms grid. Both state variables pause during refractory time; incoming
synaptic events can still accumulate. Spiking resets v and g.
Each tick: integrate and detect threshold crossings, deliver due synaptic events,
apply independent Bernoulli approximations to Poisson input, reset spikes, and
schedule outgoing events. This implementation is not claimed bit-for-bit
equivalent to Brian2 or validated as a reproduction of the full published model.

SENSORY ENCODING (APPLICATION ASSUMPTION)
At physical contact in the 2D arena (distance <23 canvas units), all 21 input
neurons receive independent 150 Hz drives of 68.75 mV. Input cells have zero
refractory duration, matching the authors' stimulation convention. This is an
idealized stimulation protocol, not a measured taste-concentration response.
No stimulation is applied while the fly is merely approaching sugar.
Seeded xorshift32 provides reproducible input timing (default seed 1).

BEHAVIORAL DECODING (APPLICATION ASSUMPTION)
After at least 300 ms of contact and three MN9 spikes during that contact, the
application consumes one sugar portion unless learned avoidance inhibits feeding.
MN9 is used as an output proxy; these
thresholds and the translation to consumption were chosen for the demo.
This is not a validated ingestion, proboscis-muscle, or feeding mechanics model.
Silencing MN9 clamps its state and blocks its spikes, preventing consumption.
Upstream cells remain active. Unchecking restores its dynamics.

WHAT IS STILL SCRIPTED
Walking, attraction to sugar, escape from shocks, pleasure/"dopamine", stress,
health loss and death remain the earlier fictional browser rules. No dopamine
concentration, electrical injury, hunger or whole-body neuroscience is simulated.
The circuit only gates the contact-taste feeding response. Learning is modeled
in a separate connectome-constrained rate model with plastic KC->MBON weights.

SYNAPTIC LEARNING MODULE
See LEARNING-NOTES.txt for the additional 910-neuron KC/PPL1/MBON model.
It replaces the scalar aversion memory used in v0.4. Taste-circuit anatomical
counts, LIF dynamics and MN9 gating remain unchanged.

VALIDATION AND LIMITATIONS
Tests check exact IDs/counts, baseline silence, input-driven MN9 activity, loss
of output when edges are removed or MN9 is silenced, causality of consumption,
seed/time-slicing reproducibility, and analytic passive decay. The full original
dataset is used only for reproducible extraction; the deployed graph is reduced.
Omitting external pathways changes the network dynamics. These tests establish
software behavior, not agreement with living flies or the full-model results.
Learning occurs in the separate mushroom-body module; in this taste circuit,
the synaptic gain is inherited from the published model. Anatomical connectivity is not a complete digital
copy of a living brain, and computed spikes are not experimental recordings.

Reproduce extraction from repository root:
uv run --with pyarrow==23.0.1 python web/scripts/extract-taste-circuit.py
Tests: node --test web/tests/*.test.js