Virtual Fly Lab v0.6 — connectome-constrained synaptic learning and A/B choice

WHAT IS IMPLEMENTED
An additional, deliberately reduced rate-neuron model of the left mushroom-body
alpha2/alpha3 motif. It contains 905 KCab Kenyon cells, 3 MBONs (one MBON18 and
two MBON14), and 2 PPL1 DANs (PPL105 and PPL106): 910 annotated FlyWire neurons.
There are 3,287 selected anatomical directed edges, representing 15,062 synapses.
Of these, 2,575 KC->MBON edges have individual plastic efficacy multipliers.
The remaining DAN->KC/MBON edges provide anatomical projection evidence. We do
not simulate their fast synaptic currents: dopamine acts as a compartmental
modulator. Matching DAN->MBON projection evidence is required in the code.
Recurrent edges, other compartments, the other hemisphere and upstream/downstream
pathways are omitted. This is a selected motif, not the complete mushroom body.

DATA AND REPRODUCIBILITY
Anatomical counts: FlyWire FAFB v630 distributed by Shiu et al.
https://github.com/philshiu/Drosophila_brain_model
Commit 91bdd1e7dcf193f3e7ca5a8933497fcef63b7960
Annotations: Schlegel et al., flyconnectome/flywire_annotations v1.1.0
https://github.com/flyconnectome/flywire_annotations/tree/df6bb136f5b3d91c3992df4e8de2642329e2a384
Commit df6bb136f5b3d91c3992df4e8de2642329e2a384
Each cell retains its exact root ID as a string, annotated type, side and
transmitter. Every exported edge retains its unmodified integer synapse count.
Extraction selects left KCab cells connected to MBON14/18, plus PPL105/106.
The JSON records source URLs and SHA-256 checksums; the extraction script checks
them before reading. No anatomical edges or neuron identities are invented.
Run: uv run --with pyarrow==23.0.1 python web/scripts/extract-learning-circuit.py
Raw source caches are not published. The derived graph is learning-circuit.json.

SCIENTIFIC MOTIVATION, NOT AN EXACT REPRODUCTION
Masek et al. (2015), A dopamine-modulated neural circuit regulating aversive
taste memory in Drosophila: https://pubmed.ncbi.nlm.nih.gov/25981787/
This study motivates the involvement of alpha-lobe KCs, PPL1 and MBONs in taste
memory. Its taste protocol paired fructose with quinine or manipulated PPL1;
it did NOT validate our virtual electrical stimulus, sensory encoding, timing,
learning coefficients or free-walking avoidance decoder.
Li et al. (2020), The connectome of the adult Drosophila mushroom body provides
insights into function: https://elifesciences.org/articles/62576
This provides type/compartment mapping: PPL105 alpha'2alpha2, PPL106 alpha3,
MBON18 alpha2sc, MBON14 alpha3, and motivates compartment-specific plasticity.
We model only the alpha2 component of PPL105 modulation. We use the same simple
depression rule in both compartments; biological differences and stages of
short-/long-term consolidation are not reproduced.

SENSORY AND REINFORCEMENT ENCODING — EXPLICIT ASSUMPTIONS
The arena presents the cue of the nearest available sugar within 180 canvas units,
including during escape. This is an abstract recognition signal, NOT a claim
that sugar has an odor or that a verified taste/vision pathway was reconstructed.
Sort KC IDs by 32-bit FNV-1a hash and activate the first 10% (90 cells) for cue 0.
The next 90 cells encode disjoint cue 1. In A/B mode feeder A presents cue 0
and feeder B presents cue 1. Loose sugar in the original mode presents cue 0.
These ensembles are synthetic, not experimentally identified sugar-responsive
KCs. Their disjointness makes the cue-specificity control deliberately simple.
Shock drives BOTH selected PPL1 neurons directly for 0.4 s at a target normalized
activity of 1, 1.5, or 2. The electrical sensory pathway is not reconstructed.
No hardcoded sugar-shock pairing boolean changes weights: current KC traces
and DAN activity determine each synaptic update. Recent KC activity can therefore
be reinforced after the cue disappears; this is a decaying eligibility trace.

DYNAMICS — ALL NUMERICAL PARAMETERS CHOSEN FOR THIS DEMONSTRATION
Integration grid dt=0.005 s. Rates are normalized relative units, not measured Hz
and not spikes. First-order relaxation uses exact exponential updates:
  KC time constant: 0.05 s; DAN: 0.1 s; MBON: 0.08 s.
  KC eligibility e_i <- max(r_i, e_i * exp(-dt/1.5 s)).
Each real KC->MBON edge has immutable anatomical count c_ij and a learned
efficacy q_ij, initialized to 1. During plasticity:
  q_ij <- max(0.05, q_ij * exp(-4 * e_i * d_compartment * dt)).
The dopamine factor is the mean rate of compartment-matched DANs with an
observed projection to that MBON. This is a phenomenological LTD rule, not a
fitted molecular model or a claim that counts measure conductance.
MBON drive is sum_i(c_ij*q_ij*r_i) / sum_{i in cue0}(c_ij).
This count-based normalization sets the naive cue-0 steady response to 1.
Counterfactual reference output uses the same KC rates and MBON dynamics with
all q=1; it is an analysis instrument, not a second anatomical neuron population.
The two memory percentages are count-weighted depression of cue-0 and cue-1 edges.
The changed-edge counter uses q<0.999. Neither display controls behavior.

BEHAVIORAL DECODER — EXPLICIT ASSUMPTION
When mean reference output >0.05, suppression is 1 - mean(MBON)/mean(reference),
clamped to [0,1]; otherwise suppression is zero. Suppression >=0.35 prevents
sugar seeking/consumption and steers the agent away from nearby sugar.
Thus the behavior depends on the response of the plastic network. This relative
decoder and its threshold are application choices, not measured motor circuitry.
The original 447-neuron LIF taste circuit remains unchanged and still gates
consumption through MN9. The two modules receive separate surrogate inputs;
there is no claim of a reconstructed anatomical link between those modules.
Motor movement, shock escape, pleasure, stress, health and death remain game
rules. Displayed PPL1 activity is unrelated to the fictional pleasure meter.

MEMORY AND CONTROLS
No automatic forgetting or weight recovery is currently modeled. In-memory
weights persist until reset or page reload. Pausing stops all neural updates.
Blocking plasticity prevents new weight changes but preserves learned weights.
Silencing PPL1 clamps the selected DAN rates to zero; old memory remains.
Restoring weights sets q=1 and clears rates, traces and reinforcement in the
learning module. Health, stress and the original taste circuit remain intact.
Full experiment reset restores all state and disables the two control blocks.

TWO-FEEDER ASSAY (APPLICATION PROTOCOL)
Both feeders provide identical sugar, at (220,310) and (680,310). A/B label
different synthetic KC ensembles, not measured odors or visual pathways.
Only the nearest available source is encoded; there is no mixture encoding.
Eating makes that feeder unavailable for 8 seconds. The motor controller
commits to a feeder until depletion or avoidance; upon avoidance it selects
the other cue if available. This explicit motor rule is not inferred anatomy.
Starting a test returns the fly to (450,310), angle -pi/2, refills feeders,
clears learning rates/traces/pending reinforcement but preserves synaptic weights,
health, stress and taste-circuit state. Shocks and plasticity are blocked in
test mode. Optional side swapping moves cue identity along with the feeder.
Each visit is entry within 40 units of an available feeder, counted again only
after exiting beyond 65 units. Visits are distinct from consumption. Counts
are kept separately for observation/training and test. A repeated test replaces
its counters; reinstalling the pair clears both phases but preserves weights.
Bars share a scale and zero-visit phases have no percentage. Unequal duration,
position, heading, refill delays and physiological game state affect counts.
The software checks are not statistical evidence from independent animals.

VALIDATION AND LIMITS
Automated tests cover exact graph counts/types, compartment matching, paired
versus cue-only/shock-only/delayed inputs, DAN silencing, plasticity blockade,
cue specificity, memory after activity has decayed, deterministic integration,
weight-transfer recall, and erasure by restoring weights. Arena tests establish
avoidance and feeding recovery through the model and actual UI event handlers.
These are causal SOFTWARE controls. They do not validate against measured fly
behavior, establish realistic retention times, or reproduce a published model.
The small network was chosen to demonstrate a mechanism and is not a complete
digital brain. Direct plasticity in the original Shiu taste model is not claimed.

ATTRIBUTION AND DATA TERMS
Credit: FlyWire Consortium, Dorkenwald et al. for reconstruction; Schlegel et al.
for annotations and cell typing; Shiu et al. for the distributed connectivity;
Masek et al. and Li et al. for the scientific motivation and compartment mapping.
Browser model and selection are adaptations by Virtual Fly Lab, without author
endorsement. SHIU-LICENSE.txt retains the upstream MIT code license.
FlyWire anatomical-data terms are separate: https://flywire.ai/guidelines
Treat this educational data derivative as noncommercial (CC BY-NC 4.0).